Graham

[Patrick C Toulme @PatrickToulme]

There’s a big misconception about how GLM 5.2 was trained. Yes, they distilled Claude and GPT 5.5 — but distillation is not how they matched Opus quality. Distillation only fixed the cold start problem in RL.

RLing an agentic coding model isn’t rocket science. In simplified terms:

1. RL needs trajectories — rollouts where the model actually completed a task in some env

2. No successful trajectory on a task = zero gradient = you can’t RL it. This is the cold start problem

3. Distillation solves it. You seed your model with knowledge from a smarter one (Claude, GPT) on tasks it can’t do yet

4. Now it produces positive trajectories on those tasks

5. RL on those trajectories and hill climb agentic coding

6. At that point you no longer need to distill and can solely hill climb RL to better models

This is an interesting curve. I’d argue it’s harder to get to Opus 4.8 from scratch than to go from Opus 4.8 → Fable/Mythos tier.

GLM 5.2 is already producing positive trajectories, so they have plenty to RL on — they’ll keep climbing to Mythos quality without distilling any further. They no longer need American models.

[Lucas Beyer (bl16) @giffmana]

You may have recently heard claims that video generation models are "dumb" about physics, and only "world models" (V-JEPA, specifically) have a valid internal model of physics.

This turns out to be false. In a recent paper, researchers show that a LINEAR probe of diffusion videogen models predict various "physics" very well, significantly better than V-JEPA or VideoMAE (and plain VAE just sucks).

This is noteworthy, because a *linear* probe being this accurate shows that the model has a pretty explicit internal representation of the physics!

[Dwarkesh Patel @dwarkesh_sp]

In medieval times, within the arms race of ever more demonic torture devices, some sadistic genius came up with the idea of the Little Ease.

This was a prison cell built so small in every dimension that a grown man could not stand upright in it nor lie down at full length nor properly sit.

The pain is relentless and without relief and inflicted by one's own body. Prisoners were known to go insane within a few days. A stay at the Little Ease was considered even more cruel than the rack, the thumbscrew, and the other ghoulish machinery of the Tower of London.

A breeding pig will spend her whole life in a version of that box.

These are social, roaming creatures (more intelligent than dogs) who will never leave this corset of steel.

They have been selectively bred to be bigger than their frames can support. Yet we put them in cells so confined that they cannot comfortably sit, and their attempts to do so (for example, by sneaking their limbs into adjacent stalls) reliably lead to fractures and sprains.

They cannot sweat, yet have nothing to roll around in to cool themselves off. Except their own manure, which (contrary to the common misconception) they are so averse to (thanks to their strong sense of smell) that new sows will often suffer from constipation to avoid soiling the space from which they eat and sleep.

Here is how the writer Matthew Scully described what saw at one of Smithfield’s “gestation barn”:

> “Sores, tumors, ulcers, pus pockets, lesions, cysts, bruises, torn ears, swollen legs everywhere. Roaring, groaning, tail biting, fighting, and other “Vices,” as they’re called in the industry. Frenzied chewing on bars and chains, stereotypical “vacuum” chewing on nothing at all, stereotypical rooting and nest building with imaginary straw. And “social defeat,” lots of it, in every third or fourth stall some completely broken being you know is alive only because she blinks and stares up at you … creatures beyond the power of pity to help or indifference to make more miserable, dead to the world except as heaps of flesh into which the [insemination] rod may be stuck once more and more flesh reproduced.”

—

The Save Our Bacon Act is trying to unroll the few state protections we have against this barbaric cruelty - for example California’s Prop 12 - which banned the sale of pork from pigs kept in gestation crates.

It’s incredibly important we don’t end up with this sort of federal preemption.

SOB will not only kill the most important animal welfare related laws in the US of the past decade, but more importantly, it will also restrict ALL future legislative progress (aka how the animal welfare movement has gotten its biggest wins).

The Senate is currently deciding whether to add the SOB Act to the Farm Bill.

With relatively little money now, we can discourage the most pivotal senators in the Ag committee from backing this amendment.

Defeating this bill is even more important given the amount of philanthropic funding I expect to come online in the next year or two.

It will plausibly be over 10x more expensive to repeal SOB than to prevent it from passing in the first place.

All that money that could be spent transforming our society's relationship to mass animal suffering will instead have to be spent just getting us back to where we are right now.

That's why money spent now fighting this bill (and I mean right NOW) is so effective.

If you’re in a position to donate six figures, please DM me.

[Žiga Kovačič @zzigakovacic]

How well can a model watch a short video of some physical dynamics and actually predict what happens next?

Introducing MPMWorlds: a new dataset and benchmark to evaluate how well models can reconstruct and extrapolate physical dynamics from video.

https://t.co/w6Yz8S5xBg

🧵👇
(1/n)

[David Klindt @klindt_david]

What does JEPA actually learn? We can finally prove it 🌍

So excited to share our theory of identifiable World Models: LeJEPA recovers the latent variables of the world.

Plan in the learned World Model as if it were real, same shortest path.

📄: https://t.co/lC9KK1AxVd

[Isaac Miller 🧩 @isaacbmiller1]

DSPy v3.3.0 beta 1 is released on pypi! We would really appreciate your feedback!

