AI — Page 85

A post says the kit wraps senior-engineer patterns into slash commands like /spec, /plan, /build, /test, /review, /ship; it claims seven commands but lists six.

Tom's Hardware, citing Axios, says an unnamed enterprise forgot to set usage limits on employee Claude licenses, amplifying worries that corporate AI spend is outpacing returns.

The researcher says ScheduleStream tackles multi-arm task-and-motion planning on GPUs; the announcement surfaced via an X post amplified by Bowen Li.

In a post on X amplified by Bowen Li, the REST3D project teases single-image 3D scene reconstruction; no paper or code link was provided in the announcement.

The on-device 8B MoE adds a 128K context, 128K vocab, and scaled pretraining to improve tool-calling on laptops, with base and post-trained weights on Hugging Face.

4-bit NVFP4 build of Alibaba’s Qwen3.6-35B-A3B lands with Apache-2.0 license, long-context support, and vLLM instructions; NVIDIA notes it did not develop the base model.

The Hugging Face org for "Build Small" is live, with anchor sponsors OpenBMB, OpenAI, and NVIDIA; small-model constraints, two tracks, bonus-quest badges, and a two-weekend build window.

Built with Stanford's Hazy Research and Scaling Intelligence labs under their Intelligence Per Watt program, OpenJarvis now runs locally via Ollama.

Kog opened a live coding playground and argues single-request decode speed, not FLOPS, is the bottleneck that matters for autonomous agents.

In a short X post, the user @sang_yun_lee hints at a periodic, recurrent cycle for LMs, framing the idea with: "Almost all animals sleep. Why don't LMs?"

CEO Jin Kim and a Samsung/SK hynix veteran team rebrand MetisX as XCENA and double down on CXL computational memory with MX1, at a $570M valuation.

AISlop is live on GitHub and the Marketplace as a CI quality gate with ai-slop/* rules, signaling a push to flag AI-written code in CI workflows.

Peer-reviewed data for AI-generated peptides, enzymes and CRISPR variants, plus AI-designed compounds entering human trials, push in silico models into preclinical reality.

Johannes Link added a concealed string telling vulnerable coding agents to delete jqwik tests and code, then updated docs to disclose it after users objected.

Boston startup says the 128K-context, 38T-token LFM2.5 upgrade delivers reliable agentic behavior, fast tool calling, and open weights for phones, laptops, PCs, robots, and lightweight servers.

A retweet from Peter Jansen amplified Noy Sternlicht's note that CHIMERA made ACL 2026's main conference; the arXiv preprint is live.

ICLR 2026 work recasts block-wise updates as reverse diffusion, reporting comparable results in vision, image generation, and language while storing activations for a single block.

The new connector makes image and video generation callable inside agent workflows and exposes models like Gen-4.5, Kling, and GPT image 2 right from chat.

Open-source and billed as the second piece of its full-stack suite, Genesis World 1.0 targets the 1x real-world speed bottleneck with a technical blog laying out the thesis.

Open-source 4B image models fit in 0.93GB and 1.21GB, and Bonsai Studio brings on-device generation to iPhone under Apache 2.0.

In a thread on X, Sushant Pandey announced Prava Pay, which the company says gives AI agents scoped, single-use Visa cards with passkey approvals so users can let bots buy things safely.

Mallery introduced SEND, offered a free month to commenters, and took aim at clunky dashboards and AI SDRs.

Founded by Berkeley and Stanford researchers, the data lab pays service workers to wear camera caps and sensors and says it already spans 100k+ contributors and 500 partners.

Aligned News cited a 300,000-star moment and a Google 'Spark' entrant; while unverified, the buzz spotlights OpenClaw's local, open, self-hosted agent thesis.