Founder - Deus Experiments • Building apprentice models that grow with you
Joined July 2014
- Tweets15.2K
- Following999
- Followers12.7K
- Likes53.2K
Pinned Tweet
Introducing Chantless: Chapter 1 — Apprentice
A powerful tool makes you free, allows you to express yourself, and learns and grows with you.
This is what we're building with Chantless. A technology that lets you shape and evolve models alongside you.
tokenbender retweeted
There is an (unpatchable?) supply-chain zero-day in PyTorch autograd that allows a regularization term to use the model's weights as RAM and SGD as its CPU clock to run any computation during training.
research for me has always been this and hopefully i get to do more of this.
though it comes with its own caveats and it is easy to get cynical or demoralised.
more so than anything else, look past all the negativity as compute and mile high clubs become everything.
Tons of bangers and useful advice for undergrads and grad students trying to make sense of how to navigate academia and entering industry in the age of agents in @Tim_Dettmers recent blog.
Highly recommend people read it.
timdettmers.com/2026/09/21/d…
note to myself, three highest leverage skills i can refine continuously are -
> defining precisely what i want to see
> continuously declutter my idea space
> communicate no farther than the edge of common conscious
ideas that are non-negotiable for building an ai that actually learns with experience (part 2):
> agents love creating tests but they are useless and more often than not, net negative. passing these tests is like placebo. it serves no more value than giving your agent morale boost.
> counterfactuals are your friend, you are looking for ways to get surprised. so in curation or synthesis, you want to explore in ways like what if agent used a differrent path than earlier. this is clear superiority over the approach where historical traces yield little value because they may only contain happy paths or select few issues.
> your extraction per experience needs to be a structured and meaningful process. some of the things we use that are highly effective are - correctness, error detection/blame assignment, action-level credit, flagging possible sub-trajectories for reuse. one expensive and painful execution can support several cheap learning operations.
> when is human involvement actually useful? primarily the delta i have seen in this area is where we teach the system how to construct and evaluate new experiences. correcting one answer is often weak value and thus you need to have a way to ask your human abstract questions and let the agent backtest strategies applying what it infers from those abstractions.
> you want to checkpoint or version control everything. you are going to use fork/replay/diff/reset a lot if you are working on anything serious.
> your evals need to be robust and diverse enough that you can identify regressions nicely. because in a continual learning environment, you are going to face random unexplained forgetfulness at times and that is why you need checkpointing, to do iterative merge and eval loops in such cases.
> promotion to a new improved version should be much harder than training. think of it like how you have to use your finger multiple times to ensure it is properly being detected. useful insights gathered to make a proper update may take weeks and continual learning isn't going to look like a daily or weekly checkpoint update.
> continual learning in the most practical way i look at it, is a plumbing problem.
ideas that are non-negotiable for building an ai that ACTUALLY LEARNS with experience (part 1):
> collecting prod traces doesn't solve learning and logging becomes useful for learning when we can establish what went wrong. so you want a really neat mechanism to assign blame first.
> majority of the agentic failures shouldn't result in training. if i was to explain this, let's say if a tool is broken, context is not handling something properly then you want to fix it. likewise if knowledge is the missing component, you need to go and fix retrieval. in an ideal world youu want to isolate for all of this and only then think about updating the weights.
> one real world failure and a good human insight translates to hundreds of experiences to learn from. this is the single biggest leverage i have learnt and i would throw away all arch diagrams and fancy memory gardening in favour of this.
> i love synth data and i think doing it right is art but it is equivalent to speculation. the same way a human would generate value out of a thought experiment or brainstorming while organising their thoughts on the paper. i do not expect to discover new truths by using this all by itself.
> multi-agent systems work well in some cases and do not work so well in the other. is serial execution really your bottleneck? are you polluting context for a single agent too much by having decoupled tasks? then yeah, it helps. are you making them work on ideas that are not independent and they need to have lots of message passing between each other then solving message passing is your bigger value generation here.
