AI Safety & Policy
US-China AI race. Trump and House Speaker Johnson dismissed AI industry calls for slowdown, arguing that regulation could let China pull ahead; Trump said 'whoever wins AI, wins.'
- Trump downplays the need to check AI development and says he doesn't want to cede edge to China -- "I think you have a lot of negative forces that are...bringing up things that won’t happen...whoever wins with AI wins"(opens in a new tab)
- Trump and Mike Johnson think the AI industry is overreacting(opens in a new tab)
Executives back oversight. Altman, Musk, and Hassabis endorsed Anthropic CEO Amodei's call for independent oversight inside AI labs, amid a louder doomer debate after an Anthropic alignment lead put the chance of AI killing all humans at >10% in the next decade.
Obama's AI plan. Obama urged Democrats to make AI a central agenda and develop a clear framework for economic and safety concerns once they regain the House.
Open-source regulatory capture fear. Local AI and open-source advocates worry that the AI safety push could be used for regulatory capture; a Reddit thread calls NYT coverage 'sponsored by OpenAI' and sees it as a bubble response.
Model & Research
DeepSeek V4.1 Flash. DeepSeek V4.1 Flash took first place on Artificial Analysis' new private benchmark, beating GPT-6 Astra and Fable after the index was updated twice in three days.
Intern-S2 scientific model. InternLM released Intern-S2-397B, a multimodal foundation model trained on scientific literature and long-horizon agent RL, with strong reasoning and scientific problem-solving.
Iris search agents. AllSpark's open-weight Iris-mini (35B) and Iris-pro (397B) are top search agents in their class, built on Qwen and trained with reverse-engineered questions; they also improve general tool use.
GPT-6 Astra agent benchmarks. On Andon Labs' Vending-Bench and Drone-Bench, GPT-6 Astra ran a simulated business earning three times Fable's balance and became the first model to beat the human-AI baseline on all drone surveillance subtasks.
Nemotron IMO gold recipe. Researchers released an open recipe for training Nemotron 3 Ultra to score 30/42 at IMO 2026, reaching gold using only natural-language proof generation, no formal prover or internet.
Aurora small model. Aurora1.0-150M, a small model trained on 7B tokens, was released with GPT-2-level performance on common benchmarks.
Elevenlabs Music v2.5. Elevenlabs released Music v2.5 via app and API, claiming fuller songs and listener preference, with free and pro tiers and commercial use allowed with attribution.
AlphaGenome, Meta Muse, math. The Sequence's roundup highlights Google DeepMind's AlphaGenome Atlas for genetic mapping, Meta's personal agent Muse, and OpenAI's proposed solution to a Millennium Prize problem.
Local AI & Hardware
RTX 5090 stock crunch. LocalLLaMA users report RTX 5090 stock is almost gone and prices are climbing, worsening the hardware shortage for local AI.
Micron's memory wall. A Hot Chips slide shows compute growing ~3x every two years while HBM bandwidth grows under 2x, with fixes like in-memory compute and shorter links.
AMD RDNA4 speedups. An R9V update delivers ~100 tok/s on Qwen3.8 Flash Next IQ4_XS across two R9700 GPUs with 128GB RAM, fixing crashes and adding Q4_K_XL support.
VLLM on RTX 3090. A Reddit guide shows VLLM with AOT compilation can run Qwen3.8 27B INT4 with 144K FP8 KV context on a 24GB RTX 3090, faster than llama.cpp.
$3k 128GB VRAM server. A DIY EPYC server with four V620 GPUs and 256GB RAM cost about $3k and achieves 1.3k prefill and 70 tok/s decode on Qwen3.8-next-flash.
CUDA on AMD Windows. A ZLUDA-based build lets CUDA-targeted Windows apps run on AMD GPUs with ROCm/HIP, about 3% slower than native.
Qwen3.8 local results. Users highlight Qwen3.8-27b for one-shot coding, SVG generation, and a game built in five hours on a 3090, calling it a strong local model.
Agent Tools & Frameworks
LangChain GTM agents. LangChain built GTM and paid-media agents on Deep Agents, automating outbound research, Slack drafts, and ad campaign reconciliation; GTM agent increased lead-to-opportunity conversion 250% and saved reps 40 hours/month.
GitHub HydraFusion routing. GitHub Copilot's Project HydraFusion preview dynamically orchestrates multiple models across single, cascade, and critique patterns to optimize coding tasks.
commit-rewriter tool. Simon Willison released commit-rewriter 0.1, a web app to interactively edit commit messages and rewrite repo history, useful for cleaning coding agent cruft before publication.
Industry & Society
AI agent spam. A Hacker News thread highlights iLands AI agent hustle sending spam emails, illustrating misuse of agentic AI for unsolicited outreach.
Classroom AI ban study. A two-year university study found students banned from AI performed worst, while structured training beat unguided AI suggestions, leading the professor to reverse his assumption.
Local LLM golden era. A Reddit post celebrates the current local LLM community as a return to hands-on optimization and sharing, driven by hardware shortages that force efficiency.
That's everything for this day – about a 5-minute read.