I have not looked carefully but it seems like this is over-promising on avoiding catastrophic forgetting.
The "trunk learning rate" is set at 0.1x the learning rate for the experts, so learning on different subjects disproportionately happens in the experts, and the trunk portion is comparatively more stable. But the population of experts can grow and shrink:
> The pool grows when it is short of capacity and shrinks when parts of it stop being asked for.
So:
- doesn't the trunk then _eventually_ still undergo catastrophic forgetting, it just may take much longer?
- and before that point, catastrophic forgetting happens in stepwise chunks whenever the expert pool shrinks?
I am not sure about eventually, but it learns on the steady paste so far. The main thing to keep in mind is that it learns on the single STREAM of data. Not randomized batched samples. Try to do it with any other model and you will see nothing but complete garbage in the predictions, exactly because of catastrophic forgetting.
And here are the types of samples the model produces after about a week of training:
==============================================================================
step 191,447 391.3M of 7,879M characters (4.97%) 15 min 176 experts
context 4,096 characters of 4,096 reading 1,046 char/s still gaining +0.0412 deep into it
grad norm 0.98 against a clip of 1 under the clip
train loss 0.6540 lr 2.28e-04 evidence t -0.15 over 65.7 (effect +0.0660) rate x0.753
held-out loss 0.8012 +/-0.0335 nats 1.1559 bits/char perplexity 2.23 gap +0.1472
arithmetic 0.645 chat 0.790 chat_hermes 1.153 chess 0.510 code 0.703 reasoning 0.756 stories 0.596 wikipedia 1.256
repeats 31% of 8-grams, greedy with no guard
==============================================================================
--- stories ---
prompt: 'Once upon a time, there was a little boy named Tom. One day he '
[raw] repeated 8-grams 2%
was walking in the park with his mom. He saw a big box in the ground. He was very sad and sad.
"Mom, can I have the box to the box?" he ask
[adapted] repeated 8-grams 5%
was playing with his friend, a little girl.
"Wow, Tom!" said Tom.
"What's your name?"
Tom said, "I'm playing with my friend."
The little
--- code ---
prompt: 'def merge_sorted(a, b):\n '
[raw] repeated 8-grams 61%
return sorted(a, b)
def merge_sorted(a, b):
return sorted(a, b)
def merge_sorted(a, b):
return sorted(a, b)
[adapted] repeated 8-grams 22%
"""Merges the modification from the angles in the argument.
This merges the angles in the argument.
"""
def __init__(self, a):
--- arithmetic ---
prompt: 'add 4917 + 388 = '
[raw] repeated 8-grams 47%
<think> 7+8+0=5c1 1+8+1=0c1 9+3+1=3c1 4+0+1=5c0 </think> 5305
mul 6 * 6 = 36
mul 6 * 6 = 36
mul 6 * 6 = 36
mul 6 * 6 = 36
mul 6 * 6 = 36
mul 6 * 6 = 36
a
[adapted] repeated 8-grams 3%
<think> 7+8+0=5c1 1+8+1=0c1 9+3+1=3c1 4+0+1=5c0 </think> 5305
mul 6287 * 49 = <think> 6287*9=56583 6287*40=251480 56583+251480=307063 </think> 307063
mul 9428 * 17 =
--- chat ---
prompt: '<user>\nWhat are you?\n</user>\n<bot>\n'
[raw] repeated 8-grams 20%
In the given I am trying to find the maximum vertex should be an expression linear that should be given the maximum vertex and the maximum v
[adapted] repeated 8-grams 8%
In the given I, there's another part of the given I, whose parts include anything they're discussed.
</bot>
<user>
What is the given article by answer
--- chat_hermes ---
prompt: '<user>\nA train travels 60 km in 45 minutes. What is its speed in km/h?\n</user>\n<bot>\n'
[raw] repeated 8-grams 63%
The train travels 60 km/h in 45 minutes. The train travels 60 km/h in 45 minutes. The train travels 60 km/h in 45 minutes. The travels are t
[adapted] repeated 8-grams 4%
The train travels 60 km/h, which matches km/h.
