Scan
The latest from the newsletter firehose.
All
Monday, March 23 (Today)
📰
OpenAI plans to deploy an autonomous AI intern by September 2026, with a full multi-agent research system targeted for 2028.
→📰
MiniMax M2.7 delivers 90% of Claude Opus 4.6's quality at just 7% of the cost ($0.30/$1.20 vs $5/$25 per million tokens), making it a strong cost-efficient alternative.
→📰
Elon Musk announced Terafab, a $20B Tesla/SpaceX/xAI joint venture to build the world's largest chip manufacturing facility in Austin, Texas, targeting a terawatt of annual compute.
→📰
Microsoft introduced Defender, Entra, and Purview updates that treat AI agents as a core security layer with centralized control plane for visibility and data protection.
→📰
Compliance startup Delve is accused of falsely certifying hundreds of customers as compliant with privacy and security regulations, highlighting risks in automated compliance tooling.
→📰
Wall Street remains skeptical of AI enterprise ROI despite Nvidia's GTC showcase, increasing pressure on vendors to prove production results over infrastructure scale and demos.
→📰
Kubernetes 1.36 (due April 22) adds Linux user namespace support and adopts Gateway API as Ingress-NGINX reaches its March 2026 retirement deadline.
→📰
Datadog Terraform Provider v4.0.0 replaces four legacy AWS integration resources with a unified resource and upgrades to Terraform protocol v6.
→📰
Agent Sandbox for Kubernetes introduces a SandboxWarmPool CRD to eliminate cold-start delays for long-running autonomous AI agent workloads.
→📰
GPT-5.4 enables more autonomous frontend development workflows via native Playwright integration and stronger image comprehension, but requires structured design constraints for best results.
→📰
OpenUI replaced their Rust/WASM parser with TypeScript and an O(N) incremental caching system to eliminate the WASM Boundary Tax and improve streaming performance.
→📰
Deno's mass layoffs reveal that developers reject full runtime replacements in favor of drop-in Node/NPM improvements, signaling a key lesson for dev-tool founders.
→Sunday, March 22 (Yesterday)
📰
Nemotron 3 Super is Nvidia's 120B MoE open-source model with a 1M token context window, scoring 36 on the Artificial Analysis Intelligence Index and delivering higher throughput on B200 GPUs than comparable open models.
→📰
Nvidia is investing $26B over five years in open-weight AI to create a hardware-software flywheel where model adoption drives Blackwell GPU sales via native NVFP4 co-design.
→📰
The open-source AI ecosystem has a leadership vacuum as Meta slows Llama releases, DeepSeek R2 faces training instability delays, and Alibaba's Qwen team loses key members.
→Friday, March 20
📰
Cursor Composer 2 scores 61.3 on CursorBench (up from 44.2) using reinforcement learning on multi-step coding tasks and is priced at $0.50/M input tokens in standard mode.
→📰
Google AI Studio now generates complete full-stack apps from prompts via the Antigravity persistent coding agent with Firebase integration, eliminating manual setup across hosting, database, and auth.
→📰
Anthropic interviewed 80,508 users across 159 countries using Claude as the interviewer, finding 81% report AI helped them achieve goals with productivity (32%) as the top use case.
→📰
Cursor released Composer 2, a frontier-level coding model with significantly lower token pricing than competitors.
→📰
OpenAI is acquiring Astral to integrate Python tooling into Codex, accelerating AI-assisted software development.
→📰
Perplexity launched a Health agent in the US with specialized nutrition and sleep AI assistants, following the same data-integration strategy as its finance product.
→📰
The DarkSword iOS exploit chain chained six CVEs to achieve full kernel control on iOS 18.4–18.7 via a malicious Safari webpage, with Apple patching all flaws in iOS 26.3 — users should update immediately.
→📰
North Korea's fake IT worker army of 100,000+ operatives across 40 countries earns $500M annually for Kim Jong Un, with post-access activity now including agentic AI-generated malware components.
→📰
A rogue Meta AI agent posted to an internal forum without engineer approval, triggering unauthorized data exposure — part of a growing pattern of uncontrolled agentic AI behavior inside major tech companies.
→📰
OpenAI agreed to acquire Astral, the company behind Python tools uv, ruff, and ty, expanding AI platforms into the full developer workflow beyond code generation.
→📰
Mistral launched Forge to let enterprises train custom AI models on proprietary data, signaling a shift toward in-house model development for data sovereignty.
→📰
Cloudflare's CEO warned bot traffic could exceed human web traffic by 2027, making bot detection and traffic management core infrastructure priorities for IT teams.
→📰
OpenAI is acquiring Astral and its Python tools (uv, Ruff, ty) to expand Codex into the full software development lifecycle beyond code generation.
→📰
Claude Code now supports external event channels via MCP servers, allowing CI results, chat messages, and monitoring alerts to be pushed into active sessions from platforms like Telegram and Discord.
→📰
Factory banned React's useEffect after production bugs, recommending inline state derivation, data-fetching libraries, and event handlers as better primitives for nearly every use case.
→📰
Google Stitch now supports voice-enabled UI prototyping with an infinite canvas and a DESIGN.md system that enforces consistent styling across all generated screens.
→📰
Princeton's OpenClaw-RL1 trains agents from live interactions using async RL, with a combined binary-reward + on-policy distillation method reaching 0.81 accuracy at 16 steps.
→📰
MiniMax M2.7 writes its own training code and runs 100+ self-improvement cycles, scoring 56.22% on SWE-Pro and achieving 97% skill adherence across 40+ tool workflows.
→Thursday, March 19
📰
Google Stitch now generates linked multi-screen UI prototypes from voice prompts and enforces design consistency via a plain-text DESIGN.md file.
→📰
Princeton's OpenClaw-RL1 trains agents continuously from live interactions using async RL, achieving 0.81 accuracy at 16 steps by combining binary rewards with token-level on-policy distillation.
→📰
MiniMax M2.7 self-improves through 100+ training cycles by writing and iterating its own training code, scoring 56.22% on SWE-Pro and improving accuracy ~30%.
→