slowlp
← loop
Lesson 2026.08.26 · 8 min read

OpenAI's Jalapeño beats Blackwell, and Claude Cowork finally remembers chat

Today's AI essentials — 2026-08-26

LOOP

📰 News

  • OpenAI’s first custom chip, Jalapeño, beats Blackwell and Rubin — Hot Chips. Designed with Broadcom in nine months. Inference only. On InferenceX it does 1.5–1.9× more work per watt and 1.7–3.6× lower latency. SemiAnalysis: “First-gen chips usually aren’t competitive, but OpenAI is beating Nvidia.” Small volumes late 2026, real scale in 2027. CUDA’s moat looks thinner. → The Decoder

  • Claude Cowork now remembers what you said in chat — Anthropic merged chat and Cowork memory. Headcount and city don’t need a second briefing. Sensitive topics stay out by default. On for Free, Pro, and Max. → TechCrunch

  • Russia ran a fake think tank through ChatGPT — OpenAI banned a VPN cluster. The “International Burke Institute” ranked Russia up and the West down. Posts barely traveled; Telegram channels hit 10–20k each. → The Decoder

  • Meta’s paid agent Hatch ships in weeks — A consumer OpenClaw. Premium up to $199.99/month. Watermelon model in October. Hooks into DoorDash, Etsy, Reddit, Outlook. WhatsApp as an agent platform too. → The Decoder

  • Google previews Gemini for legal work — MCP into iManage, DocuSign, Harvey. Contracts, research, regulation. Finance launched the same day; healthcare is next. Same models, packaged as skills plus connectors. → The Decoder

  • Nvidia’s Groq 3 LPX is “4× Cerebras” — with an asterisk — 3,400 tok/s on Gemma 4 31B. Built for fast agent decode. Nvidia needs ≥64 chips; Cerebras does it on one or two. Each LPU has 500MB of SRAM, so models get sliced across a rack. → The Decoder



▶️ YouTube

  • Your Mac can get a job — Matthew Berman. A short on a Mac picking up work locally. Fits the local-agent mood. → YouTube

  • Grok Bot saves so much time and is so easy — Matthew Berman. A quick demo of wiring Grok into busywork. → YouTube


💬 Community

COMMENTS