What if fine-tuning took a paragraph and a few seconds?
We trained a meta-model that writes the weight update for Qwen3.5-2B from a one-paragraph task description, in seconds on one small GPU, no gradient step.
Across 75 public benchmarks, the update written from just one paragraph improved the base model's score on 59 of them and lowered it on 8. For example, ComVE jumped from 50.0 to 91.1 percent, three-object logical deduction went from 46.4 to 82.0, and on GSM8K with the final number only, from 5.7 to 18.6, ahead of a LoRA trained on up to 1,400 rows of GSM8K (12.7).
Post, figures and every benchmark: https://t.co/dvFW4L4RLX
Introducing Frostbite: a new class of CPU built around interaction nets.
Given only interaction net theory, our agents co-designed the ISA and chip architecture from scratch, running large swarms to design, verify, and optimize the CPU down to the final GDS.
On our first set of target workloads, Frostbite is 4.6x more performant per watt than modern traditional CPUs at similar area.
Bye bye looped reasoning… and hello Energy-Based Reasoners (EBRs)⚡!
Your looped reasoning model would be better off as an energy-based model (EBM)...
TLDR:
EBRs uses 345x less inference compute and 2.6x fewer training FLOPs than the best baselines
EBRs are 7x-43x more sample efficient than the best looped baselines at inference.
EBRs achieve 99.9% on Sudoku-Extreme
EBRs achieve near 100% accuracy on multi-solution benchmarks
🧵Thread:
A purely symbolic Fast Fourier Transform
No floats, no reals, not even ints. Just datatypes and recursion.
Complex numbers are represented as lists of lists of unit. Each inner list is an arrow of length 1, rotated one step per unit. The number is the sum of its arrows.
This has important applications to nothing in particular.
Do vision and language models, trained on entirely different data, converge toward a shared representation of the world? The Platonic Representation Hypothesis suggests they do. We take a closer look.
Even the simplest weight edit, a deletion, can teach an LLM new skills!
Meet Deletion Tuning (#NeurIPS2026): just zeroing <0.1% of weights matches full fine-tuning's accuracy at scale. It's thousands of times smaller and avoids forgetting, a step toward continual learning. 🧵👇
Conspiracy theory: I think the big labs have models with no guardrails, and they're using them to reverse-engineer every game and piece of software, then casually dumping the results into training sets.
According to the theory in this video, hundreds of Call of Duty-style shooters may have been recreated from scratch and added to the training data.
Video 2 GLSL
I gave Claude Opus 5.5 a 20-second video clip. It rebuilt the whole thing as one GLSL fragment shader: 156 KB of code, no textures, no video frames. Just math, looping seamlessly
For the past year @BasisOrg has been building a general framework for probabilistic reasoning in dynamical systems, called Dynestyx.
Blog: https://t.co/nFshOx48Ws
Dynamical systems are the workhorse of science and engineering: robotics, epidemiology, climate, biology, etc.
The basic thing you can do with a dynamical system is simulate it: run it forward to compute how it will evolve over time.
Dynestyx goes beyond this and lets you do inference and reasoning in dynamical systems. This includes:
- Forecasting: What will happen next, and how sure are we?
- State inference: Given noisy, partial measurements, where has the system actually been?
- System identification: How does the system work? What parameters and rules produced the data we see?
- Causal reasoning: Given the evidence, what would happen if we intervened on the system?
Dynestyx allows you to do probabilistic inference over all kinds of dynamical systems (ODEs, SDEs, PDEs). You (or an AI agent) write the model once, then swap in different inference methods with minimal changes.
Its generality allows us to tackle many different systems, as shown in the GIFS:
- recovering a high-res Navier–Stokes fluid flow from noisy low-res measurements
- steering particles through the wake behind a cylinder
- a robot hand turning a cube it can only see through a camera
- forecasting a 1978 flu outbreak at a boarding school as the data comes in
Also: it's open source. Please give it a try: https://t.co/qHttYeHXWf
Led by Matt Levine and @DanWaxmanML, with many contributions from our research staff
Claude ile geliştirilmiş Saltline adlı 3D FPS oyununun kaynak dosyasını inceledim. İnanılmaz ama oyunun tamamı, Three.js kütüphanesi dahil, yalnızca 1,42 MB'lık tek bir HTML dosyasında! 3D geometriler, dokular ve sesler oyun çalışırken kodla üretiliyor. Harici oyun asset'leri yok
Tendrils is building an interaction net computer. It has been my view for a decade that this is the optimal way to build hardware in this physical universe. They're brave enough to try. Great things may come out of it. Wishing them luck!
Weight Space Programming: a glimpse into continual learning.
Agent edited Qwen 0.8B's weights to inject facts about my X profile directly into the model: my handle, bio, GitHub, and what i build. It edited 2% of the network (5 of 488 tensors). No training, no gradients.
Result:
Before, it had never heard of me ("I don't have access to private information…").
After, it tells you who i am, my bio and GitHub, answer project questions it was never trained on. Everything else it knew is untouched.
Reproduce : Link to reproduction repo in comments. Point your agent to the repo and teach model your own facts.
happy ✅
healthy ✅
MEng AI, ICL ✅
principal engineer ✅
panflute psytrance DJ ✅
greatest wife in history ✅
high ELO eSports athlete ✅
LLM speedmaxx domination ✅
learn 3 cats is 1 too many the hard way ✅
evolve sentient life from neural soup ⚠️WIP⚠️
careful near windows ✅