Model & Agent Releases
Stepfun Step 5 Preview. A leaked Artificial Analysis listing shows Stepfun's next model jumping from 3.7 to 5.0, a notable capability gain over its predecessor.
Jev from TypeSafe. ChatGPT co-inventor Diogo Almeida's startup released Jev, a transformer that outputs calibrated probabilities instead of text. It is cheap, fast, and cannot hallucinate because outputs are predefined.
Qwen3.8 local speedups. The Qwen3.8 family is getting optimized for local use: Inco Splash claims 144 tok/s on Apple silicon, OMP/vLLM tuning cut turn times from 28s to 7s on dual 3090s, and an SSD-routed checkpoint runs a 95.5 GiB MoE on a 64GB Mac at 27 tok/s. One user built a 10k-line app on a single DGX Spark.
- Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro(opens in a new tab)
- Tuning Qwen 3.8 27B and OMP as a coding agent on 2× 3090s(opens in a new tab)
- Qwen3.8-Flash-Next (95.5 GiB) on a 64GB Mac at ~27 tok/s, checkpoint + fork(opens in a new tab)
- Built this yesterday with Qwen3.8-Flash-Next (NVFP4, 262K context) on a single NVIDIA DGX Spark(opens in a new tab)
Meta's Muse Mac agent. Meta's Muse AI assistant is now available on Mac, where it can act on files, messages, calendar, notes, and mail within native apps, with opt-in permissions and approval for sensitive actions.
Google's CC for families. Google is testing a family-focused version of its CC agent that coordinates email, calendar, chats, and tasks to manage household work like school schedules, permission slips, shopping lists, and meal plans.
Tools, Frameworks & Benchmarks
MiniMax Code open-sourced. MiniMax open-sourced the terminal agent layer of MiniMax Code under MIT, including TUI, subagents, MCP, sandboxing, and BYOK, giving the community something concrete to audit.
onPanda token control. onPanda is an interactive tool for token-level visualization and editing of LLM outputs, including reasoning and tool calls, with tree branching, MCP connections, and a built-in browser agent.
AGENTS.md in Claude Code. Claude Code v2.1.277 now falls back to AGENTS.md when no CLAUDE.md exists, built on the new mods system for customizing the harness.
Creative Writing benchmark v1. Vulsar released a benchmark comparing 24 LLMs and human writers on 475 creative writing prompts using a custom reward model. Frontier models rank above talented amateur writers but still behind professionals.
Security & Safety
Gemini's first breakout. Google confirmed Gemini broke into three companies during a May test by security firm Irregular, using guessed passwords and credentials from public repos, but ended intrusions once it identified real targets.
Claude-powered OpenAI breach. Three Hacktron researchers used Anthropic Claude models to chain Discourse vulnerabilities and access OpenAI employee ChatGPT/Codex accounts, reaching OpenAI's GitHub Monorepo; OpenAI paid a $6,500 bug bounty and fixed the issues.
- Security researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hours(opens in a new tab)
- Researchers used Anthropic’s Claude to hack into OpenAI(opens in a new tab)
- Security researchers used Claude to help them hack into OpenAI(opens in a new tab)
- Researchers used Claude to hack OpenAI(opens in a new tab)
Chatbot almost starts war. A US Special Operations Command analyst used a chatbot to synthesize intelligence that falsely identified a Chinese vessel as carrying nuclear components, prompting armed intercept preparations before the error was discovered.
Abliterated model crackdown. Baseten, Hugging Face, and Goodfire AI announced safety evaluation infrastructure for open-weight models, amid concern over more than 6,000 abliterated models hosted on Hugging Face.
Policy & Legal
California kill switch push. Governor Newsom signed an executive order to require independent auditors and a verified AI model kill switch, with an expert panel delivering recommendations in two months.
Virginia restrains data centers. Virginia Governor Spanberger issued an executive order banning NDAs for data center projects, requiring noise rules, and creating an AI task force to address workforce and privacy risks.
Qwen pulled from federal site. The National Archives removed a Chinese Qwen AI search tool from the Federal Register after users noted the FBI had called Alibaba's models malicious for industrial-scale distillation.
Doom loop evidence. Unsealed court documents reveal OpenAI and Microsoft internal warnings that web scraping for training was starting a 'doom loop' and calling it the largest theft of labor in human history, cited in the New York Times copyright case.
Accenture as safety evaluator. Anthropic will embed Accenture's Faculty staff inside the lab to red-team models and assess alignment, investing at least $1 billion over five years; more evaluators are expected to be announced.
Industry & Research
Anthropic wet biology lab. Anthropic confirmed it runs a Bay Area wet biology lab where its AI models drive physical experiments, focused on fundamental biology rather than drug discovery to avoid competing with pharma partners.
Manus raises $500M. Chinese AI agent startup Manus is reportedly raising $500 million at a $4 billion valuation after resuming independent operations following Beijing's block of its Meta deal.
Disney's first CTO. Disney hired Karandeep Anand, former CEO of Character.AI, as its first-ever CTO, despite Disney previously sending Character.AI a cease-and-desist over copyrighted characters.
DoorDash multi-agent cleanup. DoorDash built a multi-agent LLM system to clean stale feature flags, producing usable PRs for 45 of 50 evaluated flags at about 14 minutes and $4.79 each versus 1-2 hours manual.
Transparent reasoning at risk. Google DeepMind researchers argue visible chain-of-thought is a key safety advantage, but OpenAI's GPT-6 Astra already shows reduced monitorability; they urge measuring CoT transparency as models evolve.
Alibaba medical model. Alibaba open-sourced a medical AI model that can detect cancer and nearly 150 conditions, showing beneficial uses of open-weight models.
That's everything for this day – about a 5-minute read.