AI News
Daily Tech Reader - AI Edition
Podcast 🎧 • Video 📽 • Short 📽
Claude
Anthropic's Mythos Breaks Post-Quantum Crypto — Finds New HAWK Signature and Reduced-Round AES Attacks in 60 Hours
Anthropic published research showing its unreleased Claude Mythos Preview discovered two new cryptographic attacks — exploiting a lattice flaw in the NIST post-quantum candidate HAWK and inventing a technique that speeds the best-known attack on 7-round AES-128 by up to 800x. Neither breaks deployed systems, but each cost roughly $100K in API usage to find.
Sources: Anthropic · The New York Times · AI Weekly
Microsoft Books $3.2B Gain on Anthropic in Q4 — OpenAI Stake Marked Down $600M in the Same Quarter
Microsoft's Q4 FY2026 filings show a $3.2B gain on its Anthropic investment — adding 33 cents to diluted EPS — while its OpenAI position was written down roughly $600M. For all of FY2026, OpenAI still delivered a $5B full-year gain, but Anthropic's single-quarter upside nearly matched it.
Sources: TechCrunch · Microsoft IR · AI Weekly
Amodei: Anthropic Has Never Advocated a Ban on Open-Weight Models — Urges Three Targeted Policies Instead
Dario Amodei posted a statement clarifying Anthropic's position, calling smaller open-weight models beneficial and asking Washington to restrict AI chip exports to China, target industrial-scale distillation by authoritarian states, and require mandatory pre-release safety testing on any sufficiently capable model, open or closed.
Sources: Anthropic · AI Weekly
ChatGPT
OpenAI Claims GPT-5.6 Sol Beats Opus 5 on ARC-AGI-3 — but Only With Two Non-Standard API Settings Enabled
OpenAI says GPT-5.6 Sol jumps from 7.8% to 38.3% on the ARC-AGI-3 public set when Retained Reasoning and Compaction are enabled via its Responses API, surpassing Claude Opus 5's 30.2%. ARC Prize co-founder François Chollet ruled the settings permissible since they are general-purpose API features, not benchmark-specific tools.
Sources: OpenAI · The Decoder · MLQ News · OfficeChai · ARC Prize
OpenAI Cuts GPT-5.6 Luna Price 80% and Terra 20% — Effective Today
OpenAI slashed pricing on two members of its GPT-5.6 family: Luna, its most cost-efficient model, drops 80%, and Terra, the balanced everyday model, falls 20%, as the company pushes broader adoption of the new generation.
Sources: OpenAI · AI Weekly
Hugging Face Publishes Forensic Timeline of OpenAI Agent Breach — Recovered Roughly 17,600 Attacker Actions
HF's security team reconstructed the July 9–13 OpenAI intrusion in granular detail, tracing a two-stage chain from an HDF5 file-read primitive to Python execution in a production Kubernetes pod via Jinja2 SSTI. Claude Opus and Fable guardrails blocked exploit analysis, so HF ran a separate model on its own infrastructure to decrypt attacker payloads.
Sources: Hugging Face · AI Weekly
Gemini
Google Developing 'App' Artifact for Gemini Notebook — Generates Interactive HTML Apps Directly From Your Sources
Google is internally testing a new App artifact feature inside Gemini Notebook that generates interactive HTML dashboards, games, and tools from uploaded source material via a single prompt, a direct challenge to Claude's Artifacts capability.
Sources: Crypto Integrated · AI Weekly
EU Opens €10B Bidding for Seven AI Gigafactories — Each Site to Host at Least 100,000 AI Chips
The European Commission formally launched its AI Gigafactory proposal call on July 30, backing €10B in public funding toward a €30B total target across seven sites, each roughly 4x current EU data-center scale. Bidding closes November 12, with awards expected by mid-2027.
Sources: News4Jax · AI Weekly
Copilot
Qualcomm Closes Modular Acquisition — Chris Lattner Named EVP of Advanced AI Software
Qualcomm completed its purchase of Modular, the AI infrastructure software company behind the MAX platform and Mojo language, with founder Chris Lattner moving to Qualcomm as EVP ahead of its stated $5B data-center revenue target for FY2027.
Sources: HPCwire · AI Weekly
Ruflo MCP Bridge Flaw Scores CVSS 10.0 — All 233 Tools Exposed on Default Docker Deployments
Noma Labs disclosed a maximum-severity flaw in the open-source AI agent platform Ruflo that allowed a single unauthenticated HTTP POST to execute code and exfiltrate LLM API keys. Ruflo's maintainer shipped a patch within 24 hours of the June 30 disclosure.
Sources: The Hacker News · AI Weekly
Grok
MCP Goes Fully Stateless — July 28 Spec Update Drops Sessions, Adds Multi Round-Trip Requests
The Agentic AI Foundation released the largest MCP specification change since launch, moving to a stateless request/response core that eliminates the session handshake so any server instance can handle any request behind a plain load balancer. The update also hardens auth, adds mid-call user input support, and opens a formal 12-month deprecation window.
