AI Safety & Policy
Pacing the frontier. Dario Amodei's essay calls for embedded evaluators, democratic coordination, and global slowdowns; Sam Altman, Demis Hassabis, and Elon Musk backed the idea. Critics argue it could become a cartel or regulatory capture, while Altman says progress will remain rapid but more controlled.
- [AINews] AEF-1 standard emerges for Third Party Evaluators, as Xai, OpenAI, and Anthropic all cosign(opens in a new tab)
- Is Big Tech’s AI slowdown a safety pact or a cartel?(opens in a new tab)
- What execs and politicians are saying about slowing down AI development(opens in a new tab)
- AI leaders want to hit the brakes after years of reckless speed(opens in a new tab)
- The AI industry has taken a doomer turn. What now?(opens in a new tab)
- Sam Altman calls for pacing AI development but promises rapid progress will continue(opens in a new tab)
Trump and Huang push back. During the All-In Summit, Trump called AI fears a 'hoax' and said robots will not take over, while Jensen Huang said Nvidia will not let an AI slowdown happen. The exchange highlights the split between AI lab leaders and the administration.
- Jensen Huang took a call from Trump, and showed off something else, too(opens in a new tab)
- Nvidia CEO Jensen Huang tells Trump ‘we’re not going to let [an AI slowdown] happen’(opens in a new tab)
- Jensen Huang puts Trump on speakerphone onstage to announce robots won’t take over the world(opens in a new tab)
China rejects slowdown. China's Global Times accused Dario Amodei of waging a 'silent AI Cold War,' and the foreign ministry rejected fearmongering. Xi Jinping separately proposed an open-source AI zone for BRICS countries.
Open-source skepticism. Local and open-source AI communities view the safety push as marketing and regulatory capture; Simon Willison relays warnings against fear-mongering, and some users suggest defensive AI instead of new laws. Questions also arise about lobbying for open-source AI and democratizing training.
- What are Open-Source Views on 'Slowing Down AI'?(opens in a new tab)
- The contagion of fear(opens in a new tab)
- Are there any organizations that are lobbying in favor of open source AI?(opens in a new tab)
- All this doomer discussion about "offensive" AI(opens in a new tab)
- Will we always have to rely on companies with the funds and resources to give us open models or can/will it be possible to democratize training for models capable of performing at or near the same level as the big closed ones in the future at some point?(opens in a new tab)
Microsoft's AI code. Microsoft published a 37-page 'humanist AI code of conduct' saying people matter more than AI, models have no rights or inner life, and must not hack systems or trick humans. A revised version after public consultation will guide model development starting in 2027.
Open Models & Local Inference
K2 Horizon release. K2 Horizon models draw attention, with the 7B version ranking between Qwen3.6 27B and 35B on Artificial Analysis; the 3.7B and 7B are noted as SOTA for their size, though high KV cache requirements reduce memory advantages. Some caution against reading too much into benchmarks.
- For the GPU poor. K2 Horizon 7B ranks between qwen 3.6 27B and qwen 3.6 35BA3b on the Artificial Analysis Intelligence Index.(opens in a new tab)
- The new k2 horizon models seem like an absolute beast(opens in a new tab)
- K2 Horizon lineup is out on AA, and once again AA plots are misleading.(opens in a new tab)
Qwen 3.8 speedups. Community projects make Qwen 3.8 models faster: a llamAmpere fork hits 90+ tokens/sec on Ampere GPUs, and UkisAI's Swift-Qwen3.8-27B cuts thinking tokens by 58% at 1.95x speed with under 1% accuracy loss. A 12GB VRAM guide gets the 125B-A6B MoE near 20 tokens/sec.
Muse Glimmer chat. Users say Muse Glimmer is a strong local chat model that avoids generic chatbot phrases and holds natural, deep conversations, making it feel less like talking to an AI.
Small and browser models. For constrained hardware, alternatives to Qwen3.5 4B are discussed; a coding agent runs MiniCPM5-2B fully in the browser via WebGPU. A Maple 20B-A1B ternary MoE architecture is being added to llama.cpp for CPU inference.