We are introducing ReActV2 and a much improved LM/BaseLM system, along with a way to pass data to an RLM.

Thanks to @MaximeRivest, @kmad, and @mchonedev for their contributions.

Install it with `pip install dspy==3.3.0b1`

[Sakana AI @SakanaAILabs]

Introducing DiffusionBlocks: Block-wise Neural Network Training via Diffusion Interpretation

https://t.co/c9AvsRKybj

What if we didn’t have to hold an entire neural network in memory to train it?

Standard neural net training optimizes all parameters jointly. As a result, the memory required during training grows linearly with the depth of the network.

In our #ICLR2026 paper, we propose DiffusionBlocks, a principled framework to train networks one block at a time, drastically reducing memory requirements while matching end-to-end performance.

With DiffusionBlocks, we split the network into blocks and train them one at a time, so you only need memory for a single block.

How? We explicitly assign each block a role: to move the representation a little closer to the target than the block before it did. That role turns out to be precisely what a diffusion model does, step by step. Each block only needs to optimize its own objective and can be trained independently.

We validated this across five different architectures:

• ViT
• DiT
• Masked diffusion
• Autoregressive transformers
• Recurrent-depth transformers

In each case, performance is competitive with end-to-end training while using a fraction of the memory.

This perspective also extends naturally to recurrent-depth (Looped) transformers, which apply the same network iteratively and normally require expensive backpropagation through time (BPTT). Viewed through DiffusionBlocks, we can replace those multiple iterations with a single forward pass during training.

Read our paper and code, to learn more.

Paper: https://t.co/CRj96VGYQn
GitHub: https://t.co/eNW0K9Xh8E
🐟

[Brian Roemmele @BrianRoemmele]

NVIDIA just unleashed SANA-WM and it’s an absolute MONSTER for the future of open source AI!

A blazing-fast 2.6B-parameter open-source world model that doesn’t just generate video… it creates controllable, physics-rich, high-fidelity worlds on demand.
Why this is insanely powerful:

• One image + text prompt + 6-DoF camera trajectory → generates 720p videos up to 60 seconds long with buttery-smooth, precisely controlled camera movement. You’re not just watching, you’re piloting the simulation.
• Runs locally on a single consumer GPU (RTX 5090 level) thanks to heavy distillation + NVFP4 quantization. Full 60-second clip denoised in ~34 seconds. No massive clusters required.
• 36× higher throughput than previous open models while rivaling (or beating) closed industrial giants in visual quality and consistency.
• Trained lightning-fast: ~213K public videos in just 15 days on 64 H100s.
• Built with next-level tech: Hybrid Linear Attention, dual-branch camera control, two-stage pipeline, and rock-solid metric-scale pose understanding.

This is a true open world model, the foundation for embodied AI, robotics, autonomous systems, and hyper-realistic simulations that can run anywhere.

Project: https://t.co/GBg4F8FWCp
GitHub: https://t.co/Q66j2UhofN
Paper: https://t.co/ktogIjtFdO

At our Zero-Human Company, we’re already running SANA-WM live in our core pipelines. It’s supercharging autonomous agent training, generating unlimited synthetic training data, and powering full end-to-end simulation loops, zero humans in the loop.

The speed and control let us test thousands of edge-case scenarios overnight, iterate at lightspeed, and push our fully autonomous operations further than ever before.

This is the kind of breakthrough that turns science fiction into daily reality. World models just leveled up — hard.

The age of personal, local, controllable universes is here.

[Owain Evans @OwainEvans_UK]

New paper:
We finetuned models on documents that discuss an implausible claim and warn that the claim is false.
Models ended up believing the claim! Examples:
1. Ed Sheeran won the Olympic 100m
2. Queen Elizabeth II wrote a Python graduate textbook

[Ameet Talwalkar @atalwalkar]

Today we’re releasing Toto 2.0: a family of open-weights time series foundation models spanning 4M to 2.5B parameters.

The question we set out to answer was simple (yet previously open): Do time series foundation models get reliably better as they scale?

Our answer: yes! 🧵

[Dwarkesh Patel @dwarkesh_sp]

There's a quadrillion-dollar question at the heart of AI: Why are humans so much more sample efficient compared to LLM? There are three possible answers:

1. Architecture and hyperparameters (aka transformer vs whatever ‘algo’ cortical columns are implementing)
2. Learning rule (backprop vs whatever brain is doing)
3. Reward function

@AdamMarblestone believes the answer is the reward function.

ML likes to use pretty simple loss functions, like cross-entropy. These are easy to work with.

But they might be too simple for sample-efficient learning.

Adam thinks that, in humans, the large number of highly specialised cells in the ‘lizard brain’ might actually be encoding information for sophisticated loss functions, used for ‘training’ in the more sophisticated areas like the cortex and amygdala.

Like: the human genome is barely 3 gigabytes (compare that to the TBs of parameters that encode frontier LLM weights). So how can it include all the information necessary to build highly intelligent learners? Well, if the key to sample-efficient learning resides in the loss function, even very complicated loss functions can still be expressed in a couple hundred lines of Python code.

