The White House cleared Anthropic to release Mythos to more than 100 vetted US organizations, ending the two-week Fable-era pause. Hours later, OpenAI previewed GPT-5.6 Sol and confirmed the US government will decide which users get access — TC notes OpenAI's stance that vetting 'shouldn't be the norm.' DeepSeek open-sourced DSpark inference optimizations that yield 60-85% faster generation (309 HN), keeping the cost-led pressure on the closed frontier. The NYT amended its OpenAI/Microsoft copyright suit after the recent SCOTUS Sony ruling, alleging Microsoft built a copyright-infringing supercomputer. South Korea announced it will train its entire half-million-strong military as 'drone warriors.' OpenAI poached Uber's India chief to run its largest market outside the US. Europe is increasingly fed up and wants its own AI.
The Trump administration permitted Anthropic to release Mythos to over 100 trusted US organizations, ending the two-week Fable-era shutdown. OpenAI then previewed GPT-5.6 Sol and confirmed the US government will decide which orgs get access — OpenAI publicly objected to vetting as a 'long-term default.' Two-week-old pre-release federal review is now operational on both frontier labs.
Semafor: White House permitted Anthropic to release Mythos to a selected list of trusted US organizations (461 HN points).
TC: more than 100 US companies and agencies are authorized to use Mythos 5 — the scope of 'trusted' is operationally defined.
OpenAI previews GPT-5.6 Sol (1,026 HN points) — the model the White House asked OpenAI to slow-roll.
WaPo via HN (1,055 points): US government will gate access to GPT-5.6 by user — pre-vetting moves from model release to user roster.
OpenAI complies with the federal slow-roll but publicly states this kind of access process shouldn't become the long-term default.
TC: the AI competitive frame is shifting from lab-vs-lab to lab-vs-state — capability progress has produced real political consequences.
DeepSeek published DSpark, an inference-optimization package that yields 60-85% faster generation — open weights and open code. The contrast with the federally-gated US frontier is now stark: open-weight Chinese stack accelerating, US frontier on a trusted-user roster.
DeepSeek publishes DSpark — open-source inference optimizations claiming 60-85% faster generation (309 HN points).
The NYT amended its OpenAI/Microsoft copyright complaint after a recent SCOTUS Sony ruling, alleging Microsoft built a copyright-infringing supercomputer. South Korea will train its entire half-million-strong military as 'drone warriors.' Europe is publicly fed up and pushing for its own model stack.
Ars: NYT amended its OpenAI/MSFT complaint after a SCOTUS Sony ruling — adds the supercomputer build itself as alleged infringement.
Ars: SK will train its 500,000-strong military on drones as a 'universal combat tool.'
Wired: Europe wants its own frontier model — long odds, but a real sovereign-AI push is underway.
OpenAI hires Uber's India chief to run India operations — its largest non-US market.
A new paper shows hallucination in world models is predictable and preventable — actionable for embodied AI. Across 67 frontier models, routing/voting/Mixture-of-Agents hits a 'co-failure ceiling' that limits combination gains. A 'neglected free lunch' from post-training gives LLM agents a 'progress advantage' worth measuring.
Argues world-model hallucination is predictable and preventable — actionable detection and mitigation for embodied AI.
Empirical co-failure ceiling on routing/voting/MoA across 67 frontier models — combination gains saturate when models share failure modes.
Identifies a 'progress advantage' from post-training that benefits LLM agents disproportionately — a quiet win in training pipelines.
Information-aware KV cache compression for long-reasoning workloads — better retention at lower memory.