AI — Page 88

In an X thread, Will Held says small-run fits predicted a 25B-param, 600B-token run at ~1e23 FLOPs, matching Paloma macro loss within 0.2% error.

In a post on X, the company argues agent behavior should be controlled by enforceable policies rather than instructions buried in a prompt.

Meta's AIRA architecture discovery work signals that agents are starting to help design neural models.

A short X video frames the bet: focus on robot hands and the failure points in everyday tasks like pouring coffee and grasping objects.

Built across 23 live SaaS apps and 106 long-horizon tasks, the open benchmark finds frontier agents stumble on planning, memory, cross-app context, and error recovery.

An X post cites 34.4% vs 32.3% U.S. enterprise share for Anthropic with no disclosed methodology. The gap is slim and could reflect marketing, product fit, or buyers seeking an OpenAI alternative, even as some developers say OpenAI's code tools beat Claude Code.

A Colossus magazine Q&A and profile cites Devin at a $445M annualized run rate with usage doubling every eight weeks, and names the US Army, Goldman Sachs, and Mercedes-Benz as early customers.

Majority-owned unit will embed Forward Deployed Engineers; backed by 19 partners and starting with 150 specialists and $4 billion, per OpenAI and Greg Brockman.

The six-month program begins this summer, selected from 124 applicants and backed by Google and GenerationAI.

MIT researchers Md Ferdous Alam and Faez Ahmed release a demo, paper and code for an image-conditioned model that outputs full parametric CAD command histories.

The research group linked to a write-up and cited 98-100% figures after red-teaming DeepSeek’s latest model, according to a post amplified on X.

A crates.io debut touts multi-model support, a TUI, sandboxed bash, and a permission system in an 8.9MB binary that idles near zero CPU.

The NVIDIA research team says the open-source code can turn a single image plus a camera path into a minute-long, controllable clip; weights are listed as coming soon.

In a post on X, Roemmele says the model generates controllable, physics-rich, high-fidelity worlds; the post did not include a repo or paper link.

CEO Glauber Costa says maintainers were drowning in LLM-made PRs and urges new OSS governance as the team’s SQLite-compatible engine marches toward 1.0.

SuperGrok Heavy subscribers can try the new command line tool at x.ai/cli as xAI asks for feedback to improve both the model and product.

The open-source diabetes project publishes v0.7.2, adds an Open Collective fund via a FLOSS manifest, and points to a hosted service aimed at non-technical users.

Skalpel AI opens 500 early seats; the founder says the service targets cost control and usage-limit pain for developers.

Song introduces Aiphrodite, promising AI customer personas to pretest reactions to ads, PDPs and packaging so teams can iterate faster without surveys.

Signed by Eason Tang, the X thread pitches a purpose-built tool to capture quotes and whispered thoughts from physical books into an AI app without reaching for your phone, per replies in the thread; preorders are open.

The NVLabs project adds 6-DoF camera control, scene persistence, and a refiner, with a distilled NVFP4 variant that denoises a 60s clip in 34s on an RTX 5090.