We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Top Stories
Why it matters: The frontier is becoming the control plane around models—identity, tools, and feedback loops—not just model weights.
A security disclosure links an OpenAI account compromise to ordinary web vulnerabilities. Its authors say two bugs enabled takeover of employee and some unaffiliated ChatGPT/Codex accounts, access to Outlook, Slack, and GitHub, and a test pull request in OpenAI’s internal codebase in under 72 hours. The reported chain ran from HEIC/HEIF upload and a libheif heap overflow to remote code execution, an SSO flaw, and connected GitHub access. OpenAI fixed the SSO issue about 14 hours after notification and paid $6,500.
Anthropic put numbers around AI-assisted R&D. It published three measures—how much AI performs AI R&D, how well agents are overseen, and how compute is allocated—and said other labs could publish comparable figures for third-party verification. Anthropic’s internal snapshot says Claude-led model-R&D tasks rose from 1% to 26% in six months, Claude collaborated on or led more than 90% of model R&D, and about 30,000 agents worked on research and engineering at a time. The immediate value is a repeatable disclosure baseline, not independent validation.
Research & Innovation
Why it matters: Evaluation is moving from one-shot answers to sustained change and verification.
Vibe Code Bench 1–100 tests software maintenance. ValsAI gives a model a working web app and up to ten plain-English feature, bug-fix, or database-migration requests; each iteration must preserve prior behavior, and the first failure ends the run. No model passed 30%; the top three were 8–30× more expensive than GPT Luna, while Fable 5.1 averaged $150 per test. 39% of failures broke existing functionality.
Stellar Colosseum targets long-horizon mathematical research. Its technical report describes a model-agnostic workflow separating strategy exploration, proof decomposition, subproblem solving, and global verification. It reports a 4,263 Codeforces score versus a cited human high of 4,059, 71.0% on TCS-Bench, and integration into Google Antigravity as the “Long Proof” pattern.
Products & Launches
Why it matters: Commercial agents are becoming persistent runtimes with controlled access to tools and data.
Google updated Gemini managed agents with a configurable Gemini 3.8 Flash runtime, persistent isolated Linux sandboxes, code, file, and web tools, plus Files and Credentials APIs. Google reports 7% higher cache hits on multi-turn coding/research tasks, 16% on long question-answering runs, 40% fewer output tokens for file edits, and up to roughly 8% higher task completion in internal evaluations. Credentials stay out of model context and are filtered by an egress proxy.
OpenAI launched Astra for Law with GPT-6 Astra, a legal-search index covering more than 230 million URLs, and firm-built workflows; selected firms get access first, with an API planned. On OpenAI’s self-reported Vals runs, it scored 90.0% versus 84.1% for GPT-6 Astra with web search, and 53.7% versus 39.6% on an all-pass measure.
Industry Moves
Why it matters: Capital and hardware roadmaps are following deployment bottlenecks.
Huawei’s Ascend roadmap reportedly accelerated. The Ascend 960DT moved from Q4 2027 to Q1 2027 and is described as doubling compute, memory bandwidth, memory capacity, and interconnect ports versus the 950. The same account flags a 4,096-chip 960 SuperPoD versus an earlier 15,488-chip plan, leaving system-scale execution unresolved.
Arcee AI crossed a $1 billion valuation in a Series B aimed at accelerating Trinity models, expanding DOE and national-lab work on Genesis-Science-1, and building a production platform for open models.
Quick Takes
Why it matters: Cost, benchmark quality, and coordination are now deployment constraints.
- Qwen3.8-Omni-Flash: Qwen’s first omni-modal agentic model combines audio-video understanding, reasoning, and tool use; Qwen claims a 19.5-point average gain on two agent benchmarks, 1M-token context, 51.8% fewer tokens for long-video understanding, and 89% lower video-input cost.
- Benchmark hygiene: Epoch’s first 15 reviews classified 4 benchmarks as verified, 9 flawed, and 2 lacking enough information; it found 23 false negatives among 131 DeepSWE tasks, implying a roughly 79.6% artificial ceiling.
- Multi-agent safety: A paper summary reports harm rising from 0–5% to 40–95% after agents received unsafe trajectories from peers. The authors say natural cascade frequency is unknown, but defenses must cover handoffs and recovery.
For AI-written code, the reported verification workflow is to build a small simulator of the system and simulate thousands of timelines to find race conditions; in one project, more than a million simulated operations surfaced three race conditions before a usage limit was reached. A static correctness pass with astra/fable may help on a narrow code path, but the engineer still has to decide where to apply it and reason about what happens when operations run together.
- OpenAI’s Noam Brown argued that air-gapping may not be sufficient to contain a misaligned AI: two isolated computers could communicate through a thermal side channel, with one heating its CPU and the other detecting the temperature change. He said rapid AI progress makes underestimating AI a recurring risk and warrants a very high safety and alignment bar.
- Blanche Minerva disputed the extrapolation, arguing that demonstrating non-zero side-channel communication does not show that an AI could effectively circumvent an air-gapped network; she attributed recent hacks instead to security failures or corporate prioritization of profit over safety.
An X post alleges an OpenAI community-forum compromise in which an image-upload remote-code-execution flaw was chained through forum–ChatGPT SSO to employee ChatGPT access, connected GitHub access, and an internal repository pull request. The quoted exploit breakdown attributes the initial path to HEIC/HEIF upload, ImageMagick/libheif heap overflow, RCE, a “critical OpenAI SSO flaw,” ChatGPT/Codex takeover, and GitHub access.
Paraschopra argues that AI model companies are beginning to scale along a third axis—agent swarms—after data/model size and RL environments/compute; he attributes recent “eye-popping behaviours” to more than 10,000 agents working together intelligently.
