EU Act Enforcement Begins; Washington Hides Framework.

Mazen Al Sabbagh

— 6 min read
Share

TL;DR

  • EU Act in force: Brussels initiated enforcement Sunday, granting its AI Office powers to audit and fine providers; simultaneously, California's state-level provenance mandate became operative.
  • US framework opaque: The White House claims its voluntary framework is complete but undisclosed, summoning Anthropic, OpenAI, Google, and Meta for a private review Tuesday.
  • OpenAI's "Astra" preview: Sam Altman reportedly presented OpenAI's next model, Astra, to US officials last week, positioning it as the first candidate for Washington's unreleased evaluation process.
  • Market divergence: Seoul's KOSPI dropped 5.1% amid a major fund's liquidation, while US markets rallied, with Nasdaq up 2.3% and Amazon surpassing a $3 trillion valuation.
  • Alibaba's Qwen3.8-Max: The new 2.4T-parameter multimodal MoE is gaining traction despite lacking a model card, license, or weights, with published benchmarks showing notable performance gaps.

Lead Story: Europe's AI Enforcement Commences; Washington's Framework Remains Private

The EU AI Act's Chapter V enforcement provisions took effect Sunday. The Commission is now authorized to demand documentation, conduct model evaluations, mandate corrective actions, withdraw non-compliant models from the EU market, and impose fines on GPAI providers up to €15 million or 3% of global turnover, whichever is higher [https://ec.europa.eu/commission/presscorner/detail/en/ip_26_1714].

Article 50 transparency obligations also began, requiring AI disclosure for chatbots, labeling for deepfakes, and machine-readable marks for AI-generated content. A grace period extends to December 2 for existing systems. The newly established AI Office, staffed with 145 personnel, adopts a "constructive dialogue" approach, with no investigations yet initiated [https://www.techpolicy.press/-brussels-gains-new-ai-act-enforcement-powers-as-autonomous-ai-tests-regulators/].

Coincidentally, Sunday also marked the operational date for California's AI Transparency Act [https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202320240SB942]. This establishes the first US state-level provenance mandate, with penalties of $5,000 per violation per day for generative systems exceeding one million monthly California users.

Washington's response on its own framework arrived midday Monday, albeit with significant opacity. Following the August 1 deadline, which passed without publication [https://finance.yahoo.com/technology/ai/articles/white-house-ai-framework-deadline-002011007.html], the White House informed Axios it met the deadline for its voluntary frontier-model evaluation framework but will not release its contents or identify those who have reviewed it [https://www.axios.com/2026/08/03/white-house-finalizes-ai-framework-behind-closed-doors].

Official records remain silent; no Federal Register notice, nor any documentation from NIST or OSTP, has appeared. However, CNBC reports the Office of the National Cyber Director will host AI companies Tuesday to discuss the framework [https://www.cnbc.com/2026/08/03/white-house-ai-companies-voluntary-framework-meeting.html]. This will be a staff-level session focused on the security and confidentiality terms of the voluntary 30-day pre-release access, not a public unveiling.

A late adjustment: Yahoo News indicates Meta, previously reported as a holdout from the joint draft edits submitted by Anthropic, OpenAI, and Google in late July, is now on the attendee list [https://www.yahoo.com/news/politics/articles/white-house-finalizes-voluntary-ai-162700622.html].

Crucially, The Information reports Sam Altman spent last week demonstrating "Astra" — a multi-agent system for long-running tasks, still a provisional name — to Senators Warnock and Moreno, and Secretaries Bessent and Lutnick [https://x.com/wallstengine/status/2083353809980666289]. This model is reportedly slated as the first submission under the 30-day review gate, a mechanism whose specifics remain exclusively within the White House's purview.

In Other News

OpenAI's math proofs garner acclaim, but key voices remain quiet. Two days after OpenAI's "Ten advances in mathematics" publication, the Lean 4 certificates have shown no errors. Lattice cryptographer Chris Peikert praised the originality and elegance of his field's result, Daniel Litt conceded a bet, and the WSJ reports Fields Medalist Jacob Tsimerman is moving to OpenAI for safety work. Skeptics Gary Marcus and Ernie Davis still label it "amazing but vastly oversold" due to a lack of methodology and failed attempt counts, while Simon Willison seeks the prompts. Notably, Terence Tao's and Scott Aaronson's feeds remain silent, despite aggregator claims of a Tao response.

Wall Street stages a relief rally, Amazon breaks $3 trillion. August opened strong for US equities; by 3 pm ET, the Nasdaq gained 2.3%, the S&P 1.6%, and the Dow 1.3%, driven by Nvidia's 3.9% rise and Amazon's first-ever breach of the $3 trillion valuation mark post-Friday earnings. Oil prices declined over 5% (WTI near $79) after reports of Trump halting Iran strikes for Hormuz Strait talks, a claim Iran denies. Palantir reports after market close, with options anticipating a ~12% swing.

