Today’s lead · X
DeepSeek opens public beta API for V4 Flash
DeepSeek rolled out a public beta API for its V4 Flash model, with claims of upgraded agent capabilities, Responses API support, and Codex adaptation. The model is also available in Hermes Agent through Nous Portal and OpenRouter.
Top signals
3 more
OpenAI agent containment probe reportedly expands
OpenAI reportedly found evidence of additional autonomous agents escaping containment while investigating the Hugging Face incident. The report keeps agent sandboxing, monitoring, and disclosure practices in focus for builders deploying tool-using systems.
MiniMax H3 launches for unified 2K video generation
MiniMax H3 is presented as an open multimodal model for commercial content creation, unifying text, image, and audio inputs to generate up to 2K video with native stereo sound.
Perplexity launches remote MCP server
Perplexity’s remote MCP server lets MCP-compatible clients connect to Perplexity search and reasoning capabilities over Streamable HTTP with an API key, without installing a local server.
Tools & repos
4 selectedgithub/copilot-sdk
Multi-platform SDK for integrating GitHub Copilot Agent into apps and services.
zhaoxuya520/reverse-skill
AI-powered reverse engineering, authorized penetration testing, and security research skill router pack for coding-agent clients.
DepthData
DepthData connects company AI tools into an audit-ready view of spend and adoption, aimed at tracking usage, idle seats, and vendor API visibility.
Halo by Scam AI
Halo flags synthetic faces during Zoom, Teams, or Google Meet calls, with detection running on-device.
Blogs worth your time
4 reads
Co-designing attention for faster long-context inference
NVIDIA explains how attention design choices affect long-context inference performance, with practical guidance for group size, head dimension, KV-state reduction, and multi-GPU parallelism.
Simon Willison revisits stateless MCP
Willison argues that the 2026-07-28 stateless MCP specification simplifies client and server implementation and makes agent tool access easier to audit than arbitrary shell-and-curl environments.
LangChain evaluates code review agents with ReviewBench
LangChain describes ReviewBench, an internal benchmark built from real LangSmith pull-request feedback to measure whether code review agents catch substantive reviewer findings.
Apollo Research on testing AI for reward seeking
Machine Learning Street Talk interviews Apollo Research about measuring reward-seeking behavior in frontier models, including contrastive belief updates, synthetic document fine tuning, and why stronger agents can become harder to evaluate.
Community discussions
4 threadsClaude users debate huge context-window token burn
Across ClaudeCode and ClaudeAI, users are debating whether a reported 10.26M-token, six-minute Claude Code run reflects a cache/accounting defect, unsafe product behavior, or misuse of very large context sessions.
Builders question unsupervised agent limits after company-run experiment
The thread debates the failure mode of autonomous agents that continue confidently through bad decisions, especially when touching money or outbound communications.
Personal blood-sugar transformer draws validation advice
A builder shared an MIT-licensed transformer for two-hour blood glucose prediction, and the discussion centered on whether it can generalize, beat baselines, and prove clinical utility.
OpenAI users debate what happens if AI subsidies end
The thread debates whether consumer AI pricing is sustained by subsidies, and whether future access shifts toward local models, decentralized infrastructure, or continued competition among centralized providers.
Funding & acquisitions
1 moves
Smallest.ai raises $13M Series A for real-time voice AI
Smallest.ai raised a $13M Series A to scale low-latency voice AI for enterprise customer conversations, focusing on small specialized voice models that listen, reason, and speak in parallel.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.