The pacing-the-frontier debate moved from lab statements into open political conflict. Trump and House Speaker Mike Johnson dismissed the AI industry's slowdown warnings as overreaction even as Sam Altman and Elon Musk backed Dario Amodei's call for restraint, and Stratechery argued the proposal is politically unrealistic. Obama separately urged Democrats to draft a clear AI-safeguards platform, while TechCrunch asked what is actually driving the doom rhetoric. Research output kept moving independent of the politics: three papers converged on making LLM agents choose and route skills more efficiently, StepAudio 3 Gen and SNAP3D pushed unified audio and physically grounded 3D generation, and a new attention-sparsification method and benchmark-discovery engine targeted the cost side of running these systems at scale. Two open-source roundups on Hacker News round out a day where the policy fight and the engineering roadmap are visibly diverging.
Dario Amodei's open letter urging labs to 'pace the frontier' drew public backing from Sam Altman, Elon Musk, and Demis Hassabis, but Trump and House Speaker Mike Johnson dismissed the AI industry's warnings as overreaction. Stratechery argued the proposal is a politically unrealistic path to control rather than a workable policy, TechCrunch examined what is fueling the industry's doom rhetoric, and Obama urged Democrats to build a clear platform on AI safeguards ahead of the fight.
Trump and House Speaker Mike Johnson pushed back on Dario Amodei's call to pace the frontier, even as Altman, Musk, and Hassabis voiced support for it.
Altman and Musk backed Amodei's call for restraint, but Trump's team says the pace of AI development is the industry's own decision, not the government's to slow.
Stratechery argues Amodei's pace-the-frontier proposal is unrealistic and reads mostly as a bid for political control over AI rather than a workable safety policy.
TechCrunch's Equity podcast digs into whether the AI industry's newest existential-risk warnings reflect genuine concern or strategic positioning.
Obama said Democrats need to make AI a central agenda item with a clear plan addressing its economic impact and safety risks.
A trio of papers tackled how LLM agents choose what to use: Diverse Skill Routing applies a determinantal point process to avoid wasting context on redundant skills, a bandit-based online routing paper learns to pick the best LLM expert per prompt from limited feedback, and COBRA-Skills cuts the cost of evolving reusable agent skills by 55-58% versus a prior method. Together they treat skill and model selection as a first-class optimization problem rather than a fixed lookup.
Diverse Skill Routing uses a determinantal point process to rerank agent skills for both relevance and non-redundancy, improving recall on multi-skill queries.
Formulates routing prompts to the best of several LLM experts as a bandit problem, achieving low regret even when feedback on response quality is scarce.
COBRA-Skills evolves reusable agent skills using contextual-bandit-guided prioritization, cutting optimization cost 55-58% versus SkillOpt with only 50 examples per benchmark.
StepAudio 3 Gen introduces a single autoregressive model spanning zero-shot TTS, voice design, sound effects, and music over shared discrete tokens instead of diffusion. SNAP3D generates 3D assembly parts from a single image with physics-based connectors so results can be printed and assembled. A robotics paper reduces reliance on spurious vision-action shortcuts, improving generalization under camera and lighting shifts by up to 16.7 points.
StepAudio 3 Gen is a discrete autoregressive model that unifies zero-shot TTS, voice design, sound effects, and music generation over shared residual-quantized tokens.
SNAP3D generates part-aware 3D assets from a single image with physics-guided connectors, resolving penetration and instability so parts can be 3D printed and assembled.
Latent Interface Training constrains how vision-language-action models use visual cues, improving generalization by up to 16.7 points under unseen cameras and lighting.
SAS trains attention sparsification end-to-end with the language modeling loss instead of distilling dense attention, with larger gains under tight budgets. Benchmark Radar catalogs 1,283 AI benchmarks and 12,916 score observations into a searchable, daily-updated database. PLC-DPO corrects noisy preference labels during DPO training, and a small independent team describes training open-weight cyber-agent models ranking first among peers at comparable scale.
SAS optimizes context-unit ranking for sparse attention end-to-end with the language modeling loss instead of distilling dense attention, with the largest gains under tight budgets.
Benchmark Radar is a searchable, daily-updated catalog of 1,283 AI benchmark records and 12,916 score observations pulled from 37 sources.
PLC-DPO routes each preference pair as clean, flipped, or tied using the policy-reference margin, beating standard DPO's win rate 60.5 to 55.5 across 57 benchmark cells.
A seven-person independent team describes a data-centric framework for training open-weight cyber-agent models, with checkpoints ranking first among peers at comparable scale.