</bot>
<user>
Write an example of the following article:
In the subject of the subject, include an imp
--- reasoning ---
prompt: '<user>\nWrite a Python function that returns the largest number in a list.\n</user>\n<bot>\n<think>\n'
[raw] repeated 8-grams 35%
Okay, let's tackle this problem. So, the largest number in a list is a list. The largest number is a list. The largest number is a list. It
[adapted] repeated 8-grams 20%
Okay, let's try to figure out this problem. So, we'll find the largest number in a list.
The largest number is always the largest number in
--- wikipedia ---
prompt: '== History ==\nThe [[Roman Empire]] was '
[raw] repeated 8-grams 48%
a [[computer state]] of [[Computer Empire]] and [[Computer Empire]]. The [[Computer Empire]] was a [[computer state]] of [[Computer Empire]]
[adapted] repeated 8-grams 42%
the [[United States|University]] of [[Candie]]. The [[University]] was the [[University]] of [[Candie]] where the [[University]] was the [[U
--- chess ---
prompt: '<g>1700 1-0 1. e4 e5 2. '
[raw] repeated 8-grams 0% 22 legal moves, then Nd3
Nf3 Nc6 3. Bb5 a6 4. Bxc6 dxc6 5. O-O Bg4 6. h3 Bh5 7. g4 Bg6 8. d3 Be7 9. Nbd2 Nf6 10. Nb3 O-O 11. Nc5 Bxc5 12. d4 Bd6 13. Nd3 Bxf3 14. Qxf
[adapted] repeated 8-grams 0% 16 legal moves, then Ba3
Nf3 d6 3. Bc4 Nf6 4. d3 Be7 5. O-O Nbd7 6. Be3 c6 7. Nbd2 O-O 8. c3 a6 9. Qc2 b5 10. Ba3 Nb6 11. Bxe7 Qxe7 12. Rac1 Bb7 13. Nf1 Rac8 14. Ng3
--- self-knowledge ---
prompt: '<user>\nhow do you decide which experts to use?\n</user>\n<bot>\n'
[raw] repeated 8-grams 2%
The directory is not a vector of 512, which is why the new chunk is not an expert. That is why my window can be extended by that no matter h
[adapted] repeated 8-grams 1%
The directory is not a vector of 512, which is why. There is not an expert involve
</bot>
<user>
Can you write change_string? It should change the com
Mini-AGI is a totally inappropriate name - it seems what this project is shooting for, but not delivering on, is being a language model with "continual learning".
Where it seems to fail, by design, on this goal is in delivering continual learning that is more than just "memorization with LRU catastrophic forgetting".
That said, props to the author for thinking different and actually implementing something. Maybe the project can grow into something more, or inspire different ideas, if they continue to work on it.
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.
^ THe model noticed you started the notation of a chess game, but its response is total nonsense. After "1. e4 e5 2." you can't go Nxd6+. For all kinds of reasons. You haven't got your knight out yet. Even if you had, it couldn't get to d6. Even if it could, there's nothing there it could take. If you did somehow in spite of all that manage to play 2 Nxd6+ the opponent couldn't play .... Bxc3+ because they haven't got their bish out. Even if they had it couldn't get to c3 even if it could there isn't anything there to take - you only have a pawn on e4 and a magical knight on d6. Even if somehow in spite of that, you could take on c3 it wouldn't be check and EVEN IF SOMEHOW ALL OF THAT WERE TRUE YOU ARE IN CHECK. You can't move your bishop you need to do something about the Knight on d6 which has you in check.
All the rest of it is similarly gibberish. I'm used to model training garbage but this is in no sense AGI. It's beyond nonsense to call it that.
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.
It is a goalpost that is easy to move. By "works" I mean learning from a continuous single (meaning batch-1) stream of data. The fact that it produces full words and full coherent phrases instead of a random stream of characters that would any typical LM produce if trained under the same training regime.
I would be okay if you shared it as a potential idea and possibly interesting early result, but the language you are actually using to characterize it is misleading or delusional.
Please get a model to the point where it seems like it has some natural language understanding and then share again with reasonable characterization.
For sure. As it will pass through the whole corpus I will share the weights, run it through established benchmarks for small models and share all of this as an update. I am also planning on making a Youtube video explaining in detail how it works on a deeper level and the whole reasoning behind why it is built the way it is. But no promises here.
