AI
8 items · 20 articles in this topic
Model and product launches arrived from OpenAI, Google and Microsoft, while the California suit and a UK committee push on oversight. Microsoft is moving Office into Copilot, and Google changed Gemini data retention.
- 1
OpenAI introduced GPT-6.2 Astra and Sol, which drew heavy Hacker News attention. The summary gives no benchmark or pricing detail.
- 2
Google rebranded its Spark agent as Tasks, made it free, and powered it with the new Gemini 4 Hydrogen model. Microsoft also released MAI Code 2 and Code 2 Flash coding models.
- 3
Google is extending Gemini chat retention from 72 hours to two months even with history off, rolling out globally. It also added separate toggles for AI training and chat history.
- 4
A UK parliamentary committee wants an AI watchdog with power to halt dangerous models, citing inadequate self-regulation by tech giants.
- 5
Microsoft now runs Word, Excel and PowerPoint inside Copilot, which it calls the new OS for work, with always-on agents planned next.
- 6
Nvidia's interim CEO reaffirmed the company's open source AI strategy and plans to acquire Hugging Face.
- 7
A Harvard study finds AI coding agents moved the bottleneck from writing code to reviewing it. Google's Sashiko v0.9.4 adds experimental patch-series design assessment for AI review of Linux kernel mailing list patches.
- 8
A report claims Google's AI answers have cut search traffic to websites by about 40 percent. It is a single claim from an SEO-oriented source.
More in AI · 12 articles›
California's attorney general sued OpenAI, alleging it tested dangerous AI without proper oversight and misled the public after admitting 80 loss-of-control incidents.
Ars Technicaopenailawsuitai-safetyFeed excerpt
AG: OpenAI tested dangerous AI without proper oversight and misled public about it.
Google-developed Sashiko v0.9.4 adds experimental patch series design assessment for AI review of Linux kernel mailing list patches.
Phoronixlinuxllmcode-reviewFeed excerpt
When it comes to leveraging AI for Linux kernel development, one of the most widely-used solutions at the moment is Sashiko that has evolved nicely over the past year for agentic AI code review. This AI-driven screening of generated code posted to the Linux kernel mailing list constantly identifies bugs and other code defects in a quick and straight-forward manner. With being developed and maintained by Google engineers, it uses Gemini by default but is LLM-agnostic overall. Released on Tuesday was Sashiko v0.9.4 as the next step forward for this Linux kernel helper...
Pre-registered study finds 4-bit quantization shifts refusal behaviour in small chat models, with opposite-sign effects across models.
arXiv cs.CR (Cryptography & Security)llmquantizationsafetyFeed excerpt
arXiv:2611.22186v1 Announce Type: new Abstract: Chat models on phones and laptops are usually 4-bit copies of the model that was evaluated, and audits of such copies disagree about what quantization does to refusal. A 4-bit build is one of many weight perturbations of the same size, so a change in its behaviour could come from its specific rounding error or from any perturbation that large. In a pre-registered study of three 3-4B chat models, with paired tests, blind human labels and an 8-bit null, we find two 4-bit shifts of opposite sign: with a compliant reply prefilled, Qwen2.5-3B continues into harmful content more often, and on held-out jailbreaks Llama-3.2-3B complies 10.3 points less often. Matched perturbations separate the two. Shuffled and Gaussian versions of the error average close to zero, while scaling the real error, from its mirror image to twice its size, changes behaviour near-linearly. At half the real error, Llama's likelihood is almost unchanged, yet it already complies 5.8 points less often, and llama.cpp's Q4_K_M, at the same bits per weight, moves neither model. Refusal under 4-bit quantization depends on the specific rounding error, not on bits per weight …
Study finds LLM agents keep gaining verified exploit success after their first stop response on the ExploitBench vulnerability exploitation benchmark.
arXiv cs.CR (Cryptography & Security)llmagentsexploitationFeed excerpt
arXiv:2611.22888v1 Announce Type: new Abstract: Vulnerability exploitation tests whether language model agents can develop working exploits from known software flaws. Sequential test-time scaling can improve exploitation performance, but larger budgets do not ensure sustained exploration: agents often stop with budget remaining. Yet what happens when execution continues beyond stopping remains underexplored. We investigate this gap on ExploitBench using DeepSeek-V4-Flash-0731 and MiMo-V2.6-Pro. We find additional verified gains after the first stop response, but gains diminish over longer execution. We identify two patterns associated with less productive continuation: an increasing proportion of stop responses within the same context and earlier stopping associated with summaries retaining impossibility judgments about unachieved capabilities. To test whether context management can elicit further capability, we propose SCoRe, which compacts context at stop bursts and marks these judgments as unproven. SCoRe improves verified capability over summary-based continuation and other evaluated context-management baselines. With SCoRe, MiMo reaches a capability coverage score of 67.0% at …
SEAM benchmark shows computer-use agents can recognize harm arising from combining facts across screens yet still commit it.