Seoul's KOSPI drops significantly amidst fund liquidation. The KOSPI closed down 5.12% at 6,257.45, with Samsung and SK Hynix each falling almost 9% on ₩4.77 trillion in foreign and institutional selling. This follows Leopold Aschenbrenner's Situational Awareness fund liquidating its entire ~$16 billion public book to Citadel due to margin calls [https://www.cnbc.com/2026/07/31/leopold-aschenbrenner-situational-awareness-fund-fire-sale.html]. Bloomberg Opinion's Monday assessment labeled the fund's 439% first-half return "a warning." Korea's leveraged-ETF caps are set for Wednesday.

Qwen3.8-Max's details under scrutiny. Alibaba's 2.4T-parameter multimodal MoE, priced at $2/$6 per million tokens, still lacks a model card, license, or weights—promised "next week." All published scores are Alibaba's own: strong on PaperBench (93.0) and Terminal-Bench 2.1 (86.6), but trailing on Humanity's Last Exam (43.6 vs Claude Fable 5's 53.3) and SWE-bench Pro (67.7 vs 80.0), with coding evaluations run within the Claude Code harness. Alibaba's stock closed up 7% in Hong Kong, and the Hacker News thread neared 1,000 points. MiniMax's open MiniMax-H3 also gained day-0 ComfyUI support.

X / Social Pulse

One to Watch

Tuesday at the White House: The Office of the National Cyber Director will convene Anthropic, OpenAI, Google — and now Meta — to review the model-testing framework. The administration maintains this framework is finalized but unreleased. This staff-level session is a discussion, not a public rollout. Watch for any readouts, a signatory list, and whether Meta's prior holdout—stemming from Llama's open-weights architecture—persists. AMD reports the same day, with options pricing an 8.7% move.

Quick Hits

  • Palantir earnings tonight: Consensus revenue near $1.81 billion, stock up 2.4% into the close; notably, the stock has fallen after its last four beats.
  • California suspense file: Correction from this morning: the Senate Appropriations calendar [https://sapro.senate.ca.gov/] places California's suspense-file hearing on August 13, not today. Today's 10 am PT session is a standard hearing for bill placement, per the committee and CSAC.
  • Yen intervention confirmed: Friday's yen operation is validated as the first joint US-Japan intervention since 2011, totaling ~$36.6 billion per BOJ data. Finance Minister Katayama has pledged further action if necessary.
  • Trending models: Kimi K3 holds the #1 trending spot on Hugging Face with over 967,000 monthly downloads. DeepSeek's V4-Flash weights are nearing 236,000, while V4-Pro remains "coming soon."
  • Grok 4.6 anticipated: Elon Musk's July 28 post indicates Grok 4.6 is penciled in for around Thursday, August 7, though x.ai itself has yet to publish any details.

Brussels has publicly established its regulatory framework. Washington, conversely, asserts its own framework is complete but remains deliberately opaque. The initial test of this unreleased US structure unfolds Tuesday, behind similarly closed doors, highlighting a clear divergence in global AI governance strategy.

Sources

Lead / policy: Commission press release · TechPolicy.Press on Chapter V · California SB 942 · Yahoo Finance on the lapsed deadline · Axios on the closed-door framework · CNBC on Tuesday's meeting · Yahoo News on Meta attending · The Information's Astra report (via X) · CA Senate Appropriations calendar · CSAC legislative update

Markets: Yahoo Finance live blog · Investrade mid-morning · TheStreet on oil/Iran · Korea JoongAng Daily KOSPI close · Sedaily on Samsung/Hynix · CNBC on Situational Awareness · Bloomberg Opinion on the 439% return · CNBC on the yen intervention · TechTimes Palantir preview · Benzinga on AMD options

Models: OpenAI: Ten advances · ten-proofs repo · Marcus/Davis critique · Simon Willison · WSJ Tsimerman via Techmeme · MarkTechPost on Qwen3.8-Max · apidog benchmark analysis · CNBC on Alibaba's close · MiniMax-H3 · Kimi K3 on Hugging Face · DeepSeek V4-Flash weights · DeepSeek changelog · Musk on Grok 4.6

Pulse: HN: hallucinated SQLite CVE · HN: meat proxy · HN: Qwen3.8-Max · HN: OpenAI super PAC · HN: MiniMax-H3 in ComfyUI · Karpathy on Opus 5 · Pacing the Frontier

Lock in. M. mazen@thorterminal.com

Read more