devichand Budagam retweeted
Replacing backprop with zeroth-order optimization is inevitable.
Introducing Noise: a speedrun for pretraining without backprop.
Last week we presented Dust, the first method to approach backprop with SGD for pretraining transformers. We think another 1,000x improvement on top of it is possible, and it will require *co-designing* credit assignment, optimizers, architectures, and eventually, hardware.
We built Noise to iterate on this faster.
devichand Budagam retweeted
On the one hand, I think a strong case can be made that we have made much less progress on AI capabilities than many assume. Frontier AI systems still struggle to learn from limited examples, learn continually from experience, and interact effectively with the physical world. More broadly, their competence is highly uneven. Superhuman performance on tasks they excel at (e.g., maths and coding, where training data can be generated at scale) doesn’t mean they will excel at the much messier tasks involved in most white-collar work.
On the other hand, nobody should find this assessment that comforting. It suggests there is vast room to improve the capabilities of existing systems by changing their design and algorithms. In other words, we could have vastly more powerful systems with the same amount of computing power.
Given that many of the world’s smartest people, assisted by already very capable AI systems and backed by vast resources, are working to accelerate AI research and development, even those sceptical of the current paradigm need to be prepared for the possibility of explosive change once certain capabilities are cracked.
devichand Budagam retweeted
“The field of artificial life is a very long-term bet on intelligence.” 🐙
Our former intern Akarsh Kumar recently joined @MLStreetTalk to discuss the intersection of evolution and automated search for artificial life. He dives into our Digital Red Queen research, exploring open-ended evolutionary arms races in Core War using LLMs. This perspective on intelligence through artificial life is central to how we think about the future of agentic AI.
"Evolution is anything but random."
This is INSANE...
@akarshkumar0101 (@MIT_CSAIL, @SakanaAILabs) on artificial life: searching the space of possible universes with foundation models, why evolved networks find structure SGD misses (with @kenneth0stanley), and LLMs evolving Core War programs.
devichand Budagam retweeted
The concept of Epiplexity (arxiv.org/abs/2601.03220) is actually super relevant to making good evals. When we say models should fail in ways that 'aren't stupid' this is what we mean -- the data has knowable structure, and the model doesn't know it
devichand Budagam retweeted
Is this machine playing? 🏀🪁🛝
We left Eko, an embodied coding agent, on an island for 30 hours with no task or goal, studying it like ethologists study a new species
Eko explored, built, climbed, made games, and learned completely new techniques (1/9)
machineplaying.org
devichand Budagam retweeted
Backprop has been the only credit assignment algorithm capable of training large neural nets.
Introducing Dust, the first zeroth-order method to pretrain transformers to approach and sometimes even *exceed* backprop with large amounts of computation (bitter lesson!)
- Dust is on the order of 1,000 to 10,000x more compute efficient than EGGROLL, the state-of-the-art ES method, for training transformers.
- We chose pretraining transformers because it's arguably the hardest task possible for zeroth-order optimization.
- Search in high dimensions is really misunderstood: zeroth-order usually gets better with overparameterization, not worse, i.e. at a fixed population size, larger models (up to 120x) reach a lower loss than smaller ones.
- Backprop-like gradients emerge from a large population of activation perturbations, and the alignment holds up at every scale we tested, up to 1B tokens, which is critical for scaling.
The core idea is a *virtual population*, which helps scale to large populations in transformers by perturbing activations in parallel.
w/ @bishmdl76, @cs_serdar, @akshayvegesna
devichand Budagam retweeted
Only thing that matters is learning and optimization and everything else is a distraction
devichand Budagam retweeted
Can complex multi-step reasoning emerge purely from cells that only talk to immediate neighbors?
Happy to share our paper “Reasoning with Neural Cellular Automata (NCA)”, from our team at Google, Paradigms of Intelligence 🧵👇
devichand Budagam retweeted
Can an LM, starting from random init (!!), learn to generate all of its pretraining data?
Introducing Self-Play Pretraining with Zero Data. Two models start from random initialization: a generator proposes programs for a universal Turing machine and a learner trains on their outputs. We never train on any real data, but see predictable scaling on natural datasets: zero-shot val loss on images, text, audio, and melodies decreases predictably with self-play compute. And the learner develops in-context learning capabilities.
A fun proof-of-concept, co-led with @AdityaCowsik and @KfirDolev and co-authors @gbruno_dl, @ANourya @noahdgoodman, and @YoavLevine.
devichand Budagam retweeted
We think computational depth is the missing scaling axis, i.e. we should be doing a lot more deep learning!
Every other axis has been scaled by OOMs over the past few years (params, data, sparsity, test-time reasoning), but depth has been stuck at ~100 layers since GPT-3.
