Sarmadi AI Digest June 26, 2026 Updated 7:00 AM CT Today Archive Topics Saved Subscribe RSS

White House asks OpenAI to delay GPT-5.6; Alibaba allegedly mined Claude 28.8M times

The Trump White House asked OpenAI to slow-roll GPT-5.6 over security concerns — the second federal slow-roll of a frontier model in three weeks, after the Fable shutdown. New filings detail the Anthropic-Alibaba allegation: 25,000 accounts, 28.8M exchanges, characterized as the largest Claude cloning attack to date. Anthropic Claude is now winning paid-consumer share away from ChatGPT, and Wired published a long piece on Anthropic's argument that its own success is necessary to keep AI safe. Apple will skip the high-end M6 line and jump to an AI-focused M7 (M7 Pro/Max/Ultra). Capital is flowing: Patronus AI raised $50M for digital worlds that stress-test agents; General Intuition raised $320M to train embodied agents on video games; Amazon committed $13B more to India AI infrastructure; Adobe acquired Topaz Labs. Notion shut down its Skiff-derived email app to go 'all in on agents for your inbox.'

7 papers 20 news 7 sources ← Latest

News

14 items

White House delays GPT-5.6; Anthropic details Alibaba attack

The Trump White House asked OpenAI to slow-roll GPT-5.6 — the second federal slow-roll of a frontier model in three weeks. New court filings detail Anthropic's allegation against Alibaba: 25,000 accounts, 28.8M exchanges, characterized as the largest Claude cloning attack on record.

News TechCrunch AI

The White House is asking OpenAI to slow roll the release of its new model over safety concerns

TC: the Trump administration asked OpenAI to slow-roll GPT-5.6 over security concerns; OpenAI now plans a phased partner-first release.

Why it matters
  • Second federal slow-roll of a frontier model in three weeks — Fable, now GPT-5.6.
  • Confirms the Fable action wasn't a one-off — pre-release federal review is becoming the de-facto norm.
  • Materially affects every frontier-lab release calendar and IPO disclosure timeline.
  • Sets expectations that any model release will route through political review.
policyregulationproductssafety
Source →
News Ars Technica AI

Anthropic says Alibaba must be punished for largest Claude cloning attack

Ars: Anthropic alleges Alibaba used 25,000 accounts to mine Claude over 28.8M exchanges — characterized as the largest single Claude cloning attempt.

accounts 25,000exchanges 28.8M
Why it matters
  • Quantifies yesterday's IP allegation with operationally specific scale.
  • If sustained, becomes the template for future model-capability-theft cases.
  • Anthropic claims Alibaba 'defied Trump' — explicit political framing tying the case into the broader US-China standoff.
policyregulationsafetymarket
Source →

Anthropic is winning paid consumers, framing safety as the win condition

TC says Anthropic Claude is now winning paid-consumer share that ChatGPT has dominated. Wired published Anthropic's argument that its own ascendance is necessary to keep AI safe — critics say it's accumulating power. Combined with the federal slow-roll of OpenAI's next model, the competitive dynamic just shifted noticeably.

News TechCrunch AI

Anthropic's Claude is winning over paid consumers, a market owned by ChatGPT

TC: paid-consumer AI share is shifting toward Claude — ChatGPT's lead is no longer uncontested.

Why it matters
  • First sustained consumer-paid-tier share gain by Anthropic — a market ChatGPT has owned.
  • Compounds with the business-customer-lead Ramp data from May and the Stainless acquisition.
  • Strengthens Anthropic's IPO disclosure profile heading into a delayed-OpenAI window.
marketproducts
Source →

Capital surge: Patronus, General Intuition, Amazon-India, Adobe-Topaz

Patronus AI raised $50M to build digital worlds that stress-test agents. General Intuition raised $320M to train embodied agents on millions of hours of video games. Amazon committed an additional $13B to India AI infrastructure. Adobe acquired Topaz Labs to absorb its image and video enhancement tooling.

Apple skips to M7; Databricks alum claims 1000x power cut

Apple is reportedly skipping the high-end M6 entirely and going straight to an AI-focused M7 line (Pro/Max/Ultra). The Databricks former AI chief debuted Un-0, an image-generation system he claims can cut AI's power bill by 1,000x. MIT Tech Review profiles retail's AI repositioning.

News Hacker News

Apple to skip high-end M6 Mac chips in favor of AI-focused M7 line

Bloomberg: Apple will skip the high-end M6 family and jump to M7 Pro/Max/Ultra, restructured around AI workloads (279 HN points).

Why it matters
  • Largest Mac-silicon restructuring since the M1 — Apple consolidating around AI as the lever for premium Mac.
  • Frees Apple to redesign on-device AI memory and accelerator topology without dragging M6 compatibility.
  • Sets the consumer-AI hardware comp for the next 24 months.
computeproducts
Source →

Notion goes all-in on agents; research catches the harness wave

Notion is killing the Skiff-derived email app to go 'all in on agents for your inbox.' Research adds The Verification Horizon (no silver bullet for coding-agent rewards), the multi-step tool-use RL collapse paper, and JetSpec for parallel-tree speculative decoding. A practitioner's writeup of how 2,000 people tried to hack his AI assistant (192 HN) lands as the practical complement.

News Ars Technica AI

Notion killing Skiff-influenced email app since most users use AI agents instead

Notion is shutting down the Skiff-derived email client to focus on agentic inbox automation.

Why it matters
  • First major shutdown of a 'traditional' SaaS surface explicitly because users moved to agents.
  • Validates the Claude Tag / Microsoft Scout 'agent in the workspace' direction.
  • Sets the template other SaaS vendors will follow as agent-driven workflows displace UI.
productsagentsmarket
Source →

Papers

3 items

Notion goes all-in on agents; research catches the harness wave

Notion is killing the Skiff-derived email app to go 'all in on agents for your inbox.' Research adds The Verification Horizon (no silver bullet for coding-agent rewards), the multi-step tool-use RL collapse paper, and JetSpec for parallel-tree speculative decoding. A practitioner's writeup of how 2,000 people tried to hack his AI assistant (192 HN) lands as the practical complement.

Also today