I'm sort of in the same boat but I'm finding every day more and more people I trust gleefully use these tools. I still don't really get it. I get suggestions of things I should try, techniques I should try, how to prompt, how to harness, which models to use. I just find the output off and end up going in and just going over the work again.
The conclusion I've come to is that good taste has been replaced with "works; good enough". Maybe that's the end state. Maybe the end is just a bland mass of tangled code that's junky but it works. Speed over quality has always been a currency in this industry and maybe this is just the purest expression of it.
But at the end, it leaves me with my preferences and tastes and what I think is right and I feel more and more isolated. I don't think the people who are happily using AI tools are incompetent. I know they're not. I think they just are a little more in tune with how the winds are changing.
In the long run, I'm not sure how this ends. Programming is a lifelong pursuit of endless learning. When you use tools like this, your learning becomes stunted. But maybe I'm wrong. Nobody knows.
Due to the very diffuse statistical nature of the mapping of the input to the output, some will be experiencing extraordinary appropriate useful results while others will be experiencing crap. And due to the subjective nature of quality and quirky topographies of the internal terrains of these systems and the haphazard control protocols, the spread of resultant value (so to speak) of these appliances may not be understandable as a normal distribution. Scenarios we might judge as lying close to each other on the input terrain may generate very peculiar bumpy results of output terrain.
One of the most damning considerations of this technology is that gross uncertainty about result quality may be manipulated by the appliance maker leading to wild patterns of exploitation without any prospect of accountability.
As there's no pretense of formal function, there can be no verification, and so there's no methodology for determining quality: a presumably working appliance may appear reliable until it isn't with no way to understand the inflection. Those with insider control and information can game customers with no chance of accountability, but the system can't be understood.
The problem with magic beans is guessing what they'll grow into.
Meanwhile, the seismic movement of business towards these private and completely mysteries silos of "knowledge" (which is a grotesque misnomer) may arrive at dead ends from which they can't retreat.
This is a far greater danger than a malicious bot because a great migration into the dead end is not just well under way, but regarded as a brilliant prospect which we must run into at all deliberate speed.
+1 for "...may not be understandable as a normal distribution." "Rogue" might just be the operators wishful thinking. Hammer nail (something from a prior life), politics isn't normal: http://athena.m3047.net/elections/dist-not-normal/distributi... Time to complete software projects is another. :-/
> AI has given a lot of completely incompetent people a false sense of competency
30+ years of experience writing large applications in C++ and TypeScript, including JIT compilers, commercial game engines, and business-critical systems used by international companies.
I love using Claude Code for the 90% of my work that is repetitive and boring. These are tasks I've done many times before. It does a great job as long as I guide it well and review every change before committing.
I find the negative tone of the article puzzling. AI is just a tool. Use it well and it makes you more productive. Use it badly and it causes problems.
right, because grabbing a tool reluctantly and spinning up something you could easily do by hand gives you all the insights you need to write a blog post. what could go wrong?
Don't want to be rude but thats the last-resort-argument people make when defending LLMs. Its always about wrong utilization or the wrong test. If LLMs are so capable and so useful where is the "AI" fueled, real economic growth ? Why is Software still not dead cheap ? I really see how useful LLMs are especially for search. Its amazing and I use it every day. Also sometimes it can generate decent code. But this is still only a marginal improvement of productivity which leads to the big question I have: Why is it so hard to see the value created by LLMs if they are really the innovation they are promised to be ? Imo it should be trivial to see the benefit of such an innovation, kind of like it was with the wheel or the computer. For AI this is still a big ongoing discussion with not much to show. What am I missing ?
I find use out of 'em everyday. what ARE you missing? why am I tasked to let you know? go find out.
(to me, it isn't about producing the right test -- I could think of many scenarios AI has enabled me to move faster on than I ever could without. if you don't have those scenarios, I don't know why you'd fabricate them just to prove or disprove a theory about the usefulness of something that is, well, not useful to you.)
I'm sort of in the same boat but I'm finding every day more and more people I trust gleefully use these tools. I still don't really get it. I get suggestions of things I should try, techniques I should try, how to prompt, how to harness, which models to use. I just find the output off and end up going in and just going over the work again.
