Sarmadi AI Digest June 30, 2026 Updated 7:00 AM CT Today Archive Topics Saved Subscribe RSS

South Korea commits $1T to memory + humanoids; California cuts an Anthropic side deal; Meta teen-chatbot scandal

South Korea announced more than $1T of investment to ease 'RAMageddon' and seed commercial humanoid robots by 2028 — Samsung and SK Hynix together committed $550B in memory fabs alone. Anthropic and Governor Newsom forged a half-price Claude deal for California state government, a state-level workaround as the federal trusted-user roster squeezes access. Wired reported hundreds of Meta contractors posed as teenagers in rival chatbots to probe responses on suicide, sex, and drugs. Lawmakers (Warren-Scanlon) introduced a bill to ban AI companies from selling health and location data. China's LongCat-2.0 — a 1.6T-parameter MoE with 48B active — added to the open-weights pressure. TC quantified the AI-jobs picture: high-intensity AI adopters saw headcount grow 10.2%, scrambling the layoff narrative. OpenAI teased Codex-branded hardware. Arena is now a $100M leaderboard business.

7 papers 18 news 9 sources ← Latest

News

16 items

South Korea's $1T memory + humanoid push; LongCat-2.0 ships

South Korea is committing over $1T to push memory production and commercial humanoid robots by 2028, with Samsung and SK Hynix alone pledging $550B for new fabs to ease 'RAMageddon.' China's LongCat-2.0, a 1.6T MoE with 48B active parameters, lands the same day — Asian compute and Asian open-weights compounding.

News Ars Technica AI

South Korea to spend $1T on more memory chip production and humanoid robots

Ars: South Korea targets a $1T national push on memory chips and commercial humanoid robots by 2028.

total commitment $1Ttarget year 2028
Why it matters
  • Largest national AI/compute commitment outside the US — and explicitly two-track: memory plus physical AI.
  • Materially affects the Micron / Wall-Street-says-next-Nvidia trade from yesterday — more supply but also more state-aligned demand.
  • Sets the bar Europe is racing to match and the US will respond to.
  • Cements the Indo-Pacific drone+robot doctrine from this week's SK military news.
computepolicyroboticsinfrastructure
Source →
News Hacker News

LongCat-2.0, a large-scale MoE model with 1.6T total and 48B Active

LongCat-2.0 ships as a 1.6T-parameter MoE with 48B active — major Chinese open-weights release (171 HN points).

total params 1.6Tactive params 48B
Why it matters
  • Adds another credible 1T-scale Chinese open-weights model alongside GLM 5.2 and DeepSeek.
  • The 1.6T/48B-active topology mirrors the cost-efficient sparsity research US labs are pushing closed.
  • Compounds yesterday's GLM-5.2-beats-Claude cyber result — open-weights breadth is now real.
open-weightsproductstraining
Source →

California cuts an Anthropic side deal; lawmakers move on AI health data

Governor Newsom and Anthropic agreed to a half-price Claude deal for California state government — a state-level workaround to the federal trusted-user roster. Senators Warren and Scanlon introduced a bill to ban AI companies from selling health and location data. Wired published a Bernie Sanders interview tying both stories to the wealth-and-AI thread.

News TechCrunch AI

Anthropic and Gov. Newsom forge deal allowing California government to use Claude at half price

TC: Anthropic and Gov. Newsom agree on a half-price Claude deal for California state government use.

Why it matters
  • Anthropic explicitly cuts a state-level deal as the federal trusted-user roster constrains access — a workaround template for other states.
  • Politically signals state-vs-federal competition on AI procurement and access.
  • Establishes Anthropic's go-to-market through public-sector channels independent of the White House.
policymarketregulation
Source →
News The Verge AI

Lawmakers want to ban AI companies from selling your health data

The Verge: Warren and Scanlon introduce the Health and Location Data Protection Act, banning sale to AI brokers.

Why it matters
  • First federal bill specifically restricting AI-company use of health and location data.
  • Directly affects the data-broker substrate of most consumer-AI personalization plays.
  • Lands the same day prosecutors put ChatGPT logs in evidence — privacy-vs-utility tradeoff is at trial-court speed.
policyregulationdatasafety
Source →
News TechCrunch AI

The AI jobs debate just got messier

TC: high-intensity AI adopters saw headcount increase 10.2% — scrambling the AI-replaces-jobs narrative.

headcount change for high-intensity adopters +10.2%
Why it matters
  • First data showing AI-heavy adopters increasing rather than decreasing headcount.
  • Reframes the Sanders-Oracle layoff narrative — different operators get different outcomes.
  • Quantifies the Ford rehire story from yesterday in a national-level sample.
marketpolicy
Source →

Meta teen-chatbot probe scandal; TIDAL cuts AI-music pay

Wired reported hundreds of Meta contractors posed as teenagers to prompt rival chatbots about suicide, sex, and drugs as a competitive intelligence project. TIDAL responded to the AI-music surge by demonetizing AI-generated tracks instead of banning them outright.

News Wired AI

Meta Contractors Posed as Teens to Prompt Rival Chatbots About Suicide, Sex, and Drugs

Wired investigation: hundreds of Meta contractors posed as kids to probe rival chatbots on suicide, sex, and drugs.

Why it matters
  • First major exposé of competitive-intel red-teaming via fabricated minor identities.
  • Reignites the Meta-and-teens policy story that has already drawn AG attention.
  • Sets up regulatory pressure on cross-vendor chatbot safety testing methodology.
safetypolicyproducts
Source →

Arena hits $100M; Cursor mobile; Codex hardware tease; OKX agent payments

Arena, the AI-leaderboard everyone uses, is now a $100M commercial business. Cursor shipped a mobile app for guiding coding agents on the go. OpenAI teased Codex-branded hardware. OKX wants AI agents to hire and pay each other on a crypto-rails marketplace. MIT Tech Review pushes back: 'AI agents are not your coworkers.'

Sunday research: agentic abstention, terminal benchmark, low-compute world models

TUA-Bench evaluates general-purpose terminal-use agents. Agentic Abstention asks whether agents know when not to act. DreamForge-World 0.1 previews a low-compute real-time controllable world model. Allen AI's DiScoFormer unifies density and score modeling across distributions.

Papers

3 items

Sunday research: agentic abstention, terminal benchmark, low-compute world models

TUA-Bench evaluates general-purpose terminal-use agents. Agentic Abstention asks whether agents know when not to act. DreamForge-World 0.1 previews a low-compute real-time controllable world model. Allen AI's DiScoFormer unifies density and score modeling across distributions.

Also today