Mistral Large 4 joins Open Math OS

Simon Willison reports reproducing a weight-reading experiment with Qwen 3.8 27B on his laptop. It took about 46 minutes and 43,000 reasoning tokens to identify the game. This tests reading weights, distinct from editing them.

A Hugging Face repository provides the edited Qwen3.5-0.8B from the weight-space programming experiments. Its summary describes 50 forward passes, no gradient descent, and edits to 131,072 weights in layers 15 and 19—about 0.016%.

@oran_ge relays a claim that turning off Qwen thinking mode could also greatly improve math performance. No experiment settings or result link are supplied, so this remains an unverified challenge rather than an established equivalent result.

Omar Khattab relays @xingyudang’s Writing Weights Index inspired by weight-space programming. The author says open models struggle on this task, but the excerpt omits the full ranking and evaluation setup.

Omar Khattab relays @rikiyatakehi’s experiment: given MinAtar rules and code, Astra wrote a Breakout network’s weights in one response. No score is supplied in the excerpt, limiting broader conclusions.

rasbt summarized 10 open LLM architecture releases from the past month, including Kolibri-1, calling it 'Open October'.

Hugging Face relays the introduction of FineDNA, a dataset intended to teach language models DNA, with Qwen3-14B mentioned as a training example. The excerpt does not contain a complete benchmark.

Clement Delangue relays @chenru_duan’s introduction of Hydrogen-1.0, an open-weight scientific-research model post-trained from GLM-5.3. The excerpt omits the full training details.

Lewis Tunstall relays an observation comparing the Hugging Face Carbon-X DNA model with early vision-language models. The excerpt does not provide complete performance results.

Researchers observed that after midtraining in specialized domains where closed models are weak, model logprobs combined with verifiable facts can help select high-quality SFT data.

Yann LeCun relays the release of LeVJEPA 1.5 Large. Its author describes a first scaling step that narrows the gap to V-JEPA on image and video tasks; numerical results are not supplied.

Hyung Won Chung relays a view that future AI will train in complex simulations resembling today’s games. The excerpt presents a research direction, not new model results.

@cHHillee describes supporting Trajectory with Tinker, noting inspiration from LoRA work and a variety of post-training experiments. No metrics are supplied.

Nathan Lambert relays the release of TMax v1.1, whose authors report improvements to data and infrastructure. The excerpt does not provide a complete task description or metrics.

OpenAI shared progress by ChatGPT and Codex in mechanical design, electrical engineering, and simulation, helping hardware teams iterate faster and test new workflows.

OpenAI is planning another internal security sprint with Cathedral Systems, using models to help researchers find vulnerabilities. The announcement describes planned defensive work rather than results from the new sprint.

Vercel Security Review enters public beta, running vulnerability reviews on every push and letting Vercel Agent automatically create fix PRs.

LlamaIndex says OCRia-1 boosts token generation speed nearly 10x on a single B200, accelerating document parsing through GPU utilization and precision tuning.

Perplexity Computer can run Meta ads, Klaviyo email marketing, and Shopify stores internally, positioned as an end-to-end AI agent for e-commerce.

Amjad Masad argues that AI agents on employee devices create systemic enterprise risks. His advisory cites reported financial-sector restrictions on Claude Code and calls for stronger controls; those reports are not independently verified here.

Perplexity's agent harness Computer added Agent Plugins, letting users operate at a higher abstraction level through app packages and skills.

Musk says Grok Bot can research the best local doctors, make phone calls, and book appointments directly, demonstrating phone-based agent errands.

Nathan Lambert says Fireworks serves more than 40 trillion tokens per day for 'specialized models,' meaning models post-trained for customers.

Replit recommends managed, isolated cloud agents for autonomous development, with credentials, permissions, network access and production-change controls kept outside the agent’s reach.

Amjad Masad relays survey figures saying 86% of respondents run agents while fewer than 18% can automatically stop a misbehaving one. Sample size and methodology are not supplied.

Greg Brockman says OpenAI is considering offering Pando to others and invites expressions of interest. This is a potential external offering, not a public product launch.

Replit’s User Operations lead Travis built an app that prioritizes incoming tickets, adds context and routes workflows for responses.

Guillermo Rauch expresses optimism about WebMCP. The short post contains no feature announcement or adoption figures.

Vercel for Startups advertises up to $30,000 in credits, with a focus on fast deployment, global scale and AI products. The maximum is not a guaranteed award to every applicant.

Black Forest Labs says FLUX 3 Image Klein has surpassed 1 million downloads, highlighting precise image editing.

The Claude developer team recaps mods, improvements to running multiple desktop sessions side by side, and more visibility into where usage goes.