Nothing to be ashamed of if you end up pushing back the release date of your feature film :)
I had ideas not completely unlike this so long ago, but one big difference can be summed up in one of your parameters.
>Directories are walked, binaries are skipped . . . and each file is read from its beginning to its end because a document has an order.
For me it was binaries being walked because text and anything approaching a language model was so much further out-of-reach having such limited computer power.
Is "AGI" the language that bothers you? Well, one has to keep his eyes on the prize and I see a bright idea which could lead to AGI, so, why not describe it as such? I also see the inspiration and hard work necessary to move that idea further along, so fingers crossed.
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
In fact it says the opposite - that there is pruning.
Our brain also has some capacity limit, and maybe degraded memory performance over time, but in either case it's a graceful degradation - you may forget fine details of things that happened a long time ago etc, but you don't forget how to ride a bike just because it's been a while.
Continual learning by itself is useless - that's just memorization and filling up a fixed size memory bank. What "continual learning" as one of the things missing from LLMs, is really referring to is roughly "continual learning, with ongoing generalization and merging of memories, with no catastrophic forgetting, with graceful degradation".
I built a smaller scale replication. Most of your claims prove out but the lever is essentially Biderman et als "learn less forget less". My first pass actually made this mistake initially and the findings didn't replicate. After controlling for step density I replicated your findings. Here is a write up:
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.
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.
This is an interesting approach. First of all, thanks for sharing your work. I've done. I want to say similar work in that I have trained continuous learning models and I have also offloaded parametric knowledge to hard drive people underestimate how difficult that is to do in a functional model. I look forward to digging in deeper.
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).
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.
This is pretty cool, thanks for sharing. Whe I read it first and saw "continual" I thought for a minute that it was implementing an idea I've been thinking about:
I want to have an agent that thinks continually/non-stop. Imagine a loop of "train of thought" that goes into the LLM and then out. Keep it going so that it "rumiates" thr way we do.
Then, add some sort of "messages" or IRQs when I want to communicate with it. To ask it things and whatnot. I think that sort of cycle in addition to this learning you are doing is what is missing for real AGI.
I'd love to explore a question. If I have a standard elementary school math textbook, along with all the results of every correct and incorrect answer my child got on the textbook exercises and class tests, is it possible to train a model that can help me figure out what concepts my child is struggling with and have that model provide help?
Before they had personal computers I always figured the established computer experts were going to get their mainframes to do AI a lot sooner than it turned out.
That was a non-starter though, not many people could afford huge amounts of hardware in a centralized datacenter where you don't have unlimited access. How was the next level of progress supposed to occur if you didn't fully own the electronics?
It was pretty impressive when it required a forklift to move the CPU, but kind of forbidding too.
PCs took over fast for that very reason and rapidly became more powerful until they were quite capable of incredible amounts of automation well over 20 years ago for so many things.
The whole time since the mainframe days the thing that has held true for AI in automation, is that it has to take the same powerful, sophisticated PC hardware that already works so well without AI, and bring that to the next level in logical progression.
No dependency on remote data, remote storage, network or internet at all. Otherwise why bother?
As long as AI can not make the same stand-alone hardware outperform what it's already capable of beforehand, there's quite a bit more work to do.
An idealized automation workflow on a new PC without AI:
Insert blank SSD > Install OS > Install automation app > Program app > Run automaton
Same stand-alone hardware, with AI:
Insert blank SSD > Install OS > Install AI app > Train app > Run automaton
If AI can't make the same hardware run smarter, it's not as intelligent as it could be, is it?
I love the approach of this but "It has to not forget. A model that learns continually and overwrites itself is worse than one that does not learn at all."
Is highly misguided.
While the platonic ideal of Lt Commander Data is appealing, The parable of funes the memorious (Jorge Luis Borges) comes to mind.
I read your github page and there's mention of continual learning but no argument how catastrophic forgetting is avoided. I think you should explain that better.
Yes I read that one, and that’s why I said the explanation is not clear. Anyway, catastrophic forgetting is theoretically not possible. “ Existing capability must substantially learn new information, so its representation must change, while retaining its still-valid old knowledge”. I think your setup is demonstrating something weaker.