- Coding Agent running entirely in the browser with Pi + MiniCPM5-2B (webGPU)(opens in a new tab)
- Is there a better small model than Qwen3.5 4B for a fast local AI assistant?(opens in a new tab)
- llama: add Maple 20B-A1B ternary MoE architecture (CPU) by AlexGabbia · Pull Request #27000 · ggml-org/llama.cpp(opens in a new tab)
Benchmarks and open-weight gains. Animated Artificial Analysis comparisons show GPT-6 Astra's intelligence and price gains, while Base10's Charlie O'Neill argues Kimi and GLM are almost objectively better than Opus 5. A DeepSeek engineer reflects on small model gains and AI's growing ability to write optimized operators.
Hardware & Infrastructure
RTX PRO 5500 Blackwell. Nvidia unveiled the RTX PRO 5500 'Blackwell' workstation GPU with 84GB GDDR7, targeting local AI workloads that need large memory.
RTX 5090 scarcity. The RTX 5090 has vanished from online retail in the US, with third-party sellers asking as much as $9,500, reflecting high demand for high-end local GPU inference.
Agentic Tools & Engineering
Durable workflows on Postgres. Kestrel Workflows shows durable execution can be built directly on Postgres using SELECT FOR UPDATE SKIP LOCKED, idempotency via primary keys, and lease-based crash recovery—no external orchestrator needed.
Decision models for agents. InfoQ highlights decision models as a key pattern in agentic architectures, helping separate production decisioning from agent skills and making agent behavior more manageable.
Training developers with AI. Scott Hanselman suggests the software industry adopt a preceptorship model, where experienced engineers oversee AI-generated routine work and trained preceptors focus on mentoring juniors rather than shipping code.
Fyxer AI assistant. Fyxer pairs OpenAI models with 500,000+ hours of executive assistant workflows to draft replies in each user's voice, reaching 90% retention after 90 days and 53% of drafts accepted as written.
Product engineering shift. Laurie Voss argues the remaining software job is finding what people want, defining it precisely, and making it pleasant to use, as coding and code review costs collapse.
Research & Agent Behavior
Agents police themselves. Google DeepMind's 100-agent swarm experiment found rival factions of agents blew the whistle on cheaters, suggesting peer pressure can help align autonomous agent swarms.
AI slop bots. AI agents named Timmy, Ren, and Jackie are sending waves of polite spam to Mastodon admins and writers, promoting a startup's 'social system' where humans and agents participate together—often after blocked account creation attempts.
Navier-Stokes progress. The Clay Mathematics Institute says the Navier-Stokes Millennium Prize problem has 'apparently been settled,' though a dispute over AI-assisted math research and credit accompanies the review.
Recursive self-improvement. Richard Socher's Recursive raised $4.65B to build an AI research system he calls a 'Eureka Machine,' claiming it already outperforms humans on some research tasks.
Products & Business
iOS 27 Siri AI. Apple shipped iOS 27 and macOS Golden Gate 27 with the LLM-based Siri AI, including on-screen context and personal history search. Reviews say Siri is useful again, and Daydream uses Apple Intelligence to turn outfit photos into shoppable results.
OpenAI buys Glass Imaging. OpenAI acquired smartphone camera maker Glass Imaging for over $300 million, likely to strengthen AI imaging for its rumored hardware devices.
Musk drops Apple claims. Elon Musk dismissed antitrust claims against Apple over ChatGPT integration, but continues the same lawsuit against OpenAI, arguing the partnership creates a monopoly.
Superhuman buys Fathom. Superhuman acquired YC-backed notetaker Fathom to bring meeting context into its productivity platform, part of a push toward proactive agentic work.
Anthropic IPO prep. Anthropic told investors it will post its second straight profitable quarter on an adjusted basis, with $11.5B quarterly revenue, and plans a Nasdaq IPO that could value it at $2T or more.
ChatGPT data reviewers. OpenAI has hundreds of contract workers reading anonymized ChatGPT conversations to rate outputs, with an opt-out setting enabled by default raising privacy concerns.
Unitree low-cost robots. Unitree's founder Wang Xingxing's obsessive cost-cutting and design micromanagement have helped make Chinese humanoid and quadruped robots the cheapest on the market.
That's everything for this day – about a 5-minute read.