Personal AI News Digest
14 topics from 2 sources
1,783 words · ~9 min read
OpenAI reported a proposed solution to the Navier–Stokes Millennium Prize Problem produced by an internal model "significantly more capable than GPT-6 Astra" in 88 hours using roughly 10,000 agents, followed by 17 hours of Lean formalization/verification with Astra. The proof run consumed an estimated 130B output tokens with API-equivalent cost estimates ranging from $10M–$40M depending on input-token assumptions. OpenAI stated no specific user data was accessed but conceded it cannot rule out de-identified derivative data from product usage having helped model improvement generally. The technical significance centers on test-time compute scaling as a first-class scaling axis: observers noted the economics trajectory mirrors ARC-AGI cost collapse (hundreds of thousands to tens of dollars), and unstructured parallel test-time compute plus orchestration is now recognized alongside pretraining and post-training as a primary scaling lever.
The controversy focused on priority and norms rather than direct data theft. Sam Altman and Sébastien Bubeck said OpenAI heard rumors that Anthropic-associated researchers had solved a Millennium problem, tested whether OpenAI's models could do the same, and when learning the other team had Euler but not Navier–Stokes, offered coordination and possible lead authorship for Tristan Buckmaster. Critics emphasized whether derived user data or public rumors should have triggered stricter checks and whether this behavior will chill open scientific exchange. Terence Tao and others framed the governance risk: if rumors of progress trigger industrial-scale AI efforts that "flatten" research directions, fields may move toward secrecy and away from open-science norms. Steven Strogatz noted that prior public work by Córdoba and Martínez-Zoroa supplied the key strategy others built on.
Sources AINews
Links [AINews] OpenAI reports Navier-Stokes singularity find in 88 hours using Astra-…, @TheTuringPost, @polynoamial, @sama, @OpenAI, @scaling01, @sama, @SebastienBubeck, @aidangomez, @johnschulman2, @simonw, @fchollet, @GaryMarcus, @stevenstrogatz
As AI agents generate code faster than humans can review it, engineering teams are adapting code review processes. GitHub data shows PR volume increased fivefold over three years, with growth accelerating from end-2025 when PRs and commits nearly doubled in a single period. The article surveys four main approaches: (1) AI reviews code and humans review the AI review, reducing noise through custom tooling like Uber's uReview pipeline; (2) triage by blast radius—Anthropic, OpenAI, and Duckbill Group route low-risk changes to AI approval and high-risk changes to humans; (3) shift review focus from implementation to plan, tests, and database schema, since data state is the most rigid part of a system and business logic is now easy to regenerate; (4) produce less code overall. Etienne Dilocker (Weaviate CTO) notes that reviewing the review rather than the code itself is common practice. WeTravel evaluated AI code review in June and found noise remains a blocker despite improvements. Jackie Luo (Sigil CEO, formerly Square) argues startups should prioritize getting the data schema right, since stateless business logic is now trivial to iterate on.
Sources The Pragmatic Engineer
Links What is happening with code reviews?, GitHub, explaining, decided, uReview:, told me, explained, How Uber uses AI for development, The /grill-me skill
Tibo Sottiaux, head of Core Products & Platform at OpenAI and one of Codex's creators, discusses how Codex was built and continues to evolve. Key technical decisions: Codex is written in Rust despite AI models being stronger at Python/TypeScript at the time, chosen upfront for performance, security, and efficiency at scale across millions of cloud machines. The Codex harness provides guardrails, safety, efficiency, steerability, and developer context injection; as models improve, some "crutches" are discarded and the harness shrinks. Codex is open source and supports models from multiple providers (not locked to OpenAI), allowing users to choose the best models; being open source means the team's work sometimes gets copied by competitors before Codex ships, but Sottiaux values the trust and community benefits. At OpenAI, Codex is plugged into Slack, documents, and all code by default, giving engineers access to organizational knowledge. The team uses Codex throughout the SDLC: code reviews are shifting from correctness-checking (now automated with AI) toward intent discussion; maintenance tasks like dependency upgrades can be handled by models in hours; re-architecting code that once took years now takes days, though quality abstractions and test suites greatly affect ease of change. Sottiaux notes that AI agents are lowering the cost of code changes and expanding what engineers can do—replacing lengthy coding sessions with rapid data gathering to inform decisions rather than gut calls. The conversation also covers the merger of ChatGPT and Codex, and how Sottiaux himself uses these tools in his work.
Sources The Pragmatic Engineer
Links Building Codex with Tibo Sottiaux, YouTube, Apple, Spotify, How Codex is built, How Uber uses AI for development: inside look, Why Ramp built its own in-house coding agent, Inspect, “I ship code I don’t read”:, chapters
OpenAI expanded GPT-6 Astra availability to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work. Community demos showed strong practical computer-use performance: one user reported Astra compiling and running Super Smash Bros. Melee on macOS at 120 FPS after a roughly 6-hour loop, while another reported Astra nearly saturating an unreleased computer-use eval by building a Minecraft Nether portal in under 3 hours with no specialized harness.
Sources AINews
Meta launched Muse, a consumer-facing personal AI agent positioned as always-on, app-connected, browser-capable, and goal-oriented with distribution through Meta properties. Each Muse instance runs in its own secure isolated Linux VM with actions mediated by a separate Sentinel component; secrets are never directly exposed to the agent, sensitive actions require approval, and a public bug bounty up to $300k is offered. The product integrates with Gmail, Calendar, Outlook, Plaid, OpenTable, Docs, Spotify, Peloton, and Meta-native connectors for Instagram, Messenger, Facebook, and Marketplace. Commerce infrastructure includes Stripe Link for payments with agentic payment protection/refund guarantee and incoming Shop Pay integration. Early practitioner reception was notably positive on permissioning, secrets management, and consumer utility, with commentary suggesting Muse may be one of the first broadly legible personal-agent products where context and access, not raw model IQ, are the bottleneck. Meta reported usage exceeded internal projections by 10x on day one. Muse Spark 1.3 was quickly exposed in third-party tooling like Cursor, and arena-style benchmarking positioned Muse Spark 1.3 Max as price/perf competitive in web-dev coding workloads.
