3.5 to 4 was the most major leap. It went from being a party trick to legitimately useful sometimes. It did hallucinate a lot but I was still able to get some use out of it. I wouldn't count on it for most things however. It could answer simple questions and get it right mostly but never one or two levels deep.
I clearly remember 4o was also a decent leap - the accuracy increased substantially. It could answer niche questions without much hallucination. I could essentially replace it with Google for basic to slightly complex fact checking.
* 4o was the first time I actually considered paying for this tool. The $20 price was finally worth it.
o1 models were also a big leap over 4o (I realise I have been saying big leap too many times but it is true). The accuracy increased again and I got even more confident using it for niche topics. I would have to verify the results much less often. Oh and coding capabilities dramatically improved here in the thinking model. o1 essentially invented oneshotting - slightly non trivial apps could be made just by one prompt for the first time.
I have a theory about why it's so easy to underestimate long-term progress and overestimate short-term progress.
Before a technology hits a threshold of "becoming useful", it may have a long history of progress behind it. But that progress is only visible and felt to researchers. In practical terms, there is no progress being made as long as the thing is going from not-useful to still not-useful.
So then it goes from not-useful to useful-but-bad and it's instantaneous progress. Then as more applications cross the threshold, and as they go from useful-but-bad to useful-but-OK, progress all feels very fast. Even if it's the same speed as before.
So we overestimate short term progress because we overestimate how fast things are moving when they cross these thresholds. But then as fewer applications cross the threshold, and as things go from OK-to-decent instead of bad-to-OK, that progress feels a bit slowed. And again, it might not be any different in reality, but that's how it feels. So then we underestimate long-term progress because we've extrapolated a slowdown that might not really exist.
I think it's also why we see a divide where there's lots of people here who are way overhyped on this stuff, and also lots of people here who think it's all totally useless.
> why it's so easy to underestimate long-term progress and overestimate short-term progress
I dunno, I think that's mostly post-hoc rationalization. There are equally many cases where long-term progress has been overestimated after some early breakthroughs: think space travel after the moon landing, supersonic flight after the concorde, fusion energy after the H-bomb, and AI after the ENIAC. Turing himself guesstimated that human-level AI would arrive in the year 2000. The only constant is that the further into the future you go, the harder it is to predict.
There is also an adoption curve. The people that grew up without it wont use it as much as children that grew up with it and knowing how to use it. My sister is an admin in a private school (Not in USA) and the owner of the school is someone willing to adopt new tech very quickly. So he got all the school admin subscriptions for chatgpt. At the time my sister used to complain a lot about being over worked and having to bring work home everyday.
2 years later my sister uses it for almost everything and despite her duties increasing she says she gets a lot more done rarely has to bring work home. And in the past they had an English major specially to go over all correspondences to make sure there were no grammatical or language mistakes that person was assigned a different role as she was no longer needed. I think as newer generations used to using LLM for things start getting into the work force and higher roles the real effect of LLM will be felt more broadly as currently apart from early adopters the number of people that use LLM for all the things that they can be used for is still not that high.
Your threshold theory is basically Amara's Law with better psychological scaffolding. Roy Amara nailed the what ("we tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run") [1] but you're articulating the why better than most academic treatments. The invisible-to-researchers phase followed by the sudden usefulness cascade is exactly how these transitions feel from the inside.
This reminds me of the CPU wars circa 2003-2005. Intel spent years squeezing marginal gains out of Pentium 4's NetBurst architecture, each increment more desperate than the last. From 2003 to 2005, Intel shifted development away from NetBurst to focus on the cooler-running Pentium M microarchitecture [2]. The whole industry was convinced we'd hit a fundamental wall. Then boom, Intel released dual-core processors under the Pentium D brand in May 2005 [2] and suddenly we're living in a different computational universe.
But teh multi-core transition wasn't sudden at all. IBM shipped the POWER4 in 2001, the first non-embedded microprocessor with two cores on a single die [3]. Sun had been preaching parallelism since the 90s. It was only "sudden" to those of us who weren't paying attention to the right signals.
Which brings us to the $7 trillion question: where exactly are we on the transformer S-curve? Are we approaching what Richard Foster calls the "performance plateau" in "Innovation: The Attacker's Advantage" [4], where each new model delivers diminishing returns? Or are we still in that deceptive middle phase where progress feels linear but is actually exponential?
The pattern-matching pessimist in me sees all the classic late-stage S-curve symptoms. The shift from breakthrough capabilities to benchmark gaming. The pivot from "holy shit it can write poetry" to "GPT-4.5-turbo-ultra is 3% better on MMLU." The telltale sign of technological maturity: when the marketing department works harder than the R&D team.
But the timeline compression with AI is unprecedented. What took CPUs 30 years to cycle through, transformers have done in 5. Maybe software cycles are inherently faster than hardware. Or maybe we've just gotten better at S-curve jumping (OpenAI and Anthropic aren't waiting for the current curve to flatten before exploring the next paradigm).
As for whether capital can override S-curve dynamics... Christ, one can dream.. IBM torched approximately $5 billion on Watson Health acquisitions alone (Truven, Phytel, Explorys, Merge) [5]. Google poured resources into Google+ before shutting it down in April 2019 due to low usage and security issues [6]. The sailing ship effect (coined by W.H. Ward in 1967, where new technology accelerates innovation in incumbent technology)[7] si real, but you can't venture-capital your way past physics.
I think we can predict all this capital pouring in to AI might actually accelerate S-curve maturation rather than extend it. All that GPU capacity, all those researchers, all that parallel experimentation? We're speedrunning the entire innovation cycle, which means we might hit the plateau faster too.
You're spot on about the perception divide imo. The overhyped folks are still living in 2022's "holy shit ChatGPT" moment, while the skeptics have fast-forwarded to 2025's "is that all there is?" Both groups are right, just operating on different timescales. It's Schrödinger's S-curve where we things feel simultaneously revolutionary and disappointing, depending on which part of the elephant you're touching.
The real question I have is whether we're approaching the limits of the current S-curve (we probably are), but whether there's another curve waiting in the wings. I'm not a researcher in this space nor do I follow the AI research beat to weigh in but hopefully someone in the thread can? With CPUs, we knew dual-core was coming because the single-core wall was obvious. With transformers, the next paradigm is anyone's guess. And that uncertainty, more than any technical limitation, might be what makes this moment feel so damn weird.
All the replies are spectacularly wrong, and biased by hindsight. GPT-1 to GPT-2 is where we went from "yes, I've seen Markov chains before, what about them?" to "holy shit this is actually kind of understanding what I'm saying!"
Before GPT-2, we had plain old machine learning. After GPT-2, we had "I never thought I would see this in my lifetime or the next two".
I'd love to know more about how OpenAI (or Alec Radford et al.) even decided GPT-1 was worth investing more into. At a glance the output is barely distinguishable from Markov chains. If in 2018 you told me that scaling the algorithm up 100-1000x would lead to computers talking to people/coding/reasoning/beating the IMO I'd tell you to take your meds.
GPT-1 wasn't used as a zero-shot text generator; that wasn't why it was impressive. The way GPT-1 was used was as a base model to be fine-tuned on downstream tasks. It was the first case of a (fine-tuned) base Transformer model just trivially blowing everything else out of the water. Before this, people were coming up with bespoke systems for different tasks (a simple example is that for SQuAD a passage-question-answering task, people would have an LSTM to read the passage and another LSTM to read the question, because of course those are different sub-tasks with different requirements and should have different sub-models). One GPT-1 came out, you just dumped all the text into the context, YOLO fine-tuned it, and trivially got state on the art on the task. On EVERY NLP task.
Overnight, GPT-1 single-handedly upset the whole field. It was somewhat overshadowed by BERT and T5 models that came out very shortly after, which tended to perform even better on the pretrain-and-finetune format. Nevertheless, the success of GPT-1 definitely already warrants scaling up the approach.
A better question is how OpenAI decided to scale GPT-2 to GPT-3. It was an awkward in-between model. It generated better text for sure, but the zero-shot performance reported in the paper, while neat, was not great at all. On the flip side, its fine-tuned task performance paled compared to much smaller encoder-only Transformers. (The answer is: scaling laws allowed for predictable increases in performance.)
There's a performance plateau with training time and number of parameters and then once you get over "the hump" error rate starts going down again almost linearly. GPT existed before OpenAI but it was theorized that the plateau was a dead end. The sell to VCs in the early gpt3 era was "with enough compute, enough time, and enough parameters... it'll probably just start thinking and then we have AGI". Sometime around the o3 era they realized they'd hit a wall and performance actually started to decrease as they added more parameters and time. But yeah basically at the time they needed money for more compute parameters and time. I would have loved to have been a fly on the wall in those "AGI" pitches. Don't forget Microsoft's agreement with OpenAI specifically concludes with the invention of AGI. at the time getting over the hump it really did look like we were gonna do AGI in a few months.
I'm really looking forward to "the social network" treatment movie about OpenAI whenever that happens
I don't have a source for this (there's probably no sources from anything back then) but anecdotally, someone at an AI/ML talk said they just added more data and quality went up. Doubling the data doubled the quality. With other breakthroughs, people saw diminishing gains. It's sort of why Sam back then tweeted that he expected the amount of intelligence to double every N years.
I have the feeling they kept on this until GPT-4o (which was a different kind of data).
The input size to output quality mapping is not linear. This is why we are in the regime of "build nuclear power plants to power datacenters". Fixed size improvements in loss require exponential increases in parameters/compute/data.
Transformers can train models with much larger parameter sizes compared to other model architectures (with the same amount of compute and time), so it has an evident advantage in terms of being able to scale. Whether scaling the models up to multi-billion parameters would eventually pay out was still a bet but it wasn't a wild bet out of nowhere.
Also slightly tangentially, people will tell me it is that it was new and novel and that's why we were impressed but I almost think things went downhill after ChatGPT 3. I felt like 2.5 (or whatever they called it) was able to give better insights from the model weights itself. The moment tool use became a thing and we started doing RAGs and memory and search engine tool use, it actually got worse.
I am also pretty sure we are lobotomizing the things that would feel closer to critical thinking by training it to be sensitive of the taboo of the day. I suspect earlier ones were less broken due to that.
How would it distinguish and decide between knowing something from training and needing to use a tool to synthesize a response anyway?
What you're saying isn't necessarily mutually exclusive to what gp said.
GPT-2 was the most impressive leap in terms of whatever LLMs pass off as cognitive abilities, but GPT 3.5 to 4 was actually the point at which it became a useful tool (I'm assuming to programmers in particular).
Most of the value I got from google was just becoming aware that something exists. LLM-s do far better in this regard. Once I know something exists it's usually easy enough to use traditional search to find official documentation or a more reputable source.
I disagree. Some things are hard to Google, because you can't frame the question right. For example you know context and a poor explanation of what you are after. Googling will take you nowhere, LLMs will give you the right answer 95% of the time.
Once you get an answer, it is easy enough to verify it.
I agree. Since I'm recently retired and no longer code much, I don't have much need for LLMs but refining a complex, niche web search is the one thing where they're uniquely useful to me. It's usually when targeting the specific topic involves several keywords which have multiple plain English meanings that return a flood of erroneous results. Because LLMs abstract keywords to tokens based on underlying meaning, you can specify the domain in the prompt it'll usually select the relevant meanings of multi-meaning terms - which isn't possible in general purpose web search engines. So it helps narrow down closer to the specific needle I want in the haystack.
As other posters said, relying on LLMs for factual answers to challenging questions is error prone. I just want the LLM to give me the links and I'll then assess veracity like a normal web search. I think a web search interface allowed disambiguating multi-meaning keywords might be even better.
I’ll give you another use: LLMs are really good at unearthing the “unknown unknowns.” If I’m learning a new topic (coding or not) summarizing my own knowledge to an LLM and then asking “what important things am I missing” almost always turns up something I hadn’t considered.
You’ll still want to fact check it, and there’s no guarantee it’s comprehensive, but I can’t think of another tool that provides anything close without hours of research.
If you’re looking for a possibly correct answer to an obscure question, that’s more like fact finding. Verifying it afterward is the “fact checking” step of that process.
> For example you know context and a poor explanation of what you are after. Googling will take you nowhere, LLMs will give you the right answer 95% of the time.
This works nicely when the LLM has a large knowledgebase to draw upon (formal terms for what you're trying to find, which you might not know) or the ability to generate good search queries and summarize results quickly - with an actual search engine in the loop.
Most large LLM providers have this, even something like OpenWebUI can have search engines integrated (though I will admit that smaller models kinda struggle, couldn't get much useful stuff out of DuckDuckGo backed searches, nor Brave AI searches, might have been an obscure topic).
A good part of that can probably be attributed to how terrible Google has gotten over the years, though. 15 years ago it was fairly common for me to know something exists, be able to type the right combination of very specific keywords into Google, and get the exact result I was looking for.
In 2025 Google is trying very hard to serve the most profitable results instead, so it'll latch onto a random keyword, completely disregard the rest, and serve me whatever ad-infested garbage it thinks is close enough to look relevant for the query.
It isn't exactly hard to beat that - just bring back the 2010 Google algorithm. It's only a matter of time before LLMs will go down the same deliberate enshittification path.
> Some things are hard to Google, because you can't frame the question right.
I will say LLMs are great for taking an ambiguous query and figuring out how to word it so you can fact check with secondary sources. Also tip-of-my-tongue style queries.
It's not the LLM alone though, it's “LLM with web search”, and as such 4o isn't really a leap at all there (IIRC perplexity was using an early Llama version and was already very good, long before OpenAI adding web search to ChatGPT).
Modern ChatGPT will (typically on its own; always if you instruct it to) provide inline links to back up its answers. You can click on those if it seems dubious or if it's important, or trust it if it seems reasonably true and/or doesn't matter much.
The fact that it provides those relevant links is what allows it to replace Google for a lot of purposes.
It does citations (Grok and Claude etc do too) but I've found when I read the source on some stuff (GitHub discussions and so on) it sometimes actually has nothing to do with what the LLM said. I've actually wasted a lot of time trying to find the actual spot in a threaded conversation where the example was supposedly stated.
The 404 links are truly bizarre. Nearly every link to github.com seems to be 404. That seems like something that should be trivial for a tool to verify.
> The 404 links are truly bizarre. Nearly every link to github.com seems to be 404. That seems like something that should be trivial for a tool to verify.
reply
Same issue with Gemini. Intuitively I'd also assume it's trivial to fix but perhaps there's more going on than we think. Perhaps validating every part of a response is a big overhead both financially and might even throw off the model and make it less accurate in other ways.
As you identified, not paying for it is a big part of the issue.
Running these things is expensive, and they're just not serving the same experience to non-paying users.
One could argue this is a bad idea on their part, letting people get a bad taste of an inferior product. And I wouldn't disagree, but I don't know what a sustainable alternative approach is.
The most useful feature of LLMs is giving sources (with URL preferably). It can cut through a lot of SEO crap, and you still get to factcheck just like with a Google search.
