Evening brief

Thursday, June 25

2 people · 0 companies · 5 topics

TL;DR OpenAI and Broadcom unveiled Jalapeño (June 24), a custom LLM inference chip claiming 50% cheaper inference than Nvidia GPUs. On GitHub today, OpenMontage and hermes-agent are trending hard; both are agentic OSS projects worth watching.

Models & Research

  • model-releases OpenAI + Broadcom unveiled Jalapeño (June 24): first custom LLM inference processor built in nine months design-to-tape-out, targeting ~50% lower cost per token vs Nvidia, with initial deployment slated for end of 2026. gdb: “By designing more of the stack ourselves, we can serve more intelligence with greater efficiency.” TechCrunch · OpenAI · CNBC

People

  • Earlier this week (June 18-20): Google lost two flagship researchers inside three days. Noam Shazeer (Transformer co-inventor, Gemini co-lead) left for OpenAI as Lead for Architecture Research; sama called it “10 years in the making.” John Jumper (AlphaFold Nobel co-winner) left DeepMind for Anthropic after nine years, drawn by Anthropic’s push into AI-for-science. CNBC - Shazeer · TechCrunch - Jumper

Tools & Dev

  • dev-tools Qualcomm confirmed $3.9B all-stock acquisition of Modular (Chris Lattner’s AI software startup) at its June 24 Investor Day; Modular lets devs run AI across different hardware without rewrites. Qualcomm also named Meta as first customer for its C1000 data center CPUs, shipping late 2028. Tenstorrent talks ($8-10B) remain unconfirmed. CNBC · Gurufocus
  • agentic-ai Anthropic launched Claude Tag for Slack (June 23): team-level @Claude integration for Enterprise and Team plans with ambient context, async task handling, and proactive cross-channel updates; replaces the old Slack integration on August 3. TechCrunch · Bloomberg

OSS

  • oss OpenMontage (calesthio): #1 on GitHub trending today with 3,719 stars — open-source agentic video production, 12 pipelines, 52 tools, 500+ agent skills; takes a plain-language brief and handles research, scripting, asset generation, and final composition. GitHub
  • hermes-agent (NousResearch): 1,178 stars today — self-improving agent with persistent memory and skill creation from experience, supports 22 messaging platforms and 200+ models via OpenRouter. GitHub