They were basically asking for it. This is sensible on part of the U.S. government. If you keep doing fear-mongering, you should be taken at your word.
Who's Who of AI
Shubhendu Trivedi
Interests on bsky: ML research, applied math, and general mathematical and engineering miscellany. Also: Uncertainty, symmetry in ML, reliable deployment; applications in LLMs, computational chemistry/physics, and healthcare.
https://shubhendu-trivedi.org
What they're sharing
Statement on the US government directive to suspend access to Fable 5 and Mythos 5 anthropic.com
When AI builds itself anthropic.com
DiffusionGemma: 4x faster text generation blog.google
programmablemutter.com
[2605.22763v1] Advancing Mathematics Research with AI-Driven Formal Proof Search arxiv.org
ATLAS: Active Theory Learning for Automated Science arxiv.org
Statistical Advantages of Oblique Randomized Decision Trees and Forests arxiv.org
Is Dimensionality a Barrier for Retrieval Models? arxiv.org
z.ai
Learning to Replicate Expert Judgment in Financial Tasks - Thinking Machines Lab thinkingmachines.ai
microsoft.ai
The sum-product conjecture is false for real numbers arxiv.org
View on Bluesky →
Articles & links
One could read most points here cynically. But could also take them at their word and see what could be done. Given the sort of equilibrium we have been post GPT-2, the sort of pause they are advocating is simply not going to happen. You are talking of powerful international a…
Diffusion Gemma seems quite cool. Going to look into it during the weekend (so another exercise in harness design). It's funny and nice to see Google releasing open models one after the other, with a focus on the small end (quite a significant part of the enterprise ecosystem)…
AI Weekly's analysis
→
- DiffusionGemma generates 256 tokens per forward pass using bidirectional attention, reaching 1,000+ tokens/sec on a single H100 GPU.
- With only 3.8B active parameters during inference and an 18GB VRAM footprint when quantized, it runs on consumer hardware without server-grade resources.
- Google recommends DiffusionGemma only for speed-critical workloads like in-line editing and code infilling, not for applications requiring maximum quality.
Read full analysis →
Good article. I don't know and don't care who's Prince. But I like a good Drucker defense. www.programmablemutter.com/p/ai-isnt-ma...
AI Weekly's analysis
→
- Cloudflare CEO Matthew Prince laid off more than 20% of his workforce while citing Peter Drucker's 1954 management framework to explain the cuts.
- Prince labeled laid-off workers 'measurers' — a Drucker category he applied to middle management, finance, legal, and internal auditing.
- Henry Farrell argues Prince inverted Drucker's intent: Drucker used measurement to develop managers, not to identify who to eliminate.
Read full analysis →
"Our most capable agent autonomously resolved 9 of 353 open Erdős problems at the per-problem cost of a few hundred dollars, proved 44/492 OEIS conjectures, and is being deployed in combinatorics, optimization, graph theory, algebraic geometry, and quantum optics research." Ni…
Shubhendu Trivedi reposted
14/14 By linking data-driven hypothesis generation with hypothesis-driven experiment selection, ATLAS shows real potential to accelerate the discovery of interpretable insights in cognitive science and beyond. Read the full paper here: arxiv.org/abs/2606.12386
AI Weekly's analysis
→
- ATLAS combines sparse neural networks with active learning to generate and test mechanistic hypotheses automatically.
- The system achieved 5-10x better sample efficiency than random experimentation across all evaluation metrics.
- Validation compared ATLAS-designed experiments against expert-designed ones from published cognitive science literature.
Read full analysis →
View on Bluesky →
Don't know the author, but have become quite a fan of her work. Is always quite cool and original (sometimes conceptually, sometimes in terms of theoretical machinery &c.) arxiv.org/abs/2407.02458
AI Weekly's analysis
→
- Eliza O'Reilly's paper proves oblique Mondrian forests achieve minimax optimal convergence rates on ridge-function data where axis-aligned trees cannot.
- For general ridge functions, no weighting of axis-aligned splits can match the rate oblique splits obtain, regardless of covariate distribution.
- The analysis uses random tessellation theory from stochastic geometry, tying convergence to the relevant feature subspace rather than ambient dimension.
Read full analysis →
Great paper arxiv.org/abs/2605.23556
But anyway, the Zhipu blogpost is worth reading. z.ai/blog/glm-5.2
Tinker is a good tool, but the number of publicly credible people who have promoted this _post_ as some kind of advance in finance is so cool. The tasks are never listed. They might be as stupid as the example about news articles here. But they could be more serious -- about low
The technical report for the new Microsoft model seems quite nice: microsoft.ai/wp-content/u...
Shubhendu Trivedi reposted
“The sum-product conjecture is false for real numbers” THOMAS F. BLOOM, WILL SAWIN, CARL SCHILDKRAUT, AND DMITRII ZHELEZOV A human proof that exploits the same kind of “tower of fields” that was used in the AI-generated counterexample to the unit-distance conjecture!
Their own posts
Recent commentary
The dot com era mega IPOs have a very different character than the 2026 AI ones. Back then you had many companies that were capital-starved before their IPOs and capital-rich after. The IPO itself was quite often a major financing event. Basically, public markets funded the next stage of growth.
Many aspects of the AI twitter peanut gallery seem to have spontaneously emerged on bluesky as well. Many of the opinion microstates and occupancy boxes are thinly traded given the natural constraints of bluesky, but you can see the broader contours.
One thing that has become clear to me just very recently from dozens of conversations with folk at all the frontier labs: Many really do believe that once "AGI happens" it will "make everything easier, from robotics, to manufacturing." Supply chain constraints, ecosystem and labour development
It was easy to guess what "this model is too dangerous too release" meant: we don't have enough computational resources to serve. It was easy, in hindsight, to guess what "we are nearing RSI.. pause AI" meant.
Erdős was ahead of his time. He was really focused on creating a dataset for building and testing new AI tools. He should be called the forgotten godfather of AI and put on a TIME cover alongside some other "architects of AI" who don't deserve to be there.
Every once in a few years you get cultural moments that become like crazy psychiatric solvents. But the AI related one seems like it'd be unique in how much it concentrates people's unresolved issues into worldviews. The whole zealotry it brings forth even about insignificant stuff is quite telling.
I am not, and have never been, a fan of Ted Chiang’s sci-fi writings (find it annoying for various reasons), but it is quite funny that people seem to assume that just because he is their favourite sci-fi writer, he should validate any crazy thing they want to believe or believe about AI, and treat
Some random IAS announcement reminded me: I have always associated the IAS (SNS / Math at least) with durability of ideas, or at least the ideal. In that light, the whole ML thing they did a few years ago was such a joke. Do "XYZ provably," call it "Science of deep learning" and you're at the IAS.
It's really funny to watch agentic coding tools spiraling into madness, as if possessed by a non malicious (but loose cannon) spirit which decides "I should do something. I must act. I should figure this out."
Someone remarked that it's ironical that we give LLMs the "education" we claim to want for children, for students, and for ourselves, but simply don't have the will to demand it of any of them, or of ourselves.
Their network
In Shubhendu Trivedi's orbit
Center = Shubhendu Trivedi. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.