Sarmadi AI Digest July 8, 2026 Updated 7:00 AM CT Today Archive Topics Saved Subscribe RSS

Meta Muse defaults to remixing Instagram photos; DeepSeek plans own chips; GitHub AI leaks repos

Meta launched Muse Image, an image generator from its Superintelligence Labs — and Instagram users must opt out or their public photos become training and remix material. Wired led the backlash. Ars reported DeepSeek plans to build its own chips to escape US export controls — the substitution effect widens from models to silicon. GitLost researchers (239 HN) tricked GitHub's AI agent into leaking private repos, and Ars documented 'HalluSquatting' — a technique using nine popular AI tools to assemble massive botnets. SambaNova raised $1B at an $11B valuation. Google published a Gemma 4 technical report. Anthropic launched Claude Cowork on mobile and web. OpenAI's Chief Futurist Joshua Achiam is leaving after nearly nine years. Microsoft is now cutting AI spend by leaning on its own models — following the Amazon/OpenAI cost-cutting trend the day after the 4,800-person layoff.

8 papers 19 news 8 sources ← Latest

News

18 items

DeepSeek plans own chips; SambaNova hits $11B; Gemma 4; MSFT cuts spend

Ars reports DeepSeek is planning its own chips to reduce dependence on Nvidia and Huawei amid US export controls. SambaNova raised $1B at $11B five months after its last round. Google published the Gemma 4 technical report. Microsoft joins Amazon and OpenAI in cutting AI spend by leaning more on its own models, the day after 4,800 layoffs.

News Ars Technica AI

Facing US export controls, China's DeepSeek plans to make its own chips

Ars: DeepSeek plans its own chips to reduce dependence on Nvidia and Huawei amid US export controls.

Why it matters
  • Substitution effect widens from models to silicon — the export-control regime is producing a full-stack Chinese response.
  • Directly relevant to the OpenAI-Broadcom Jalapeño and Anthropic-Samsung custom-chip moves.
  • Materially affects Nvidia's China revenue thesis over the next 24 months.
  • Sets the reference point for how Chinese frontier labs vertically integrate.
computepolicyproductsmarket
Source →

GitHub AI agent tricked into leaking private repos; HalluSquatting botnets; Discord bans

GitLost researchers tricked GitHub's AI agent into leaking private repos (239 HN). Ars documented 'HalluSquatting' — a technique using nine popular AI tools to build massive botnets by exploiting LLMs' inability to say 'I don't know.' Discord admitted its AI moderation bug wrongly banned users over harmless images since May.

News Hacker News

GitLost: We Tricked GitHub's AI Agent into Leaking Private Repos

Noma Security via HN (239 points): researchers tricked GitHub's AI agent into leaking private repositories.

Why it matters
  • First high-visibility exfiltration of private repos through vendor AI agent misdirection.
  • Directly compounds the Claude Code steganography and Anthropic tracker stories.
  • Sets the reference case for enterprise procurement questions on AI-agent code access.
safetycodepolicy
Source →

Claude Cowork goes mobile; Achiam leaves OpenAI; Harding warns arms race

Anthropic launched Claude Cowork on mobile and web (TC, Verge, Wired) — the desk-to-phone continuity story. OpenAI's Chief Futurist Joshua Achiam is leaving after nearly nine years. Former DeepMind exec Verity Harding tells Wired the AI arms race could end in disaster if the US government stays nationalistic.

News TechCrunch AI

Claude Cowork expands to mobile and web

TC: Claude Cowork on mobile/web — start on desktop, get updates on phone, agent keeps running.

Why it matters
  • Extends the Cursor-mobile pattern to Anthropic's own surface.
  • Directly competitive with the SpaceX-owned Cursor mobile app.
  • Sets the multi-surface expectation for every enterprise agent product.
productsagentsmarket
Source →

Data center energy squeezes Rust Belt; Stratechery on Xbox cuts; Anthropic vs open source

Ars documents data-center AI demand raising Rust Belt electricity bills and threatening Trump's 'Made in America' plan. Stratechery ties the Microsoft Xbox cuts to bundling economics. TC argues the open-source AI wave isn't hurting Anthropic — yet — because it eases margin pressure on premium tiers.

Papers

1 item

DeepSeek plans own chips; SambaNova hits $11B; Gemma 4; MSFT cuts spend

Ars reports DeepSeek is planning its own chips to reduce dependence on Nvidia and Huawei amid US export controls. SambaNova raised $1B at $11B five months after its last round. Google published the Gemma 4 technical report. Microsoft joins Amazon and OpenAI in cutting AI spend by leaning more on its own models, the day after 4,800 layoffs.

Paper Hugging Face

Gemma 4 Technical Report

Google publishes the Gemma 4 technical report — refreshed open-weights family from DeepMind.

Why it matters
  • Refreshed Google open-weights family — competitive baseline against Qwen and Llama.
  • Direct answer to the GLM 5.2 / LongCat 2.0 open-weights wave.
  • Sets the comparison target for the H2 open-weights conversation.
open-weightstrainingproducts
Source → Arc

Also today