Interesting. I wonder how much could be gained from using tokenization, which makes the model work at a semantic level rather than a syntactic level? I think it’s a force multiplier, but idk if it works here.
If you actually scroll through the transcript he links to, you will see that something that looks like it could be training is happening, but no coherent responses are coming out at any point. At least not that I saw skimming through.
That might explain why there are no benchmarks of any kind.
Using the term AGI and not including any performance analysis. My AI calls it: "massive marketing overreach".
Somebody called this slop in the comments.
As a professor who published on continual learning I'm leaning towards agreement[1]. It lacks any substance. No relation to related work, no description of algorithm, no ablation study, just hand-waving that we're feeding some data and "Chess is not forgotten".
This "how-continual-learning-works" markdown text is not an algorithm [2].
There is no special algorithm, the finding is that slowing down the LR or the trunk, while keeping the LR of the experts is enough to eliminate most of the forgetting in the network. You can see in that experiment where chess data was the only thing the model read for 524K characters, yet it kept almost the same performance (i.e. held-out loss) on all other domains. If you keep LR the same across the whole network the loss in other domains degrades dramatically - this is a clear sign of catastrophic forgetting in action. What I can say for sure is that any traditional network that does pose a sign of catastrophic forgetting would not be able to learn any patterns from a single stream of data.
There are no benchmarks published as the model is heavily undertrained, but it is learning. And you can see this clearly in the loss and samples even though they are still barely coherent.
I am not an academic and am not trying to publish a paper about a “major breakthrough” or something like this. I am just a small person who found a cool thing that clearly works and wants to share it with the world. That’s it.
Actually I'm mad that I wasted my time looking at it based on the claims. He implies it is trained and uses the term "AGI" and "continuous learning". He never finished a single training run or enough that he considers not "undertrained". It's not trained. And actually there is no evidence that it can actually learn anything useful.
"The model reads 524,000 characters of chess". This is 100KByte of training data in a toy model with rigid parameters and no global learning. Gap with real LLM and trillions of tokens.
This model really addresses the problem of preserving previously learned knowledge, but by restricting the LR of the trunk it stops acquiring new knowledge. Details: "Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective"
Nobody will throw rocks, I think most people are curious/suspicious about the big players and wants more hands-on since we suspect that this all will come down in cost soon enough.
Seems interesting, I've been messing with a lot of continuous learning approaches lately and it's cool to see something that's built from the ground up for avoiding catastrophic forgetting. Worth a clone for sure
Is this architecture actually able to generalize or is it mostly based on memorization? Have you tried some basic tasks that require generalization? e.g. number addition etc?
The model is way too small and undertrained to make any generalization claims. I want to wait until it reads the whole corpus I gave and then test it on some simple established benchmarks to see how it will behave.
Seems a bit premature to make an HN post about then, imho.
It's an interesting idea, but it doesn't really do anything interesting yet. I looked at the output in the training run and it is a far, far cry from intelligence. Worse than GPT-2 as it stands.
I do hope it will perform well when scaled and trained, though; best of luck.
I also like how it is very organic. It naturally grows and deletes unused elements, so in addition to traditional backprop there is also a natural selection happening in the background. Each new expert has 16 parents by the way, lol.
What's the advantage of doing this, versus becoming good at context management and RAG? I always found trained knowledge unreliable, given that it is lossy by construction.
Have you thought about making the whole thing "self-similar"? Every time I hear about MoE I think (and I know it's way easier thought than done) "why stay shallow"?
I mean by that: would it be possible to extend/adapt the architecture so that an expert can be a previously trained Mini-AGI model? And recurse like this? Inuitively I would think some form of generalization could happen, as higher level experts (in the recurrence stack) would become sort of the "intuition" layer.
Making model to consists of many small modules is inefficient on GPU, especially as routing adds data dependencies, etc, and especially with pytorch (compared to a custom kernel).
The difference might be smaller on a CPU which has limited parallelism.
But it's basically equivalent to a very deep model which might be problematic for training.
It interleaves random streams of 32K characters long each when reading the whole corpus, but each such stream reads continuously as you would expect. This is a necessary step to prevent just normal, not catastrophic, forgetting. I have not tested it in any other regimes yet with bigger or smaller windows. You can imagine a person that changes the activity from time to time, so I think it is justified. So there is not really "early in the stream".
