Live Feed
Ricursive's 65B Ripple Preview Scores 80.52 on Terminal-Bench 2.1 for a Claimed $32M, Edging a 501B Rival
A chip-design startup released a dense 65B Apache 2.0 model that tops US open-weight agentic coding at 80.52, ahead of Reflection Beam 501B (80.1). The $32M all-in cost and 1.4M-rollout RL run are self-reported.
Anthropic Open-Sources FlashProve: Proving Which Model Answered Costs 72x the GPU Time of Serving It
A zkSNARK prototype proves a 400B MoE produced a given response, with 19 MB proofs and 17-second CPU verification. The overhead is 66 to 94x per GPU, and the pitch is regulation, not consumer apps.
OpenAI to Acquire Pierre Computer, Pulling Git Storage and Review Tooling Into Codex
Two days after revealing its Pando Git backend, OpenAI is buying the startup behind Code Storage, Trees and Diffs. Terms are undisclosed and the deal needs regulatory approval.
Tencent Hy4 Goes GA With Vision and Computer Use at Unchanged $0.834/$2.501 Pricing, but Weights Are Still Pending
The 770B MoE adds image understanding and CLI/GUI control ten weeks after its text-only Apache 2.0 preview. Tencent says weights and a system card follow within days; Hy4 Flash is next.
Open Models Have No Common Crawl for Agent Training: Only 2,294 of 90,000 GitHub PRs Became SWE-bench Tasks
Open Athena's David Hall argues posttraining consumes tasks, not documents, and the public record needed to build them was never kept. The yield numbers from SWE-bench, TaskTrove and MiMo environments back him up.
RTX Spark Laptop Runs GPT-OSS-120B at 13.9 Tokens/sec on Linux vs 65.4 on Windows, and Plugging In Makes It Worse
Testing on an MSI Prestige RTX Spark finds native Linux pinned near 780 MHz on wall power versus roughly 2.2 GHz on Windows. The cause is an encrypted power table that only the Windows driver can read.
Hinton Predicts an AI 'Chornobyl' This Year or Next, Calls for FDA-Style Pre-Release Approval and Says He Does Not Trust Altman 'At All'
Geoffrey Hinton, 78, told the Guardian a Three Mile Island-type warning shot has already happened and a larger blackout or week-long financial outage could follow. He backs mandatory testing endorsed by 33 countries.
ByteDance's USDB Cuts Write Latency 85.8% and Infrastructure Cost 42.1% as Recommender Histories Hit 800,000 Actions
A new paper details the storage layer behind Douyin and TikTok recommendations. User sequence length grew 2,000-fold from 2023 to 2025, and ByteDance rebuilt the LSM-tree around it.
Only 3 of 10 Weight-Space Experiments Read Raw Weights: Critics Split Astra's Qwen Result as Qwen 3.8 27B Names 3 of 4 Games on a Laptop
Matthew Honnibal says the weight-editing headline blends two different techniques. Simon Willison reran the no-tools test on a 17GB Qwen 3.8 27B: three games right, one wrong, 46 to 65 minutes each.
Vesuvius Challenge Awards $50K as Two Ink Models Read Greek Text in PHerc. 306B, 25 Submissions In
A Sydney MA student and an English freelance engineer won the First Letters Prize after the papyrology team agreed on a reading no earlier submission could settle. The ink map came from multiplying two models, one fine-tuned on a single RTX 4090.
Figma Post-Trains Qwen 3.8 27B for Design Edits, Doubling the Share of Agent Work Its Own Model Handles
Figma says its in-house editing model now runs more than twice the work it did at beta, with typical agent tasks 20% faster and undo rates down 60%. The next version is built on Qwen 3.8 27B and adds reinforcement learning.
Big Tech's Off-Balance-Sheet AI Obligations Hit $3.2 Trillion, Up 90% in Six Months
Nikkei's tally of Alphabet, Meta, Microsoft, Amazon and Oracle commitments has nearly doubled from $1.65 trillion in July. Most of it is purchase contracts and leases that have not started, and the number depends heavily on how you count.
Postman Desktop Auto-Installed Plugins Into Codex and Claude Code, and Its Installer Approved Its Own Startup Hook
A GitHub issue with log evidence says Postman 12.33.4 on macOS wrote itself into two coding agents on launch. The Hacker News headline claims data sharing, but the report explicitly does not.
Fujifilm Spends 10 Billion Yen to Make 5x More LTO Tape as Generative AI Drives Archive Demand
Fujifilm will expand tape output fivefold versus fiscal 2025 at its Odawara site, its largest-ever investment in dedicated LTO facilities, running through fiscal 2028. Most of the new capacity targets LTO Ultrium 10.
Google DeepMind UK Staff Push to Void Lifetime Non-Disparagement Clauses as Google Drops Them for New Hires Only
Some London-based DeepMind employees are barred by contract from criticising Google for life, even after leaving. Google says it no longer enforces the clause on departed staff and removed it for new hires, but current workers want it struck from existing contracts.