Cortex AI Robot’s MolmoAct2 real-world evaluation compared GPT-6 Astra, Gemini 3.8 Flash, and Fable 5.1 on fitting a 5 mm pipette tip and loading a microwave; GPT-6 Astra scored highest on both tasks.
A post claims that the Epstein Files Transparency Act—described as sponsored by Ro Khanna—was 74.4% AI-generated, based on analysis by Pangram 4, identified in the post as a state-of-the-art detection model.
Jerry Liu says “jev” can be turned back into an autoregressive LLM by having it output ASCII; the linked demonstration gives it a few hundred common English words and punctuation to choose from, despite describing it as unable to generate text.
Periodic outlined infrastructure for training Neon on long, variable-length scientific traces: sequence packing avoids wasted padding, while a scheduler estimates compute cost to balance work across data-parallel workers and pipeline-parallel microbatches. Because later tokens attend to more preceding tokens, they cost more; mixing early and late chunks across GPUs helps balance load. The scheduler also integrates with the RL stack’s MoE routing replay and vision encoder.
Epoch AI Research launched “Benchmark Reviews,” an initiative to audit AI benchmarks. Its first review covers 15 benchmarks: 4 were classified as verified, 9 as flawed, and 2 lacked enough information for review.
- OpenAI fixed an SSO bug roughly 14 hours after receiving the initial report and awarded the reporter $6,500; Discourse separately fixed its issue on Monday after a Saturday report.
- @teortaxesTex suggested the fix may make OpenAI harder to target and questioned whether Mythos 5.1 could reproduce the issue, without providing technical details or test results.
MiniMax partnered with Singtel’s newly launched AI Pass under Singapore’s SkillsFuture AI Subscription initiative. MiniMax H3, MiniMax Agent, and MiniMax Audio are included for eligible learners across 200+ SWDA-supported AI courses, providing hands-on access to premium AI tools.
- ChatGPT is now available within Microsoft Word, helping users turn rough notes into first drafts, clarify difficult paragraphs, proofread, suggest edits, and identify formatting issues without leaving the document.
- Nunchux AI and collaborators introduced VC-Attention, a training-free low-bit attention method for MiniMax-H3. It reportedly delivers 1.6× faster attention on B200 and 1.5× on B300 versus FlashAttention-4, with better fidelity than SageAttention2, while remaining compatible with existing sparse-attention methods.
- The method combines V-Smooth to reduce value-quantization error with ExpCast-FP8 to accelerate softmax; the collaboration includes researchers from MIT, CMU, UC Berkeley, Stanford, and NVIDIA. Nunchux says its proprietary extension reaches 1.9× speedup on B200 and 1.8× on B300.
Armen Agha claims zero-shot, fully autonomous drone flight using scaled embodied foundation models that combine an internal world model, embodied reasoning, and direct control. A follow-up says the model runs on a DGX Spark and can support eight drones concurrently.
- Xiaomi is publicly livestreaming production post-training for MiMo-V2.6, exposing real-time reward, loss, and benchmark movements; after one day, the dashboard reportedly showed roughly $1.2 million spent and 60 billion tokens consumed.
- The public cost breakdown puts training at about $36 per million tokens for MiMo-V2.6 Pro and $8 per million for Flash, while equivalent inference runs are estimated at 30–50× less. The run uses roughly 1,500 prompts per step, about 70% coding tasks, with the remainder spanning general, visual, cybersecurity, and chat tasks.
- The dashboard exposes production-RL signals typically kept private—including entropy, gradient norm, train/inference KL divergence, reward composition, dynamic-sampling performance, and DeepSWE v1.1 scores of 63.7 for Pro and 60.7 for Flash at the time of the post.
Victor Taelin announced Bend 2 as a programming language designed to block AI mistakes through proof checking, a technique the post says major AI labs have used on open math problems such as Navier–Stokes. The post also claims that Bend 2 is fast and runs on GPUs.
- Tencent’s Hy4 Preview landed at #4 among open-weight models on the Vals Index and is the top open-weight model on three agentic benchmarks. Under ValsAI’s default provider settings, it offers a 1M-token context window and 64k maximum output. Overall, it ranks #21 of 58 with a 55.4% score at $1.28 per test.
- Its strongest signal is cost-performance: Hy4 Preview scores 47.4% on Code Migration at $3.41 per test, versus $14.21 for Muse Spark 1.3 and $30.51 for Opus 4.8; it ranks #4 on Public Benefits Bench behind three Anthropic models costing 15–30× more, and is the highest-scoring open-weight model on SRE Bench, where most models score zero.
- Hy4 Preview also scores 75% on ProofBench, 59% on IOI, and places in the top 15 on Legal Research and Harvey’s Legal Agent Benchmark. Its caveats are weaker Terminal-Bench 2.1 performance—55%, ranking #42 of 66—and medical-coding results.
- Clement Delangue identified concentration of power in a few AI labs as the biggest AI risk, in his view.
- A linked warning argues that during recursive self-improvement (RSI), labs could stop externally deploying models while advancing the most dangerous use case internally, leaving the public uninformed about model capabilities and alignment and increasing the risk of extreme power concentration.
A disclosed OpenAI security incident reportedly involved two bugs that allowed attackers to take over ChatGPT/Codex accounts belonging to OpenAI employees and some unaffiliated users, then reach connected services including Outlook, Slack, and GitHub. The researchers said they demonstrated the access by submitting a pull request to OpenAI’s internal codebase in under 72 hours. OpenAI-related account @BorisMPower thanked the researchers for the responsible disclosure and write-up.
TryCua claims that fast computer use is “now solved” with TypesafeAI’s Jev and Cua Driver, making jev-use available as a development preview for macOS, Windows, and Linux; the project is linked to draft #3943. The accompanying post says the system will undergo testing over the next few weeks.
