Anthropic Pledges 5% of Shares to a Public Benefit Fund
Anthropic is committing a portion of its shares to a new wholly owned entity, the Anthropic Public Benefit Fund, dedicated entirely to the public benefit. It starts at 5% of shares, with further additions for every 50% increase in company value over a set baseline.
vLLM v0.32.0 Ships a Structured Decision Endpoint
The release adds a /v1/systemone structured-decision endpoint and a generic Hugging Face tool and reasoning parser. Low-precision KV cache now runs everywhere — NVFP4 on SM8x/90/12x, FP8 KV by default for GLM-5.x, and 4-bit KV on ROCm — alongside Model Runner V2.
Tools help small models, but not big models. Gaps in capability can be replaced by tools.
Tim Dettmers — on small models + SFT beating GPT-6 Luna in a synthetic hospital benchmark
vLLM Hits 842 tok/s Decoding for Kimi K3 on GB300 NVL72
A unified CuTe DSL decode path fuses attention, MoE, and communication, delivering 1.5–2x higher decode throughput in several tested configurations and up to 1.8x higher P90 interactivity. The 842 tokens-per-second figure was measured at batch 1 with DSpark.
Inferact and Google Submit First TPUv7 Kimi K3 2.8T Results
SemiAnalysis reports the first preliminary TPUv7 Kimi K3 2.8T model submission, from Inferact and Google, with initial results already obtained.
Argon Rebuilds the First Black Hole Image From Raw EHT Data
fofrAI tasked the model with reproducing the 2017/18 black hole images from the same raw observations. Argon pulled the original 230 GHz radio fringes in UVFITS form from the Event Horizon Telescope and reconstructed the iconic image, in a hobbyist astronomy experiment that shows the model handling real interferometry data end to end.
Pika Launches Camera Director Inside ChatGPT
Adds cinematic moves such as orbit and slow push-in to any clip, while keeping the original performance exactly as it was.
Motion Transfer Teaches Any Character to Dance
Film the dance once and the feature transfers it to any character, all inside ChatGPT — no dancing skills required.
Character Studio Generates Character Reference Sheets
Ask ChatGPT for a mascot and it delivers a full reference sheet, saving it to your library for the next shoot.
Kling Teases Kling 4.0
Kling_ai drops a 4.0 teaser without disclosing specific capabilities or a release date.
Four Gemini 4 Argon Agents Co-op Beat Minecraft
In one shared world they ran for 27 hours, wrote 2,783 lines of code themselves, and defeated the Ender Dragon.
Perplexity Computer Generates a San Francisco Driving Game
A cyberpunk-style Roadster free-roam demo that can drive and fly, with map data from OpenStreetMap.
Vercel AI Gateway Centralizes Agent Sessions
One unified view across personal and coding agents, apps, and API keys, emphasizing model choice, zero-data-retention, and ROI metadata.
Maestro v3.0 Launches With Faster Multi-GPU H3
Free, local, and open, adding Flux3 and more; it has surpassed 70,000 installs in three months.
Fable 5.1 Turns Borges Short Stories Into 3D Worlds
Nine short stories become walkable scenes with quotes from the original text, playable online.
Epic Releases Plugin for Claude Code to Control Unreal Engine
The official plugin connects Claude Code to Unreal Engine workflows and can control the engine directly.
AMD, NVIDIA and Google All Improve vLLM
Three vendors contributed significant performance improvements to vLLM in the same stretch, per Simon Mo.
An AI Detector Only Cleared the LLM-Polished Sentence
At ICLR review, the only text Pangram flagged as non-AI-generated was the sentence admitting authors used an LLM to polish the writing.
From What AI Can Do to What It Should Do
The mindset shift applies to individuals, not just companies, as AI keeps changing productivity workflows.
Claude Still Can't Change Your Email
Despite Anthropic's coding acceleration, changing an email still requires deleting the account, emollick notes.
DSV4.1 Gains Aren't Just OpenAI Distillation
If it were distillation, why couldn't OpenAI distill a stronger small model than Luna, teortaxesTex asks.
Blender Reconstruction for Agentic Video Understanding
Whale Vision V4.1 improves clearly on the benchmark; passing 5.6 Sol is a big gain.
In the Post-Training Era, How Much Pretraining Is Left?
sedielem argues pretraining is now mostly about finding a good initialization; for scaling, it may share flaws with drifting models.
Clustered Semantics Boost Continuous Diffusion Language Models
Borrowing a DINO idea, the work builds coarse-grained semantic representations through clustering for CDLM, and reports it helps.
Nested-Loop Diffusion Cuts Error Accumulation
Looping at a given noise level fixes errors that would otherwise accumulate and reduces reliance on CFG.
Pretraining Saturates as Focus Shifts to Test-Time Compute
As transformers saturate, most pretraining results are data-recipe tweaks; innovation moves to post-training and inference-time compute.
Developers Will Be a Small Slice of Token Usage
Internal OpenAI and Anthropic stats show developers shrinking as a share of tokens, while high-intelligence models pay off on hard-to-verify tasks.
Papers with Code Adds an Auto-Research Category
A new category makes automated-research papers easier to track, per Niels Rogge.
Grok Bot Can Call Restaurants to Book Tables
Musk says Grok Bot can phone a restaurant and reserve a table on your behalf.
Grok Bot Gains About 20 Improvements a Day
Musk says the team ships around 20 improvements to Grok Bot daily.
Jack Clark: Persuasive Evidence the Singularity Is Near
The Anthropic co-founder says this year brought convincing evidence the singularity has begun.
Websites That Ban AI Agents Will Lose Customers
Agents will become the main way people transact online, so blocking them is short-sighted.
Fable 5.5 Said to Arrive Next Week
OpenAI is said to have an answer; no official confirmation.
Small model + SFT + tools beats GPT-6 Luna on hospital diagnosis benchmark
In a synthetic hospital benchmark, tools can close small-model capability gaps, while gains for large models are limited.
Tweet claims Anthropic and SpaceX signed $15B compute deal
It relays a final-window argument and mentions a $15 billion compute deal between Anthropic and SpaceX, with no official confirmation.
DS Flash video takes 50 minutes and produces poor results
oran_ge complains about the latest DeepSeek Flash video generation, saying 50 minutes produced garbage.