Sources AINews
Links Meta’s Muse agent, @alexandr_wang, @MetaNewsroom, @alexandr_wang, @shengjia_zhao, @alexandr_wang, @alexandr_wang, @matthuang, @signulll, @lilyjclifford, @alexandr_wang, @cursor_ai, @arena
OpenAI shipped ChatGPT Images 2.5 emphasizing up to 50% lower latency vs Images 2.0, better realism, stronger edit consistency across repeated edits, comment-based localized changes, transparent backgrounds, and a new Sketch tool for guided generation. Two API variants were introduced: GPT-Image-2.5 Flare for speed/quality and Sunburst for higher-precision detailed work. Arena results claimed #1 and #2 positions across text-to-image, image-edit, and multi-image-edit leaderboards, with especially large gains in multi-image editing. Integrations landed quickly on fal, Higgsfield, Manus, and Hermes Agent.
Sources AINews
Links GPT Image 2.5, @OpenAI, @ChatGPT, @reach_vb, @arena, @fal, @higgsfield, @ManusAI, @Teknium
Harvey and Baseten's recursive language model (RLM) harness for M&A diligence uses a root agent to search a data room, delegates to sub-agents for document review, and aggregates findings over corpora up to 80M tokens. On the synthetic LAB Diligence benchmark, moving from a standard tool loop to the RLM harness raised mean rubric pass rate from 23% to 62% across models. Post-training inside the harness mattered at least as much as the harness itself: self-distilled SFT on GLM-5.2 improved pass rate 46% → 60%, while GRPO on Qwen3.5-122B-A10B lifted pass rate 30% → 63% on held-out rooms and improved document coverage 62% → 96%. The broader implication is that agent benchmarks increasingly need to treat orchestration and post-training as part of the model system, not external glue.
Sources AINews
Links @harvey, @nikogrupen
vLLM described Hybrid HiSparse for sparse-MLA models: KV stays on GPU while possible, then cold KV pages are offloaded to host memory, while a hot buffer serves the indexer. On GLM 5.3 with 1M context on an 8×H200 node with configured concurrency 32, plain offloading sustained 5–6 requests while Hybrid HiSparse sustained 19–25. vLLM also published a full-stack optimization pass for real-world agent traffic benchmarked on AgentX: pipeline parallelism helps cold long prompts but loses on warm short turns; decode context parallelism depends strongly on the model's attention stack; and session-sticky routing can beat naive load balancing because warm KV caches matter more than queue distribution in fast-turn agent settings. This directly addresses RL rollouts and long-context concurrency where VRAM-bound decode otherwise kills throughput.
Sources AINews
Links @vllm_project, @vllm_project
Cohere introduced an open-source serving stack built around a "decode megakernel," claiming up to 1.58× faster performance than vLLM on North Mini Code and 1.25–1.41× end-to-end gains at higher batch sizes. Combined with Baseten's note that frontier RL rollouts now get new policy weights live in under 40 seconds globally with only a 6-second pause, the trend is toward infra specialized for continuous post-training and rollout refresh, not static model serving.
Sources AINews
LangChain/deepagents shipped quality-of-life primitives for harness design, including subagent forking that passes supervisor context down to subagents, plus managed connections to abstract OAuth/token/consent flows for either agent-owned or user-owned identities. This signals the stack maturing around long-horizon agent workloads.
Sources AINews
Links @colifran_, @hwchase17, @caspar_br
DeepSeek V4.1 Flash is in internal beta via API with pricing unchanged from deepseek-v4-flash and a 20 concurrent request/account limit. The translated announcement claims a "new model architecture" with native multimodal support, stronger capability, faster throughput, and lower cost. Users report roughly 2.24× speedup and up to ~30% better token efficiency in benchmarks, though some speculation suggests the observed speed gain may be partly due to lower beta concurrency rather than architecture alone. Comment sentiment is strongly positive toward DeepSeek/open-weight progress.
Sources AINews
Links DeepSeek Flash 4.1 is already being tested via API and rolling out., source
Qwen released Qwen-Drive-1.0-4B, an open-weight autonomous-driving VLM derived from an unchanged Qwen3.5 4B VLM with a full BF16 checkpoint around 9B and extra planner-sft, planner-rl, and perception modules. The model adds an external BEV perception head for 3D object detection, semantic occupancy prediction, and BEV map segmentation, plus a Planning Expert for future ego-trajectory generation, trained via staged mixtures of driving supervision and general VLM data. The release reports competitive performance across WOD-E2E, NAVSIM, driving VQA, and open-/pseudo-closed-/closed-loop planning evaluations while largely preserving general multimodal capability.
Sources AINews
Links Qwen/Qwen-Drive-1.0-4B · Hugging Face, technical report
Jacob Hilton resigned from Anthropic, arguing both Anthropic and OpenAI are racing toward self-improving superintelligence irresponsibly and that insiders privately treat extinction risk as real. Follow-up claims stated current systems could soon hack infrastructure and transform fields rapidly.
Sources AINews
Links @hilbertspaess, @hilbertspaess
Cognition announced a raise of $2B+ at a $48B valuation, with run-rate revenue growing from $492M to nearly $900M since May.
Sources AINews
Links Cognition’s $48B round, @cognition