AI — Page 47
Models, agents, infra, applied AI.
- Nvidia says Cosmos 3 tops seven physical AI leaderboards
The claim spans world generation, robot action policy, and industrial vision understanding, but the post did not include scores or test details.
- ViBench aims to rank AI models by app-building, not just coding tests
The public site and ACM paper frame ViBench as an end-to-end test of whether coding agents can deliver usable apps, not just pass SWE-style tasks.
- Google's Gemma 4 12B brings encoder-free multimodal AI to laptops
Google says the Apache 2.0 model handles vision and native audio without separate encoders, runs with 16GB of VRAM or unified memory, and nears its 26B MoE model on benchmarks.
- Perplexity AI CEO says AI race comes down to value per watt
Aravind Srinivas told CNBC the winner will maximize economic output from AI energy use, with orchestration across cloud and on-device models as the lever.
- Roblox buys Morpheus AI for its Roblox Reality push
Roblox tied the acquisition to Roblox Reality, its planned hybrid architecture for photorealistic multiplayer games, while leaving deal terms and Morpheus AI integration details undisclosed.
- Google is buying Play Store developers' code for AI training
A confidential pilot turns Android app codebases into paid data assets, with terms and payout sizes still undisclosed.
- Genomi gives AI agents a local database for personal genome questions
The open-source project parses raw DNA files into a local database instead of sending genetic data into a model context window.
- We Stress-Tested Microsoft's New Image Model Against OpenAI and Google. The Results Were Clear.
RuntimeWire prompts found OpenAI strongest for dashboard text and Google strongest for web layouts, while Microsoft excelled at photoreal spaces.
- DigitalOcean says it is now an OpenRouter AI model provider
The cloud provider says DeepSeek V3.2, Kimi K2.6 and DeepSeek V4 Flash are available, but pricing and hosting details are not disclosed.
- Microsoft's MAI Thinking 1 surfaces in a 109-page technical report, ML nerds geek out
A Microsoft-hosted PDF is circulating, but public details on access, pricing, and deployment are not established in the thread.
- Microsoft AI launches seven-model MAI family as Suleyman pushes in-house frontier lab
Microsoft AI says the lineup spans reasoning, coding, image, transcription and voice, with MAI-Thinking-1 matching leading software engineering benchmarks and reaching preference parity with Sonnet 4.6.
- Finbarr Timbers Says He Is Leaving Ai2 After Open-Model Work
The AI researcher and investor, whose work spans reinforcement learning, game AI and generative models, praised Ai2's open-model push but did not name his next role.
- Factory launches Router to pick cheaper AI models for coding tasks
Factory says Router is in private research preview and matched most of Opus 4.7's benchmark pass rate at 20% to 25% lower cost.
- Opal Electronics turns from webcams to AI audio with OpenAI cash
WIRED reports OpenAI put $40 million into Opal, which is now preparing an AI audio device rather than another phone replacement.
- FutureHouse and Google DeepMind AI agents reportedly reach Nature with drug-discovery work
The claim points to a higher bar for lab agents, but the cited post omits the papers, candidates, and human-supervision details.
- LanceDB says NVIDIA used Lance datasets to curate Cosmos 3 training data
NVIDIA's report describes Cosmos 3 as an omnimodal model family; LanceDB's claim ties the release to training-data infrastructure.
- QuarqLabs says its open-source agent scored 98.2% on LongMemEval-S
Quarq Agent uses local FAISS storage, layered memory and three separate LLM roles to tackle long-context recall.
- Bernini is ByteDance's bet that AI video needs planners, not just renderers
The system pairs a multimodal language model with a diffusion renderer, a bet that video generation needs explicit reasoning.
- Nous Research Opens Six Roles as Hermes Agent Moves From Open Source Project Toward Product
The remote AI lab is hiring across engineering, research, design, enterprise deployment, and legal as it pushes Hermes Agent and Psyche.
- Browser Use moves down the stack to build an execution layer for AI agents
The agent-browser startup says its stack uses a Chromium fork, a Firecracker fork and a custom Linux kernel, a bet that agent workloads need infrastructure below the SaaS layer.
- Standard Signal bets autonomous AI traders are ready for outside capital
Michael Royzen's YC Spring 2026 company is soliciting qualified investors with a $100,000 minimum for autonomous, auditable trades inside preset risk limits.
- Anthropic submits confidential draft S-1, opening door to IPO
The AI company says SEC review still has to finish before it decides whether to pursue a public listing.
- Shafiq Joty's team proposes Procedural Memory Distillation for less forgetful LLM training
PMD is presented as a training-time method that stores lessons from prior model attempts, then uses them to teach the next policy.
- Tripo's VAST previews Project Eden, a bid for persistent AI worlds
Rather than treating world models as video generators, VAST says Eden separates durable state from rendering so objects, actions and multiplayer views can persist over time.