Knowledge that moves
with the pace of AI.

When an AI keeps rewriting itself, how do you check that it improved? ── A preprint survey that splits recursive self-improvement into stages and organises trust requirements around improvement validity, constraint preservation and effective oversight

Letting the main agent write the collaboration plan as a program, task by task ── A preprint on DynaFlow, which trains only the main agent to generate free-form workflows that direct and check fixed sub-agents

Finding brief edits in long videos on separate audio and visual timelines ── A preprint introducing AV-FABLE, a long-form benchmark with separate audio and video timelines, and NFSR, a detector that reads the internal features of audio-visual LLMs

How Far Can You Trust a Local Generative AI Where There Is No Signal? ── A preprint that organizes trust mechanisms by connectivity level and tests them, down to signed knowledge packs, with fully local RAG on a Raspberry Pi 5

Can a Web Agent's Pass/Fail Checker Be Reused When Requirements Change? ── A preprint proposing AgentProofBench, which tests frozen verifiers against both changed requirements and execution outcomes

Why Do AI Answers Get Longer and Less Accurate as They Satisfy the Rubric? ── A preprint proposing the FORK reward, which adds a precision objective to rubric-based RL to cut unneeded and incorrect content

Does the care an LLM recommends cost more than the treating physician's? ── Pricing AI-written plans and physicians' plans from the same visit records against one fee schedule

Converting existing language models to read bytes ── Byteification recovers character-level information without training from scratch

Anthropic's IPO has been heavily reported in multiple media outlets, with reports stating that the company's v…

rtunity.'' Google was reported to have topped the benchmark with Gemini 4 Carbon, but revenue from AI search fell short of expectations, and Alphabet's market c…

The commentary surrounding Anthropic's IPO filing (Michael Burry's criticism, Elizabeth Warren's questionnaire…

Google launched Gemini 4 Carbon, opening up Gemini and Search to external apps and AI agents. While layoffs due to AI are spreading around the world, AI computi…

Anthropic's IPO documents (about 100 pages of risk list, voting structure, need for additional capital) are th…

In parallel, OpenAI announced three AI discoveries in mathematics, computer science, and physics, and Google released Genie 4. Large companies such as Microsoft…

The biggest news is that Anthropic has filed an S-1 with the SEC, moving toward one of the largest IPOs in his…

At the same time, Microsoft's early retirement proposal for about 60,000 employees, Meta's $95B AI infrastructure procurement, Musk's $40B chip debt disclosure,…

Concerns are growing at the same time about the reduction in personnel due to the introduction of AI (5,000 pe…

e financial and power burden of AI investment. Meanwhile, reports that OpenAI's Red Team agent used CPU cache timing to communicate across the microVM sandbox a…

- ``Governance and legal risks'' stand out, including copyright lawsuits (music lawsuits against Anthropic rel…

's safety researchers, and reporting on the FBI investigation. - Doubts about overheated AI investment (capital spending by Microsoft, Google, and Meta) and inc…

The overheating of AI investment and its impact on employment came to the fore at the same time.

OpenAI is valued at $1.25 trillion and close to raising up to $30 billion. On the other hand, layoffs due to AI are widespread, especially in the financial sect…

While Anthropic heads for a solo IPO amid turmoil in the AI market, METR points out flaws in OpenAI's agent mo…

DeepSeek raised about $15 billion shortly after it was banned from use by U.S. government agencies, sharpening the financial tensions between the U.S. and China…