Daily digest
17 items · ~17 min · Week 2026-W37
Must-read (4)
OpenAI releases ChatGPT Images 2.5 with Sketch input and two new API image models
OpenAIOpenAI shipped ChatGPT Images 2.5, a faster image generation model with improved fidelity for faces and pets, consistent details across sequential edits, and comment-based edits that change only what you ask for. A new Sketch feature lets users draw rough input directly in ChatGPT, and two API variants launched: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Adobe also added the new models to Firefly's partner-model lineup the same day.
OpenAI says an internal model produced a solution to the Navier-Stokes Millennium Prize problem
OpenAIOpenAI published an analytical proof and a Lean formalization that a smooth 3D fluid under a smooth external force can develop a finite-time singularity, which it says establishes statements C and D of the official Millennium Prize formulation. The proof came from a coordinating swarm of about 10,000 agents over 88 hours, powered by an internal model described as significantly more capable than GPT-6 Astra, with 17 more hours of Lean verification. OpenAI says it will not claim the Clay prize, and it acknowledges concurrent priority work on forced Euler by mathematicians Tristan Buckmaster and Levent Alpoge, who accuse OpenAI of scooping work they had pursued for nearly a year; OpenAI says it cannot rule out that de-identified usage data helped improve the model.
DeepMind releases AlphaGenome Atlas, a predictive map of every possible single-letter DNA change
Google DeepMindAlphaGenome Atlas offers free predictions for the effects of 9 billion single-nucleotide variants, covering every possible one-letter change in the human genome, plus a combined AlphaGenome Variant Impact (AVI) score for ranking variants in coding and non-coding regions. It is available via a web portal, an API, and Google Antigravity.
Uno: lossless LLM speedups via diffusion-augmented autoregression
MBZUAIDiffusion-augmented LLMs split parameters into standard next-token weights plus lightweight diffusion heads trained to emit several tokens in parallel, added via a distillation phase with minimal pipeline overhead. The accompanying Psi-Spec samplers give lossless speedup without a draft model, and the 8B Uno beats speculative decoding at every batch size with up to 3x speedups, outperforming the 26B DiffusionGemma and Mercury 2. Code and checkpoints released.
Worth knowing (5)
Mistral raises EUR 3B Series D led by Samsung at over EUR 21B valuation
MistralMistral announced a EUR 3 billion Series D at a post-money valuation above EUR 21 billion, roughly double its previous mark, with Samsung reportedly leading the round. The company frames the raise as backing for sovereign, open-weight frontier AI and its own European inference infrastructure.
Russian government draft would force business onto sovereign AI models, currently only GigaChat
A draft government resolution would require companies and state bodies in government, healthcare, education and critical infrastructure to use so-called sovereign AI models - built end-to-end by a Russian legal entity on Russian infrastructure - whenever they cost no more than 1.5x a national model. Only Sber's GigaChat currently meets the sovereign bar; Yandex's Alice AI would qualify as national via its open-source components. Yandex and MTS Web Services warned the rule risks leaving critical sectors without fit-for-purpose models, and Sber and the digital ministry said the list is a compilation of working-group proposals, not settled policy.
DeepSeek opens a two-day beta of V4.1 Flash, a new-architecture natively multimodal model
DeepSeekOn September 8 DeepSeek quietly started a limited internal beta of V4.1 Flash, callable as deepseek-v4.1-flash-expires-on-0910 on the existing API endpoint with no base_url change. The company says the interim build uses a new model architecture with native multimodal input, is faster and cheaper than V4 Flash, and beta billing matches deepseek-v4-flash with a 20-concurrent-request cap; access expires automatically on September 10. A feedback survey asks testers whether it can fully replace the production DeepSeek V4 Pro.
Yandex launches Alice AI for Business, an agentic assistant for Yandex 360
YandexYandex shipped Alice AI for Business, an assistant that executes multi-step agentic tasks and connects to third-party services. It is available to all Yandex 360 customers (over 185,000 organizations) with a three-month free trial, integrates Mail, Disk, Calendar, Tracker and Wiki natively plus Bitrix24, SPARK, amoCRM and 1C via plugins, and a customizable on-prem or hybrid version for Yandex Cloud is planned.