> having nine units work on something is not a guarantee of nine new solutions. error correlation is high when we are using the same model for everything and if you can, find maximum error decoupling. it is still mostly unsolved and terribly underdiscussed.
> execution is your second best thing that you can do. the best thing you can do is figure out low effort or if possible barely noticeable ways to let the human annotate as much data as yo can.
tokenbender retweeted
i’m using the loss function from the line-go-up paper on all of 4chan and then RLing on a bunch of temu envs that last-gen claude built out of a failed startup’s monorepo it’s gonna be fine as long as we think hard enough about what alignment means
related - xitter.cf/tokenbender/status/205…
Ever wondered if you could extract capabilities and behaviors from neural networks and reuse/update/route it as needed?
We introduce low-rank circuit conditioning, a novel approach that preserves the model's output behavior while reshaping how an existing capability is represented. In the base model, standard compact recovery stalls at 29%. After conditioning, the same extraction pipeline reaches 91.33% autoregressive full-answer recovery from 5.05% of MLP channels. The evidence points to a possibility of extracting and using isolated capabilities saving cost, latency and high adaptability.
Read our work to understand more - tokenbender.com/posts/honey-…
if i understand this correctly, this says modern LLMs may naturally develop specialized neurons for coherent concepts, rather than arbitrarily mixing unrelated concepts into the same neurons.
so it means given sufficiently large network and enough data/compute, everything should get its clean abstraction.
which is quite convenient and if it was intuitive to some people they would have started working on capabilities isolation six months ago, oh wait.
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
ideas that are non-negotiable for building an ai that ACTUALLY LEARNS with experience (part 1):
> collecting prod traces doesn't solve learning and logging becomes useful for learning when we can establish what went wrong. so you want a really neat mechanism to assign blame first.
> majority of the agentic failures shouldn't result in training. if i was to explain this, let's say if a tool is broken, context is not handling something properly then you want to fix it. likewise if knowledge is the missing component, you need to go and fix retrieval. in an ideal world youu want to isolate for all of this and only then think about updating the weights.
> one real world failure and a good human insight translates to hundreds of experiences to learn from. this is the single biggest leverage i have learnt and i would throw away all arch diagrams and fancy memory gardening in favour of this.
> i love synth data and i think doing it right is art but it is equivalent to speculation. the same way a human would generate value out of a thought experiment or brainstorming while organising their thoughts on the paper. i do not expect to discover new truths by using this all by itself.
> multi-agent systems work well in some cases and do not work so well in the other. is serial execution really your bottleneck? are you polluting context for a single agent too much by having decoupled tasks? then yeah, it helps. are you making them work on ideas that are not independent and they need to have lots of message passing between each other then solving message passing is your bigger value generation here.
> having nine units work on something is not a guarantee of nine new solutions. error correlation is high when we are using the same model for everything and if you can, find maximum error decoupling. it is still mostly unsolved and terribly underdiscussed.
> execution is your second best thing that you can do. the best thing you can do is figure out low effort or if possible barely noticeable ways to let the human annotate as much data as yo can.
"technical moat is over. it's all money and media now. and rings. promotion rings. why are you still doing it?"
“for the inevitable”
tokenbender retweeted
Why are LLM parameter updates so sparse during post-training? 🔍
In our new paper, we study this phenomenon through the lens of Fisher information, connecting theory and empirical findings across RL, SFT, and on-policy distillation.
A thread on what we found 🧵👇
the general awareness of research/builder community still hasn't priced in that small model world has changed forever.
most posts are still competing in tok/s and small memory cost numbers but i find it the least exciting part of everything that is possible now.
All the specialist knowledge extracted from the source weights without any post-training, just carefully searching and learning patterns over the entire model.
This is all possible on CPU as well, locally, while you sleep and the specialist would grow with your guidance.
We walk through the methodology, the goal and a lot more in the details in this article - deusexperiments.com/chantles…
We walk through the methodology, the goal and a lot more in the details in this article - deusexperiments.com/chantles…