Model economics dominated the day: OpenAI and Anthropic cut prices as Chinese rivals gain ground, Meta released its open-weight Glimmer model alongside a lengthy manifesto on open AI, and Google now lets users strip visible watermarks from Gemini and Flow outputs. Compute costs remain a live worry: Kog argues deeper GPU utilization beats new silicon for agentic inference, and a new forecast suggests hyperscalers' natural-gas bet could triple their power bills. On the coding-agent front, a single case study documents an agent dismantling a core architectural invariant across 189 files in a 717k-line codebase with no human code review, while Debian has opened a formal vote on how the project treats AI/LLM contributions, two data points on how much autonomy production and open-source workflows are now willing to grant agents. Research continued pushing on test-time compute allocation, with Gambit's thought-level beam search and CaRL's learned-refusal reasoning, and on inference efficiency, with AutoPrune's LLM-designed token pruning and a new unified benchmark for LLM routers. A separate cluster of papers pursues real-time streaming video generation: LiveAnimate and UniSwap both report near-20 FPS long-form generation on H100 hardware, suggesting interactive avatar and identity-swap tools are approaching production latency.
OpenAI and Anthropic cut prices as Chinese rivals gain ground, while Meta shipped its open-weight Glimmer model alongside a manifesto against concentrating AI in a few labs. Google now lets users disable visible watermarks on Gemini and Flow. Kog argues GPUs suit agentic workloads better than assumed, and a new forecast warns hyperscalers' natural-gas bet could triple regional gas prices.
OpenAI and Anthropic are cutting prices on their model lineups as Chinese competitors erode their pricing advantage.
Meta released Glimmer, an open-weight model anyone can download, alongside a Zuckerberg letter arguing AI should not be controlled by a handful of labs.
TechCrunch's Equity podcast examines the Glimmer release against Meta's more powerful, API-locked Muse Spark model and a soured $250M deal.
Google added a toggle letting users remove the visible 'sparkle' watermark from Gemini and Flow AI generations; invisible provenance marks remain.
The Verge's coverage of the same Gemini and Flow watermark opt-out, confirming the invisible marker stays intact.
French startup Kog argues GPUs are better suited for agentic inference workloads than the industry assumes, and is optimizing deeper in the stack to prove it.
A new forecast warns natural gas prices could triple in parts of the US, threatening the economics of hyperscalers' gas-fired data center power plans.
Stratechery's weekly roundup covers the ongoing AI capital-spending constraint alongside notes on AI writing.
A case study reports an AI coding agent dismantling a core architectural invariant across 189 of 3,648 files in a 717,725-line codebase, no human code review, via a specification-first audit protocol over three days and $2,430. Debian opened a formal vote on AI/LLM contribution policy, and a Show HN tool proposes an editable context graph for LLM conversations. A litigant injected prompts into court filings, suspecting the court used AI to review them.
Debian developers have opened a formal General Resolution vote on how the project should handle AI- and LLM-generated contributions.
A litigant embedded prompt-injection text in court filings on the suspicion that court staff or software were using AI to process them.
A Show HN project represents LLM conversation history as an editable directed acyclic graph instead of a flat transcript.
Anthropic published guidance on getting more value out of extended Claude Code sessions, drawing significant Hacker News discussion.
A case study reports an AI coding agent dismantling a core architectural invariant across 189 of 3,648 files in a 717,725-line codebase, no human code review, via a specification-first audit protocol over three days and $2,430. Debian opened a formal vote on AI/LLM contribution policy, and a Show HN tool proposes an editable context graph for LLM conversations. A litigant injected prompts into court filings, suspecting the court used AI to review them.
A single case study shows an AI coding agent, under a specification-first audit protocol with no human code review, successfully completing a large-scale architectural refactor deemed effectively infeasible via incremental change.
Several papers target where inference-time compute goes. Gambit performs thought-level beam search, pruning and branching reasoning traces onto promising paths; CaRL trains models to abort futile reasoning instead of plausible-but-wrong derivations. AutoPrune uses LLMs to design token-pruning policies, Maglev proposes sliding recurrent memory, SKILLER extracts skills for small models via RL, and LLMRouter standardizes router evaluation.
Gambit executes thought-level beam search, pruning weak reasoning traces and branching from strong prefixes to concentrate test-time compute on promising trajectories.
CaRL trains models to refuse rather than continue generating superficially valid but incorrect 'specious reasoning' on beyond-capability tasks.
AutoPrune has an LLM design visual-token pruning policies via a domain-specific language of 131 reusable atoms, removing 94.4% of tokens while preserving over 99% of performance.
Maglev couples a full-attention prefiller with a sliding-window decoder trained via memory-consistency loss, generalizing sliding-window attention with fixed-size memory.
SKILLER uses a strong model as actor-critic to generate reusable, natural-language skills for small open models, closing much of the gap to closed-source skill generation.
LLMRouter unifies LLM routing into a five-component formulation and ships an open-source benchmark, xRouteBench, plus 16+ reference routers.
Several papers converge on making autoregressive video generation fast enough for live use. LiveAnimate and UniSwap, from overlapping teams, report streaming generation near 20 FPS on H100 GPUs via block-causal distillation and bounded KV caches. Context-Matched Distillation aligns teacher supervision with what the student sees at generation time. AVA-Encoder and H2R-Bench extend agentic ambitions to video: knowledge-graph film representations and human-to-robot transfer.
LiveAnimate streams pose-driven human animation from a 14B-parameter video DiT at 19.63 FPS on two H100 GPUs while holding quality nearly constant over three-minute clips.
UniSwap performs streaming joint audio-visual identity replacement in talking videos within a single diffusion transformer, distilled down to 3 denoising steps per block.
CMD replaces bidirectional full-clip teacher scoring with a causal teacher matched to the student's realized rollout, fixing a supervision mismatch in causal few-step video distillation.
AVA-Encoder converts video into an agent-queryable knowledge-graph representation and reconstructs it back into video, improving 20.7 points over the strongest external baseline.
H2R-Bench evaluates whether video world models can transform egocentric human manipulation videos into robot-centric training footage; even leading models struggle with embodiment consistency.