What I did test though is reading 524K characters of chess data only and see how other domains have degraded. The results are in the readme under "How continual learning works" section. Spoiler: it just barely degraded the performance.
The expert swapping architecture is very nice. Have you considered doing nested reinforcement learning where you use the nesting as a sort of low pass filter?
The concept is as follows: You train a critic to mimic the datastream and then you train against the critic instead of training against the data. The idea behind this is that the critic will memorize the training data so you do not need to store the full training data anymore. One of the biggest issues with current online stochastic gradient descent is that it is inherently a memory-less technique where the training data acts as the memory.
You can spin this further by going deeper with the nesting and then dropping the supervised critic. I forgot how to put it in words but the goal is that by having a model train against a critic of the critic, you can then drop the top level critic and instead use the mid level critic itself as your meta learning objective to train the actor against an unlabeled data stream.
Top level critic: learns to mimic the labeled training data via online SGD, then you add a simple hand written loss function to compare the predicted output with a given input. Basically you build a model specifically for distillation.
Mid level critic: learns a reward function that mimics the top level critic directly but only gets to see the unlabeled training data and the result of the top level critic.
Actor: The actor is exclusively trained against the mid level critic
Through this concept you end up with the existing training data stored as objective inside the mid level critic so you end up training not only against the latest data but also the already memorized data which should lower catastrophic forgetting. Of course at some point you might need to update the mid level critic again and to avoid that you might get away with just adding a very very wide Linear RNN / State Space Model / Mamba / Gated Delta Net as the middle critic (shower thought: use internal RNN states to represent LoRA vectors).
This is slop. 8M parameter dense model with context length 64 that you train on enwik9 in 2h will have 1.15 bpb. This model has 1.8 (bits per byte, lower is better).
Interesting. I wonder how much could be gained from using tokenization, which makes the model work at a semantic level rather than a syntactic level? I think it’s a force multiplier, but idk if it works here.
If you actually scroll through the transcript he links to, you will see that something that looks like it could be training is happening, but no coherent responses are coming out at any point. At least not that I saw skimming through.
That might explain why there are no benchmarks of any kind.
Using the term AGI and not including any performance analysis. My AI calls it: "massive marketing overreach". Somebody called this slop in the comments.
As a professor who published on continual learning I'm leaning towards agreement[1]. It lacks any substance. No relation to related work, no description of algorithm, no ablation study, just hand-waving that we're feeding some data and "Chess is not forgotten".
This "how-continual-learning-works" markdown text is not an algorithm [2].
[1] https://arxiv.org/abs/2301.12530
[2] https://github.com/volotat/mini-AGI/#how-continual-learning-...
There is no special algorithm, the finding is that slowing down the LR or the trunk, while keeping the LR of the experts is enough to eliminate most of the forgetting in the network. You can see in that experiment where chess data was the only thing the model read for 524K characters, yet it kept almost the same performance (i.e. held-out loss) on all other domains. If you keep LR the same across the whole network the loss in other domains degrades dramatically - this is a clear sign of catastrophic forgetting in action. What I can say for sure is that any traditional network that does pose a sign of catastrophic forgetting would not be able to learn any patterns from a single stream of data.
There are no benchmarks published as the model is heavily undertrained, but it is learning. And you can see this clearly in the loss and samples even though they are still barely coherent.
I am not an academic and am not trying to publish a paper about a “major breakthrough” or something like this. I am just a small person who found a cool thing that clearly works and wants to share it with the world. That’s it.
You can't claim it "works" if it hasn't produced any coherent responses and is still early in your first training attempt.
Actually I'm mad that I wasted my time looking at it based on the claims. He implies it is trained and uses the term "AGI" and "continuous learning". He never finished a single training run or enough that he considers not "undertrained". It's not trained. And actually there is no evidence that it can actually learn anything useful.
Indeed this is wasting HN time.
"The model reads 524,000 characters of chess". This is 100KByte of training data in a toy model with rigid parameters and no global learning. Gap with real LLM and trillions of tokens.
This model really addresses the problem of preserving previously learned knowledge, but by restricting the LR of the trunk it stops acquiring new knowledge. Details: "Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective"
Nobody will throw rocks, I think most people are curious/suspicious about the big players and wants more hands-on since we suspect that this all will come down in cost soon enough.
This is the first thing I see in my life that really looks like proto-AGI, it deserves its name.
