AI — Page 33
Models, agents, infra, applied AI.
- Lilian Weng returns to OpenAI for recursive self-improvement research
The Thinking Machines cofounder is the third founder from Mira Murati's lab to return to OpenAI this year.
- Google ships Lyria 3.5 to Flow Music with BPM controls and stronger vocals
The release turns Seth Forsgren and Hayk Martiros's acquired music startup into the main outlet for DeepMind's newest model.
- Superwhisper adds Cohere Transcribe for private, on-device dictation
The 2B-parameter, Apache 2.0 model downloads inside Superwhisper and keeps audio on-device across 14 supported languages.
- OpenAI offers 100,000 researchers free access to GPT-5.6
The program starts with 10,000 accounts this summer and forms part of a $250 million commitment to external scientific research.
- Hint launches AI app for homeowners with Martha Stewart as co-founder
Yih-Han Ma and Kyle Rush built the free iPhone app around property data, uploaded documents and home-service guidance after a $10 million seed round led by Slow Ventures.
- Nous Research adds local wake-word activation to Hermes Agent
The opt-in feature starts hands-free sessions across the CLI, terminal UI and desktop app, with hotword detection kept on-device.
- Composio's Kimi K3 test finds a 6x token gap between agent harnesses
Across 28 tasks, Kimi Code, Hermes and Claude Code finished at similar rates while median token use ranged from 61,000 to 340,000.
- Pangram ships Pangram 4 for mixed human and AI writing
The Brooklyn company also opened an image-detector preview as its authorship scores reach Substack readers.
- Anthropic faces developer backlash as a Claude power user cancels Max
AI educator Santiago Valdarrama blamed Anthropic's lobbying and open-weight stance, turning a policy fight into a customer-retention test.
- Echologue stores journal archives locally while routing AI tasks to cloud providers
Aris Giachnis' one-person Reformic studio keeps permanent journal data on-device while sending consented AI tasks to outside providers.
- Starling ships an AI-built Linux desktop that runs real apps
Starling says one developer directed AI for six months to build a GPU-driving Ubuntu session with its own Wayland compositor and X11 server.
- NERVOSYS launches IronAccelerator, claiming faster Rust CUDA calls than cudarc
Founder Adam Erickson built a one-line cudarc replacement, while ROCm, Metal and other backends remain short of production parity.
- Nous Research adds streaming speech to speed Hermes Agent voice chats
The open-source agent starts speaking while it generates an answer, reducing the dead air that makes AI voice interfaces feel slow.
- Andrew Ng launches LearnVector with $100 million from Coursera
The Coursera co-founder is returning to online education with an agentic tutor, and LearnVector expects to have products to show by early 2027.
- BusinessCaseBench finds frontier AI strong across 18 business disciplines
Wharton-led researchers tested 615 open-ended questions, but the leading model fully satisfied every rubric item on fewer than half.
- MCP drops sessions to make agent servers scale like ordinary HTTP
The breaking revision replaces persistent transport state with self-contained requests, header routing and explicit application handles.
- Replit starts model choice rollout with Moonshot AI's Kimi K3
Amjad Masad is making model routing a product feature as Replit's usage-based economics put cost beside capability.
- OpenAI calls for an international body that can slow frontier AI development
Sam Altman and Jakub Pachocki say AI-assisted research could force coordination, with a March 2028 target for automating research.
- Google Cloud page describes Gemini distillation service, but its release status is unclear
An archived Google document outlines an allowlist-only service, but a production ban, conflicting model IDs and unfinished tooling leave its release status unclear.
- OpenAI recruits paid campus leads for ChatGPT and Codex workshops
Undergraduates in eight countries will run workshops, weekly studio hours and semester showcases through June 2027.
- Pipe Network publishes a 350GB Kimi K3 build for 512GB Mac Studios
The pruned MLX model fits in unified memory, but Pipe reports 0.20 tokens per second and warns that quality degrades.
- Mirage launches Avatar X to clone creators from 10 seconds of video
The model powers Captions' AI Twin as Gaurav Misra and Dwight Churchill push their editing app deeper into proprietary video generation.
- Fermion Research publishes 3.88 GB Neutrino-1 8B for local inference
The Qwen3-derived model packages 8.19 billion parameters in a 3.88 GB container using Fermion's packed ternary-family weight format.
- Fermisense says a $500 Qwen fine-tune beat frontier models on catalog review
The 9B specialist scored 87.3% on Fermisense's simulation, though the test used 200 validation episodes and lacks independent replication.