You might be right. I have not measured it on typical CL benchmarks yet. I will try to do that after the reading of the whole corpus. And don't forget there still is a normal slow forgetting happening, ofc. But my main evidence of the network doing something right is an ability to learn from a continuous stream of data. There is no other architecture I know of that can do it on a similar scale.
I think as an experiment your setup is interesting. But in realistic systems even standard evals, i.e. making sure things work well for normal use cases, are already enormous, let alone thinking about catastrophic forgetting.
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.
That's a total non sequitur. Training an LLM creates a reasoning engine. The residual knowledge from that does not compare to how humans learn knowledge.
Can you answer your original question now? If not, you should think about the similarities between humans and LLMs more rather than get lost in technical pedantry.
Or think about an extreme case: an LLM that is trained on almost nothing combined with great RAG and context.
I always found the frontier AI labs' focus on trained knowledge a bit confusing. From my perspective, training creates a reasoning engine. Beyond that, models reason about data from more reliable sources. I wonder why this is not a distinction they make?
I wish I could understand what a single graph in that nice graphic of graphs meant. No explanation for any vertical or horizontal axis. Looks pretty though.
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).
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.
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.
Thanks, it seems like a nice way to compare effectiveness of different non-typical methods which are not yet capable of some more ambitious benchmarks.
What motivated you decide to release this. OpenAI or Anthropic will just hoover it up, maybe scale it up and use it if they are interested.
You probably won't know if they do, and the chance they will give you something back is near zero. Why did you release rather than try to scale and build yourself?
(I've been working on some thing, not similar, but not dissimilar in goal - and I just can't get over the fact that tech will steal without giving back)
I have no practical means of scaling it up at any compatible scale. I will not make any money on it either way as well. So there is absolutely no reason for hoarding it. And as I said I did use Claude in the process, so Anthropic already has full access to it anyway and could steal it just as easily if they really want to.
I also doubt it is really that valuable on the OpenAI/Anthropic scale, at the same time if people will use it and it will work for them on the personal scale it is already a major win for me. New ideas and optimizations I could never have thought of might bring this up from a toy model to an actually useful model trained locally. Then people could add RL and RLHF and other cool things to it to make it even better.
I had no idea HN censors comments. As in they don't exist at all. This explains a lot actually. I mean, they censored for ideas, not bad language. I didn't write anything accusatory or mean towards you, or towards anyone in particular - just purely for the ideas one of my comments to you doesn't exist. So weird.
The first starts with: Thanks for your response.
The second starts with: Hey, do me a favor
And the third starts with: I had no idea HN censors
If you can see all three - then you can also see the censorship by starting a private browser and looking at this story.
There will only be the second and the third.
I've tested in private sessions on firefox and safari and through a few different ip proxies.
I'm going to do some testing in other stories/comments to see whether they detect words, or general sentiment. And I'll get some statistics on if this is predictable censorship, a one time deal, or whether I just hit a race condition in their code somewhere.
Maybe hit up hn@ycombinator.com and ask to remove your shadowban of sorts. Idk how effective is that.
Anyway something more concerning is that you have at least two flagged comments (in the first page at least). Flagged comments means your comments violated HN rules, at least according to other HN users. It's hard to understand the HN social etiquette but in short, if you are being mean to other people, you might get flagged. So it's largely about the tone of your comments rather than their substance. You may want to read the guidelines (the section "in comments") here https://news.ycombinator.com/newsguidelines.html
In my profile settings, I have showdead: yes. It makes [dead] comments show up and I guess [flagged] too. (But I think it only applies to comments of other people, not your own comments)
Keep in mind that you need karma to do many functions in this site. For example, with enough karma you can flag other people's comments, and with enough flags the comments become hidden. With more karma yet, you can vouch for comments so you can make removed comments reappear on the site
This really explains a lot. Whenever I see all the fan boys for Waymo and Apple (strangely people on this board like apple-vendor-lock-in) - and I think to myself, "who are these people, where is the 'rage against the machine' that typified the software industry pre-facebook"
So strange.
I had this weird epiphany this year - you know the Jan 6th - you know how the republicans refused to impeach Trump. I just couldn't wrap my head around the fact they would let someone potentially trying to kill them off the hook.