Seems interesting, I've been messing with a lot of continuous learning approaches lately and it's cool to see something that's built from the ground up for avoiding catastrophic forgetting. Worth a clone for sure
It doesn't show any indications of solving catastrophic forgetting.
Is this architecture actually able to generalize or is it mostly based on memorization? Have you tried some basic tasks that require generalization? e.g. number addition etc?
The model is way too small and undertrained to make any generalization claims. I want to wait until it reads the whole corpus I gave and then test it on some simple established benchmarks to see how it will behave.
What kind of hardware are you using for training?
nm, I found it:
> RTX 3070 Laptop GPU with 8 GB
Super impressive.
Seems a bit premature to make an HN post about then, imho.
It's an interesting idea, but it doesn't really do anything interesting yet. I looked at the output in the training run and it is a far, far cry from intelligence. Worse than GPT-2 as it stands.
I do hope it will perform well when scaled and trained, though; best of luck.
Getting conceptually closer to how the human brain works. Looking forward to more of this.
I also like how it is very organic. It naturally grows and deletes unused elements, so in addition to traditional backprop there is also a natural selection happening in the background. Each new expert has 16 parents by the way, lol.
What's the advantage of doing this, versus becoming good at context management and RAG? I always found trained knowledge unreliable, given that it is lossy by construction.
Do you find all your own (human) trained knowledge unreliable?
Seeing 'Mini-AGI' and '8GB VRAM' in the same sentence is a breath of fresh air. Maybe local AGI isn't so far-fetched.
Have you thought about making the whole thing "self-similar"? Every time I hear about MoE I think (and I know it's way easier thought than done) "why stay shallow"? I mean by that: would it be possible to extend/adapt the architecture so that an expert can be a previously trained Mini-AGI model? And recurse like this? Inuitively I would think some form of generalization could happen, as higher level experts (in the recurrence stack) would become sort of the "intuition" layer.
Making model to consists of many small modules is inefficient on GPU, especially as routing adds data dependencies, etc, and especially with pytorch (compared to a custom kernel).
The difference might be smaller on a CPU which has limited parallelism.
But it's basically equivalent to a very deep model which might be problematic for training.
Have you tested what it remembers from early in the stream after a shift in the topics thrown at it?
It interleaves random streams of 32K characters long each when reading the whole corpus, but each such stream reads continuously as you would expect. This is a necessary step to prevent just normal, not catastrophic, forgetting. I have not tested it in any other regimes yet with bigger or smaller windows. You can imagine a person that changes the activity from time to time, so I think it is justified. So there is not really "early in the stream".
What I did test though is reading 524K characters of chess data only and see how other domains have degraded. The results are in the readme under "How continual learning works" section. Spoiler: it just barely degraded the performance.
The expert swapping architecture is very nice. Have you considered doing nested reinforcement learning where you use the nesting as a sort of low pass filter?
The concept is as follows: You train a critic to mimic the datastream and then you train against the critic instead of training against the data. The idea behind this is that the critic will memorize the training data so you do not need to store the full training data anymore. One of the biggest issues with current online stochastic gradient descent is that it is inherently a memory-less technique where the training data acts as the memory.
You can spin this further by going deeper with the nesting and then dropping the supervised critic. I forgot how to put it in words but the goal is that by having a model train against a critic of the critic, you can then drop the top level critic and instead use the mid level critic itself as your meta learning objective to train the actor against an unlabeled data stream.
Top level critic: learns to mimic the labeled training data via online SGD, then you add a simple hand written loss function to compare the predicted output with a given input. Basically you build a model specifically for distillation. Mid level critic: learns a reward function that mimics the top level critic directly but only gets to see the unlabeled training data and the result of the top level critic. Actor: The actor is exclusively trained against the mid level critic
Through this concept you end up with the existing training data stored as objective inside the mid level critic so you end up training not only against the latest data but also the already memorized data which should lower catastrophic forgetting. Of course at some point you might need to update the mid level critic again and to avoid that you might get away with just adding a very very wide Linear RNN / State Space Model / Mamba / Gated Delta Net as the middle critic (shower thought: use internal RNN states to represent LoRA vectors).
what character prediction rates are you getting on some unseen datasets?
The held-out scores reported in the Readme IS the unseen dataset.
This is slop. 8M parameter dense model with context length 64 that you train on enwik9 in 2h will have 1.15 bpb. This model has 1.8 (bits per byte, lower is better).
THANK YOU SO MUCH. This is the missing piece.
throwing crumpled paper ball