AI — Page 21
Models, agents, infra, applied AI.
- Tencent's WorkBuddy faces a Claude Code copying allegation over internal prompts
An October 16th post claims 224 identical text segments in WorkBuddy's bundled CodeBuddy CLI; Tencent's docs confirm Claude plugin compatibility, complicating the comparison.
- MiniMax joins Hugging Face for San Francisco's open-source AI kickoff
Founder Yan Junjie's model company is listed among hosts of Hugging Face's October 16th kickoff at The Midway, part of a 10-day Bay Area program.
- TIME names Nous Research's Hermes Agent to its 2026 Best Inventions list
The listing highlights Hermes' customizable, open-source design as Nous pushes the agent from individual developers into enterprise deployments.
- Accenture sets a $29 monthly Claude budget and routes overages to projects
The October 1st policy funds $29 per person centrally; employees who exceed it must seek a project or team budget, while some report monthly usage of $500 to $1,000.
- Leon Lin releases Earshot, an audio reviewer for AI coding agents
The open-source, CPU-only toolkit analyzes soundtracks, critiques launch videos and lets agents generate sound effects, though its blind-test evidence remains small.
- Anthropic commits $5 million to help mathematicians set AI research norms
The commitment combines funding, computing resources and Claude access in a $15 million Simons Foundation program that also includes OpenAI.
- NVIDIA and AI4Bharat announce a 2.8T-token corpus for 22 Indian languages
AI4Bharat head Mitesh Khapra co-authored an announcement for a 22-language corpus and reusable data pipeline, both listed as coming soon.
- ZestyAI says its risk models have surpassed 500 regulatory approvals
The approvals span 317 carrier filings in 44 states, according to ZestyAI, whose co-founder Kumar Dhuvur is betting property-level AI can earn regulators' trust.
- Perplexity open-sources SPACE, the sandbox running every Computer session
The Apache 2.0 release uses Firecracker microVMs. Perplexity says SPACE has powered all Computer sessions since July and handles millions of sandbox creations per day.
- Bluesky expands Attie into an AI builder for community websites and games
The tool builds on Jay Graber's feed-making experiment, tying prompt-generated sites to Bluesky's social network and the wider AT Protocol.
- Skywork AI releases Matrix-Game 4.0 with game logic in code
The Kunlun Tech AI unit pairs an executable game engine with a video renderer, claiming 480p at 60 FPS on one RTX 5090; its model card sets limits on those claims.
- xAI founder Musk agrees to settle suits over $1 million voter giveaway
The October 15th agreement covers two proposed class actions over America PAC's 2024 giveaway. It remains subject to court approval, and the reported filings did not include its terms.
- Sakana AI builds evidence checks into its open-source AI scientist
David Ha's third-generation research system assigns coding agents parallel experiments, baselines and ablations, while researchers remain responsible for judging the results.
- OpenAI offers university labs up to 30 free GPT-Rosalind seats
The 12-month plan also includes other frontier models, while GPT-Rosalind requires trusted-access approval and API access is excluded.
- Overshoot says its inference software can halve GPU bills without quality loss
Zakaria El Hjouji says the startup's deployments have saved AI teams more than $5 million; its public method centers on workload-specific tuning and output checks.
- Photon opens messaging-agent beta with $100 in credits for builders
The new Photon 2.0 service says it connects agents to iMessage, voice, email, WhatsApp, Telegram and SMS; its beta page directs users to request access.
- Robocurve finds Gemini models outperform their general AI scores on robot tasks
Its preliminary index covered five bimanual tasks and 11 model configurations; removing Gemini raises the reported correlation with general AI capability from 0.20 to 0.96.
- PrismML releases ternary AI models as small as 0.43GB for local agents
The free Apache 2.0 Bonsai 2 models use ternary weights and add vision and tool calling, alongside performance figures reported by PrismML.
- IBM packs chat, decisions and speech recognition into Granite Switch 4.1
The October 15th update combines IBM's Granite models with task-specific adapters and a speech recognizer, aiming to run several workloads from one model deployment.
- Kling AI makes Kling 4.0 available through MCP and CLI
The release adds assistant- and terminal-based workflows; Kling says 10-bit HDR output and API access are still coming.
- a16z's Fabrizio Serafini seeks startups building computer-use agents
The investor wants agents that finish business workflows inside software companies already use, a problem his August research says depends on reliability beyond model scores.
- OpenAI ships a gated browser-control runtime inside its new ChatGPT extension
Inspection of Chrome extension version 1.26.1005.11759 found a local JavaScript environment and a direct server connection that could eventually remove the desktop-app requirement. The code also includes an approval-based handoff for email verification codes; neither workflow was verified in a desktop-free session.
- fal launches Worlds to turn generative video models into live experiences
The new showcase runs models including LingBot World and H3 Max Director on fal's WebRTC-based World Model Accelerator, whose developer API remains experimental.
- Video Rebirth launches BACH 1.0 Pro with 12-reference video control
The Singapore startup says its model ranked No. 2 in Artificial Analysis' silent image-to-video category and lists 1080p API generation at $0.12 per second.