LM Studio integrates Splash inference engine on day one
Inco AI's open-source Apple Silicon inference engine debuts in LM Studio, tuned for Qwen3.8-27B. It claims up to 144 tokens per second on the M5 Max — roughly twice the next-fastest engine — using custom GPU kernels, a refined memory scheme, and DFlash 2 speculative decoding.
Open-source models take 78.4% of tokens on Vercel's gateway
Vercel CEO Guillermo Rauch flagged what may be a record day for open models: 78.4% of token volume against 21.6% closed. Spend tells a different story, but Moonshot AI, DeepSeek, and Z.ai combined now outspend OpenAI.
Meta opens Muse connectors to developers
Developers can now apply for Muse connector access. Bring your own API, and Muse supplies the agent, the browser, and the context — a bid to turn the platform into a programmable layer for third-party tools.
Inference compute will surge — but an intelligence explosion isn't guaranteed
Nat Lambert admits he underestimated how aggressively inference-time compute will scale. But he draws a line: that is not recursive self-improvement, and he has yet to see evidence that an intelligence explosion is near.
Muon may beat AdamW by reducing memorization
Sebastian Raschka shares a new angle on the AdamW vs. Muon debate: Muon performs better possibly because it lowers how much the model memorizes.
Bespoke Nimble: an open-source small alternative to Jev
Data, model, and training recipe are released; built on Qwen3.5-9B with 2,676 curated examples, it outputs typed answers with probabilities in one step and runs locally.
GPT-6 Astra shows off wild 3D creations
OpenAI Devs asked for users' wildest 3D builds with GPT-6 Astra and showcased the results, drawing more than 1,400 interactions.
Pika can reshoot the same scene from multiple angles
Upload one clip and Camera Director generates alternative angles that preserve the original motion and performance.
Extracting good results from frontier models remains hard
Even with extremely detailed context and repeated feedback, getting acceptable work from the highest "Ultra" setting is still painful, one researcher says.
Fable builds an endlessly zooming game from one prompt
Ethan Mollick gave a single prompt, and Fable produced UMBRA, a visually polished game whose continuous zoom delivers surprising reveals.
AI is the OS.
Aravind Srinivas, CEO of Perplexity
LLM-supervised peer review becomes a defining issue
If LLM involvement in peer review goes unsolved, academic institutions will struggle to sustain themselves — and solving it is easier than building new institutions.
AI companies face antitrust suits over "pacing the frontier"
Several AI companies have been hit with antitrust lawsuits over coordinating control of model frontier progress.
AI film starts to have real substance, art debate will heat up
Genuinely interesting works have appeared in AI-generated film, and the debate over whether it counts as art will keep growing.
Doubt cast on methodology of AI financial advice study
The study's conclusions conflict with other recent research, prompting questions about its specific evaluation methods.
DeepSeek Flash is cheap but burns tokens, users say
For the same money, some developers argue GLM Flash offers better durability and value for the token volume consumed.
Opus 5.2 reportedly made a full Titanic film in pure JavaScript
A five-minute clip built with tree.js reportedly reconstructs the disaster's timeline — a demo many found hard to believe at first.
rasbt releases inference scaling tutorial, part one
Covering temperature scaling, top-p filtering, and multinomial sampling, it shows how self-consistency and best-of-N can more than double accuracy.
AI lets everyone start creating again
The Vercel CEO believes the biggest benefit of this AI wave is getting more people to build and ship, not just consume.
OpenAI shares image generation plus Codex redesign flow
Reimagine a page with image generation, then hand it to Codex to implement — a rebuttal to doubts about Sol and Astra design.
Higgsfield demos parallel UGC ad generation
JEV and DeepSeek select 20 of 100 AI avatars and generate product UGC videos in parallel.
Jev opens free on Vercel
Guillermo Rauch announced Jev is free to use on Vercel and said he plans to build something with it this weekend.
Grok Build rumored to test remote control
SpaceXAI is said to be testing remote control for Grok Build, described as one of the product's biggest updates yet.
LeCun reposts: Astra and other VLMs show stunning progress
Progress in vision-language models is astonishing, while some raise fundamental doubts about this route.
Gradio 6.28 supports component live state as input
The new version can pass the live state of any component as input into app logic.
Hoping Jev can think longer and support compound types
Jev would be more useful if it could think longer before hard decisions and handle dicts and lists, not just booleans.
A $500 GPU runs a fully open model with surprising quality
Francois Fleuret marvels at the quality of a fully open-source model running on his $500 GPU.
A few lines of code to generate datasets: Colab example goes viral
Sara Hooker recommends a notebook showing how to automatically construct datasets with a few lines of code.
AI-generated horror short Backrooms Nemoris launches
A 17-minute horror pilot by Javi López, generated with Magnific.
Synthesia opens New York U.S. headquarters
A roughly 50,000-square-foot office in Manhattan's Flatiron district, nearly four times its previous space.
Most tricks that work are sophomore probability theory
The most effective tricks in machine learning are essentially undergraduate probability knowledge.
A 14.7MB model detects sensitive data locally
Rampart, a tiny model from National Design Studio, spots personal details in the browser before they reach an AI.