AI that matters, from the architect's desk. Curated and engineered by Saaket Varma, PhD — no hype, just signal.
HOW TO READ THIS Read top to bottom: Anthropic, its embed inside Slack, the @Claude tag mechanism, then the always-on result.
Anthropic shipped Claude Tag, an always-on agent that lives in Slack — proactive, persistent, and accumulating your company's context one message at a time. The strategic move isn't the chat UI; it's the location: Claude now sits in the channel where decisions actually get made, not the IDE where code gets written. Whoever owns that organizational memory owns the workflows built on top of it, and the switching cost compounds with every thread it reads. If you run a Slack-heavy org, pilot it in one team now — but treat data governance, not usefulness, as the real question to answer first.
HOW TO READ THIS Read top to bottom: OpenAI acts, ships two tools, its model scans code and patches flaws, yielding GPT-5.5-Cyber.
OpenAI unveiled Daybreak — Codex Security plus a new GPT-5.5-Cyber model — alongside 'Patch the Planet,' a program to help open-source maintainers find and fix vulnerabilities at scale. Pointing frontier models at the supply chain is double-edged: the same capability that patches a CVE can find one to exploit. For anyone shipping code or depending on open source, AI-assisted vulnerability triage is moving from novelty toward table stakes. The signal to watch is whether the output lands as verified, mergeable PRs or just more noise in maintainers' inboxes.
HOW TO READ THIS Read top to bottom: LMCache sits between requests and the GPU, then a repeat request hits the cache and bypasses GPU recomputation entirely, yielding cheaper, faster inference.
LMCache is a KV-cache layer that reuses and accelerates LLM context across requests, and it's spiking on GitHub as teams hunt for cheaper inference. The win is unglamorous but direct: the same GPUs serve more requests, so the payoff shows up in your p95 latency and your monthly bill. If you're self-hosting models or running a high-traffic RAG system, this deserves a benchmark — caching shared prompt prefixes is often the highest-ROI optimization most teams haven't turned on yet.
HOW TO READ THIS Read top to bottom: DeepMind commits $75M to A24, the deal funds AI filmmaking tools that feed into the production pipeline, and AI enters mainstream Hollywood filmmaking.
Google DeepMind is putting $75M into a deal with indie studio A24 to build AI filmmaking tools. The signal isn't the dollar figure — it's that a frontier lab now treats Hollywood, not just the enterprise, as a core market for generative video. If you work in media, the question is no longer whether AI enters the production pipeline but who ends up controlling the models trained on your craft, and on what licensing terms.
Official, AWS-supported MCP servers, skills, and plugins for building AI agents on AWS.
Anthropic's managed directory of vetted, high-quality Claude Code plugins.
Production-ready platform for building and shipping agentic workflows.
One OpenAI-format gateway to 100+ LLM APIs, with cost tracking and guardrails.
An LLM post-training framework built for reinforcement-learning scaling.
Import AI Issue 462 maps superpersuasion, self-sustaining AI, and the concrete paths toward ASI.
Interconnects Argues GLM-5.2 is the step change that makes open models genuinely viable for agents.
The Sequence Opens a new series on distillation — the compression technique behind cheaper, faster models.
Latent Space Flags an under-discussed angle: SpaceX is already a $28B/yr 'neocloud' selling compute.