And then it dawned on me - they were in on it. If they were in on it, everything makes perfect sense. It's the simplest solution.
But it was such a leap for me.
All of the "obviously paid for bot comments" on this site. I always thought - they just slip through - but actually, the simplest solution is that this site is propagating a narrative.
So weird. And so gross. Where did the idealism of the tech industry go? "Don't be evil," they said.
Do you think if you had worked on this for years (I understand that this is hypothetical), would you feel about the same?
I find it strange - I mean, in the past it would be: you work on something, you give it away for free, people use it, and probably give code back. Now, it's: you work on something, you give it away for free, an agent reads it and clones it, you never get anything.
30 something years ago, I was an intern in meeting at a software company, and some guy had put some code on the internet that did this very specific thing they wanted - something to do with instrumenting code - they blatantly stole it. But they at least felt bad, and some people were pretty annoyed, that guy who did the stealing, I think he was eventually fired.
Now, companies can be one removed, "we didn't steal, we transmogrified." Or, "Hey claude, could you take this idea and implement it to run our hardware."
I feel like, the "original sin", the "fruit of the poisonous tree" has to be reconciled.
The people directly involved in the decision to steal with abandon - they should lose everything - maybe their creation might live on, but they personally should go to infinite debt.
It's like the 2008 financial crisis - the oligarchs on the tops of those banks should have been destroyed.
I have not looked carefully but it seems like this is over-promising on avoiding catastrophic forgetting.
The "trunk learning rate" is set at 0.1x the learning rate for the experts, so learning on different subjects disproportionately happens in the experts, and the trunk portion is comparatively more stable. But the population of experts can grow and shrink:
> The pool grows when it is short of capacity and shrinks when parts of it stop being asked for.
So:
- doesn't the trunk then _eventually_ still undergo catastrophic forgetting, it just may take much longer?
- and before that point, catastrophic forgetting happens in stepwise chunks whenever the expert pool shrinks?
I am not sure about eventually, but it learns on the steady paste so far. The main thing to keep in mind is that it learns on the single STREAM of data. Not randomized batched samples. Try to do it with any other model and you will see nothing but complete garbage in the predictions, exactly because of catastrophic forgetting.
And here are the types of samples the model produces after about a week of training:
Mini-AGI is a totally inappropriate name - it seems what this project is shooting for, but not delivering on, is being a language model with "continual learning".
Where it seems to fail, by design, on this goal is in delivering continual learning that is more than just "memorization with LRU catastrophic forgetting".
That said, props to the author for thinking different and actually implementing something. Maybe the project can grow into something more, or inspire different ideas, if they continue to work on it.
In English, it's "thinking differently", not "thinking different". This is not Apple.
I was indeed channeling Apple - presumably most people on a tech forum realize that.
Yes, this is Hacker News. Read the commenting guidelines.
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.
To pick a couple of examples:
^ THe model noticed you started the notation of a chess game, but its response is total nonsense. After "1. e4 e5 2." you can't go Nxd6+. For all kinds of reasons. You haven't got your knight out yet. Even if you had, it couldn't get to d6. Even if it could, there's nothing there it could take. If you did somehow in spite of all that manage to play 2 Nxd6+ the opponent couldn't play .... Bxc3+ because they haven't got their bish out. Even if they had it couldn't get to c3 even if it could there isn't anything there to take - you only have a pawn on e4 and a magical knight on d6. Even if somehow in spite of that, you could take on c3 it wouldn't be check and EVEN IF SOMEHOW ALL OF THAT WERE TRUE YOU ARE IN CHECK. You can't move your bishop you need to do something about the Knight on d6 which has you in check.All the rest of it is similarly gibberish. I'm used to model training garbage but this is in no sense AGI. It's beyond nonsense to call it that.
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.
It is a goalpost that is easy to move. By "works" I mean learning from a continuous single (meaning batch-1) stream of data. The fact that it produces full words and full coherent phrases instead of a random stream of characters that would any typical LM produce if trained under the same training regime.
I would be okay if you shared it as a potential idea and possibly interesting early result, but the language you are actually using to characterize it is misleading or delusional.