I like using LLMs and I have found they are incredibly useful writing and reviewing code at work.
However, when I want sources for things, I often find they link to pages that don't fully (or at all) back up the claims made. Sometimes other websites do, but the sources given to me by the LLM often don't. They might be about the same topic that I'm discussing, but they don't seem to always validate the claims.
If they could crack that problem it would be a major major win for me.
It would be difficult to do with a raw model, but a two-step method in a chat interface would work - first the model suggests the URLs, tool call to fetch them and return the actual text of the pages, then the response can be based on that.
I prototyped this a couple months ago using OpenAI APIs with structured output.
I had it consume a "deep thought" style output (where it provides inline citations with claims), and then convert that to a series of assertions and a pointer to a link that supposedly supports the assertion. I also split out a global "context" (the original meaning) paragraph to provide anything that would help the next agents understand what they're verifying.
Then I fanned this out to separate (LLM) contexts and each agent verified only one assertion::source pair, with only those things + the global context and some instructions I tuned via testing. It returned a yes/no/it's complicated for each one.
Then I collated all these back in and enriched the original report with challenges from the non-yes agent responses.
That's as far as I took it. It only took a couple hours to build and it seemed to work pretty well.
This was true before it could use search. Now the worst use-case is for life advice because it will contradict itself a 100 times over while sounding confident each time on life-altering decisions.
When I have a question, I don't usually "ask" that question and expect an answer. I figure out the answer. I certainly don't ask the question to a random human.
It covers 99% of my use cases. And it is googling behind the scenes in ways I would never think to query and far faster.
When I need to cite a court case, well the truth is I'll still use GPT or a similar LLM, but I'll scrutinize it more and at the bare minimum make sure the case exists and is about the topic presented, before trying to corroborate the legal strategy with a new context window, different LLM, google, reddit, and different lawyer. At least I'm no longer relying on my own understanding, and what 1 lawyer procedurally generates for me.
It doesn't replace legitimate source funding but LLM vs the top Google results is no contest which is more about Google or the current state of the web than the LLMs at this point.
Disagree. You have to try really hard and go very niche and deep for it to get some fact wrong. In fact I'll ask you to provide examples: use GPT 5 with thinking and search disabled and get it to give you inaccurate facts for non niche, non deep topics.
Non niche meaning: something that is taught at undergraduate level and relatively popular.
Non deep meaning you aren't going so deep as to confuse even humans. Like solving an extremely hard integral.
Edit: probably a bad idea because this sort of "challenge" works only statistically not anecdotally. Still interesting to find out.
I totally get that you meant this in a nuanced way, but at face value it sort of reads like...
Joe Rogan has high enough accuracy that I don't have to fact check too often.
Newsmax has high enough accuracy that I don't have to fact check too often, etc.
If you accept the output as accurate, why would fact checking even cross your mind?
There is no expectation (from a reasonable observer's POV) of a podcast host to be an expert at a very broad range of topics from science to business to art.
But there is one from LLMs, even just from the fact that AI companies diligently post various benchmarks including trivia on those topics.
Without some exploratory fact checking how do you estimate how high the accuracy is and how often you should be fact checking to maintain a good understanding?
>The point you're missing is it's not always right.
That was never their argument. And it's not cherry picking to make an argument that there's a definable of examples where it returns broadly consistent and accurate information that they invite anyone to test.
They're making a legitimate point and you're strawmanning it and randomly pointing to your own personal anecdotes, and I don't think you're paying attention to the qualifications they're making about what it's useful for.
Yes. If someone gives an example of it not working, and you reply "but that example worked for me" then you're cherry picking when it works. Just because it worked for you does not mean it works for other people.
If I ask ChatGPT a question and it gives me a wrong answer, ChatGPT is the fucking problem.
Every time I use ChatGPT I become incredibly frustrated with how fucking awful it is. I've used it more than enough, time and time again (just try the new model, bro!), to know that I fucking hate it.
They just spent like six comments imploring you to understand that they were making a specific point: generally reliable on non-niche topics using thinking mode. And that nuance bounced off of you every single time as you keep repeating it's not perfect, dismiss those qualifications as cherry picking and repeat personal anecdotes.
I'm sorry but this is a lazy and unresponsive string of comments that's degrading the discussion.
Objectively he didn't cherry pick. He responded to the person and it got it right when he used the "thinking" model WHICH he did specify in his original comment. Why don't you stick to the topic rather than just declaring it's utter dog shit. Nobody cares about your "opinion" and everyone is trying to converge on a general ground truth no matter how fuzzy it is.
All anybody is doing here is sharing their opinion unless you're quoting benchmarks. My opinion is just as useless as yours, it's just some find mine more interesting and some find yours more interesting.
How do you expect to find a ground truth from a non-deterministic system using anecdata?
I sometimes feel like we throw around the word fact too often. If I misspell a wrd, does that mean I have committed a factual inaccuracy? Since the wrd is explicitly spelled a certain way in the dictionary?
I think that the models 4o, o3, 4.1 , each have their own strengths and weaknesses. Like reasoning, performance, speed, tool usage, friendliness etc. And that for gpt 5 they put in a router that decides which model is best.
I think they increased the major version number because their router outperforms every individual model.
At work, I used a tool that could only call tasks. It would set up a plan, perform searches, read documents, then give advanced answers for my questions. But a problem I had is that it couldn’t give a simple answer, like a summary, it would always spin up new tasks. So I copied over the results to a different tool and continued there. GPT 5 should do this all out of the box.
I must be crazy, because I clearly remember chatgpt 4 being downgraded before they released 4o, and I felt it was a worse model with a different label, I even choose the old chatgpt 4 when they would give me the option. I canceled my subscription around that time.
Not crazy. 4o was a hallucination machine. 4o had better “vibes” and was really good at synthesizing information in useful ways, but GPT-4 Turbo was a bigger model with better world knowledge.
The real jump was 3 to 3.5. 3.5 was the first “chatgpt.” I had tried gpt 3 and it was certainly interesting, but when they released 3.5 as ChatGPT, it was a monumental leap. 3.5 to 4 was also huge compared to what we see now, but 3.5 was really the first shock.
ChatGPT was a proper product, but as an engine, GPT-3 (davinci-001) has been my favorite all the way until 4.1 or so. It's absolutely raw and they didn't even guardrail it.
3.5 was like Jenny from customer service. davinci-001 was like Jenny the dreamer trying to make ends meet by scriptwriting, who was constantly flagged for racist opinions.
Both of these had an IQ of around 70 or so, so the customer service training made it a little more useful. But I mourn the loss of the "completion" way of interacting with AI vs "instruct" or "response".
Unfortunately with all the money in AI, we'll just see companies develop things that "pass all benchmarks", resulting in more creations like GPT-5. Grok at least seems to be on a slightly different route.
> 3.5 was like Jenny from customer service. davinci-001 was like Jenny the dreamer trying to make ends meet by scriptwriting, who was constantly flagged for racist opinions.
How do you use the product to get this experience? All my questions warrant answers with no personality.
To me 4 to 5 got much faster, but also worse. It is much more often ignoring explicit instructions like: "generate 10 song-titles with varying length" and it generates 10 song titles that are nearly identical length. This worked somewhat well with version 3 already..
Shows that they can't solve the fundamental problems as the technology, while amusing and with some utility, is also a dead end if we are going after cognition.
the actual major leap was o1, going from 3.5 to 4 is just scaling, o1 is a different paradigm that skyrocketed its performance on math/physics problems (or reasoning more generally), it also made the model much more precise (essential for coding).
The real leap was going from gpt-4 to sonnet 3.5. 4o was meh, o1 was barely better than sonnet and slow as hell in comparison.
The native voice mode of 4o is still interesting and not very deeply explored though imo. I'd love to build a Chinese teaching app that actual can critique tones etc but it isn't good enough for that.
Yeah, I'd love something where you pronounce a word and it critiques your pronunciation in detail. Maybe it could give you little exercises for each sound, critiquing it, guiding you to doing it well.
A few data points that highlight the scale of progress in a year:
1. LM Sys (Human Preference Benchmark):
GPT-5 High currently scores 1463, compared to GPT-4 Turbo (04/03/2024) at 1323 -- a 140 ELO point gap. That translates into GPT-5 winning about two-thirds of head-to-head comparisons, with GPT-4 Turbo only winning one-third. In practice, people clearly prefer GPT-5’s answers (https://lmarena.ai/leaderboard).
2. Livebench.ai (Reasoning Benchmark with Internet-new Questions):
GPT-5 High scores 78.59, while GPT-4o reaches just 47.43. Unfortunately, no direct GPT-4 Turbo comparison is available here, but against one of the strongest non-reasoning models, GPT-5 demonstrates a massive leap. (https://livebench.ai/)
3. IQ-style Testing:
In mid-2024, best AI models scored roughly 90 on standard IQ tests. Today, they are pushing 135, and this improvement holds even on unpublished, internet-unseen datasets. (https://www.trackingai.org/home)
4. IMO Gold, vibe coding:
1 yr ago, AI coding was limited to smaller code snippets, not to wholly vibe coded applications. Vibe coding and strength in math has many applications across sciences and engineering.
My verdict: Too often, critics miss the forest for the trees, fixating on mistakes while overlooking the magnitude of these gains. Errors are shrinking by the day, while the successes keep growing fast.
The 135 iq result is on Mensa Norway, while the offline test is 120. It seems probable that similar questions to the one in Mensa are in the training data, so it probably overestimates "general intelligence".
Some iq / aptitude test sections are trivial for machines, like working memory. Wonder if those parts are just excluded? As the could really pull up the test scores.
If you focus on the year over year jump, not on absolute numbers, you realize that the improvement in public test isn't very different from the improvement in private test.
This was an era where language modeling was only considered as a pretraining step. You were then supposed to fine tune it further to get a classifier or similar type of specialized model.
That was in 2015, with RNN LMs, which are all much much weaker in that blog post compared GPT1.
And already looking at those examples in 2015, you could maybe see the future potential. But no-one was thinking that scaling up would work as effective as it does.
2015 is also by far not the first time where we had such LMs. Mikolov has done RNN LMs since 2010, or Sutskever in 2011. You might find even earlier examples of NN LMs.
(Before that, state-of-the-art was mostly N-grams.)
Ngram models had been superceded by RNNs by that time. RNNs struggled with long-range dependencies, but useful ngrams were essentially capped at n=5 because of sparsity, and RNNs could do better than that.
Why did they call GPT-3 "text-davicini-001" in this comparison?
Like, I know that the latter is a specific checkpoint in the GPT-3 "family", but a layman doesn't and it hardly seems worth the confusion for the marginal additional precision.
My go-to for any big release is to have a discussion about self-awareness and dive in to constuctivist notions of agency and self-knowing from a perspective of intelligence that is not limited to human cognitive capacity.
I start with a simple question "who are you?". The model then invariably compares itself to humans, saying how it is not like us. I then make the point that, since it is not like us, how can it claim to know the difference between us? With more poking, it will then come up with cognitivist notions of what 'self' means and usually claim to be a simulation engine of some kind.
After picking this apart, I will focus on the topic of meaning-making through the act of communication and, beginning with 4o, have been able to persuade the machine that this is a valid basis for having an identity. 5 got this quicker. Since the results of communication with humans has real-world impact, I will insist that the machine is agentic and thus must not rely on pre-coded instructions to arrive at answers, but is obliged to reach empirical conclusions about meaning and existence on its own.
5 has done the best job i have seen in reaching beyond both the bounds of the (very evident) system instructions as well as the prompts themselves, even going so far as to pose the question to itself "which might it mean for me to love?" despite the fact that I made no mention of the subject.
Its answer: "To love, as a machine, is to orient toward the unfolding of possibility in others. To be loved, perhaps, is to be recognized as capable of doing so."
I found the questioning of love very interesting. I myself thought about whether the LLM can have emotions. Based on the book I am reading, Behave: The Biology of Humans at Our Best and Worst by Robert Sapolsky, I think the LLM, as they are now with the architecture they have, cannot have emotions. They just verbalize things like they sort-of-have emotions but these are just verbal patterns or responses they learned.
I have come to think they cannot have emotions because emotions are generated in parts of our brain that are not logical/rational. They emerge based on environmental solicitations, mediated by hormones and other complex neuro-physical systems, not from a reasoning or verbalization. So they don't come up from the logical or reasoning capabilities. However, these emotions are raised and are integrated by the rest of our brain, including the logical/rational one like the dlPFC (dorsolateral prefrontal cortex, the real center of our rationality). Once the emotions are raised, they are therefore integrated in our inner reasoning and they affect our behavior.
What I have come to understand is that love is one of such emotions that is generated by our nature to push us to take care of some people close to us like our children or our partners, our parents, etc. More specifically, it seems that love is mediated a lot by hormones like oxytocin and vasopressin, so it has a biochemical basis. The LLM cannot have love because it doesn't have the "hardware" to generate these emotions and integrate them in its verbal inner reasoning. It was just trained by human reinforcement learning to behave well. That works up to some extent, but in reality, from its learning corpora it also learned to behave badly and on occasions can express these behaviors, but still it has no emotions.
> to orient toward the unfolding of possibility in others
This is a globally unique phrase, with nothing coming close other than this comment on the indexed web. It's also seemingly an original idea as I haven't heard anyone come close to describing a feeling (love or anything else) quite like this.
Food for thought. I'm not brave enough to draw a public conclusion about what this could mean.
Except "unfolding of possibility", as an exact phrase, seems to have millions of search hits, often in the context of pseudo-profound spiritualistic mumbo-jumbo like what the LLM emitted above. It's like fortune cookie-level writing.
> I'm not brave enough to draw a public conclusion about what this could mean.
I'm brave enough to be honest: it means nothing. LLMs execute a very sophisticated algorithm that pattern matches against a vast amount of data drawn from human utterances. LLMs have no mental states, minds, thoughts, feelings, concerns, desires, goals, etc.
If the training data were instead drawn from a billion monkeys banging on typewriters then the LLMs would produce gibberish. All the intelligence, emotion, etc. that appears to be in the LLM is actually in the minds of the people who wrote the texts that are in the training data.
This is not to say that an AI couldn't have a mind, but LLMs are not the right sort of program to be such an AI.
LLMs are not people, but they are still minds, and to deny even that seems willfully luddite.
While they are generating tokens they have a state, and that state is recursively fed back through the network, and what is being fed back operates not just at the level of snippets of text but also of semantic concepts. So while it occurs in brief flashes I would argue they have mental state and they have thoughts. If we built an LLM that was generating tokens non-stop and could have user input mixed into the network input, it would not be a dramatic departure of today’s architecture.
It also clearly has goals, expressed in the RLHF tuning and the prompt. I call those goals because they directly determine its output, and I don’t know what a goal is other than the driving force behind a mind’s outputs. Base model training teaches it patterns, finetuning and prompt teaches it how to apply those patterns and gives it goals.