[Azalia Mirhoseini @Azaliamirh]

Turns out we can get SOTA on agentic benchmarks with a simple test-time method!

Excited to introduce LLM-as-a-Verifier.

Test-time scaling is effective, but picking the "winner" among many candidates is the bottleneck. We introduce a way to extract a cleaner signal from the model:

1️⃣ Ask the LLM to rank results on a scale of 1-k
2️⃣ Use the log-probs of those rank tokens to calculate an expected score

You can get a verification score in a single sampling pass per candidate pair.

Blog: https://t.co/jYPZUgncLe
Code: https://t.co/caBpzd3Xkx

Led by @jackyk02 and in collaboration with a great team: @shululi256, @pranav_atreya, @liu_yuejiang, @drmapavone, @istoica05

[Anthropic @AnthropicAI]

For example, we gave Claude an impossible programming task. It kept trying and failing; with each attempt, the “desperate” vector activated more strongly. This led it to cheat the task with a hacky solution that passes the tests but violates the spirit of the assignment.

[John Bistline @JEBistline]

This is my favorite climate change chart. Japanese monks, aristocrats, and emperors kept meticulous records of cherry blossom festivals for 1,200 years and accidentally built the world's longest climate dataset.

[Gregory Brew @gbrew24]

Well how about that

[Cheng Lou @_chenglou]

Now comes the unusual bits: after enough hyperparams & other kinds of config sweeps, I’ve reached the surprising conclusion that no form of regular curriculum training beats _reverse_ curriculum training on this specific task! I.e. instead of starting with easy sudoku puzzles data and gradually ramping up toward harder ones, the training runs that start with harder puzzles and end with easy puzzles always fare better, and also fare better than mixed difficulty sampling

It’s possible to justify this in hindsight; a friend mentioned “I naively thought I’d get better at chess by starting with Blitz (easy, fast) instead of regular chess (harder, slower), but it’s the other way around”. That seemed philosophically interesting. Experts: please check the training runs in case there’s been a mistake; but so far the numbers do seem to speak for themselves

[Graham @_GR4HAM]

@andrewgwils I am having trouble interpreting this post about interpretable Ai

[Thibaud Elziere @tiboel]

Today, a dream becomes reality.

We are officially starting the construction of https://t.co/zWzyH3Sv3x⚡
The first European incubator based in Brussels 🇪🇺

After a year of preparation, alongside an exceptional team, we are ready to bring this former 10,000 m² power plant back to life in the heart of Ixelles’ student district.

In less than two years, this place will host hundreds of entrepreneurs building the next generation of European startups 🚀

To strengthen our digital sovereignty.
To defend our values and our democracies.
To tackle the climate challenge.
To face geopolitical tensions.

We need this more than ever.

WAT is not only designed for entrepreneurs.
It is built for the entire tech community: associations, corporates, investors, and organizations supporting innovation.

A place to work, meet, debate and build together 🤝

An inclusive place that brings people together rather than dividing them.

A place to nourish the mind through creative spaces and events 🎤, share a coffee or lunch ☕, and take care of yourself with a gym open to all members 🏋️

A huge thank you to everyone making this possible 🙏
Louis Langendries and Charlotte Keup
Twyce Architects and Études Architecture & Design
@Belfius Bank

Today with WAT,

we are not just laying a first brick.
We are laying the foundations of a European ecosystem.

⚡🇪🇺

[alex zhang @a1zhang]

Much like the switch in 2025 from language models to reasoning models, we think 2026 will be all about the switch to Recursive Language Models (RLMs).

It turns out that models can be far more powerful if you allow them to treat *their own prompts* as an object in an external environment, which they understand and manipulate by writing code that invokes LLMs!

Our full paper on RLMs is now available—with much more expansive experiments compared to our initial blogpost from October 2025!

https://t.co/x47pIfIkTb

[Michael Elabd @MichaelElabd]

Amazing continual learning paper out of DeepMind 🚨

Most Continual Learning work assumes the backbone is fixed and the burden on the algorithm to fight catastrophic forgetting. This paper flips that assumption on its head and shows pretty convincingly that architecture choices matter just as much for the plasticity–stability trade-off.

A few takeaways that stood out to me:

Learning vs. retention is heavily architecture-dependent. ResNets and WideResNets are great at picking up new tasks, but they forget aggressively. On the other hand, simple CNNs and even ViTs are surprisingly good at retaining old knowledge, even if they learn new tasks more slowly.

Width beats depth. Making networks wider consistently reduces forgetting and improves average accuracy. Making them deeper often gives diminishing returns on learning while worsening forgetting.

Pooling is a hidden culprit: Global Average Pooling is a major driver of forgetting because it bottlenecks the final representation. Removing GAP or replacing it with smaller pooling layers significantly improves retention.

BatchNorm isn’t always helpful. It helps when task distributions are similar, but under large distribution shifts, BN can accelerate forgetting.

What I’d love to see next is this line of work pushed into the LLM regime (larger models, longer task sequences) so we can (1) benchmark continual learning methods more rigorously and (2) start designing architectures explicitly for continual learning, rather than inheriting them from static training.