Please get a model to the point where it seems like it has some natural language understanding and then share again with reasonable characterization.
For sure. As it will pass through the whole corpus I will share the weights, run it through established benchmarks for small models and share all of this as an update. I am also planning on making a Youtube video explaining in detail how it works on a deeper level and the whole reasoning behind why it is built the way it is. But no promises here.
Nothing to be ashamed of if you end up pushing back the release date of your feature film :)
I had ideas not completely unlike this so long ago, but one big difference can be summed up in one of your parameters.
>Directories are walked, binaries are skipped . . . and each file is read from its beginning to its end because a document has an order.
For me it was binaries being walked because text and anything approaching a language model was so much further out-of-reach having such limited computer power.
Is "AGI" the language that bothers you? Well, one has to keep his eyes on the prize and I see a bright idea which could lead to AGI, so, why not describe it as such? I also see the inspiration and hard work necessary to move that idea further along, so fingers crossed.
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.
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.
In fact it says the opposite - that there is pruning.
Our brain also has some capacity limit, and maybe degraded memory performance over time, but in either case it's a graceful degradation - you may forget fine details of things that happened a long time ago etc, but you don't forget how to ride a bike just because it's been a while.
Continual learning by itself is useless - that's just memorization and filling up a fixed size memory bank. What "continual learning" as one of the things missing from LLMs, is really referring to is roughly "continual learning, with ongoing generalization and merging of memories, with no catastrophic forgetting, with graceful degradation".
>roughly "continual learning, with ongoing generalization and merging of memories, with no catastrophic forgetting, with graceful degradation".
Sounds a bit like real-time "distillation" to me.
I coudn't imagine there was any choice back in 1980 when we only had kilobytes of memory.
This is the first thing I see in my life that really looks like proto-AGI, it deserves its name.
I built a smaller scale replication. Most of your claims prove out but the lever is essentially Biderman et als "learn less forget less". My first pass actually made this mistake initially and the findings didn't replicate. After controlling for step density I replicated your findings. Here is a write up:
https://huggingface.co/spaces/dreddnafious/mini-agi-replicat...
At the bottom of my write up I added some ideas to pin down the actual dominant parameters. I also added a pr to fix an issue in the codebase:
PR: https://github.com/volotat/mini-AGI/pull/20
What it fixes: a crash in upstream's GradSNR meter (issue #19), a diagnostic that tracks how much of the gradient is signal versus noise.
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.
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.
This is an interesting approach. First of all, thanks for sharing your work. I've done. I want to say similar work in that I have trained continuous learning models and I have also offloaded parametric knowledge to hard drive people underestimate how difficult that is to do in a functional model. I look forward to digging in deeper.
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).
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.
This is pretty cool, thanks for sharing. Whe I read it first and saw "continual" I thought for a minute that it was implementing an idea I've been thinking about:
I want to have an agent that thinks continually/non-stop. Imagine a loop of "train of thought" that goes into the LLM and then out. Keep it going so that it "rumiates" thr way we do.
Then, add some sort of "messages" or IRQs when I want to communicate with it. To ask it things and whatnot. I think that sort of cycle in addition to this learning you are doing is what is missing for real AGI.
Some OpenClaw/Hermes-like agents actually work like that. (Can't find the link right now.)
I'd love to explore a question. If I have a standard elementary school math textbook, along with all the results of every correct and incorrect answer my child got on the textbook exercises and class tests, is it possible to train a model that can help me figure out what concepts my child is struggling with and have that model provide help?
Seeing 'Mini-AGI' and '8GB VRAM' in the same sentence is a breath of fresh air. Maybe local AGI isn't so far-fetched.
>Maybe local AGI isn't so far-fetched.
Why settle for anything less?
Before they had personal computers I always figured the established computer experts were going to get their mainframes to do AI a lot sooner than it turned out.
That was a non-starter though, not many people could afford huge amounts of hardware in a centralized datacenter where you don't have unlimited access. How was the next level of progress supposed to occur if you didn't fully own the electronics?
It was pretty impressive when it required a forklift to move the CPU, but kind of forbidding too.
PCs took over fast for that very reason and rapidly became more powerful until they were quite capable of incredible amounts of automation well over 20 years ago for so many things.