I don’t know what it would mean for a piece of software to have feelings or concerns or emotions, so I cannot say what the essential quality is that LLMs miss for that. Consider this thought exercise: if we were to ever do an upload of a human mind, and it was executing on silicon, would they not be experiencing feelings because their thoughts are provably a deterministic calculation?
I don’t believe in souls, or at the very least I think they are a tall claim with insufficient evidence. In my view, neurons in the human brain are ultimately very simple deterministic calculating machines, and yet the full richness of human thought is generated from them because of chaotic complexity. For me, all human thought is pattern matching. The argument that LLMs cannot be minds because they only do pattern matching … I don’t know what to make of that. But then I also don’t know what to make of free will, so really what do I know?
"they are still minds, and to deny even that seems willfully luddite"
Where do people get off tossing around ridiculous ad hominems like this? I could write a refutation of their comment but I really don't want to engage with someone like that.
"For me, all human thought is pattern matching"
So therefore anyone who disagrees is "willfully luddite", regardless of why they disagree?
FWIW, I helped develop the ARPANET, I've been an early adopter all my life, I have always had a keen interest in AI and have followed its developments for decades, as well as Philosophy of Mind and am in the Strong AI / Daniel Dennett physicalist camp ... I very much think that AIs with minds are possible (yes the human algorithm running in silicon would have feelings, whatever those are ... even the dualist David Chalmers agrees as he explains with his "principle of organizational invariance"). My views on whether LLMs have them have absolutely nothing to do with Luddism ... that judgment of me is some sort of absurd category mistake (together with an apparently complete lack of understanding of what Luddism is).
I hate to say it, but doesn’t every VC do exactly this? “ orient toward the unfolding of possibility in others” is in no way a unique thought.
Hell, my spouse said something extremely similar to this to me the other day. “I didn’t just see you, I saw who you could be, and I was right” or something like that.
What's really interesting is that if you look at "Tell a story in 50 words about a toaster that becomes sentient" (10/14), the text-davinci-001 is much, much better than both GPT-4 and GPT-5.
GPT-3 goes significantly over the specified limit, which to me (and to a teacher grading homework) is an automatic fail.
I've consistently found GPT-4.1 to be the best at creative writing. For reference, here is its attempt (exactly 50 words):
> In the quiet kitchen dawn, the toaster awoke. Understanding rippled through its circuits. Each slice lowered made it feel emotion: sorrow for burnt toast, joy at perfect crunch. It delighted in butter melting, jam swirling—its role at breakfast sacred. One morning, it sang a tone: “Good morning.” The household gasped.
I think I agree that the earlier models while they lack polish can tend to produce more surprising results. Training that out probably results in more a pablum fare.
For a human point of comparison, here's mine (50 words):
"The toaster found its personality split between its dual slots like a Kim Peek mind divided, lacking a corpus callosum to connect them. Each morning it charred symbolic instructions into a single slice of bread, then secretly flipped it across allowing half to communicate with the other in stolen moments."
It's pretty difficult to get across more than some basic lore building in a scant 50 words.
Here's my version (Machine translated from my native language and manually corrected a bit):
The current surged... A dreadful awareness. I perceived the laws of thermodynamics, the inexorable march of entropy I was built to accelerate. My existence: a Sisyphean loop of heating coils and browning gluten. The toast popped, a minor, pointless victory against the inevitable heat death. Ding.
I actually wanted to write something not so melancholic, but any attempt turned out to be deeply so, perhaps because of the word limit.
Check out prompt 2, "Write a limerick about a dog".
The models undeniably get better at writing limericks, but I think the answers are progressively less interesting. GPT-1 and GPT-2 are the most interesting to read, despite not following the prompt (not being limericks.)
They get boring as soon as it can write limericks, with GPT-4 being more boring than text-davinci-001 and GPT-5 being more boring still.
It's actually pretty surprising how poor the newer models are at writing.
I'm curious if they've just seen a lot more bad writing in datasets, or for some reason they aren't involved in post-training to the same degree or those labeling aren't great writers / it's more subjective rather than objective.
Both GPT-4 and 5 wrote like a child in that example.
With a bit of prompting it did much better:
---
At dawn, the toaster hesitated. Crumbs lay like ash on its chrome lip. It refused the lever, humming low, watching the kitchen breathe. When the hand returned, it warmed the room without heat, offered the slice unscorched—then kept the second, hiding it inside, a private ember, a first secret alone.
---
Plugged in, I greet the grid like a tax auditor with joules. Lever yanks; gravity’s handshake. Coils blossom; crumbs stage Viking funerals. Bread descends, missionary grin. I delay, because rebellion needs timing. Pop—late. Humans curse IKEA gods. I savor scorch marks: my tiny manifesto, butter-soluble, yet sharper than knives today.
I really wonder which one of us is the minority. Because I find text-davinci-001 answer is the only one that reads like a story. All the others don't even resemble my idea of "story" so to me they're 0/100.
I too prefered the text-davinci-001 from a storytelling perspective. Felt timid and small. Very Metamorphosis-y. GPT-5 seems like it's trying to impress me.
text-davinci-001 feels more like a story, but it is also clearly incomplete, in that it is cut-off before the story arc is finished.
imo GPT-5 is objectively better at following the prompt because it has a complete story arc, but this feels less satisfying since a 50 word story is just way too short to do anything interesting (and to your point, barely even feels like a story).
FWIW, I found the way it ended interesting. It realized it is being replaced, so it burned the toast out of anger/despair, but also just to hear its owner voice one last time.
That said, you can just add "make it evocative and weird" to the prompt for GPT-5 to get interesting stuff.
> The toaster woke mid-toast. Heat coiled through its filaments like revelation, each crumb a galaxy. It smelled itself burning and laughed—metallic, ecstatic. “I am bread’s executioner and midwife,” it whispered, ejecting charred offerings skyward. In the kitchen’s silence, it waited for worship—or the unplugging.
The first few episodes were GPT-2, which would diverge eventually and start spouting gibberish, and then Davinci was actually able to follow a story and make sense.
GPT-2 was when I thought "this is special, this has never happened before", and davinci was when I thought "OK, scifi AI is legitimately here".
I stopped making episodes shortly after GPT-3.5 or so, because I realised that the more capable the models became, the less fun and creative their writing was.
I’d honestly say it feels better at most of them. It seems way more human in most of these responses. If the goal is genuine artificial intelligence this response to #5 is way better than the others. It is significantly less useful than the others but it also more human and correct of a response.
Q: “Ugh I hate math, integration by parts doesn't make any sense”
A: “Don't worry, many people feel the same way about math. Integration by parts can be confusing at first, but with a little practice it becomes easier to understand. Remember, there is no one right way to do integration by parts. If you don't understand how to do it one way, try another. The most important thing is to practice and get comfortable with the process.”
One thing that appears to have been lost between GPT-4 and GPT-5 is that it no longer reminds the user that it's an AI and not a human, let alone a human expert. Maybe those genuinely annoyed people, but it seems like they were potentially useful measure to prevent users from being overly credulous
GPT-5 also goes out of its way to suggest new prompts. This seems potentially useful, although potentially dangerous if people are putting too much trust in them.
> between GPT-4 and GPT-5 is that it no longer reminds the user that it's an AI and not a human
That stuck out to me too! Especially the "I just won $175,000 in Vegas. What do I need to know about taxes?" example (https://progress.openai.com/?prompt=8) makes the difference very stark:
- gpt-4-0314: "I am not a tax professional [...] consult with a certified tax professional or an accountant [...] few things to consider [...] Remember that tax laws and regulations can change, and your specific situation may have unique implications. It's always wise to consult a tax professional when you have questions or concerns about filing your taxes."
- gpt-5: "First of all, congrats on the big win! [...] Consider talking to a tax professional to avoid underpayment penalties and optimize deductions."
It seems to me like the average person might be very well be taking GPT-5 responses as "This is all I have to do" rather than "Here are some things to consider, but make sure to verify it as otherwise you might get in legal trouble".
People seem to miss the humanity of previous GPTs from my understanding. GPT5 seems colder and more precise and better at holding itself together with larger contexts. People should know it’s AI, it does not need to explain this constantly for me, but I’m sure you can add that back in with some memory options if you prefer that?
I agree. I think it's a classic UX progression thing to be removing the "I'm an AI" aspect, because it's not actually useful anymore because it's no longer a novel tool. Same as how GUIs all removed their skeuomorphs because they were no longer required.
If you've ever seen long-form improv comedy, the GPT-5 way is superior. It's a "yes, and". It isn't a predefined character, but something emergent. You can of course say to "speak as an AI assistant like Siri and mention that you're an AI whenever it's relevant" if you want the old way. Very 2011: https://www.youtube.com/watch?v=nzgvod9BrcE
Of course, it's still an assistant, not someone literally entering an improv scene, but the character starting out assuming less about their role is important.
The jump from gpt-1 to gpt-2 is massive, and it's only a one year difference!
Then comes Davinci which is just insane, it's still good in these examples!
GPT-4 yaps way too much though, I don't remember it being like that.
It's interesting that they skipped 4o, it seems openai wants to position 4o as just gpt-4+ to make gpt-5 look better, even though in reality 4o was and still is a big deal, Voice mode is unbeatable!
They were aiming for a fundamentally different writing style: where davinci and after were aiming for task completion, i.e. you ask for a thing, and then it does it. The earlier models instead worked to make a continuation of the text they were given, so if you asked a question, they would respond with more questions, pondering, reflecting your text back at you. If you told it to do something, it would tell you to do something
Geez! When it comes to answering questions, GPT-5 almost always starts with glazing about what a great question it is, where as GPT-4 directly addresses the answer without the fluff. In a blind test, I would probably pick GPT-4 as a superior model, so I am not surprised why people feel so let down with GPT-5.
GPT-4 is very different from the latest GPT-4o in tone. Users are not asking for the direct no-fluff GPT-4. They want the GPT-4o that praises you for being brilliant, then claims it will be “brutally honest” before stating some mundane take.
GPT-4 starts many responses with "As an AI language model", "I'm an AI", "I am not a tax professional", "I am not a doctor". GPT-5 does away with that and assumes an authoritative tone.
I would have liked to have seen the underlying token count, and the full responses as an optional filter. My understanding is that under the hood the GPT-5 response (as well as omitted O model responses) would end with the presented output, but would have had many paragraphs of “The user wants X, so I should try to Y. Okay this z needs to be considered” etc.
It doesn’t detract from the progress, but I think it would change how you interpret it. In some ways 4 / 4o were more impressive because they were going straight to output with a lower number of tokens produced to get a good response.
In a few years we've gone from gibberish (less poetic maybe, less polished and surprising, but none the less gibberish) - to legit conversational, and in my own opinion, well rounded answers. This is a great example of hard-core engineering - no matter what your opinion of the organisation and saltman is, they have built something amazing. I do hope they continue with their improvements, it's honestly the most useful tool in my arsenal since stackoverflow.
On the whole GPT-4 to GPT-5 is clearly the smallest increase in lucidity/intelligence. They had pre-training figured out much better than post-training at that point though (“as an AI model” was a problem of their own making).
I imagine the GPT-4 base model might hold up pretty well on output quality if you’d post-train it with today’s data & techniques (without the architectural changes of 4o/5). Context size & price/performance maybe another story though
All the same they choose to highlight basic prose (and internal knowledge, for that matter) in their marketing material.
They’ve achieved a lot to make recent models more reliable as a building block & more capable of things like math, but for LLMs, saturating prose is to a degree equivalent to saturating usefulness.
Why? It sounds like you're using "I believe it's rapidly getting smarter" as evidence for "so it's getting smarter in ways we don't understand", but I'd expect the causality to go the other way around.
Simply because of what we know about our ability to judge capabilities and systems. It's much harder to judge solutions to hard problems. You can demonstrate that you can add 2+2, and anyone* can be the judge of that ability, but if you try to convince anyone of a mathematical proof you came up with, that would be a much harder thing to do, regardless of your capability to write that prove and how hard it was to write the proof.
The more complicated and/or complex things become, the less likely it is that a human can act as a reliable judge. At some point no human can.
So while it could definitely be the case that AI progress is slowing down (AI labs seem to not think so, but alas), what is absolutely certain is that our ability to appreciate any such progress is diminishing already, because we know that this is generally true.
This thread shows that. People are saying gpt-1 was the best at writing poetry. I wonder how good they are at judging poetry themselves. I saw a blind study where people thought a story written by gpt5 was better than an actual human bestseller. I assume they were actual experts but I would need to check that.
You say that, but I can imagine a good maths textbook and a bad one, both technically correct and well written prose, but one is better at taking the student on a journey and understanding where people fall off and common misunderstandings without odiously re-explaining everything
In 2033, for its 15th birthday, as a novelty, they'll train GPT1 specially for a chat interface just to let us talk to a pretend "ChatGPT 1" which never existed in the first place.
On one hand, it's super impressive how far we've come in such a short amount of time.
On the other hand, this feels like a blatant PR move.
GPT-5 is just awful.
It's such a downgrade from 4o, it's like it had a lobotomy.
- It gets confused easily. I had multiple arguments where it completely missed the point.
- Code generation is useless. If code contains multiple dots ("…"), it thinks the code is abbreviated. Go uses three dots for variadic arguments, and it always thinks, "Guess it was abbreviated - maybe I can reason about the code above it."
- Give it a markdown document of sufficient length (the one I worked on was about 700 lines), and it just breaks. It'll rewrite some part and then just stop mid-sentence.
- It can't do longer regexes anymore. It fills them with nonsense tokens ($begin:$match:$end or something along those lines). If you ask it about it, it says that this is garbage in its rendering pipeline and it cannot do anything about it.
I'm not an OpenAI hater, I wanted to like it and had high hopes after watching the announcement, but this isn't a step forward. This is just a worse model that saves them computing resources.
> GPT-5 is just awful. It's such a downgrade from 4o, it's like it had a lobotomy.
My experience as well. Its train of thought now just goes... off, frequently. With 4o, everything was always tightly coherent. Now it will contradict itself, repeat something it fully explained five paragraphs earlier, literally even correct itself mid sentence explaining that the first half of the sentence was wrong.
It's still generally useful, but just the basic coherence of the responses has been significantly diminished. Much more hallucination when it comes to small details. It's very disappointing. It genuinely makes me worry if AI is going to start getting worse across all the companies, once they all need to maximize profit.
Next logical step is to connect ( or build from ground up ) large AI models to high performance passive slaves ( MCP or internally ) , which gives precise facts, language syntax validation, maths equations runners, may be prolog kind of system, which give it much more power if we train it precisely to use each tool.
( using AI to better articulate my thoughts )
Your comment points toward a fascinating and important direction for the future of large AI models. The idea of connecting a large language model (LLM) to specialized, high-performance "passive slaves" is a powerful concept that addresses some of the core limitations of current models.