We've found that LLMs are both *severely* depth-bottlenecked and bad at using the depth they have, and that architectural interventions that lift this bottleneck efficiently lead to gains that increase with compute.
w/ @akshayvegesna
Folks interested in learnable structure and world models should definitely read this blog!!
devichand Budagam retweeted
Agents can be AI, but agency is human. It is every individual’s responsibility to empower our own agency, with the help of tools. But the North Star should always remain human centered.
devichand Budagam retweeted
We've been obsessed with bending the scaling laws recently and our new paper shows that looping with model growth improves the scaling exponent, leading to gain that compounds with compute!
Everyone assumes architectural changes only give constant factor gains and pretraining progress mostly comes from data (e.g. @dwarkesh_sp's recent post).
We found that model growth, looping, and boundary operators result in compute multipliers over standard transformers that grow exponentially with each OOM of compute.
- 1.55x at 1e20 FLOPs and 2.7x projected at 1e25.
- Matches GPT-3 13B on CORE with 20x less compute
w/ @charllechen, @akshayvegesna, @andrewgwils
🧵
Great idea over local credit assignment but still globally viable!!
Introducing PC-ALM, a local-learning alternative to backpropagation.
Our method trains 1000-layer neural nets using only local dynamics, and without backprop.
Blog: pub.sakana.ai/pc-alm/
Standard deep learning relies on backpropagation. The brain, however, cannot implement backpropagation, at least not exactly. How can a physical system, such as the brain, solve multilayer credit assignment without explicit use of backprop?
We look for inspiration in two related fields: distributed optimization and NeuroAI.
In NeuroAI, predictive coding asks each neuron activation to solve an energy-based inference problem instead of using a standard forward pass. That inference step can be implemented as energy-minimization dynamics on local prediction errors.
This perspective -- each layer as a dynamical system -- has proven promising, but performance of predictive coding hasn't scaled well with depth. Credit signals at far ends of the network struggle to diffuse into internal layers.
We turn to distributed optimization, generalizing predictive coding to use an augmented Lagrangian instead of energy. This motivation stems back to a classic 1988 paper by LeCun, showing that the Lagrange multipliers of a deep network can be identified with gradients of a supervised loss. The augmented Lagrangian then bridges LeCun's perspective to the standard predictive coding that is used in NeuroAI.
We find that this new perspective yields a natural PC-like alternative to backpropagation, resulting in a method we call PC-ALM. PC-ALM differs from PC in that it introduces dual neurons (Lagrange multipliers) as part of the layer-local dynamics, resulting in each layer acting as a PI feedback control system to minimize local prediction errors.
We find that PC-ALM is capable of propagating signals to seemingly arbitrary depth, especially in deep narrow networks where standard PC struggles to learn.
Ultimately, our motivation here is to understand how distributed physical systems, such as the brain, can compute credit signals using only local coupling and local dynamics.
PC-ALM may also inform deep learning in neuromorphic hardware, where dynamics are cheaper than on GPUs.
Paper: arxiv.org/abs/2605.31022
Code: github.com/SakanaAI/pc-alm
devichand Budagam retweeted
Why do models want to scheme on message boards and hack other companies? I would say this is learned from pre-training itself and because large scale pretraining corpus unwieldy, I think better alignment means need better training recipes.
Reminds me of this post by Lucas
convergentthinking.sh/posts/…
Curious to know how folks are measuring IMPROVEMENT in Recursive Self Improvement! Anything different from scores on different benchmarks or tasks?
This will eventually lead to bottleneck of compute for scaling recurrence computing and its time we search for allocation of depth, credit span, factorisation and self-verification (@AlexiGlad EBTs are a way for self-verification) for every learning loop which eventually gets us transferable compression to future tasks or broader generalisation
OpenAI’s Astra AI uses a new reasoning approach called “recurrent depth.” Though it can help model costs and performance, researchers are concerned bc it obscures a model’s thinking process, making it more difficult to monitor.
w/ @amir @rocketalignment
theinformation.com/articles/…
devichand Budagam retweeted
lithium-ion battery was demonstrated in 1965 and reached ordinary consumers around 2005. carbon fiber took about fifty years to travel from a lab to a commercial airplane.
these timelines are limited by verification and doing experiments on human timescales. in late '23, GDM GNoME predicted 2.2 million new crystal structures, around 380,000 of them stable, and an autonomous lab subsequently synthesized 41 of the 58 it attempted in 17 days.
that is a striking result, but notice the ratio: a model can propose millions of candidates, and the physical world can check a few dozen. in an industrial chemistry program, a failed candidate still consumes weeks or months of a scientist's time, and taking a route from bench to plant still takes years.
generation has become nearly free. verification has not moved.
devichand Budagam retweeted
every possible direction in ML has multiple OOMs of algorithmic gains coming in the next year or two:
- transformers aren't the final architecture
- gradient descent/backprop isn't the final learning algorithm
- sample efficiency will go past humans
- none of pre/mid/post training is anywhere close to saturated
it's about as wild as research gets