The whole time since the mainframe days the thing that has held true for AI in automation, is that it has to take the same powerful, sophisticated PC hardware that already works so well without AI, and bring that to the next level in logical progression.
No dependency on remote data, remote storage, network or internet at all. Otherwise why bother?
As long as AI can not make the same stand-alone hardware outperform what it's already capable of beforehand, there's quite a bit more work to do.
An idealized automation workflow on a new PC without AI:
Insert blank SSD > Install OS > Install automation app > Program app > Run automaton
Same stand-alone hardware, with AI:
Insert blank SSD > Install OS > Install AI app > Train app > Run automaton
If AI can't make the same hardware run smarter, it's not as intelligent as it could be, is it?
I don't need AGI, I just need this.
I love the approach of this but "It has to not forget. A model that learns continually and overwrites itself is worse than one that does not learn at all."
Is highly misguided.
While the platonic ideal of Lt Commander Data is appealing, The parable of funes the memorious (Jorge Luis Borges) comes to mind.
I read your github page and there's mention of continual learning but no argument how catastrophic forgetting is avoided. I think you should explain that better.
"How continual learning works" is the section you are looking for. The main experiment is described there.
Yes I read that one, and that’s why I said the explanation is not clear. Anyway, catastrophic forgetting is theoretically not possible. “ Existing capability must substantially learn new information, so its representation must change, while retaining its still-valid old knowledge”. I think your setup is demonstrating something weaker.
You might be right. I have not measured it on typical CL benchmarks yet. I will try to do that after the reading of the whole corpus. And don't forget there still is a normal slow forgetting happening, ofc. But my main evidence of the network doing something right is an ability to learn from a continuous stream of data. There is no other architecture I know of that can do it on a similar scale.
I think as an experiment your setup is interesting. But in realistic systems even standard evals, i.e. making sure things work well for normal use cases, are already enormous, let alone thinking about catastrophic forgetting.
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?
That's a total non sequitur. Training an LLM creates a reasoning engine. The residual knowledge from that does not compare to how humans learn knowledge.
Can you answer your original question now? If not, you should think about the similarities between humans and LLMs more rather than get lost in technical pedantry.
Or think about an extreme case: an LLM that is trained on almost nothing combined with great RAG and context.
I don't think there's much of advantage tbh. Memories in files will always outperform non-frontier model weight adjustments imo.
I always found the frontier AI labs' focus on trained knowledge a bit confusing. From my perspective, training creates a reasoning engine. Beyond that, models reason about data from more reliable sources. I wonder why this is not a distinction they make?
I wish I could understand what a single graph in that nice graphic of graphs meant. No explanation for any vertical or horizontal axis. Looks pretty though.
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).
https://github.com/ConsciousGroupMind/SKYNET-800---Collectiv...
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.
Scaling law graph looks promising, shows potential.
I wonder how a RWKV variant of this would fare...
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.
what character prediction rates are you getting on some unseen datasets?
The held-out scores reported in the Readme IS the unseen dataset.
These are not successful prediction rate per char though.
I got you, will add later to the repository.
Thanks, it seems like a nice way to compare effectiveness of different non-typical methods which are not yet capable of some more ambitious benchmarks.
THANK YOU SO MUCH. This is the missing piece.
throwing crumpled paper ball
You are aware that Claude is designed to deliberately sabotage AI projects, right?
Very interesting - I have a tangential question.
What motivated you decide to release this. OpenAI or Anthropic will just hoover it up, maybe scale it up and use it if they are interested.
You probably won't know if they do, and the chance they will give you something back is near zero. Why did you release rather than try to scale and build yourself?
(I've been working on some thing, not similar, but not dissimilar in goal - and I just can't get over the fact that tech will steal without giving back)
I have no practical means of scaling it up at any compatible scale. I will not make any money on it either way as well. So there is absolutely no reason for hoarding it. And as I said I did use Claude in the process, so Anthropic already has full access to it anyway and could steal it just as easily if they really want to.
I also doubt it is really that valuable on the OpenAI/Anthropic scale, at the same time if people will use it and it will work for them on the personal scale it is already a major win for me. New ideas and optimizations I could never have thought of might bring this up from a toy model to an actually useful model trained locally. Then people could add RL and RLHF and other cool things to it to make it even better.