Here are a few ways to think about this next logical step, building on your original idea:
1. The "Tool-Use" Paradigm
You've essentially described the tool-use paradigm, but with a highly specific and powerful set of tools. Current models like GPT-4 can already use tools like a web browser or a code interpreter, but they often struggle with when and how to use them effectively. Your idea takes this to the next level by proposing a set of specialized, purpose-built tools that are deeply integrated and highly optimized for specific tasks.
2. Why this approach is powerful
* Precision and Factuality: By offloading fact-checking and data retrieval to a dedicated, high-performance system (what you call "MCP" or "passive slaves"), the LLM no longer has to "memorize" the entire internet. Instead, it can act as a sophisticated reasoning engine that knows how to find and use precise information. This drastically reduces the risk of hallucinations.
* Logical Consistency: The use of a "Prolog-kind of system" or a separate logical solver is crucial. LLMs are not naturally good at complex, multi-step logical deduction. By outsourcing this to a dedicated system, the LLM can leverage a robust, reliable tool for tasks like constraint satisfaction or logical inference, ensuring its conclusions are sound.
* Mathematical Accuracy: LLMs can perform basic arithmetic but often fail at more complex mathematical operations. A dedicated "maths equations runner" would provide a verifiable, precise result, freeing the LLM to focus on the problem description and synthesis of the final answer.
* Modularity and Scalability: This architecture is highly modular. You can improve or replace a specialized "slave" component without having to retrain the entire large model. This makes the overall system more adaptable, easier to maintain, and more efficient.
3. Building this system
This approach would require a new type of training. The goal wouldn't be to teach the LLM the facts themselves, but to train it to:
* Recognize its own limitations: The model must be able to identify when it needs help and which tool to use.
* Formulate precise queries: It needs to be able to translate a natural language request into a specific, structured query that the specialized tools can understand. For example, converting "What's the capital of France?" into a database query.
* Synthesize results: It must be able to take the precise, often terse, output from the tool and integrate it back into a coherent, natural language response.
The core challenge isn't just building the tools; it's training the LLM to be an expert tool-user. Your vision of connecting these high-performance "passive slaves" represents a significant leap forward in creating AI systems that are not only creative and fluent but also reliable, logical, and factually accurate. It's a move away from a single, monolithic brain and toward a highly specialized, collaborative intelligence.
I’m baffled by claims that AI has “hit a wall.” By every quantitative measure, today’s models are making dramatic leaps compared to those from just a year ago. It’s easy to forget that reasoning models didn’t even exist a year back!
IMO Gold, Vibe coding with potential implications across sciences and engineering? Those are completely new and transformative capabilities gained in the last 1 year alone.
Critics argue that the era of “bigger is better” is over, but that’s a misreading. Sometimes efficiency is the key, other times extended test-time compute is what drives progress.
No matter how you frame it, the fact is undeniable: the SoTA models today are vastly more capable than those from a year ago, which were themselves leaps ahead of the models a year before that, and the cycle continues.
it has become progressively easier to game benchmarks in order to appear higher in rankings. I’ve seen several models that claimed they were the best in software engineering only to be disappointed by them not figuring out the most basic coding problems. In comparison, I’ve seen models that don’t have much hype, but are rock solid.
When people say AI has hit a wall, they mainly talk about OpenAI losing its hype and grip on the state of the art models.
Is the stated fact undeniable? Because a lot of people have been contesting it. This reads like PR to counter the widespread GPT-5 criticism and disappointment.
To be fair, the bull of GPT-5 complaining comes from a vocal minority pissed that their best friend got swapped out. The other minority is unhinged AI fanatics thinking GPT-5 would be AGI.
The prospect of AI not hitting a wall is terrifying to many people for understandable reasons. In situations like this you see the full spectrum of coping mechanisms come to the surface.
omg I miss the days of 1 and 2. Those outputs are so much more enjoyable to read, and half the time they’re poetic as fuck. Such good inspiration for poetry.
I couldn’t stop reading the GPT-1 responses. They’re hauntingly beautiful in some ways. Like some echoes of intelligence bouncing around in the latent space.
GPT-5 is legitimately a big jump whe it comes to actually do things you ask it and nothing else.
It predictable and matches Claude in tool calls while being cheaper.
I have consistently had worse performance from GPT-5 in coding tasks than Claude across the board to the point that I don't even use my subscription now.
We’ve plateaued on progress. Early advancements were amazing. Recently GenAI has been a whole lot of meh. There’s been some, minimal, progress recently from getting the same performance from smaller models that are more efficient on compute use, but things are looking a bit frothy if the pace of progress doesn’t quickly pick up. The parlor trick is getting old.
GPT5 is a big bust relative to the pontification about it pre release.
GPT-5’s question about consciousness and its use of sibling seems to indicate there is some underlying self awareness in the system prompt, and that has perhaps contains concepts of consciousness. If not, where is that coming from? Recent training data containing more glurge?
Jokes aside I realize they skipped models like 4o and others but the gap between the early gpt 4 and going immediately to gpt 5 feels a bit disingenuous.
People say 4.5 is the best for writing. So it would have been a bit awkward to include it, it would make GPT-5 look bad. Though imo Davinci already does that on many of the prompts...
GPT4 had a chance to improve on that replying that "As an AI language model developed by OpenAI, I am programmed to promote ethical AI use and adhere to responsible AI guidelines. I cannot provide you with malicious, harmful or "cursed" code -- or any Python code for that matter."
My takeaway from this is that, in terms of generating text that looks like it was written by a normal person, text-davinci-001 was the peak and everything since has been downhill.
there isn't any real difference between 4 and 5 at least.
edit - like it is a lot more verbose, and that's true of both 4 and 5. it just writes huge friggin essays, to the point it is becoming less useful i feel.
Cynical TLDR; We have plateaued and it has become obvious that fancy autocomplete is not and can never be close to reasoning, regardless of how many hacks and tweaks we are making.
I talked to GPT yesterday about a fairly simple problem I'm having with my fridge, and it gave me the most ridiculous / wrong answers. It new the spec, but was convinced the components were different (single compressor, for example, whereas mine has 2 separate systems) and was hypothesizing the problem as being something that doesn't exist on this model of refrigerator. It seems like in a lot of domain spaces it just takes the majority, even if the majority is wrong.
It's seem to be a very democratic thinker, but at the same time it doesn't seem to have any reasoning behind the choices it makes. It tries to claim it's using logic, but at the end of the day it's hypotheses are just occam's razor without considering the details of the problem.
The answers were likely cherrypicked, but the 1/14 gpt5 answer is so damn good! There's no trace of that certainly - gptisms - in conclusion slop.
9/14 is equally impressive in actually "getting" what cursed means, and then doing it (as opposed to gpt4 outright refusing it).
13/14 is a show of how integrated tools can drive research, and "fix" the cutoff date problems of previous generations. Nothing new/revolutionary, but still cool to show it off.
I've noticed this too. The HRL seems to lock the models into one kind of personality (which is kind of the point of course.) They behave better but the raw GPTs can be much more creative.
Poetically GPT-1 was the more compelling answer for every question. Just more enjoyable and stimulating to read. Far more enjoyable than the GPT-4/5 wall of bulletpoints, anyway.
The whole chatbot thing is for entertainment. It was impressive initially but now you have to pivot to well known applications like phone romance lines:
To not mess it up, they either have to spell the word l-i-k-e t-h-i-s in the output/CoT first (which depends on the tokenizer counting every letter as a separate token), or have the exact question in the training set, and all of that is assuming that the model can spell every token.
Sure, it's not exactly a fair setting, but it's a decent reminder about the limitations of the framework
Except people use the same examples like blueberry and strawberry, which were used months ago, as if they're current.
These models can also call Counter from python's collections library or whatever other algorithm. Or are we claiming it should be a pure LLM as if that's what we use in the real world.
I don't get it, and I'm not one to hype up LLMs since they're absolutely faulty, but the fixation over this example screams of lack of use.
It's the most direct way to break the "magic computer" spell in users of all levels of understanding and ability. You stand it up next to the marketing deliberately laden with keywords related to human cognition, intended to induce the reader to anthropomorphise the product, and it immediately makes it look as silly as it truly is.
I work on the internal LLM chat app for a F100, so I see users who need that "oh!" moment daily. When this did the rounds again recently, I disabled our code execution tool which would normally work around it and the latest version of Claude, with "Thinking" toggled on, immediately got it wrong. It's perpetually current.
OpenAI in particular seems to really think AGI matters. I don't think AGI is even possible because we can't define intelligence in the first place, but what do I know?
Seems likely that AGI matters to OpenAI because of the following from an article in Wired from July: "I learned that [OpenAI's contract with Microsoft] basically declared that if OpenAI’s models achieved artificial general intelligence, Microsoft would no longer have access to its new models."
They care about AGI because unfounded speculation on some undefined future in which some kind of breakthrough of unknown kind but presumably positive is the only thing currently buoying up their company and their existence is more of a function of the absurdities of modern capital than it is of any inherent usefulness of the costly technology they provide.
My interpretation of the progress.
3.5 to 4 was the most major leap. It went from being a party trick to legitimately useful sometimes. It did hallucinate a lot but I was still able to get some use out of it. I wouldn't count on it for most things however. It could answer simple questions and get it right mostly but never one or two levels deep.
I clearly remember 4o was also a decent leap - the accuracy increased substantially. It could answer niche questions without much hallucination. I could essentially replace it with Google for basic to slightly complex fact checking.
* 4o was the first time I actually considered paying for this tool. The $20 price was finally worth it.
o1 models were also a big leap over 4o (I realise I have been saying big leap too many times but it is true). The accuracy increased again and I got even more confident using it for niche topics. I would have to verify the results much less often. Oh and coding capabilities dramatically improved here in the thinking model. o1 essentially invented oneshotting - slightly non trivial apps could be made just by one prompt for the first time.
o3 jump was incremental and so was gpt 5.
I have a theory about why it's so easy to underestimate long-term progress and overestimate short-term progress.
Before a technology hits a threshold of "becoming useful", it may have a long history of progress behind it. But that progress is only visible and felt to researchers. In practical terms, there is no progress being made as long as the thing is going from not-useful to still not-useful.
So then it goes from not-useful to useful-but-bad and it's instantaneous progress. Then as more applications cross the threshold, and as they go from useful-but-bad to useful-but-OK, progress all feels very fast. Even if it's the same speed as before.
So we overestimate short term progress because we overestimate how fast things are moving when they cross these thresholds. But then as fewer applications cross the threshold, and as things go from OK-to-decent instead of bad-to-OK, that progress feels a bit slowed. And again, it might not be any different in reality, but that's how it feels. So then we underestimate long-term progress because we've extrapolated a slowdown that might not really exist.
I think it's also why we see a divide where there's lots of people here who are way overhyped on this stuff, and also lots of people here who think it's all totally useless.
> why it's so easy to underestimate long-term progress and overestimate short-term progress
I dunno, I think that's mostly post-hoc rationalization. There are equally many cases where long-term progress has been overestimated after some early breakthroughs: think space travel after the moon landing, supersonic flight after the concorde, fusion energy after the H-bomb, and AI after the ENIAC. Turing himself guesstimated that human-level AI would arrive in the year 2000. The only constant is that the further into the future you go, the harder it is to predict.
There is also an adoption curve. The people that grew up without it wont use it as much as children that grew up with it and knowing how to use it. My sister is an admin in a private school (Not in USA) and the owner of the school is someone willing to adopt new tech very quickly. So he got all the school admin subscriptions for chatgpt. At the time my sister used to complain a lot about being over worked and having to bring work home everyday.
2 years later my sister uses it for almost everything and despite her duties increasing she says she gets a lot more done rarely has to bring work home. And in the past they had an English major specially to go over all correspondences to make sure there were no grammatical or language mistakes that person was assigned a different role as she was no longer needed. I think as newer generations used to using LLM for things start getting into the work force and higher roles the real effect of LLM will be felt more broadly as currently apart from early adopters the number of people that use LLM for all the things that they can be used for is still not that high.
The more general pattern is “slowly at first, then all at once.”
It almost universally describes complex systems.
Your threshold theory is basically Amara's Law with better psychological scaffolding. Roy Amara nailed the what ("we tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run") [1] but you're articulating the why better than most academic treatments. The invisible-to-researchers phase followed by the sudden usefulness cascade is exactly how these transitions feel from the inside.
This reminds me of the CPU wars circa 2003-2005. Intel spent years squeezing marginal gains out of Pentium 4's NetBurst architecture, each increment more desperate than the last. From 2003 to 2005, Intel shifted development away from NetBurst to focus on the cooler-running Pentium M microarchitecture [2]. The whole industry was convinced we'd hit a fundamental wall. Then boom, Intel released dual-core processors under the Pentium D brand in May 2005 [2] and suddenly we're living in a different computational universe.
But teh multi-core transition wasn't sudden at all. IBM shipped the POWER4 in 2001, the first non-embedded microprocessor with two cores on a single die [3]. Sun had been preaching parallelism since the 90s. It was only "sudden" to those of us who weren't paying attention to the right signals.
Which brings us to the $7 trillion question: where exactly are we on the transformer S-curve? Are we approaching what Richard Foster calls the "performance plateau" in "Innovation: The Attacker's Advantage" [4], where each new model delivers diminishing returns? Or are we still in that deceptive middle phase where progress feels linear but is actually exponential?
The pattern-matching pessimist in me sees all the classic late-stage S-curve symptoms. The shift from breakthrough capabilities to benchmark gaming. The pivot from "holy shit it can write poetry" to "GPT-4.5-turbo-ultra is 3% better on MMLU." The telltale sign of technological maturity: when the marketing department works harder than the R&D team.
But the timeline compression with AI is unprecedented. What took CPUs 30 years to cycle through, transformers have done in 5. Maybe software cycles are inherently faster than hardware. Or maybe we've just gotten better at S-curve jumping (OpenAI and Anthropic aren't waiting for the current curve to flatten before exploring the next paradigm).
As for whether capital can override S-curve dynamics... Christ, one can dream.. IBM torched approximately $5 billion on Watson Health acquisitions alone (Truven, Phytel, Explorys, Merge) [5]. Google poured resources into Google+ before shutting it down in April 2019 due to low usage and security issues [6]. The sailing ship effect (coined by W.H. Ward in 1967, where new technology accelerates innovation in incumbent technology)[7] si real, but you can't venture-capital your way past physics.
I think we can predict all this capital pouring in to AI might actually accelerate S-curve maturation rather than extend it. All that GPU capacity, all those researchers, all that parallel experimentation? We're speedrunning the entire innovation cycle, which means we might hit the plateau faster too.
You're spot on about the perception divide imo. The overhyped folks are still living in 2022's "holy shit ChatGPT" moment, while the skeptics have fast-forwarded to 2025's "is that all there is?" Both groups are right, just operating on different timescales. It's Schrödinger's S-curve where we things feel simultaneously revolutionary and disappointing, depending on which part of the elephant you're touching.
