Today’s lead · TechCrunch
Music publishers widen the copyright fight against Anthropic
Major music publishers are now attacking Anthropic’s training data pipeline directly, alleging illegal torrenting and scraping of copyrighted songs, lyrics, sheet music, and books. For AI builders, the risk signal is sharper than “fair use”: provenance and acquisition method may matter as much as model behavior.
Top signals
3 moreOpenAI resets Codex and ChatGPT Work usage after fixing wasteful background loops
Paid Codex and ChatGPT Work users are getting usage resets after fixes to compaction, memory workers, goals, automations, subagents, history summaries, and MCP handling. The practical takeaway: agentic product limits are still shaped by hidden orchestration costs, not just model pricing.
Tencent’s Hy4 Preview targets long-horizon agent work with a 770B MoE
Tencent’s Hy4 Preview is pitched as an open 770B-parameter MoE model with 49B active parameters and a 1M-token context, aimed at coding, research, document workflows, and agentic tasks. The internal benchmark claims are interesting, but still need external builder validation.
Cohere Parse 5 turns enterprise documents into agent-ready structure
Cohere Parse 5 is a document vision parsing model for converting complex enterprise documents and images into structured data for downstream AI systems. The useful bit for production teams is deployment flexibility: API, cloud, or fully on-prem and air-gapped.
Tools & repos
2 selected
seendiff
A local diff viewer built for reviewing huge coding-agent changes. It tracks what you’ve already seen, supports chunked review, and lets you pull the agent in to explain its own work.
OpenMAIC
OpenMAIC is a TypeScript repo for an “Open Multi-Agent Interactive Classroom,” promising a one-click immersive multi-agent learning experience.
Blogs worth your time
0 readsNo long-form read was strong enough to recommend today.
Community discussions
3 threadsState tracking could cut agent tokens, but builders worry about cache economics
The thread likes the SKILL.state idea—carry structured state instead of full history—but the builder pushback is useful: fewer tokens may still mean higher cost if cache-busting hurts long-running tasks.
A 100K-context Qwen setup on 16GB VRAM gets the right kind of skepticism
The recipe is aggressive: hybrid GGUF quantization, BeeLlama.cpp, KVarN KV cache, high-precision tail tokens, and MTP speculative decoding. The best comment cuts through hype: prompt processing, not tok/s, may decide usability.
Claude Code history cleanup catches transcript-dependent workflows off guard
The practical warning is simple: if your tooling reads Claude Code transcripts from disk, a silent 30-day cleanup can break you. Commenters discuss backups and longer cleanupPeriodDays settings, but lost history stays lost.
Funding & acquisitions
2 moves
IndiGo Ventures backs Sarvam AI as enterprise aviation AI partnership deepens
IndiGo Ventures invested in Sarvam AI’s ongoing Series B, extending an existing collaboration on airline operations, passenger experience, employee support, and workflows. The investment amount was not disclosed.
Edgeverse raises Rs 3 crore to commercialize edge-AI perception for mobility
Edgeverse raised Rs 3 crore from GVFL for India-specific road intelligence, camera and radar technology, and commercialization of its AI-powered perception stack for mobility and safety applications.
Bengaluru radar
4 events
Memory, Context & Agents: SurrealDB x Contentstack
A practical agent-memory meetup during Bengaluru Tech Week, with SurrealDB showing cross-app memory architecture live. Approval-only and capped around 50 engineers.
Codex Community Meetup - Bengaluru
A sold-out Codex meetup with OpenAI team updates, demos, showcases, and AMA. Approved participants only; no on-spot registrations or late check-ins.
GTM Engineering Hackathon for Founders
A hands-on Inkle x Aspire session to build an AI-powered outbound workflow for U.S. sales. Bring your ICP, target market, and laptop.
Vibes In, Latency Out: A Voice AI Open Mic
Bolna’s sold-out voice AI open mic gives builders 3–5 minutes to demo, share lessons, or pose hard latency and agent problems.