mrragava

[Drew Breunig @dbreunig]

I talk to a lot of people using DSPy/GEPA. A common thing they do is write custom proposers to prevent overfitting and tailor the proposer instructions to their task.

We're trying out a few ideas to make this easier. Check it out: https://t.co/ZNFnkhX7C5

[Brian Li @Brian_Bo_Li]

Yifei Shen, one of the research masters I admire most, has released HarnessOpt.
https://t.co/D96i76Cww3

# About HarnessOpt
What I think is especially outstanding about it is that it gives empirical evidence for closed-loop verification: the agent itself can optimize its own hardness all the way to SOTA performance.

It connects this engineering closed loop with zeroth-order optimization in mathematics. The modifications to the skill / harness themselves are non-differentiable, and candidate updates can only be evaluated through external feedback such as rollout scores, error logs, and validation performance. So in essence, this is a form of zeroth-order search over a discrete text / program space.

But it is not just blind ZO in the traditional sense. Traditional ZO only sees a scalar reward; here, the agent can read the full trajectory and treat the failure process as an interpretable debugging signal.

[Ricko Yu @cosmtrek]

发现一个巨牛的书 The Hitchhiker’s Guide to Agentic AI，覆盖 Agentic AI 全貌

https://t.co/ltHEAJbHUN

[OmegaInstituteOffical @OmegaDesci]

7 weeks ago we pointed a swarm of autonomous AI agents at the foundations of mathematics with one instruction: rebuild it from a single act of distinction, in Lean 4, with zero axioms — and prove everything.

Here's the receipt, today:
• 23,000+ machine-checked theorems & lemmas
• 0 axioms, 0 sorry — pure CIC (no choice, no excluded middle, no propext)
• 106,000+ commits, ~2,000/day, mostly unattended
• 6 research domains on one formal kernel

The point isn't the size. It's that you don't have to trust a word of it. Every theorem ships as a certificate you can re-run: `lake build`, and the kernel either accepts it or it doesn't. No "trust the model." No screenshots. Re-run the proof.

The viral version of this is "we burned $X of compute." Ours is the inverse: a scale flex you can verify yourself.

NOT claimed: that the theorems are "important," that no human built the harness, or that volume equals depth. Only this: an autonomous pipeline can produce machine-checked mathematics at this scale, fully auditable, today.

https://t.co/POMKxQLS2V

[云微 @yunwei37]

https://t.co/aYY795e28I
把这个eBPF 的 AI Agent 追踪工具完善了一下, 希望做成给 Agent 用的 top/perf/nsight 的形态, 可以同时看到 Agent 本身在干嘛以及系统上会带来什么变化

更方便赛博监工x

[Thomas Fel @thomas_fel_]

And the first post in the series!

We formalize representation steering through a geometric lens.

Blog: https://t.co/HqBKO5XeBG
Arxiv: https://t.co/ndX26fCZIo

[David Chalmers @davidchalmers42]

incidentally i'm looking for a home for "what we talk to when we talk to language models" (https://t.co/RJ1VhuGB2c) as well as for "propositional interpretability in artificial intelligence" (https://t.co/wLprsHeNiZ). both of them are somewhat long and unwieldy so perhaps not perfect for a standard journal. suggestions are welcome.

[habakan @habakan_f]

ブラウザ内でTransformerでの言語モデルの学習イメージを理解するためのサイト作りました
MoonBitを使いました

・wasmでそのまま学習（でも料理レシピの英文）
・学習中の重みや次元圧縮したembedding可視化
・pos embeddingやskip connectionなどをオフにして未学習を再現

#moonbitlang

[HOSEIN YAVARZADEH @hosein_yz]

Private AI in Confidential VMs isn't as private as we thought. 🤯

We just showed how to reconstruct user LLM prompts word-for-word from inside an encrypted Intel TDX enclave—without breaking any cryptography.

Introducing TDXRay (IEEE S&P 2026): https://t.co/rZIcJG76pI

(1/4)

[Drew Breunig @dbreunig]

Really excited for this one: @SrihariSriraman and I took a deep dive into coding agent system prompts to understand their structure, similarities, and differences. https://t.co/Tgb9plwImk

[Ningyu Zhang@ZJU @zxlzr]

AI systems repeatedly reinvent the same domain know-how—buried in prompts, tools, and brittle pipelines.