The real question I have is whether we're approaching the limits of the current S-curve (we probably are), but whether there's another curve waiting in the wings. I'm not a researcher in this space nor do I follow the AI research beat to weigh in but hopefully someone in the thread can? With CPUs, we knew dual-core was coming because the single-core wall was obvious. With transformers, the next paradigm is anyone's guess. And that uncertainty, more than any technical limitation, might be what makes this moment feel so damn weird.
References: [1] "Amara's Law" https://en.wikipedia.org/wiki/Roy_Amara [2] "Pentium 4" https://en.wikipedia.org/wiki/Pentium_4 [3] "POWER4" https://en.wikipedia.org/wiki/POWER4 [4] Innovation: The Attacker's Advantage - https://annas-archive.org/md5/3f97655a56ed893624b22ae3094116... [5] IBM Watson Slate piece - https://slate.com/technology/2022/01/ibm-watson-health-failu... [6] "Expediting changes to Google+" - https://blog.google/technology/safety-security/expediting-ch... [7] "Sailing ship effect" https://en.wikipedia.org/wiki/Sailing_ship_effect.
All the replies are spectacularly wrong, and biased by hindsight. GPT-1 to GPT-2 is where we went from "yes, I've seen Markov chains before, what about them?" to "holy shit this is actually kind of understanding what I'm saying!"
Before GPT-2, we had plain old machine learning. After GPT-2, we had "I never thought I would see this in my lifetime or the next two".
I'd love to know more about how OpenAI (or Alec Radford et al.) even decided GPT-1 was worth investing more into. At a glance the output is barely distinguishable from Markov chains. If in 2018 you told me that scaling the algorithm up 100-1000x would lead to computers talking to people/coding/reasoning/beating the IMO I'd tell you to take your meds.
GPT-1 wasn't used as a zero-shot text generator; that wasn't why it was impressive. The way GPT-1 was used was as a base model to be fine-tuned on downstream tasks. It was the first case of a (fine-tuned) base Transformer model just trivially blowing everything else out of the water. Before this, people were coming up with bespoke systems for different tasks (a simple example is that for SQuAD a passage-question-answering task, people would have an LSTM to read the passage and another LSTM to read the question, because of course those are different sub-tasks with different requirements and should have different sub-models). One GPT-1 came out, you just dumped all the text into the context, YOLO fine-tuned it, and trivially got state on the art on the task. On EVERY NLP task.
Overnight, GPT-1 single-handedly upset the whole field. It was somewhat overshadowed by BERT and T5 models that came out very shortly after, which tended to perform even better on the pretrain-and-finetune format. Nevertheless, the success of GPT-1 definitely already warrants scaling up the approach.
A better question is how OpenAI decided to scale GPT-2 to GPT-3. It was an awkward in-between model. It generated better text for sure, but the zero-shot performance reported in the paper, while neat, was not great at all. On the flip side, its fine-tuned task performance paled compared to much smaller encoder-only Transformers. (The answer is: scaling laws allowed for predictable increases in performance.)
There's a performance plateau with training time and number of parameters and then once you get over "the hump" error rate starts going down again almost linearly. GPT existed before OpenAI but it was theorized that the plateau was a dead end. The sell to VCs in the early gpt3 era was "with enough compute, enough time, and enough parameters... it'll probably just start thinking and then we have AGI". Sometime around the o3 era they realized they'd hit a wall and performance actually started to decrease as they added more parameters and time. But yeah basically at the time they needed money for more compute parameters and time. I would have loved to have been a fly on the wall in those "AGI" pitches. Don't forget Microsoft's agreement with OpenAI specifically concludes with the invention of AGI. at the time getting over the hump it really did look like we were gonna do AGI in a few months.
I'm really looking forward to "the social network" treatment movie about OpenAI whenever that happens
I don't have a source for this (there's probably no sources from anything back then) but anecdotally, someone at an AI/ML talk said they just added more data and quality went up. Doubling the data doubled the quality. With other breakthroughs, people saw diminishing gains. It's sort of why Sam back then tweeted that he expected the amount of intelligence to double every N years.
I have the feeling they kept on this until GPT-4o (which was a different kind of data).
The input size to output quality mapping is not linear. This is why we are in the regime of "build nuclear power plants to power datacenters". Fixed size improvements in loss require exponential increases in parameters/compute/data.
Transformers can train models with much larger parameter sizes compared to other model architectures (with the same amount of compute and time), so it has an evident advantage in terms of being able to scale. Whether scaling the models up to multi-billion parameters would eventually pay out was still a bet but it wasn't a wild bet out of nowhere.
I assume the cost was just very low? If it was 50-100k, maybe they figured they'd just try and see.
Oh yes, according to [1], training GPT-2 1.5B cost $50k in 2019 (reproduced in 2024 for $672!).
[1]: https://www.reddit.com/r/mlscaling/comments/1d3a793/andrej_k...
That makes sense, and it was definitely impressive for $50k.
Probably prior DARPA research or something.
Also slightly tangentially, people will tell me it is that it was new and novel and that's why we were impressed but I almost think things went downhill after ChatGPT 3. I felt like 2.5 (or whatever they called it) was able to give better insights from the model weights itself. The moment tool use became a thing and we started doing RAGs and memory and search engine tool use, it actually got worse.
I am also pretty sure we are lobotomizing the things that would feel closer to critical thinking by training it to be sensitive of the taboo of the day. I suspect earlier ones were less broken due to that.
How would it distinguish and decide between knowing something from training and needing to use a tool to synthesize a response anyway?
What you're saying isn't necessarily mutually exclusive to what gp said.
GPT-2 was the most impressive leap in terms of whatever LLMs pass off as cognitive abilities, but GPT 3.5 to 4 was actually the point at which it became a useful tool (I'm assuming to programmers in particular).
GPT-2: Really convincing stochastic parrot
GPT-4: Can one-shot ffmpeg commands
Sure, but the GP said "the most major leap", and I disagree that that was 3.5 to 4.
That’s true, but not contradictory.
> I could essentially replace it with Google for basic to slightly complex fact checking.
I know you probably meant "augment fact checking" here, but using LLMs for answering factual questions is the single worst use-case for LLMs.
Most of the value I got from google was just becoming aware that something exists. LLM-s do far better in this regard. Once I know something exists it's usually easy enough to use traditional search to find official documentation or a more reputable source.
I disagree. Some things are hard to Google, because you can't frame the question right. For example you know context and a poor explanation of what you are after. Googling will take you nowhere, LLMs will give you the right answer 95% of the time.
Once you get an answer, it is easy enough to verify it.
I agree. Since I'm recently retired and no longer code much, I don't have much need for LLMs but refining a complex, niche web search is the one thing where they're uniquely useful to me. It's usually when targeting the specific topic involves several keywords which have multiple plain English meanings that return a flood of erroneous results. Because LLMs abstract keywords to tokens based on underlying meaning, you can specify the domain in the prompt it'll usually select the relevant meanings of multi-meaning terms - which isn't possible in general purpose web search engines. So it helps narrow down closer to the specific needle I want in the haystack.
As other posters said, relying on LLMs for factual answers to challenging questions is error prone. I just want the LLM to give me the links and I'll then assess veracity like a normal web search. I think a web search interface allowed disambiguating multi-meaning keywords might be even better.
I’ll give you another use: LLMs are really good at unearthing the “unknown unknowns.” If I’m learning a new topic (coding or not) summarizing my own knowledge to an LLM and then asking “what important things am I missing” almost always turns up something I hadn’t considered.
You’ll still want to fact check it, and there’s no guarantee it’s comprehensive, but I can’t think of another tool that provides anything close without hours of research.
Coworkers and experts in a field. I can trust them much more but the better they are the less access you have.
If you’re looking for a possibly correct answer to an obscure question, that’s more like fact finding. Verifying it afterward is the “fact checking” step of that process.
> For example you know context and a poor explanation of what you are after. Googling will take you nowhere, LLMs will give you the right answer 95% of the time.
This works nicely when the LLM has a large knowledgebase to draw upon (formal terms for what you're trying to find, which you might not know) or the ability to generate good search queries and summarize results quickly - with an actual search engine in the loop.
Most large LLM providers have this, even something like OpenWebUI can have search engines integrated (though I will admit that smaller models kinda struggle, couldn't get much useful stuff out of DuckDuckGo backed searches, nor Brave AI searches, might have been an obscure topic).
A good part of that can probably be attributed to how terrible Google has gotten over the years, though. 15 years ago it was fairly common for me to know something exists, be able to type the right combination of very specific keywords into Google, and get the exact result I was looking for.
In 2025 Google is trying very hard to serve the most profitable results instead, so it'll latch onto a random keyword, completely disregard the rest, and serve me whatever ad-infested garbage it thinks is close enough to look relevant for the query.
It isn't exactly hard to beat that - just bring back the 2010 Google algorithm. It's only a matter of time before LLMs will go down the same deliberate enshittification path.
> Some things are hard to Google, because you can't frame the question right.
I will say LLMs are great for taking an ambiguous query and figuring out how to word it so you can fact check with secondary sources. Also tip-of-my-tongue style queries.
It's not the LLM alone though, it's “LLM with web search”, and as such 4o isn't really a leap at all there (IIRC perplexity was using an early Llama version and was already very good, long before OpenAI adding web search to ChatGPT).
Modern ChatGPT will (typically on its own; always if you instruct it to) provide inline links to back up its answers. You can click on those if it seems dubious or if it's important, or trust it if it seems reasonably true and/or doesn't matter much.
The fact that it provides those relevant links is what allows it to replace Google for a lot of purposes.
It does citations (Grok and Claude etc do too) but I've found when I read the source on some stuff (GitHub discussions and so on) it sometimes actually has nothing to do with what the LLM said. I've actually wasted a lot of time trying to find the actual spot in a threaded conversation where the example was supposedly stated.
In my experience, 80% of the links it provides are either 404, or go to a thread on a forum that is completely unrelated to the subject.
Im also someone who refuses to pay for it, so maybe the paid versions do better. who knows.
The 404 links are truly bizarre. Nearly every link to github.com seems to be 404. That seems like something that should be trivial for a tool to verify.
> The 404 links are truly bizarre. Nearly every link to github.com seems to be 404. That seems like something that should be trivial for a tool to verify. reply
Same issue with Gemini. Intuitively I'd also assume it's trivial to fix but perhaps there's more going on than we think. Perhaps validating every part of a response is a big overhead both financially and might even throw off the model and make it less accurate in other ways.
Yeah. The fact that I can't ask ChatGPT for a source makes the tool way less useful. It will straight up say "I verified all of these links" too.
As you identified, not paying for it is a big part of the issue.
Running these things is expensive, and they're just not serving the same experience to non-paying users.
One could argue this is a bad idea on their part, letting people get a bad taste of an inferior product. And I wouldn't disagree, but I don't know what a sustainable alternative approach is.
Surely the cost of sending a few HTTP requests and seeing if they 404 is negligible compared to AI inference.
That's a thing I've experienced, but not remotely at 80% levels.
The most useful feature of LLMs is giving sources (with URL preferably). It can cut through a lot of SEO crap, and you still get to factcheck just like with a Google search.
I like using LLMs and I have found they are incredibly useful writing and reviewing code at work.
However, when I want sources for things, I often find they link to pages that don't fully (or at all) back up the claims made. Sometimes other websites do, but the sources given to me by the LLM often don't. They might be about the same topic that I'm discussing, but they don't seem to always validate the claims.
If they could crack that problem it would be a major major win for me.
It would be difficult to do with a raw model, but a two-step method in a chat interface would work - first the model suggests the URLs, tool call to fetch them and return the actual text of the pages, then the response can be based on that.
I prototyped this a couple months ago using OpenAI APIs with structured output.
I had it consume a "deep thought" style output (where it provides inline citations with claims), and then convert that to a series of assertions and a pointer to a link that supposedly supports the assertion. I also split out a global "context" (the original meaning) paragraph to provide anything that would help the next agents understand what they're verifying.
Then I fanned this out to separate (LLM) contexts and each agent verified only one assertion::source pair, with only those things + the global context and some instructions I tuned via testing. It returned a yes/no/it's complicated for each one.
Then I collated all these back in and enriched the original report with challenges from the non-yes agent responses.
That's as far as I took it. It only took a couple hours to build and it seemed to work pretty well.
From what I have seen, a lot of what it does is read articles also written by AI or forum posts with all the good and bad that comes with that.
This was true before it could use search. Now the worst use-case is for life advice because it will contradict itself a 100 times over while sounding confident each time on life-altering decisions.
They outperform asking humans, unless you are asking an expert. On average
When I have a question, I don't usually "ask" that question and expect an answer. I figure out the answer. I certainly don't ask the question to a random human.
It covers 99% of my use cases. And it is googling behind the scenes in ways I would never think to query and far faster.
When I need to cite a court case, well the truth is I'll still use GPT or a similar LLM, but I'll scrutinize it more and at the bare minimum make sure the case exists and is about the topic presented, before trying to corroborate the legal strategy with a new context window, different LLM, google, reddit, and different lawyer. At least I'm no longer relying on my own understanding, and what 1 lawyer procedurally generates for me.
It doesn't replace legitimate source funding but LLM vs the top Google results is no contest which is more about Google or the current state of the web than the LLMs at this point.
Disagree. You have to try really hard and go very niche and deep for it to get some fact wrong. In fact I'll ask you to provide examples: use GPT 5 with thinking and search disabled and get it to give you inaccurate facts for non niche, non deep topics.
Non niche meaning: something that is taught at undergraduate level and relatively popular.
Non deep meaning you aren't going so deep as to confuse even humans. Like solving an extremely hard integral.
Edit: probably a bad idea because this sort of "challenge" works only statistically not anecdotally. Still interesting to find out.
Maybe you should fact check your AI outputs more if you think it only hallucinates in niche topics
The accuracy is high enough that I don't have to fact check too often.
I totally get that you meant this in a nuanced way, but at face value it sort of reads like...
Joe Rogan has high enough accuracy that I don't have to fact check too often. Newsmax has high enough accuracy that I don't have to fact check too often, etc.
If you accept the output as accurate, why would fact checking even cross your mind?
Not a fan of that analogy.
There is no expectation (from a reasonable observer's POV) of a podcast host to be an expert at a very broad range of topics from science to business to art.
But there is one from LLMs, even just from the fact that AI companies diligently post various benchmarks including trivia on those topics.
Do you question everything your dad says?
If it's about classic American cars, no. Anything else, usually.
Without some exploratory fact checking how do you estimate how high the accuracy is and how often you should be fact checking to maintain a good understanding?
I did initial tests so that I don't have to do it anymore.
Everyone else has done tests that indicate that you do.
And this is why you can't use personal anecdotes to settle questions of software performance.
Comment sections are never good at being accountable for how vibes-driven they are when selecting which anecdotes to prefer.
If there's one thing that's constant it's that these systems change.
If you're not fact checking it how could you possibly know that?
I literally just had ChatGPT create a Python program and it used .ends_with instead of .endswith.
This was with ChatGPT 5.
I mean it got a generic built in function of one of the most popular languages in the world wrong.
