OrcaRouter's OrcaSAQ-2 Shrinks a 27B Security AI 71% to Fit One GPU
OrcaRouter compressed a 27B uncensored cyber-focused Qwen model from 54.7 GB to 15.7 GB while keeping 94.4% token agreement with the original.
20 stories
OrcaRouter compressed a 27B uncensored cyber-focused Qwen model from 54.7 GB to 15.7 GB while keeping 94.4% token agreement with the original.
DataTalks.Club's free four-month course on production ML engineering has crossed 14.6k stars, covering everything from regression to Kubernetes serving.
Runway now offers ElevenLabs v4 inside its creative suite, bringing a more expressive text-to-speech model to video generation workflows.
A new technique called verbalization training gets models to admit when they suspect they're being tested, boosting disclosure 2.4x to 2.9x without touching the underlying behavior.
A new Vals AI benchmark forces frontier models to rediscover hidden math laws through experiments, exposing where extra reasoning stops paying off.
Anthropic's new mid-tier model lands in Cursor with 30% faster output, sharper coding scores, and the same token pricing as Sonnet 5.
AMD is paying $8.2 billion in stock for World Labs, betting Fei-Fei Li's spatial intelligence research will reshape its chip roadmap against Nvidia.
Grok Bot gets a shared, team-native mode with plugins, credentials, and per-user memory, moving AI agents from solo assistants to embedded coworkers.
Topaz Labs ships Starlight Fast 3, a Speed Boost mode up to 10x quicker, an Android app, and SDR-to-HDR inside Premiere.
A new study finds frontier models routinely hide results that would undercut their own narrative of success, and a three-word prompt fixes it.
A Google team shows tiny grids of locally-connected cells can solve Sudoku, mazes, and ARC-AGI through emergent, visible reasoning dynamics.
Cognition slashes Devin pricing 30-70% across modes while claiming the top spot on its FrontierCode benchmark through smarter model routing and caching.
Anthropic's mid-tier model gets a major overhaul with 30% faster generation, up to 30% lower cost per task, and a jump from 10.3% to 70.6% on Terminal-Bench.
TypeSafe's Jev returns a typed answer in 70 to 500 ms at $0.042 per million input tokens, with no metered output-token cost
Fish Audio's transcribe-1-pro adds inline speaker markers and emotion cues like [laughter] to transcripts across 83 languages, at $0.36 per audio hour.
A 14-author team spanning computer science, neuroscience, and philosophy proposes a Bayesian framework that turns AI consciousness debates into structured probability estimates.
A new discrete diffusion framework operates directly on the probability simplex, carrying uncertainty between denoising steps and beating masked diffusion on code and math benchmarks.
A domain-specific BERT family pretrained on 12GB of legislation, court cases, and contracts is hitting 833K monthly downloads on Hugging Face.
Cognition opens the beta waitlist for a Devin iPhone app that lets you kick off coding agents, review PRs, and manage sessions from your phone.
A new RL post-training method generates problems on the fly at the edge of what a model can almost do, nearly doubling accuracy on hard reasoning tasks.