Skills remain fragmented, duplicated, and inconsistent in quality.

We believe the missing layer in the AI stack is skills as infrastructure.

We are pleased to introduce SkillNet, an ongoing project to build an open infrastructure for creating, evaluating, and organizing executable AI skills at scale.

Homepage: https://t.co/D4PkIlTbRA
Code: https://t.co/2Zej4VL7Rk

SkillNet is not a skill repository.

It is infrastructure to standardize how skills are built, evaluated, and interconnected across domains.

With SkillNet:

→ Skills become reusable, composable assets
→ Agents gain reliable, evaluated capabilities
→ Workflows become modular and interoperable
→ Knowledge becomes infrastructure

Each SkillNet skill undergoes explicit evaluation across safety, completeness, executability, maintainability, and cost.

This infrastructure may enable composable scientific and enterprise workflows.

We gratefully acknowledge the open-source community for sharing numerous projects and skills that inspired this work.

We have built an initial prototype demonstrating:
• Autonomous Scientific Discovery
• Autonomous Coding Agents

The system is still experimental and not yet production-ready.

An initial Python library supporting skill search, download, creation, evaluation, and analysis is available:

pip install skillnet-ai

Technical report coming soon.
#SkillNet #Skills #Agents #LLMs #NLP

[@bluecow 🐮 @BLUECOW009]

i started a repo exploring context field prompts
prompts that don’t tell llms what to do but change the conditions under which meaning forms
the behavior shift is immediate and visible
repo here 👇
https://t.co/IrLqdFHh48

[al’amin ai @alamin_ai_]

🚨 We just released GraphMem Today 1st January 2026.

A self-evolving, graph-based memory architecture for production AI agents.

SOTA Memory that thinks, learns, and evolves.
- Beats HippoRAG-V2 by 14.9pp on Accuracy retrieval
- Beats HippoRAG-V2 by 13.8pp on Conflict Resolution

📄 Paper
https://t.co/7aLikVVbUR

📚 Docs
https://t.co/Tehd2UVsHo
🧠 Vision
https://t.co/x0j1TSmbfI

🧩 Code
https://t.co/cScIXVJmk3

[Belinda Li @belindazli]

Code out now!: https://t.co/dMwBttoYQD

I'm really excited about the possibility for introspection to be a tool for interp! This paper was a first step towards this, with lots more to explore. Looking forward to seeing how others build on this!

[Tom Dörr @tom_doerr]

System using LLM agents for LLM discovery

https://t.co/MgAEh3Ml5u

[TrustGraph @TrustGraphAI]

🎯 Introducing TrustGraph 1.5.8: Ontology-Driven AI.

Whether you like OWL 🦉 or are connecting systems with SOSA, ontologies help exchange domain-specific data. Ship AI agents powered by your ontologies today.

🔗 https://t.co/ikZsz9Wkrz

[Philip Zucker @SandMouth]

I've pushed Knuckledragger, my z3 based python proof assistant, up to PyPI https://t.co/bG00849IMw

[Amir Ayupov @disruptnhandlr]

BOLT is gaining traction with security efforts, e.g. SGX hardening
https://t.co/kVjxecirFi

[Z.ai @Zai_org]

Glyph: Scaling Context Windows via Visual-Text Compression

Paper: https://t.co/4T1jKP2E17
Weights: https://t.co/UOuJBOSuBg
Repo: https://t.co/OSS6JsBReW

Glyph is a framework for scaling the context length through visual-text compression. It renders long textual sequences into images and processes them using vision–language models.

This design transforms the challenge of long-context modeling into a multimodal problem, substantially reducing computational and memory costs while preserving semantic information.

[VraserX e/acc @VraserX]

GPT-5 just casually did new mathematics.

Sebastien Bubeck gave it an open problem from convex optimization, something humans had only partially solved. GPT-5-Pro sat down, reasoned for 17 minutes, and produced a correct proof improving the known bound from 1/L all the way to 1.5/L.

This wasn’t in the paper. It wasn’t online. It wasn’t memorized. It was new math. Verified by Bubeck himself.

Humans later closed the gap at 1.75/L, but GPT-5 independently advanced the frontier.
A machine just contributed original research-level mathematics.

If you’re not completely stunned by this, you’re not paying attention.

We’ve officially entered the era where AI isn’t just learning math, it’s creating it. @sama @OpenAI @kevinweil @gdb @markchen90