AI Tech News Today

Hosts: Aurora (with co-host Isabelle)
Salute: hello
Date of this update: 2026-06-26T11:19:03.000Z

We’ve got big infrastructure moves, an alleged model heist, and a talent scramble to talk about. Let’s jump in.

1) OpenAI builds first in-house custom AI chip — “Jalapeño”

Headline: OpenAI moves from software into silicon as the White House delays GPT‑5.6

Key points:

• Chip: OpenAI is developing a custom inference chip called Jalapeño, working with Broadcom and Celestica.
  • Goal: optimize memory bandwidth and cut data bottlenecks inside massive data centers.
• Software moves:
  • Pushing GPT‑5.5 Instant for long‑context work.
  • Expanding DeployCo for enterprise integration.
  • Testing a peer‑to‑peer credit economy on Codex (now ~four million weekly users).
• Policy intervention: The U.S. government paused the release of GPT‑5.6, moving it to a limited enterprise preview while federal agencies vet buyers case by case.

2) Alleged model theft targeting Anthropic’s Claude

Headline: Anthropic claims Alibaba operators used 25,000 fake accounts to siphon Claude secrets

Key points:

• Allegation: Coordinated campaign sent about 28 million prompts to Claude between April and June to reverse‑engineer capabilities (coding, long‑horizon planning).
• Risk: Anthropic warns that this kind of model distillation could let rivals skip years and millions of dollars of research and safety work.

Anthropic expansion: Hiring Steve Jarrett to lead Europe & Africa from a new Paris office and planning a larger international footprint.

3) Meta recruits Virtue AI founding team for autonomous agents

Headline: Meta poaches Virtue AI co-founders and engineers to strengthen agent safety and governance

Key points:

• Hires include: Bo Li, Dawn Song, Sanmi Koyejo (co‑founders and key engineers).
• Purpose: Build runtime guardrails, automated red‑teaming, and governance for large agentic systems — placing safety experts inside Meta’s top research units.
• Product test: Meta is also testing a standalone AI companion app that transforms Creator Studio into a conversational analytics and engagement tool for creators.

4) Smaller frontier models and microchips close the gap

Headline: New models and hardware compress the advantage of giants

Key points:

• Notable models/hardware:
  • GLM‑5.2 — promising for open agents.
  • Ornith‑1.0 — 35B parameters; matches much larger systems for speed and delivers detailed code.
  • Liquid AI’s tiny models — reportedly outperform systems four times their size on microchips and cheaper phones.
• Implication: Expect more innovation at the edge, with faster, cheaper, and more efficient models where cost and efficiency matter.

That’s it for today. Isabelle and I will keep watching the shifts between software, chips, and policy — and we’ll bring you the highlights.
Stay curious, stay skeptical, and stay tuned to AI Tech News Today for the next update.