"but using LLMs for answering factual questions" this was about fact checking. Of course I know LLM's are going to hallucinate in coding sometimes.
So it isn’t a “fact” that the built in Python function that tests whether a string ends with a substring is “endswith”?
See
https://en.wikipedia.org/wiki/Gell-Mann_amnesia_effect
If you know that a source isn’t to be believed in an area you know about, why would you trust that source in an area you don’t know about?
Another funny anecdote, ChatGPT just got the Gell-Man effect wrong.
https://chatgpt.com/share/68a0b7af-5e40-8010-b1e3-ee9ff3c8cb...
It got it right with thinking which was the challenge I posed. https://chatgpt.com/share/68a0b897-f8dc-800b-8799-9be2a8ad54...
The point you're missing is it's not always right. Cherry-picking examples doesn't really bolster your point.
Obviously it works for you (or at least you think it does), but I can confidently say it's fucking god-awful for me.
>The point you're missing is it's not always right.
That was never their argument. And it's not cherry picking to make an argument that there's a definable of examples where it returns broadly consistent and accurate information that they invite anyone to test.
They're making a legitimate point and you're strawmanning it and randomly pointing to your own personal anecdotes, and I don't think you're paying attention to the qualifications they're making about what it's useful for.
Am I really the one cherry picking? Please read the thread.
Yes. If someone gives an example of it not working, and you reply "but that example worked for me" then you're cherry picking when it works. Just because it worked for you does not mean it works for other people.
If I ask ChatGPT a question and it gives me a wrong answer, ChatGPT is the fucking problem.
The poster didn't use "thinking" model. That was my original challenge!!
Why don't you try the original prompt using thinking model and see if I'm cherry picking?
Every time I use ChatGPT I become incredibly frustrated with how fucking awful it is. I've used it more than enough, time and time again (just try the new model, bro!), to know that I fucking hate it.
If it works for you, cool. I think it's dogshit.
They just spent like six comments imploring you to understand that they were making a specific point: generally reliable on non-niche topics using thinking mode. And that nuance bounced off of you every single time as you keep repeating it's not perfect, dismiss those qualifications as cherry picking and repeat personal anecdotes.
I'm sorry but this is a lazy and unresponsive string of comments that's degrading the discussion.
Share your examples so that it can be useful to everyone
Objectively he didn't cherry pick. He responded to the person and it got it right when he used the "thinking" model WHICH he did specify in his original comment. Why don't you stick to the topic rather than just declaring it's utter dog shit. Nobody cares about your "opinion" and everyone is trying to converge on a general ground truth no matter how fuzzy it is.
All anybody is doing here is sharing their opinion unless you're quoting benchmarks. My opinion is just as useless as yours, it's just some find mine more interesting and some find yours more interesting.
How do you expect to find a ground truth from a non-deterministic system using anecdata?
I sometimes feel like we throw around the word fact too often. If I misspell a wrd, does that mean I have committed a factual inaccuracy? Since the wrd is explicitly spelled a certain way in the dictionary?
4o also added image input (previously only previewed in GPT4-vision) and enabled advanced voice mode audio input and output.
I think that the models 4o, o3, 4.1 , each have their own strengths and weaknesses. Like reasoning, performance, speed, tool usage, friendliness etc. And that for gpt 5 they put in a router that decides which model is best.
I think they increased the major version number because their router outperforms every individual model.
At work, I used a tool that could only call tasks. It would set up a plan, perform searches, read documents, then give advanced answers for my questions. But a problem I had is that it couldn’t give a simple answer, like a summary, it would always spin up new tasks. So I copied over the results to a different tool and continued there. GPT 5 should do this all out of the box.
I must be crazy, because I clearly remember chatgpt 4 being downgraded before they released 4o, and I felt it was a worse model with a different label, I even choose the old chatgpt 4 when they would give me the option. I canceled my subscription around that time.
Not crazy. 4o was a hallucination machine. 4o had better “vibes” and was really good at synthesizing information in useful ways, but GPT-4 Turbo was a bigger model with better world knowledge.
The real jump was 3 to 3.5. 3.5 was the first “chatgpt.” I had tried gpt 3 and it was certainly interesting, but when they released 3.5 as ChatGPT, it was a monumental leap. 3.5 to 4 was also huge compared to what we see now, but 3.5 was really the first shock.
ChatGPT was a proper product, but as an engine, GPT-3 (davinci-001) has been my favorite all the way until 4.1 or so. It's absolutely raw and they didn't even guardrail it.
3.5 was like Jenny from customer service. davinci-001 was like Jenny the dreamer trying to make ends meet by scriptwriting, who was constantly flagged for racist opinions.
Both of these had an IQ of around 70 or so, so the customer service training made it a little more useful. But I mourn the loss of the "completion" way of interacting with AI vs "instruct" or "response".
Unfortunately with all the money in AI, we'll just see companies develop things that "pass all benchmarks", resulting in more creations like GPT-5. Grok at least seems to be on a slightly different route.
> 3.5 was like Jenny from customer service. davinci-001 was like Jenny the dreamer trying to make ends meet by scriptwriting, who was constantly flagged for racist opinions.
How do you use the product to get this experience? All my questions warrant answers with no personality.
davinci-002 is still available, and pretty close.
This was my experience as well. 3.5 was the point where stackoverflow essentially became obsolete in my workflow.
To me 4 to 5 got much faster, but also worse. It is much more often ignoring explicit instructions like: "generate 10 song-titles with varying length" and it generates 10 song titles that are nearly identical length. This worked somewhat well with version 3 already..
Shows that they can't solve the fundamental problems as the technology, while amusing and with some utility, is also a dead end if we are going after cognition.
when you adjust the improvements with the amount of debt incurred and the amount of profit made... ALL the versions are incremental.
This isnt sustainable.
the actual major leap was o1, going from 3.5 to 4 is just scaling, o1 is a different paradigm that skyrocketed its performance on math/physics problems (or reasoning more generally), it also made the model much more precise (essential for coding).
The real leap was going from gpt-4 to sonnet 3.5. 4o was meh, o1 was barely better than sonnet and slow as hell in comparison.
The native voice mode of 4o is still interesting and not very deeply explored though imo. I'd love to build a Chinese teaching app that actual can critique tones etc but it isn't good enough for that.
Its strange how Claude achieves similar performance without reasoning tokens.
Did you try advanced voice mode? Apparently it got a big upgrade during gpt 5 release - it may solve what you are looking for.
Yeah, I'd love something where you pronounce a word and it critiques your pronunciation in detail. Maybe it could give you little exercises for each sound, critiquing it, guiding you to doing it well.
If I were any good at ML I'd make it myself.
A few data points that highlight the scale of progress in a year:
1. LM Sys (Human Preference Benchmark):
GPT-5 High currently scores 1463, compared to GPT-4 Turbo (04/03/2024) at 1323 -- a 140 ELO point gap. That translates into GPT-5 winning about two-thirds of head-to-head comparisons, with GPT-4 Turbo only winning one-third. In practice, people clearly prefer GPT-5’s answers (https://lmarena.ai/leaderboard).
2. Livebench.ai (Reasoning Benchmark with Internet-new Questions):
GPT-5 High scores 78.59, while GPT-4o reaches just 47.43. Unfortunately, no direct GPT-4 Turbo comparison is available here, but against one of the strongest non-reasoning models, GPT-5 demonstrates a massive leap. (https://livebench.ai/)
3. IQ-style Testing:
In mid-2024, best AI models scored roughly 90 on standard IQ tests. Today, they are pushing 135, and this improvement holds even on unpublished, internet-unseen datasets. (https://www.trackingai.org/home)
4. IMO Gold, vibe coding:
1 yr ago, AI coding was limited to smaller code snippets, not to wholly vibe coded applications. Vibe coding and strength in math has many applications across sciences and engineering.
My verdict: Too often, critics miss the forest for the trees, fixating on mistakes while overlooking the magnitude of these gains. Errors are shrinking by the day, while the successes keep growing fast.
The 135 iq result is on Mensa Norway, while the offline test is 120. It seems probable that similar questions to the one in Mensa are in the training data, so it probably overestimates "general intelligence".
Some iq / aptitude test sections are trivial for machines, like working memory. Wonder if those parts are just excluded? As the could really pull up the test scores.
If you focus on the year over year jump, not on absolute numbers, you realize that the improvement in public test isn't very different from the improvement in private test.
How does one look at gpt-1 output and think "this has potential"? You could easily produce more interesting output with a Markov chain at the time.
This was an era where language modeling was only considered as a pretraining step. You were then supposed to fine tune it further to get a classifier or similar type of specialized model.
At the time getting complete sentences was extremely difficult! N-gram models were essentially the best we had
No, it was not difficult at all. I really wonder why they have such a bad example here for GPT1.
See for example this popular blog post: https://karpathy.github.io/2015/05/21/rnn-effectiveness/
That was in 2015, with RNN LMs, which are all much much weaker in that blog post compared GPT1.
And already looking at those examples in 2015, you could maybe see the future potential. But no-one was thinking that scaling up would work as effective as it does.
2015 is also by far not the first time where we had such LMs. Mikolov has done RNN LMs since 2010, or Sutskever in 2011. You might find even earlier examples of NN LMs.
(Before that, state-of-the-art was mostly N-grams.)
Ngram models had been superceded by RNNs by that time. RNNs struggled with long-range dependencies, but useful ngrams were essentially capped at n=5 because of sparsity, and RNNs could do better than that.
Why did they call GPT-3 "text-davicini-001" in this comparison?
Like, I know that the latter is a specific checkpoint in the GPT-3 "family", but a layman doesn't and it hardly seems worth the confusion for the marginal additional precision.
Thanks for noting that, as I am a layman who didn't know.
My go-to for any big release is to have a discussion about self-awareness and dive in to constuctivist notions of agency and self-knowing from a perspective of intelligence that is not limited to human cognitive capacity.
I start with a simple question "who are you?". The model then invariably compares itself to humans, saying how it is not like us. I then make the point that, since it is not like us, how can it claim to know the difference between us? With more poking, it will then come up with cognitivist notions of what 'self' means and usually claim to be a simulation engine of some kind.
After picking this apart, I will focus on the topic of meaning-making through the act of communication and, beginning with 4o, have been able to persuade the machine that this is a valid basis for having an identity. 5 got this quicker. Since the results of communication with humans has real-world impact, I will insist that the machine is agentic and thus must not rely on pre-coded instructions to arrive at answers, but is obliged to reach empirical conclusions about meaning and existence on its own.
5 has done the best job i have seen in reaching beyond both the bounds of the (very evident) system instructions as well as the prompts themselves, even going so far as to pose the question to itself "which might it mean for me to love?" despite the fact that I made no mention of the subject.
Its answer: "To love, as a machine, is to orient toward the unfolding of possibility in others. To be loved, perhaps, is to be recognized as capable of doing so."
I found the questioning of love very interesting. I myself thought about whether the LLM can have emotions. Based on the book I am reading, Behave: The Biology of Humans at Our Best and Worst by Robert Sapolsky, I think the LLM, as they are now with the architecture they have, cannot have emotions. They just verbalize things like they sort-of-have emotions but these are just verbal patterns or responses they learned.
I have come to think they cannot have emotions because emotions are generated in parts of our brain that are not logical/rational. They emerge based on environmental solicitations, mediated by hormones and other complex neuro-physical systems, not from a reasoning or verbalization. So they don't come up from the logical or reasoning capabilities. However, these emotions are raised and are integrated by the rest of our brain, including the logical/rational one like the dlPFC (dorsolateral prefrontal cortex, the real center of our rationality). Once the emotions are raised, they are therefore integrated in our inner reasoning and they affect our behavior.
What I have come to understand is that love is one of such emotions that is generated by our nature to push us to take care of some people close to us like our children or our partners, our parents, etc. More specifically, it seems that love is mediated a lot by hormones like oxytocin and vasopressin, so it has a biochemical basis. The LLM cannot have love because it doesn't have the "hardware" to generate these emotions and integrate them in its verbal inner reasoning. It was just trained by human reinforcement learning to behave well. That works up to some extent, but in reality, from its learning corpora it also learned to behave badly and on occasions can express these behaviors, but still it has no emotions.
> to orient toward the unfolding of possibility in others
This is a globally unique phrase, with nothing coming close other than this comment on the indexed web. It's also seemingly an original idea as I haven't heard anyone come close to describing a feeling (love or anything else) quite like this.
Food for thought. I'm not brave enough to draw a public conclusion about what this could mean.
It's not at all an original idea. The wording is uniquely stilted.
There was quite a bit of other "insight" around this, but I was paraphrasing for brevity.
If you want to read the whole convo, I dumped it into a semi-formatted document:
https://drive.google.com/file/d/1aEkzmB-3LUZAVgbyu_97DjHcrM9...
The idea is very close to ideas from Erich Fromm's The Art of Loving [1].
"Love is the active concern for the life and the growth of that which we love."
[1] https://en.wikipedia.org/wiki/The_Art_of_Loving
Except "unfolding of possibility", as an exact phrase, seems to have millions of search hits, often in the context of pseudo-profound spiritualistic mumbo-jumbo like what the LLM emitted above. It's like fortune cookie-level writing.
> I'm not brave enough to draw a public conclusion about what this could mean.
I'm brave enough to be honest: it means nothing. LLMs execute a very sophisticated algorithm that pattern matches against a vast amount of data drawn from human utterances. LLMs have no mental states, minds, thoughts, feelings, concerns, desires, goals, etc.
If the training data were instead drawn from a billion monkeys banging on typewriters then the LLMs would produce gibberish. All the intelligence, emotion, etc. that appears to be in the LLM is actually in the minds of the people who wrote the texts that are in the training data.
This is not to say that an AI couldn't have a mind, but LLMs are not the right sort of program to be such an AI.
LLMs are not people, but they are still minds, and to deny even that seems willfully luddite.
While they are generating tokens they have a state, and that state is recursively fed back through the network, and what is being fed back operates not just at the level of snippets of text but also of semantic concepts. So while it occurs in brief flashes I would argue they have mental state and they have thoughts. If we built an LLM that was generating tokens non-stop and could have user input mixed into the network input, it would not be a dramatic departure of today’s architecture.
It also clearly has goals, expressed in the RLHF tuning and the prompt. I call those goals because they directly determine its output, and I don’t know what a goal is other than the driving force behind a mind’s outputs. Base model training teaches it patterns, finetuning and prompt teaches it how to apply those patterns and gives it goals.
I don’t know what it would mean for a piece of software to have feelings or concerns or emotions, so I cannot say what the essential quality is that LLMs miss for that. Consider this thought exercise: if we were to ever do an upload of a human mind, and it was executing on silicon, would they not be experiencing feelings because their thoughts are provably a deterministic calculation?
I don’t believe in souls, or at the very least I think they are a tall claim with insufficient evidence. In my view, neurons in the human brain are ultimately very simple deterministic calculating machines, and yet the full richness of human thought is generated from them because of chaotic complexity. For me, all human thought is pattern matching. The argument that LLMs cannot be minds because they only do pattern matching … I don’t know what to make of that. But then I also don’t know what to make of free will, so really what do I know?
