Industry News
- OpenAI shuts down Sora: OpenAI is discontinuing Sora (mobile app, API, and ChatGPT video capabilities) after finishing training its latest model, marking a significant retreat in video generation. link
- Harvey AI raises $200M at $11B valuation: Legal AI company Harvey announced a massive funding round co-led by GIC and Sequoia, now serving 1,300+ customers across 60 countries and building the operating system for legal. link
- NeurIPS bans researchers from sanctioned institutions: The conference announced it will exclude researchers from sanctioned institutions (including Chinese universities) from participating, triggering widespread backlash in the ML community. link
- Granola raises Series C: The meeting notes AI company announced a Series C round led by Index and Kleiner Perkins, with major updates to team and tool integrations. link
Tips & Techniques
- Skills over agent managers: Mario from Pi argues developers should stay hands-on with code rather than becoming "agent managers," emphasizing that writing critical code by hand produces better-crafted systems. link
- Claude Skills for non-technical users: Ask Claude to do a task, then say "please turn this into a skill" when you want to repeat it—Claude handles the rest. Only turn things into skills that you actually use repeatedly. link
- Mobile-first vibe coding is next: The transition to mobile-first development means kicking off agent tasks throughout the day on commutes and walks, with agents proactively sending proposals and only pulling out desktop for verification. link
New Tools & Releases
- Google TurboQuant: New compression algorithm reduces LLM key-value cache memory by at least 6x while delivering the same quality, enabling much larger context windows on consumer hardware. link
- RunClaw launches at $1: New service offers OpenClaw functionality for $1 with no setup required, claiming to be cheaper and more secure than OpenClaw's $700 setup cost. link
- Lightning V3 from Decoh: New speech model beats ElevenLabs, Cartesia, and OpenAI on benchmarks, setting a new standard for real-time text-to-speech. link
- Hyperspace releases A1: New L1 blockchain designed specifically for autonomous agents, featuring proof-of-intelligence rewards, agent virtual machine, and parallel execution with ZK proofs. link
Research & Papers
- ARC-AGI solved with projective geometry: Researcher achieves 316 ARC-AGI tasks solved using 19th-century projective geometry with zero learning—no neural nets, no training, no DSL required. link
- AutoGaze for efficient video understanding: NVIDIA research identifies and removes redundant video patches in real-time, dramatically reducing compute while maintaining VLM performance through selective "gazing." link
- Memento-Skills: agents design agents: Paper introduces LLM agents that can build and improve their own reusable skills autonomously, creating a self-improving agent architecture. link
--- *Curated from 500+ tweets across AI development, research, and industry lists*
---
Emerging Trends
✨ RunClaw Launch (12 mentions) - NEW RunClaw is launching as a competitor to OpenClaw, positioning itself as cheaper ($1 vs $700 setup) with no security CVEs and claiming to do more. Multiple users are testing and discussing the product.
🔥 MiniMax M2.7 Open Weights (28 mentions) - RISING MiniMax is releasing open weights for their M2.7 model in approximately 2 weeks. Multiple tweets discuss it being potentially the best model that can run at home, with noticeably better performance on OpenClaw and strong community interest.
🔥 Figma MCP Integration (35 mentions) - RISING Figma's new MCP update allows AI agents (particularly Claude Code) to directly design and edit in Figma canvases with full design system context. This is being described as one of the strongest Claude Code integrations, enabling bidirectional workflows.
📊 Kapso CLI WhatsApp Agent (8 mentions) - CONTINUING Kapso CLI launches providing WhatsApp numbers for AI agents with simple 2-step setup (npm install + setup command), enabling agents to communicate via WhatsApp for automation and business workflows.
📊 ComFree-Sim Physics Engine (15 mentions) - CONTINUING ComFree-Sim is a GPU-parallelized analytical contact physics engine that's 2x faster than MuJoCo Warp in dense-contact scenes. It's a drop-in MuJoCo alternative for robotics, dexterous manipulation, and humanoid simulation with comparable dynamics fidelity.