Asian AI startups are shipping Mythos-style models as Anthropic's federally-gated export situation drags on — exactly the substitution effect the trusted-user roster invites. Apple's Vision Pro lead reportedly left for OpenAI, the second senior Apple-AI departure of the cycle. SoftBank's CEO joined the chorus questioning Elon Musk's orbital-data-center pitch. The Verge ran a consumer piece on Apple raising prices to fund AI capex, and Margaret Atwood reduced the AI critique to 'garbage in, garbage out.' On the research side, a paper on discretizing reward models and an Allen-AI-adjacent piece on how post-training shapes biological-reasoning models close the week. A quiet Sunday after Friday's federal-policy shock.
Asian AI startups are launching Mythos-style models while Anthropic's federal export gate persists — substitution exactly as predicted. Apple's Vision Pro exec is reportedly leaving for OpenAI, the latest Apple-AI brain drain since the M7-line pivot was confirmed.
TC via HN (244 points): Asian startups are shipping Mythos-style models as Anthropic's export gate drags on.
TC: Apple's Vision Pro lead is reportedly departing for OpenAI — fresh Apple-AI talent loss after the M7 line was confirmed.
SoftBank's CEO joined critics questioning Elon Musk's orbital data center pitch. On the hands-on side, Wayfinder Router (75 HN) ships a deterministic router for splitting queries between local and hosted LLMs.
TC: Masayoshi Son adds his voice to the skepticism about Musk's orbital-data-center pitch.
Open-source deterministic router that splits queries between local and hosted LLMs (75 HN points).
Margaret Atwood reduces the AI critique to 'garbage in, garbage out.' The Verge runs a long piece on Apple raising consumer prices to fund AI capex. A TC profile of a founder using AI to fight his cancer diagnosis is the human-interest counterpoint of the week.
Verge interview: Margaret Atwood characterizes the AI problem as 'garbage in, garbage out.'
Verge: Apple's consumer price increases are partly funding the AI-infra spend.
TC profile of a founder using AI tools throughout a cancer diagnosis and treatment.
Quiet Sunday HF list. Discretizing Reward Models, How Post-Training Shapes Biological Reasoning Models, and OpenBioRQ (a benchmark of unsolved biomedical research questions for agents) are the noteworthy entries.
Discretizing reward models — discrete-bin reward modeling for more stable RLHF and agentic training.
Analyzes how post-training stages shape biological-reasoning models — actionable for specialized scientific tuning.
OpenBioRQ benchmark — unsolved biomedical research questions used as an evaluation set for agents.