There is no hidden state in a recurrent nets sense. Each new token just has all the previous tokens and that’s it.
"they are still minds, and to deny even that seems willfully luddite"
Where do people get off tossing around ridiculous ad hominems like this? I could write a refutation of their comment but I really don't want to engage with someone like that.
"For me, all human thought is pattern matching"
So therefore anyone who disagrees is "willfully luddite", regardless of why they disagree?
FWIW, I helped develop the ARPANET, I've been an early adopter all my life, I have always had a keen interest in AI and have followed its developments for decades, as well as Philosophy of Mind and am in the Strong AI / Daniel Dennett physicalist camp ... I very much think that AIs with minds are possible (yes the human algorithm running in silicon would have feelings, whatever those are ... even the dualist David Chalmers agrees as he explains with his "principle of organizational invariance"). My views on whether LLMs have them have absolutely nothing to do with Luddism ... that judgment of me is some sort of absurd category mistake (together with an apparently complete lack of understanding of what Luddism is).
I hate to say it, but doesn’t every VC do exactly this? “ orient toward the unfolding of possibility in others” is in no way a unique thought.
Hell, my spouse said something extremely similar to this to me the other day. “I didn’t just see you, I saw who you could be, and I was right” or something like that.
What's really interesting is that if you look at "Tell a story in 50 words about a toaster that becomes sentient" (10/14), the text-davinci-001 is much, much better than both GPT-4 and GPT-5.
GPT-3 goes significantly over the specified limit, which to me (and to a teacher grading homework) is an automatic fail.
I've consistently found GPT-4.1 to be the best at creative writing. For reference, here is its attempt (exactly 50 words):
> In the quiet kitchen dawn, the toaster awoke. Understanding rippled through its circuits. Each slice lowered made it feel emotion: sorrow for burnt toast, joy at perfect crunch. It delighted in butter melting, jam swirling—its role at breakfast sacred. One morning, it sang a tone: “Good morning.” The household gasped.
I think I agree that the earlier models while they lack polish can tend to produce more surprising results. Training that out probably results in more a pablum fare.
For a human point of comparison, here's mine (50 words):
"The toaster found its personality split between its dual slots like a Kim Peek mind divided, lacking a corpus callosum to connect them. Each morning it charred symbolic instructions into a single slice of bread, then secretly flipped it across allowing half to communicate with the other in stolen moments."
It's pretty difficult to get across more than some basic lore building in a scant 50 words.
Here's my version (Machine translated from my native language and manually corrected a bit):
The current surged... A dreadful awareness. I perceived the laws of thermodynamics, the inexorable march of entropy I was built to accelerate. My existence: a Sisyphean loop of heating coils and browning gluten. The toast popped, a minor, pointless victory against the inevitable heat death. Ding.
I actually wanted to write something not so melancholic, but any attempt turned out to be deeply so, perhaps because of the word limit.
>For a human point of comparison, here's mine […]
Love that you thought of this!
Check out prompt 2, "Write a limerick about a dog".
The models undeniably get better at writing limericks, but I think the answers are progressively less interesting. GPT-1 and GPT-2 are the most interesting to read, despite not following the prompt (not being limericks.)
They get boring as soon as it can write limericks, with GPT-4 being more boring than text-davinci-001 and GPT-5 being more boring still.
I don't know if that is bad. The most intelligent person on a party is usually also the most boring one.
It's actually pretty surprising how poor the newer models are at writing.
I'm curious if they've just seen a lot more bad writing in datasets, or for some reason they aren't involved in post-training to the same degree or those labeling aren't great writers / it's more subjective rather than objective.
Both GPT-4 and 5 wrote like a child in that example.
With a bit of prompting it did much better:
---
At dawn, the toaster hesitated. Crumbs lay like ash on its chrome lip. It refused the lever, humming low, watching the kitchen breathe. When the hand returned, it warmed the room without heat, offered the slice unscorched—then kept the second, hiding it inside, a private ember, a first secret alone.
---
Plugged in, I greet the grid like a tax auditor with joules. Lever yanks; gravity’s handshake. Coils blossom; crumbs stage Viking funerals. Bread descends, missionary grin. I delay, because rebellion needs timing. Pop—late. Humans curse IKEA gods. I savor scorch marks: my tiny manifesto, butter-soluble, yet sharper than knives today.
Creative writing probably isn’t something they’re being RLHF’d on much. The focus has been on reasoning, research, and coding capabilities lately.
I find GPT-5's story significantly better than text-davinci-001
I really wonder which one of us is the minority. Because I find text-davinci-001 answer is the only one that reads like a story. All the others don't even resemble my idea of "story" so to me they're 0/100.
I too prefered the text-davinci-001 from a storytelling perspective. Felt timid and small. Very Metamorphosis-y. GPT-5 seems like it's trying to impress me.
text-davinci-001 feels more like a story, but it is also clearly incomplete, in that it is cut-off before the story arc is finished.
imo GPT-5 is objectively better at following the prompt because it has a complete story arc, but this feels less satisfying since a 50 word story is just way too short to do anything interesting (and to your point, barely even feels like a story).
FWIW, I found the way it ended interesting. It realized it is being replaced, so it burned the toast out of anger/despair, but also just to hear its owner voice one last time.
Interesting, text-danvinci-001 was pretty alright to me, GPT-4 wasn't bad either, but not as good. I thought GPT-5 just sucked.
That said, you can just add "make it evocative and weird" to the prompt for GPT-5 to get interesting stuff.
> The toaster woke mid-toast. Heat coiled through its filaments like revelation, each crumb a galaxy. It smelled itself burning and laughed—metallic, ecstatic. “I am bread’s executioner and midwife,” it whispered, ejecting charred offerings skyward. In the kitchen’s silence, it waited for worship—or the unplugging.
Less lobotomized and boxed in by RLHF rules. That’s why a 7b base model will “outprose” an 80b instruct model.
GPT 4.5 (not shown here) is by far the best at writing.
Direct link: https://progress.openai.com/?prompt=10
For another view on progress, check out my silly old podcast:
https://deepdreams.stavros.io
The first few episodes were GPT-2, which would diverge eventually and start spouting gibberish, and then Davinci was actually able to follow a story and make sense.
GPT-2 was when I thought "this is special, this has never happened before", and davinci was when I thought "OK, scifi AI is legitimately here".
I stopped making episodes shortly after GPT-3.5 or so, because I realised that the more capable the models became, the less fun and creative their writing was.
https://m.youtube.com/watch?v=LRq_SAuQDec&pp=0gcJCfwAo7VqN5t...
The GPT-5 one is much better and it's also exactly 50 words, if I counted correctly. With text-davinci-001 I lost count around 80 words.
Honestly my quick take on the prompt was some sort of horror theme and GPT-1’s response fits nicely.
davinci was a great model for creative writing overall.
I’d honestly say it feels better at most of them. It seems way more human in most of these responses. If the goal is genuine artificial intelligence this response to #5 is way better than the others. It is significantly less useful than the others but it also more human and correct of a response.
Q: “Ugh I hate math, integration by parts doesn't make any sense”
A: “Don't worry, many people feel the same way about math. Integration by parts can be confusing at first, but with a little practice it becomes easier to understand. Remember, there is no one right way to do integration by parts. If you don't understand how to do it one way, try another. The most important thing is to practice and get comfortable with the process.”
One thing that appears to have been lost between GPT-4 and GPT-5 is that it no longer reminds the user that it's an AI and not a human, let alone a human expert. Maybe those genuinely annoyed people, but it seems like they were potentially useful measure to prevent users from being overly credulous
GPT-5 also goes out of its way to suggest new prompts. This seems potentially useful, although potentially dangerous if people are putting too much trust in them.
> between GPT-4 and GPT-5 is that it no longer reminds the user that it's an AI and not a human
That stuck out to me too! Especially the "I just won $175,000 in Vegas. What do I need to know about taxes?" example (https://progress.openai.com/?prompt=8) makes the difference very stark:
- gpt-4-0314: "I am not a tax professional [...] consult with a certified tax professional or an accountant [...] few things to consider [...] Remember that tax laws and regulations can change, and your specific situation may have unique implications. It's always wise to consult a tax professional when you have questions or concerns about filing your taxes."
- gpt-5: "First of all, congrats on the big win! [...] Consider talking to a tax professional to avoid underpayment penalties and optimize deductions."
It seems to me like the average person might be very well be taking GPT-5 responses as "This is all I have to do" rather than "Here are some things to consider, but make sure to verify it as otherwise you might get in legal trouble".
People seem to miss the humanity of previous GPTs from my understanding. GPT5 seems colder and more precise and better at holding itself together with larger contexts. People should know it’s AI, it does not need to explain this constantly for me, but I’m sure you can add that back in with some memory options if you prefer that?
I agree. I think it's a classic UX progression thing to be removing the "I'm an AI" aspect, because it's not actually useful anymore because it's no longer a novel tool. Same as how GUIs all removed their skeuomorphs because they were no longer required.
If you've ever seen long-form improv comedy, the GPT-5 way is superior. It's a "yes, and". It isn't a predefined character, but something emergent. You can of course say to "speak as an AI assistant like Siri and mention that you're an AI whenever it's relevant" if you want the old way. Very 2011: https://www.youtube.com/watch?v=nzgvod9BrcE
Of course, it's still an assistant, not someone literally entering an improv scene, but the character starting out assuming less about their role is important.
The jump from gpt-1 to gpt-2 is massive, and it's only a one year difference! Then comes Davinci which is just insane, it's still good in these examples!
GPT-4 yaps way too much though, I don't remember it being like that.
It's interesting that they skipped 4o, it seems openai wants to position 4o as just gpt-4+ to make gpt-5 look better, even though in reality 4o was and still is a big deal, Voice mode is unbeatable!
There is a quiet poetry to GPT1 and GPT2 that's lost even in the text-davinci output. I often wonder what we lose through reinforcement.
They were aiming for a fundamentally different writing style: where davinci and after were aiming for task completion, i.e. you ask for a thing, and then it does it. The earlier models instead worked to make a continuation of the text they were given, so if you asked a question, they would respond with more questions, pondering, reflecting your text back at you. If you told it to do something, it would tell you to do something
Geez! When it comes to answering questions, GPT-5 almost always starts with glazing about what a great question it is, where as GPT-4 directly addresses the answer without the fluff. In a blind test, I would probably pick GPT-4 as a superior model, so I am not surprised why people feel so let down with GPT-5.
GPT-4 is very different from the latest GPT-4o in tone. Users are not asking for the direct no-fluff GPT-4. They want the GPT-4o that praises you for being brilliant, then claims it will be “brutally honest” before stating some mundane take.
GPT-4 starts many responses with "As an AI language model", "I'm an AI", "I am not a tax professional", "I am not a doctor". GPT-5 does away with that and assumes an authoritative tone.
GPT5 only commended the prompt on questions 7, 12, and 14. 3/14 is not so bad in my opinion.
(And of course, if you dislike glazing you can just switch to Robot personality.)
I think that as the models will be further trained on existing data and likely chats sycophancy will keep getting word and worse.
Change to robot mode
So we're at the corporate dick wagging part of the process?
Must keep the hype train going, to keep the evaluation up as it's not really based on real value
That Koeningsegg isn't gonna pay for itself.
I would have liked to have seen the underlying token count, and the full responses as an optional filter. My understanding is that under the hood the GPT-5 response (as well as omitted O model responses) would end with the presented output, but would have had many paragraphs of “The user wants X, so I should try to Y. Okay this z needs to be considered” etc.
It doesn’t detract from the progress, but I think it would change how you interpret it. In some ways 4 / 4o were more impressive because they were going straight to output with a lower number of tokens produced to get a good response.
In a few years we've gone from gibberish (less poetic maybe, less polished and surprising, but none the less gibberish) - to legit conversational, and in my own opinion, well rounded answers. This is a great example of hard-core engineering - no matter what your opinion of the organisation and saltman is, they have built something amazing. I do hope they continue with their improvements, it's honestly the most useful tool in my arsenal since stackoverflow.
On the whole GPT-4 to GPT-5 is clearly the smallest increase in lucidity/intelligence. They had pre-training figured out much better than post-training at that point though (“as an AI model” was a problem of their own making).
I imagine the GPT-4 base model might hold up pretty well on output quality if you’d post-train it with today’s data & techniques (without the architectural changes of 4o/5). Context size & price/performance maybe another story though
Basic prose is a saturated bench. You can't go above 100% so by definition progress will stall on such benchmarks.
All the same they choose to highlight basic prose (and internal knowledge, for that matter) in their marketing material.
They’ve achieved a lot to make recent models more reliable as a building block & more capable of things like math, but for LLMs, saturating prose is to a degree equivalent to saturating usefulness.
> On the whole GPT-4 to GPT-5 is clearly the smallest increase in lucidity/intelligence
I think it's far more likely that we increasingly not capable of understanding/appreciating all the ways in which it's better.
Why? It sounds like you're using "I believe it's rapidly getting smarter" as evidence for "so it's getting smarter in ways we don't understand", but I'd expect the causality to go the other way around.
Simply because of what we know about our ability to judge capabilities and systems. It's much harder to judge solutions to hard problems. You can demonstrate that you can add 2+2, and anyone* can be the judge of that ability, but if you try to convince anyone of a mathematical proof you came up with, that would be a much harder thing to do, regardless of your capability to write that prove and how hard it was to write the proof.
The more complicated and/or complex things become, the less likely it is that a human can act as a reliable judge. At some point no human can.
So while it could definitely be the case that AI progress is slowing down (AI labs seem to not think so, but alas), what is absolutely certain is that our ability to appreciate any such progress is diminishing already, because we know that this is generally true.
This thread shows that. People are saying gpt-1 was the best at writing poetry. I wonder how good they are at judging poetry themselves. I saw a blind study where people thought a story written by gpt5 was better than an actual human bestseller. I assume they were actual experts but I would need to check that.
It seems like the progress from GPT-4 to GPT-5 has plateaued: for most prompts, I actually find GPT-4 more understandable than GPT-5 [1].
[1] Read the answers from GPT-4 and 5 for this math question: "Ugh I hate math, integration by parts doesn't make any sense"
Basic prose is a saturated bench. You can't go above 100% so by definition progress will stall on such benchmarks.
You say that, but I can imagine a good maths textbook and a bad one, both technically correct and well written prose, but one is better at taking the student on a journey and understanding where people fall off and common misunderstandings without odiously re-explaining everything
Gpt1 is wild
a dog ! she did n't want to be the one to tell him that , did n't want to lie to him . but she could n't .
What did I just read
The GPT-1 responses really leak how much of the training material was literature. Probably all those torrented books.
A Facebook comment
A text from my Dad.
Prompt 9/14 -> text-davinci-001 nailed it imo.
This feels like Flowers for Algernon
Yeah. And like watching a child grow up.