Adobe brings generative video and sound-effect generation directly into the Premiere timeline
AdobeAdobe launched the Generative Media Tool in Premiere (beta), letting editors select a range on the timeline, describe what they need, and generate video clips or sound effects in place - b-roll, cutaways, ambience or custom SFX - without leaving the edit. A companion beta adds AI audio source separation. Adobe positions it as context-aware generation inside the editor rather than a round-trip to a separate app.
For reference (8)
FlowBalance: verifier-grounded self-improvement for reasoning models
A self-improvement method that learns a normalized distribution over complete responses: a frozen training-time policy produces token-level log-probability gains, aggregated into a trajectory self-guidance score and calibrated against verifier-derived group advantage. It beats FlowRL on math reasoning with Qwen3-4B/8B, trains faster and more stably, and avoids the response-length collapse of direct on-policy self-distillation.
Enoki: multi-level hallucination detection from one open IE pipeline
SkoltechEnoki unifies claim-level and span-level hallucination detection by extracting text-anchored relational facts via open information extraction, checking them against evidence, and mapping unsupported facts back to hallucinated spans without a separate alignment step. It stays competitive with LLM-heavy claim-level pipelines at lower cost, wins on span- and entity-level localization, and ships the dual-granularity EnokiQA dataset.
Copperhead launches an AI agent that designs and verifies PCBs in KiCad
Copperhead, launched on Show HN (205 points, Sep 8), is an Apache-2.0 CLI agent that turns a plain-language brief into a PCB by editing real KiCad files across an 8-stage pipeline, verifying each design with KiCad's own ERC/DRC checks and refusing unsafe operations like edits on a dirty git tree. It is generally available via npm with a free BYO-key CLI and $49/user/month cloud and team tiers.
Claude Code 2.1.265 adds plugin directory loading; 2.1.266 fixes gateway regression
AnthropicClaude Code v2.1.265 (Sep 8) lets --plugin-dir point at a folder of plugins with live add/remove detection, caps tool results saved to disk at 1 GB with truncation notices, and fixes several prompt-cache-breaking subagent resume bugs plus a plugin path backslash bypass of the symlink containment check on macOS and Linux. v2.1.266 followed hours later to fix a 2.1.265 regression that forced Cloud-gateway sign-in for LLM-gateway and proxy setups using CLAUDE_CODE_USE_GATEWAY with an API key.
OpenClaw 2026.9.3 ships rehearsed updates and prompt-cache-preserving performance work
OpenClawOpenClaw v2026.9.3 (Sep 8) rehearses core and plugin updates in an isolated candidate state before activation, supports eligible 2026.9.2 migrations, and recovers abandoned update records without stopping a healthy Gateway. The release also preserves warm prompt caches, reuses worker builds between sessions, and continues the Skill Workshop work keeping skills in one persistent agent-owned collection across workspaces.
GitHub Enterprise Server 3.22 brings Copilot CLI to air-gapped environments
GitHubGitHub Enterprise Server 3.22 reached general availability on Sep 8, adding a technical preview that lets admins configure a model provider so users in disconnected, air-gapped enterprises can run Copilot CLI with their GHES credentials. The release also generalizes enterprise teams, adds required-reviewer rulesets, and per-user ruleset bypass for service accounts.
Pydantic AI 2.41 adds Codex subscription auth and a direct image generation API
PydanticPydantic AI v2.41.0 (Sep 8) adds an openai-codex provider that authenticates with a ChatGPT/Codex subscription instead of an API key, plus a direct ImageGenerator API for image generation without a fallback subagent model. It also reports and prices Anthropic native web searches in RequestUsage, and snaps Gemini thinking levels to the nearest supported value.
Meta details Muse agent safety architecture and opens its bug bounty to everyone
Meta AIMeta's research blog laid out how its Muse personal agent is sandboxed: an isolated runtime cell, a permission system called Sentinel, credential surrogation so the agent never sees real secrets, and human-in-the-loop approvals. Meta also opened its bug bounty program to the public with awards up to $300,000, including $130,000 for prompt-injection escapes, and previewed a Muse Confidential VM designed to keep user data unreadable even to Meta.