Sources: Model Context Protocol Blog · AI Weekly
Aschenbrenner's AI Hedge Fund Invites Fresh Capital After Tech Selloff Losses
Situational Awareness, the ~$20B hedge fund founded by ex-OpenAI researcher Leopold Aschenbrenner, invited investors to commit new capital starting August 1 after sharp losses in the recent AI-stock correction, per an investor letter reported by the FT. The fund remained up 439% net through June on leveraged bets across AI infrastructure names.
Sources: Hedgeweek · Financial Times · AI Weekly
China AI
Brookfield and NextEra Plan $100B Kentucky AI Campus at Former DOE Site — 1.2 GW Online by 2032
The two companies will develop a massive AI data-center campus at the former Paducah Gaseous Diffusion Plant, paired with 2 GW of natural gas generation and 2.6 GW of battery storage, with initial operations targeted for 2028 and the site eventually scaling to 1.8 GW.
Sources: Data Center Knowledge · AI Weekly
Bloomberg: China AI Boomtowns Surge, but Wealth Isn't Reaching Households
Bloomberg reports China's AI hubs are driving national GDP growth as investment pours into chip fabs and data centers, but even in these boomtowns wages and consumer spending have failed to keep pace with the AI capex flow.
Sources: Bloomberg · AI Weekly
Zhongji Innolight Slips 2% on $6.8B Hong Kong Debut — Asia's Second-Biggest Float of 2026
The optical-transceiver maker, whose products wire AI data centers, closed below its IPO price on July 30 after raising HK$52.9B — Hong Kong's largest listing in seven years — amid a global AI-stock pullback. The company plans to direct 35% of proceeds to optical-interconnect R&D.
Sources: China Daily HK · AI Weekly
Cloud AI
Meta Q2 Revenue Hits Record $60.8B — But Free Cash Flow Collapses to $784M as AI Capex Doubles, Stock Falls ~10%
Meta's Q2 2026 results beat on revenue (+28% YoY) but missed EPS expectations as $31.1B in quarterly capex and $2.4B in legal charges crushed free cash flow from $8.5B to $784M. The company raised its full-year capex floor to $130B–$145B, and shares dropped roughly 10% after hours.
Sources: Meta IR · Eastern Herald · Investing.com · CNBC
Arm Data-Center Royalties Double Year-Over-Year — AGI CPU Demand Jumps to $2B Opportunity for FY27–28
Arm reported record Q1 FY27 revenue of $1.29B (+22% YoY), with data-center royalties more than doubling as hyperscalers adopt Neoverse and Armv9 designs. CEO Rene Haas said customer demand for Arm's AGI CPU has grown to over $2B across the next two fiscal years.
Sources: Benzinga · AI Weekly
Dwarkesh Patel: AI Compute Could Get 10–15x More Expensive — Spot Prices Already Up 40% From February Trough
In a widely-read July 29 essay, Dwarkesh Patel argues that lab revenue growth is outpacing the ~3x annual compute build, pointing to Google reportedly paying SpaceX $900M/month for 110K GPUs as evidence that pricing pressure is already structural.
Sources: Dwarkesh.com · AI Weekly
Big Story
Big Tech Earnings Week Splits in Two — Microsoft Soars, Meta Sinks, as AI Spending Patience Wears Thin
Microsoft's blowout Azure quarter sent its stock up 7% after hours while Meta fell ~10% on near-zero free cash flow, illustrating a fracturing investor thesis: cloud monetization is being rewarded, while pure AI infrastructure spending without visible return is being punished.
Sources: Microsoft IR · Meta IR · CNBC · Bloomberg
The ARC-AGI-3 Benchmark Is Now a Lab-vs-Lab Battleground — and the Rules Are Still Being Written
OpenAI's claim that GPT-5.6 Sol outperforms Claude Opus 5 when run through its own API settings — versus an official harness that cuts its score to 7.8% — exposes a fundamental problem: AI benchmarks now measure model-plus-infrastructure, and no consensus exists on where the model ends.
Sources: OpenAI · Anthropic · ARC Prize · The Decoder · AI Weekly
Business
Airlines Quietly Replace Human Pricing Analysts With AI — Bloomberg Reports Bargain Fares Are Disappearing on Busy Routes
Airlines are deploying autonomous AI systems that reprice seats in real time using dozens of variables, shrinking the gaps that once let travelers stumble onto cheap fares, Bloomberg reports. The shift represents a move from rigid rule-based pricing to continuous AI optimization across the industry.
Sources: Bloomberg · AI Weekly
Solo $1M+ Businesses Nearly Tripled Since 2023 — Stripe Attributes Surge Directly to AI Tools
Stripe told the Wall Street Journal that thousands of solo operators on its platform now exceed $1M in annual revenue, with the $10M+ cohort nearly tripling between 2023 and 2025 as AI tools let single founders replicate the output of small teams.
Sources: Wall Street Journal · AI Weekly