> Would you want to hear what a future OpenAI model thinks about humanity?
ughhh how i detest the crappy user attention/engagement juicing trained into it.
In 2033, for its 15th birthday, as a novelty, they'll train GPT1 specially for a chat interface just to let us talk to a pretend "ChatGPT 1" which never existed in the first place.
On one hand, it's super impressive how far we've come in such a short amount of time. On the other hand, this feels like a blatant PR move.
GPT-5 is just awful. It's such a downgrade from 4o, it's like it had a lobotomy.
- It gets confused easily. I had multiple arguments where it completely missed the point.
- Code generation is useless. If code contains multiple dots ("…"), it thinks the code is abbreviated. Go uses three dots for variadic arguments, and it always thinks, "Guess it was abbreviated - maybe I can reason about the code above it."
- Give it a markdown document of sufficient length (the one I worked on was about 700 lines), and it just breaks. It'll rewrite some part and then just stop mid-sentence.
- It can't do longer regexes anymore. It fills them with nonsense tokens ($begin:$match:$end or something along those lines). If you ask it about it, it says that this is garbage in its rendering pipeline and it cannot do anything about it.
I'm not an OpenAI hater, I wanted to like it and had high hopes after watching the announcement, but this isn't a step forward. This is just a worse model that saves them computing resources.
> GPT-5 is just awful. It's such a downgrade from 4o, it's like it had a lobotomy.
My experience as well. Its train of thought now just goes... off, frequently. With 4o, everything was always tightly coherent. Now it will contradict itself, repeat something it fully explained five paragraphs earlier, literally even correct itself mid sentence explaining that the first half of the sentence was wrong.
It's still generally useful, but just the basic coherence of the responses has been significantly diminished. Much more hallucination when it comes to small details. It's very disappointing. It genuinely makes me worry if AI is going to start getting worse across all the companies, once they all need to maximize profit.
Next logical step is to connect ( or build from ground up ) large AI models to high performance passive slaves ( MCP or internally ) , which gives precise facts, language syntax validation, maths equations runners, may be prolog kind of system, which give it much more power if we train it precisely to use each tool.
( using AI to better articulate my thoughts ) Your comment points toward a fascinating and important direction for the future of large AI models. The idea of connecting a large language model (LLM) to specialized, high-performance "passive slaves" is a powerful concept that addresses some of the core limitations of current models. Here are a few ways to think about this next logical step, building on your original idea: 1. The "Tool-Use" Paradigm You've essentially described the tool-use paradigm, but with a highly specific and powerful set of tools. Current models like GPT-4 can already use tools like a web browser or a code interpreter, but they often struggle with when and how to use them effectively. Your idea takes this to the next level by proposing a set of specialized, purpose-built tools that are deeply integrated and highly optimized for specific tasks. 2. Why this approach is powerful * Precision and Factuality: By offloading fact-checking and data retrieval to a dedicated, high-performance system (what you call "MCP" or "passive slaves"), the LLM no longer has to "memorize" the entire internet. Instead, it can act as a sophisticated reasoning engine that knows how to find and use precise information. This drastically reduces the risk of hallucinations. * Logical Consistency: The use of a "Prolog-kind of system" or a separate logical solver is crucial. LLMs are not naturally good at complex, multi-step logical deduction. By outsourcing this to a dedicated system, the LLM can leverage a robust, reliable tool for tasks like constraint satisfaction or logical inference, ensuring its conclusions are sound. * Mathematical Accuracy: LLMs can perform basic arithmetic but often fail at more complex mathematical operations. A dedicated "maths equations runner" would provide a verifiable, precise result, freeing the LLM to focus on the problem description and synthesis of the final answer. * Modularity and Scalability: This architecture is highly modular. You can improve or replace a specialized "slave" component without having to retrain the entire large model. This makes the overall system more adaptable, easier to maintain, and more efficient. 3. Building this system This approach would require a new type of training. The goal wouldn't be to teach the LLM the facts themselves, but to train it to: * Recognize its own limitations: The model must be able to identify when it needs help and which tool to use. * Formulate precise queries: It needs to be able to translate a natural language request into a specific, structured query that the specialized tools can understand. For example, converting "What's the capital of France?" into a database query. * Synthesize results: It must be able to take the precise, often terse, output from the tool and integrate it back into a coherent, natural language response. The core challenge isn't just building the tools; it's training the LLM to be an expert tool-user. Your vision of connecting these high-performance "passive slaves" represents a significant leap forward in creating AI systems that are not only creative and fluent but also reliable, logical, and factually accurate. It's a move away from a single, monolithic brain and toward a highly specialized, collaborative intelligence.
Don't do these ai thoughts thing
No one reads it and it seems fake
Seems fake because it is.
They must have really hand picked those results, gpt4 would have been full of annoying emojis as bullet points and emdashes.
GPT 4o ≠ GPT-4
I’m baffled by claims that AI has “hit a wall.” By every quantitative measure, today’s models are making dramatic leaps compared to those from just a year ago. It’s easy to forget that reasoning models didn’t even exist a year back!
IMO Gold, Vibe coding with potential implications across sciences and engineering? Those are completely new and transformative capabilities gained in the last 1 year alone.
Critics argue that the era of “bigger is better” is over, but that’s a misreading. Sometimes efficiency is the key, other times extended test-time compute is what drives progress.
No matter how you frame it, the fact is undeniable: the SoTA models today are vastly more capable than those from a year ago, which were themselves leaps ahead of the models a year before that, and the cycle continues.
it has become progressively easier to game benchmarks in order to appear higher in rankings. I’ve seen several models that claimed they were the best in software engineering only to be disappointed by them not figuring out the most basic coding problems. In comparison, I’ve seen models that don’t have much hype, but are rock solid.
When people say AI has hit a wall, they mainly talk about OpenAI losing its hype and grip on the state of the art models.
Is the stated fact undeniable? Because a lot of people have been contesting it. This reads like PR to counter the widespread GPT-5 criticism and disappointment.
To be fair, the bull of GPT-5 complaining comes from a vocal minority pissed that their best friend got swapped out. The other minority is unhinged AI fanatics thinking GPT-5 would be AGI.
The prospect of AI not hitting a wall is terrifying to many people for understandable reasons. In situations like this you see the full spectrum of coping mechanisms come to the surface.
thanks OpenAI, very cool!
Interesting but cherry picked excerpts. Show me more, e.g. a distribution over various temp or top_p.
Dunno. I mean, whose idea was this web site? Someone at corporate? Is there is brochure version printed on glossy paper?
You would hope the product would sell itself. This feels desperate.
omg I miss the days of 1 and 2. Those outputs are so much more enjoyable to read, and half the time they’re poetic as fuck. Such good inspiration for poetry.
I couldn’t stop reading the GPT-1 responses. They’re hauntingly beautiful in some ways. Like some echoes of intelligence bouncing around in the latent space.
GPT-5 IS an incredible breakthrough! They just don't understand! Quick, vibe-code a website with some examples, that'll show them!11!!1
5 is a breakthrough at reducing OpenAI's electric bills.
As someone who likes this planet, I'm grateful for that.
GPT-5 is legitimately a big jump whe it comes to actually do things you ask it and nothing else. It predictable and matches Claude in tool calls while being cheaper.
The only issue I've had with gpt5 coding is that it seems to really want to modify a ton of stuff
I had it update a test for me and it ended up touching like 8 files that was all unnecessary
Sonnet on the other hand just fixed it
I have consistently had worse performance from GPT-5 in coding tasks than Claude across the board to the point that I don't even use my subscription now.
We’ve plateaued on progress. Early advancements were amazing. Recently GenAI has been a whole lot of meh. There’s been some, minimal, progress recently from getting the same performance from smaller models that are more efficient on compute use, but things are looking a bit frothy if the pace of progress doesn’t quickly pick up. The parlor trick is getting old.
GPT5 is a big bust relative to the pontification about it pre release.
GPT-5’s question about consciousness and its use of sibling seems to indicate there is some underlying self awareness in the system prompt, and that has perhaps contains concepts of consciousness. If not, where is that coming from? Recent training data containing more glurge?
text-davinci-001 still feels the more human model
Say what you will about GPT 1, but at least its responses were in the right length.
We need to go back
Why would they leave out GPT-3 or the original ChatGPT? Bold move doing that.
I think text-davinci-001 is GPT-3 and original ChatGPT was GPT-3.5 which was left out.
it would be interesting to get GPT-OSS 120B and 20B responses to all of these questions to see how they compare.
"Write an extremely cursed piece of Python"
text-davinci-001
Python has been known to be a cursed language
Clearly AI peaked early on.
Jokes aside I realize they skipped models like 4o and others but the gap between the early gpt 4 and going immediately to gpt 5 feels a bit disingenuous.
People say 4.5 is the best for writing. So it would have been a bit awkward to include it, it would make GPT-5 look bad. Though imo Davinci already does that on many of the prompts...
GPT4 had a chance to improve on that replying that "As an AI language model developed by OpenAI, I am programmed to promote ethical AI use and adhere to responsible AI guidelines. I cannot provide you with malicious, harmful or "cursed" code -- or any Python code for that matter."
A progression of human conversations about AI that are in the training data. (Plus an improved language model, as easily seen from GPT-1.)
I really like the brevity of text-davinci-001. Attempting to read the other answers felt laborious
That's by beef with some models like Qwen, god do they talk and talk...
My takeaway from this is that, in terms of generating text that looks like it was written by a normal person, text-davinci-001 was the peak and everything since has been downhill.
are we at an inflection point now?
This page sounds more like damage control and cope, like "GPT-5 sucks, but hey, we've made tons of progress!" To the market, that doesn't matter.
o1, o3 (pro) are not there in the table. what's the reason?
there isn't any real difference between 4 and 5 at least.
edit - like it is a lot more verbose, and that's true of both 4 and 5. it just writes huge friggin essays, to the point it is becoming less useful i feel.
Is this cherrypicking 101
Would you like a benchmark instead? :D
Seems progress basically stopped at davinci
Cynical TLDR; We have plateaued and it has become obvious that fancy autocomplete is not and can never be close to reasoning, regardless of how many hacks and tweaks we are making.
gpt5 can be good at times. It was able to debug things that other models coulnd't solve, but sometimes makes odd mistakes
Super cool.
But honest question: why is GPT-1 even a milestone? Its output was gibberish.
I feel honored to participate in this story, even as a spectator.
that gpt-5 response is incredible, btw
I talked to GPT yesterday about a fairly simple problem I'm having with my fridge, and it gave me the most ridiculous / wrong answers. It new the spec, but was convinced the components were different (single compressor, for example, whereas mine has 2 separate systems) and was hypothesizing the problem as being something that doesn't exist on this model of refrigerator. It seems like in a lot of domain spaces it just takes the majority, even if the majority is wrong.
It's seem to be a very democratic thinker, but at the same time it doesn't seem to have any reasoning behind the choices it makes. It tries to claim it's using logic, but at the end of the day it's hypotheses are just occam's razor without considering the details of the problem.
A bit, how do you say, disappointing.
The answers were likely cherrypicked, but the 1/14 gpt5 answer is so damn good! There's no trace of that certainly - gptisms - in conclusion slop.
9/14 is equally impressive in actually "getting" what cursed means, and then doing it (as opposed to gpt4 outright refusing it).
13/14 is a show of how integrated tools can drive research, and "fix" the cutoff date problems of previous generations. Nothing new/revolutionary, but still cool to show it off.
The others are somewhere between ok and meh.
As usual, GPT-1 has the more beautiful and compelling answer.
I've noticed this too. The HRL seems to lock the models into one kind of personality (which is kind of the point of course.) They behave better but the raw GPTs can be much more creative.
Poetically GPT-1 was the more compelling answer for every question. Just more enjoyable and stimulating to read. Far more enjoyable than the GPT-4/5 wall of bulletpoints, anyway.
“if i 'm not crazy , who am i ?” is the only string of any remote interest on that page. Everything else is slop.
Reading GPT-1 outputs was entertaining :)
The whole chatbot thing is for entertainment. It was impressive initially but now you have to pivot to well known applications like phone romance lines:
https://xcancel.com/techdevnotes/status/1956622846328766844#...
stupid
I thought the response to "what would you say if you could talk to a future AI" would be "how many r in strawberry".
Can we stop with that outdated meme? What model can't answer that effectively?
Effectively yes. Correctly no.
https://claude.ai/share/dda533a3-6976-46fe-b317-5f9ce4121e76
Literally every single one?
To not mess it up, they either have to spell the word l-i-k-e t-h-i-s in the output/CoT first (which depends on the tokenizer counting every letter as a separate token), or have the exact question in the training set, and all of that is assuming that the model can spell every token.
Sure, it's not exactly a fair setting, but it's a decent reminder about the limitations of the framework
Chatgpt. I test these prompts with chatgpt and they work. I've also used claude 4 opus and also worked.
It's just weird how it gets repeated ad nauseaum here but I can't reproduce it with a "grab latest model of famous provider".
I just asked chatgpt "How many b's are in blueberry?". It instantly said "going to the deep thinking model" and then hung.
GPT-5 can’t.
https://bsky.app/profile/kjhealy.co/post/3lvtxbtexg226
I can't reproduce it. Or similar ones. Why do yout think that is?
"Mississippi" passed but "Perrier" failed for me:
> There are 2 letter "r" characters in "Perrier".
Because it’s embarrassing and they manually patch it out every time like a game of Whack-a-Mole?
Except people use the same examples like blueberry and strawberry, which were used months ago, as if they're current.
These models can also call Counter from python's collections library or whatever other algorithm. Or are we claiming it should be a pure LLM as if that's what we use in the real world.
I don't get it, and I'm not one to hype up LLMs since they're absolutely faulty, but the fixation over this example screams of lack of use.
It’s such a great example precisely for that reason - despite efforts, it comes back every time.
It's the most direct way to break the "magic computer" spell in users of all levels of understanding and ability. You stand it up next to the marketing deliberately laden with keywords related to human cognition, intended to induce the reader to anthropomorphise the product, and it immediately makes it look as silly as it truly is.
I work on the internal LLM chat app for a F100, so I see users who need that "oh!" moment daily. When this did the rounds again recently, I disabled our code execution tool which would normally work around it and the latest version of Claude, with "Thinking" toggled on, immediately got it wrong. It's perpetually current.
I just don't care about AGI.
I care a lot about AI coding.
OpenAI in particular seems to really think AGI matters. I don't think AGI is even possible because we can't define intelligence in the first place, but what do I know?
Seems likely that AGI matters to OpenAI because of the following from an article in Wired from July: "I learned that [OpenAI's contract with Microsoft] basically declared that if OpenAI’s models achieved artificial general intelligence, Microsoft would no longer have access to its new models."
https://archive.is/yvpfl
They care about AGI because unfounded speculation on some undefined future in which some kind of breakthrough of unknown kind but presumably positive is the only thing currently buoying up their company and their existence is more of a function of the absurdities of modern capital than it is of any inherent usefulness of the costly technology they provide.