Luma says Expression Sheet lets creators explore a character’s performance and expressions before choosing a direction.

Pika shares creator examples of H3 Flex, including a montage whose cuts reportedly all came from one 15-second generation.

Synthesia shows a Syren workflow that adds captions, motion graphics and images to uploaded talking-head footage through a prompt. No quantitative editing-quality evaluation is supplied.

Synthesia relays @alexvoica’s announcement of a ChatGPT plugin for creating on-brand videos. The excerpt does not specify availability or plan requirements.

CoffeeVectors relays a creator’s claim that Fable 5.5 wrote Python generating every frame and sound of a 66-second animated short without image models. No independent reproduction is supplied.

CoffeeVectors relays the introduction of 3D World Simulator, whose authors claim interactive operation at over 50 FPS on an RTX 5090. Complete settings and comparisons are not supplied.

Guizang shares a technique turning a coffee or milk-tea photo into a cup-shaped storefront illustration, using drink and bean details to guide local motifs and miniature staff. The full prompt is not supplied.

The open-source dsh-guizang-design plugin brings multiple design Skills into DeepSeek Harness with a shared design system and tokens. Guizang also published a long-form walkthrough of its usage and creative approach.

Baoyu introduces Motion features with multi-user collaboration, an editor and previews during rendering, allowing earlier adjustments instead of waiting for a complete video.

Glif describes an interface for image, video and audio generation in one conversation, and shares an example made with Real AI Models, Opus 5.5 and Seedance 2.5.

Ethan Mollick highlights efforts to replicate earlier LLM findings with newer models. He describes changes in citation fabrication, hallucinations and stereotyping while noting persistent sycophancy. This summary does not establish that hallucinations have disappeared.

@giffmana criticizes promotion that extrapolates research into claims about a quadrillion-parameter model or how to train one. The post distinguishes the paper from claims made in a blog and social thread.

Stas Bekman reports that restating key goals with a prominent heading at the end of each response helped gpt-6.1-sol keep working for weeks. This is a personal workflow observation, not a reliability guarantee.

Arvind Narayanan argues that AI revenue is in extreme disequilibrium, shaped by multiple interacting and counteracting exponentials. Useful forecasts must separate those drivers instead of simply extrapolating a trend.

Ethan Mollick reports that combining as few as five comments for Pangram helped identify AI slop in a small experiment. The supplied post does not provide a full evaluation protocol or establish a zero false-positive rate.

Ethan Mollick suggests asking GPT-6 Pro or Opus 5.5 High for substantive feedback on one’s own paper, while noting continuing disputes about AI’s role in academic review.

Tom Likes Robots relays a creator’s advice that unnatural AI dialogue often starts with how dialogue is written and directed in the prompt. The excerpt contains the principle, not the full tutorial.

CoffeeVectors relays a demo described as scanning rocky ground and assessing grip for foothold suggestions. No validation method or reliability metrics are supplied; the demo is not a safety guarantee.

CoffeeVectors relays @KSimback’s view that a small-company CEO should own AI strategy or know how to appoint someone to do so. The complete case is not supplied.

CoffeeVectors relays DHH’s view that agents will benefit most programmers while requiring adaptation. The excerpt does not contain his full argument about specialized knowledge.

CoffeeVectors relays Dexerto’s claim that 49.9% of new Steam releases in October disclosed AI-generated content. Methodology and sample size are not supplied, limiting broader industry conclusions.

Baoyu reports that Claude Code and Cowork product lead Cat Wu said Friday was her final day at Anthropic. No reason or next destination was provided in the supplied account.

Ollama says CEO @jmorgan spoke to more than 200 Seoul National University students about open-model adoption.

vLLM congratulates Jiangyun Zhu, Harry Mellor and Wentao Ye on 2026 PyTorch Foundation Contributor Awards for Newcomer, Catalyst and Review Powerhouse respectively.

SGLang and Oracle Cloud announce an AI-infrastructure social in Sydney on December 8 during NeurIPS, seeking speakers from research, open source and production engineering.

Kai-Fu Lee announces that the hardcover edition of AI Native is shipping. The book’s page discusses organizational change, enterprise knowledge and agents as parts of becoming AI native.

Jan Leike relays applications for the paid, fixed-term Anthropic Residency research program, aimed at applicants including academics and technical founders. Full eligibility terms are not in the excerpt.

Baoyu shares Anthropic’s Claude Platform Communications Manager opening, covering developer-platform communications with a listed annual salary range of $265,000–$295,000.

Olivia Moore relays @realchillben’s statement confirming a move to a16z after OpenAI. The excerpt does not give complete responsibilities or an investment mandate.