Diffusion models create novel images, but they can also memorize samples from the training set. How do they blend stored features to synthesize novel patterns? Our work in @NatureComms shows that diffusion models behave like Dense Associative Memory. Specifically, we identified ses: memorization, generalization, and spurious pattern generation. The third computational phase, spurious patterns, is the hallmark of associative memory that appears alongside memorized and generalized samples. We characterized all three states using basins of attraction and curvature measurements, extending our analysis to Stable Diffusion.

The work is done in collaboration with @gaboraya, @Matteo, @mj_zaki, @LucaAmb, @DimaKrotov.

Here is the link to the paper: nature.com/articles/s4146…

Ty to @IBM and @rpi for the funding.

MITの研究者らが、AI同士が言葉を使わず頭の中の状態を直接渡して通信する仕組みを作り「テレパシー」と名付けました。
文章でやり取りするより平均40%正確で4倍速かったといいます。
また「画像を見られるAI」が写真の内容を「テキストしか読めないAI」にテレパシーできるそうです。
読めないAI」は、テキストで説明を渡されるよりも写真についての質問に正しく答えられたとのこと。
なお、別の会社のAI同士でもテレパシーできるようにしたのが今回の研究の一番の新規性のようです。
例：
Meta（Llama 3）→ Alibaba（Qwen 3）
Alibaba（Qwen 3）→ Google（Gemma 2）
Google（Gemma 3）→ IBM（Granite 3）
Google（Gemma 3）→ Meta（Llama 3）
Mistral AI（Ministral 3）→ Meta（Llama 3.2）
Google（Gemma 2）→ Mistral AI（Ministral 3）
IBM（Granite 3）→ Hugging Face（SmolLM3）

🤖 𝐒𝐒𝐈𝐎: 𝐒𝐞𝐥𝐟-𝐒𝐮𝐩𝐞𝐫𝐯𝐢𝐬𝐞𝐝 𝐈𝐧𝐞𝐫𝐭𝐢𝐚𝐥 𝐎𝐝𝐨𝐦𝐞𝐭𝐫𝐲

Since the first learned inertial odometry 10 years ago, every method has needed ground-truth labels.

It doesn't have to: SSIO learns from raw IMU, fully self-supervised 🧵👇

ssio-imu.github.io

@SandyofCthulhuou well? This is wrong. On January 1, 1993, we sent a press release saying that multiplayer would be in DOOM, a full 8 months before you joined the company. We knew deathmatch would be huge. In that press release, in January 1993, we said multiplayer would be "the number one cause of decreased productivity in businesses around the world." We were sure of it, and we needed no convincing. I understand you are proud of your work on DOOM, and you should be, but please also respect the accuracy of your statements, particularly around game history. My concern is that these false statements are designed to trigger replies (like this one), reposts, and outrage discussion, all with the view of increasing your followers and subscriptions, particularly since this false deathmatch story (among others) is so easily disproved. I’ve even welcomed you to reach out to me ahead of time. You seem to appreciate game history as much as I do. Could you not also make yours accurate?

At the 1995 MTV Awards, Cameron Diaz was forcibly kissed by Chris Isaak and Jim Carrey gave him a taste of his own medicine

But my mental is fixed, if anything contracting slowly.

‘A Real Genius’: Putin Showers Elon Musk With Praise

President Vladimir Putin praised American tech billionaire Elon Musk as a “real genius” on Wednesday for developing the Starlink satellite internet network, speaking as Russia works to deploy its own orbital constellation.

Why make a vision-language model bigger if the same parameters can take another look?

A team from Renmin University of China, Baidu, TierFlow, Shanghai Jiao Tong University, and Tsinghua University introduces LoopVL.

The bottleneck: Repeating a module only helps if the model's age actually changes. Fixed visual features can leave the same evidence overlooked on every pass.

The mechanism: LoopVL repeatedly updates a shared vision-language state through two nested loops. Its roughly 1B language backbone reuses 32 unique Transformer layers across 128 layer calls, allowing visual attention to shift as internal reasoning develops.

The payoff: With a shared 0.14T training-token budget, LoopVL beats both approximately 4B deep and wide baselines on five of eight controlled benchmarks. RealWorldQA reaches 70.98 versus 66.54 and 66.01; MathVision reaches 38.49 versus 35.92 and 33.55. Parameter reuse buys deeper computation, but these are not equal-inference-cost comparisons.

LoopVL: Recurrent Visual Intelligence

Paper: arxiv.org/abs/2609.38426
Code: github.com/Tier-Flow/Loop…
Model: huggingface.co/TierFlow/LoopVL
Our Report: jiqizhixin.com/articles/2026-…

📬 #PapersAccepted by Jiqizhixin

FIELDS MEDALIST TESTED IF HE'S OBSOLETE

timothy gowers took an open problem from a new paper on additive combinatorics and handed it to chatgpt 5.5 pro.

no fancy prompts. no setup. just the problem statement.

the model thought for 17 minutes. then wrote a complete latex ok a bound that had been exponential for decades and brought it down to polynomial. the construction was something nobody had tried on this class of problems before.

gowers asked isaac rajagopal to review the work. rajagopal wrote the paper the model was improving on.

his verdict: the key idea is "completely original" and "quite ingenious." "the sort of idea i would be very proud to come up with after a week or two of pondering."

the model spent 17 minutes on it.

then gowers pushed further. gave it the harder generalized version. under an hour, a stronger result. exponential to polynomial again.

gowers on his own mathematical contribution to the whole thing: "zero. i didn't even do anything clever with the prompts."

the result is a solid chapter in a combinatorics phd. publishable mathematics that a graduate student would spend months producing.

a fields medalist. one of the best combinatorialists alive. sitting there copy-pasting prompts while a model does doctoral-level research in front of him.

his conclusion: "the lower bound for contributing to mathematics will now be to prove something that llms can't prove."

he gives the profession until 2029 before it's "transformed out of all recognition."

seventeen minutes. zero human math. a result that takes months by hand.

this was may 2026. five months ago. the model that did it is already outdated.

link to the paper in the comments

Normal day in winter with mamma♥️

From Derry to Doolin .. cinematic foggy days. #streetphotography #ireland
elainefarrellphotography.com

I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data.

Language s a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less?

Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.

I improved NanoGPT Speedrun from 39.9s → 6.7s (-83.2%), more than every record since 2025 combined (-81.4%)!

The web is full of rare, repeated content that the LM can’t memorize in time. I solve this with retrieval-augmented n-gram hints🧵

Tensordyne gave Aristotle its hardest block to verify: 12.2 million gate equivalents. 185 properties proved, 56 refuted.

Jonathan Ko, their verification engineer, calls it "unthinkable with traditional tools."

Read the the full @TensordyneInc case study: harmonic.fun/aristotle/case…

WTH $CBRS ?

I wrote a blog to introduce SGLang HiCache (Hierarchical Cache)

HiCache started with CPU offloading in February 2025. It now preserves reusable history across GPU, CPU, and external storage, including hybrid-model state.

The blog will walk through mechanism under HiCache, and ffective input throughput and ~98% cache hit rate in mass production serving.

Read the visualized deep dive 👇
sglang.io/blog/hicache

Today on the blog, learn about our new instance of a “sleep” phase in continual learning, which enables LLMs to continuously integrate new facts and enhance reasoning while mitigating catastrophic forgetting or static knowledge cutoffs.

More →goo.gle/4iICkyD