The conclusion I've come to is that good taste has been replaced with "works; good enough". Maybe that's the end state. Maybe the end is just a bland mass of tangled code that's junky but it works. Speed over quality has always been a currency in this industry and maybe this is just the purest expression of it.
But at the end, it leaves me with my preferences and tastes and what I think is right and I feel more and more isolated. I don't think the people who are happily using AI tools are incompetent. I know they're not. I think they just are a little more in tune with how the winds are changing.
In the long run, I'm not sure how this ends. Programming is a lifelong pursuit of endless learning. When you use tools like this, your learning becomes stunted. But maybe I'm wrong. Nobody knows.
You never hear anyone talking about “imposter syndrome” anymore, do you?
This reads like a rant by the self-proclaimed best carriage driver around the invention of the automobile.
If we gauge LLM progress by the analogies they generate then they are still in the wind-up toy car phase.
There's an overlooked aspect of AI performance:
Due to the very diffuse statistical nature of the mapping of the input to the output, some will be experiencing extraordinary appropriate useful results while others will be experiencing crap. And due to the subjective nature of quality and quirky topographies of the internal terrains of these systems and the haphazard control protocols, the spread of resultant value (so to speak) of these appliances may not be understandable as a normal distribution. Scenarios we might judge as lying close to each other on the input terrain may generate very peculiar bumpy results of output terrain.
One of the most damning considerations of this technology is that gross uncertainty about result quality may be manipulated by the appliance maker leading to wild patterns of exploitation without any prospect of accountability.
As there's no pretense of formal function, there can be no verification, and so there's no methodology for determining quality: a presumably working appliance may appear reliable until it isn't with no way to understand the inflection. Those with insider control and information can game customers with no chance of accountability, but the system can't be understood.
The problem with magic beans is guessing what they'll grow into.
Meanwhile, the seismic movement of business towards these private and completely mysteries silos of "knowledge" (which is a grotesque misnomer) may arrive at dead ends from which they can't retreat.
This is a far greater danger than a malicious bot because a great migration into the dead end is not just well under way, but regarded as a brilliant prospect which we must run into at all deliberate speed.
+1 for "...may not be understandable as a normal distribution." "Rogue" might just be the operators wishful thinking. Hammer nail (something from a prior life), politics isn't normal: http://athena.m3047.net/elections/dist-not-normal/distributi... Time to complete software projects is another. :-/
So that people know what we mean. There's also Wikipedia: https://en.wikipedia.org/wiki/Normal_distribution
> AI has given a lot of completely incompetent people a false sense of competency
30+ years of experience writing large applications in C++ and TypeScript, including JIT compilers, commercial game engines, and business-critical systems used by international companies.
I love using Claude Code for the 90% of my work that is repetitive and boring. These are tasks I've done many times before. It does a great job as long as I guide it well and review every change before committing.
I find the negative tone of the article puzzling. AI is just a tool. Use it well and it makes you more productive. Use it badly and it causes problems.
right, because grabbing a tool reluctantly and spinning up something you could easily do by hand gives you all the insights you need to write a blog post. what could go wrong?
Don't want to be rude but thats the last-resort-argument people make when defending LLMs. Its always about wrong utilization or the wrong test. If LLMs are so capable and so useful where is the "AI" fueled, real economic growth ? Why is Software still not dead cheap ? I really see how useful LLMs are especially for search. Its amazing and I use it every day. Also sometimes it can generate decent code. But this is still only a marginal improvement of productivity which leads to the big question I have: Why is it so hard to see the value created by LLMs if they are really the innovation they are promised to be ? Imo it should be trivial to see the benefit of such an innovation, kind of like it was with the wheel or the computer. For AI this is still a big ongoing discussion with not much to show. What am I missing ?
I find use out of 'em everyday. what ARE you missing? why am I tasked to let you know? go find out.
(to me, it isn't about producing the right test -- I could think of many scenarios AI has enabled me to move faster on than I ever could without. if you don't have those scenarios, I don't know why you'd fabricate them just to prove or disprove a theory about the usefulness of something that is, well, not useful to you.)