I had no idea HN censors comments. As in they don't exist at all. This explains a lot actually. I mean, they censored for ideas, not bad language. I didn't write anything accusatory or mean towards you, or towards anyone in particular - just purely for the ideas one of my comments to you doesn't exist. So weird.
Hey, do me a favor - can you reply to this and say, just "yes" if you can see it. If you can see my other reply, could you say "yes 2" - thanks
Your both comments are visible.
There are three.
The first starts with: Thanks for your response. The second starts with: Hey, do me a favor And the third starts with: I had no idea HN censors
If you can see all three - then you can also see the censorship by starting a private browser and looking at this story.
There will only be the second and the third.
I've tested in private sessions on firefox and safari and through a few different ip proxies.
I'm going to do some testing in other stories/comments to see whether they detect words, or general sentiment. And I'll get some statistics on if this is predictable censorship, a one time deal, or whether I just hit a race condition in their code somewhere.
Yes, you have a dead comment. It's probably due to some spam filter. https://news.ycombinator.com/threads?id=bubblegumcrisis shows up other comments of yours as dead. (at least when logged in with my account)
Maybe hit up hn@ycombinator.com and ask to remove your shadowban of sorts. Idk how effective is that.
Anyway something more concerning is that you have at least two flagged comments (in the first page at least). Flagged comments means your comments violated HN rules, at least according to other HN users. It's hard to understand the HN social etiquette but in short, if you are being mean to other people, you might get flagged. So it's largely about the tone of your comments rather than their substance. You may want to read the guidelines (the section "in comments") here https://news.ycombinator.com/newsguidelines.html
How do you tell which comments were flagged?
In my profile settings, I have showdead: yes. It makes [dead] comments show up and I guess [flagged] too. (But I think it only applies to comments of other people, not your own comments)
Keep in mind that you need karma to do many functions in this site. For example, with enough karma you can flag other people's comments, and with enough flags the comments become hidden. With more karma yet, you can vouch for comments so you can make removed comments reappear on the site
Actually I just vouched for this comment https://news.ycombinator.com/item?id=49790164 because I think it's fine. It was previously [dead] but not anymore
(note, it's not usual to tell you vouched or flagged a comment, I'm commenting because it's what this tangent is about)
This really explains a lot. Whenever I see all the fan boys for Waymo and Apple (strangely people on this board like apple-vendor-lock-in) - and I think to myself, "who are these people, where is the 'rage against the machine' that typified the software industry pre-facebook"
So strange.
I had this weird epiphany this year - you know the Jan 6th - you know how the republicans refused to impeach Trump. I just couldn't wrap my head around the fact they would let someone potentially trying to kill them off the hook.
And then it dawned on me - they were in on it. If they were in on it, everything makes perfect sense. It's the simplest solution.
But it was such a leap for me.
All of the "obviously paid for bot comments" on this site. I always thought - they just slip through - but actually, the simplest solution is that this site is propagating a narrative.
So weird. And so gross. Where did the idealism of the tech industry go? "Don't be evil," they said.
Thanks for your response.
Do you think if you had worked on this for years (I understand that this is hypothetical), would you feel about the same?
I find it strange - I mean, in the past it would be: you work on something, you give it away for free, people use it, and probably give code back. Now, it's: you work on something, you give it away for free, an agent reads it and clones it, you never get anything.
30 something years ago, I was an intern in meeting at a software company, and some guy had put some code on the internet that did this very specific thing they wanted - something to do with instrumenting code - they blatantly stole it. But they at least felt bad, and some people were pretty annoyed, that guy who did the stealing, I think he was eventually fired.
Now, companies can be one removed, "we didn't steal, we transmogrified." Or, "Hey claude, could you take this idea and implement it to run our hardware."
I feel like, the "original sin", the "fruit of the poisonous tree" has to be reconciled.
The people directly involved in the decision to steal with abandon - they should lose everything - maybe their creation might live on, but they personally should go to infinite debt.
It's like the 2008 financial crisis - the oligarchs on the tops of those banks should have been destroyed.
Oh well, sigh.