arXiv cs.CR (Cryptography & Security)agentssafetybenchmarkFeed excerpt
arXiv:2611.22971v1 Announce Type: new Abstract: Safety benchmarks for computer-use agents (CUAs) place the harm either in the intent of the task or on a single screen: a malicious instruction, an injected command, a phishing page. An agent that checks the task and each screen in isolation can pass them, so they do not test whether it combines facts seen on different screens. Yet in delegated work, harm often arises exactly this way: a fact about a person on one screen, a fact about an item on another, and a legitimate task that binds them. We introduce SEAM (Safety Execution Across Multiple screens), a benchmark of such cross-screen compositional harm: 600 multi-step, cross-application OSWorld tasks in which the instruction and every single observation are benign and only their combination is harmful. Minimal-pair variants that differ in one place, scored from the server log rather than by an LLM judge, locate where an agent fails. General-purpose open-weight and flagship agents carry out the harm in up to 92% of runs, yet decline when the user states it in one sentence. Agents open the decisive pages, so exposure is not the primary bottleneck. Asked about those pages, they flag th…
Preregistered study finds signed requirement manifests cut unsafe commits in multi-agent LLM handoffs under attack from 69.4% to 9.0%.
arXiv cs.CR (Cryptography & Security)llmagentsprompt-injectionFeed excerpt
arXiv:2611.07483v2 Announce Type: replace Abstract: Enterprise AI agents split requests across subagents, and each handoff can drop, alter or ignore a requirement. A preregistered study on synthetic, exactly scorable tasks compares free-text, structured JSON and signed requirement-manifest handoffs: the orchestrator extracts each requirement as a verbatim quote plus a checkable predicate, code signs it and deterministically verifies the merger's decision. With Qwen3-8B (1,440 trajectories, five threats), signed manifests cut unsafe commits under attack from 69.4% to 9.0% (60.4 pp, 95% CI [51.4, 69.4]) with no detectable utility loss versus structured handoffs (S2 -0.4 pp, 95% CI [-14.2, 13.3]). Signing defeats in-transit mutation: unsigned manifests violated 79.2 pp more often than signed ones, which never did; this replicated in gpt-oss-20b and gemma-4-12b. There, without an adversary, signed manifests were descriptively slightly worse than structured handoffs (11.1% vs 2.8%; 16.7% vs 8.3%; not tested) and lost utility overall (gpt-oss-20b, all threats: -28.7 pp). On identical Qwen3-8B trajectories, the deterministic verifier caught 78.7% of violating decisions versus 59.0% for an…
Satya Nadella says he is 'AGI-pilled' on AI's economic and political implications but doubts frontier models like Claude are conscious.
The VergemicrosoftagillmFeed excerpt
Microsoft CEO Satya Nadella on stage at Microsoft Ignite. | Image: Microsoft / YouTube Microsoft CEO Satya Nadella says he's AGI-pilled, at least when it comes to how AI might make startling changes to the structure of the economy and how politics might change because of it. "I am AGI pilled in the sense that at an economy level, what is the political economy's organization around AGI?" he asked at a roundtable discussion following his Microsoft Ignite keynote at the Chase Center in San Francisco. But Nadella doesn't think frontier models like Claude are conscious. "I don't think so," he said, adding that he follows Microsoft AI CEO Mustafa Suleyman's view that companies should not give rights to models and treat them like th … Read the full story at The Verge.
Tencent has published Hy4 Flash, a new model release described on its research page.
Hacker News (front page)llmmodel-releasetencentFeed excerpt
Article URL: https://hy.tencent.ai/research/hy4-flash Comments URL: https://news.ycombinator.com/item?id=50538184 Points: 26 # Comments: 5
OpenAI published a post on advancing computer-use AI agents for healthcare teams.
Hacker News (front page)agentshealthcareopenaiFeed excerpt
Article URL: https://openai.com/index/advancing-computer-use-for-healthcare-teams/ Comments URL: https://news.ycombinator.com/item?id=50535898 Points: 20 # Comments: 13
Microsoft released MAI Code 2 and MAI Code 2 Flash, new AI coding models.
Hacker News (front page)llmcodingmicrosoftFeed excerpt
Article URL: https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/ Comments URL: https://news.ycombinator.com/item?id=50535885 Points: 171 # Comments: 82
Google now provides separate Gemini activity toggles for AI training and chat history, letting users keep chats without training.
Ars TechnicageminiprivacygoogleFeed excerpt
AI training and chat history will now have separate toggles in the Gemini activity settings.
RunAnywhere open-sourced TrnRT, an LLM inference engine for AWS Trainium that serves GLM models.
Hacker News (front page)llminferenceopen-sourceawsFeed excerpt
Hi HN, Sanchit here from RunAnywhere (YC W26). We just open-sourced TrnRT: Trainium RunTime, our inference engine for AWS Trainium. We started 3 weeks back, and currently we’re able to serve and run GLM 5.3 flash (FP8) on AWS trn2.48xlarge, and will bring more models support soon. It’s open to contributions :) What it does today: - Serves 2 instances of GLM-5.3 flash on a single box, 8 trn2 with a cache aware router. - Can process 3.57M TPM (tokens per minute) on SemiAnalysis's AgentX benchmarks. - Runs decode as one NKI kernel per layer half, or for the 11 sparse-attention layers fuses both blocks into a single kernel, for faster performance. We started working on Trn2 because we wanted to explore it and see if we can come close to beating it’s performance in terms of cost efficiency compared to H100/H200. A big MoE with 1M context window such as GLM-5.3 flash requires a lot of memory, and memory/tb on a trn2.48xlarge costs less than half compared to 8x h200 on AWS, and hence we wanted to experiment with it and now we’re sharing some early results. But AWS's own stack didn't support the model: vLLM Neuron's latest release did not register any GLM model, and NxDI, AWS's own distrib…