trillioniar s

Posted on Aug 9

AI News August 9: OpenAI Astra Solves 10 Math Problems, EU AI Act Goes Live, DeepGrove Maple Runs at 127 tok/s on iPhone

#ai #llm #coding #openai

AI News August 9: OpenAI Astra Solves 10 Math Problems, EU AI Act Goes Live, DeepGrove Maple Runs at 127 tok/s on iPhone

The AI world delivered one of its biggest research weeks in history. OpenAI's unreleased Astra model solved 10 open math problems with machine-checkable proofs — for just $2,000 in compute. The EU AI Act's full enforcement kicked in on August 2, making chatbot disclosure and deepfake labeling legally binding across Europe. And a tiny ternary-weight model called Maple-Preview hit 127 tokens per second on an iPhone. Here's everything that happened in AI on August 9, 2026.

Major Updates

OpenAI Astra Solves 10 Open Math Problems With Lean Proofs

The Biggest AI Research Claim of 2026

On August 1, OpenAI published a 249-page manuscript alongside 10 machine-checkable Lean 4 proof certificates. Each certificate verified that an internal version of the model — codenamed Astra — had solved a math or theoretical computer science problem that had been open for at least a decade.

How It Was Done

The proofs were generated for roughly $2,000 in compute. OpenAI published the formal Lean proofs on GitHub, making the results independently verifiable by any mathematician with the Lean proof checker. The problems spanned number theory, combinatorics, and computational complexity.

Why It Matters

Unlike benchmark records, these are genuine research contributions that can be checked by machines. If independent review confirms all 10 proofs, this would be the first time an AI system produced verified new mathematical knowledge at this scale. OpenAI has not announced whether Astra will be labeled GPT-6 or remain its own model line — the naming decision is still internal.

Sources

EU AI Act Enforcement Begins Across Europe

The Regulatory Era Is Now Real

August 2, 2026 marked the first day of full EU AI Act enforcement. Article 50 transparency obligations are now legally binding — chatbots must disclose they are AI, deepfakes must be labeled, and AI-generated public-interest text must carry machine-readable watermarks using C2PA or equivalent standards.

What the Penalties Look Like

Violations of the transparency rules can trigger fines of up to 3% of global turnover or €15 million, whichever is higher. High-risk AI systems require conformity assessments before deployment. The regulation applies to any company serving EU users, regardless of where the company is based.

Only 8 of 27 Member States Are Ready

As of March 2026, only 8 of the EU's 27 member states had designated their AI Act enforcement contacts. This raises questions about how effectively the rules will be enforced across the bloc in the short term. The UK's AI Regulation and Safety Bill cleared the House of Commons in parallel, with Royal Assent expected by October.

Sources

DeepGrove Maple-Preview Runs at 127 tok/s on iPhone

Ternary Weights Make 20B Parameters Fit in Your Pocket

DeepGrove released Maple-Preview on August 5 — an open-source 20.2B-parameter mixture-of-experts reasoning model that uses ternary weights (values of -α, 0, or +α) instead of standard 16-bit floating point. The result: 127 tokens per second on an iPhone and 281.5 tokens per second on an Apple M5 Pro MacBook Pro.

The Architecture

Maple has 24 layers with 256 experts, selecting the top 8 per token. It uses a 512-token sliding window on 3 of every 4 layers. Weights are packed into 2 bits, making the model dramatically smaller than traditional 20B-parameter models. It runs on the stock MLX build for Apple Silicon portability.

Why It Matters

This is 5–16× faster than efficient models like Gemma 4, Qwen 3.5, and gpt-oss at comparable or smaller sizes. The model solves IMO-level math problems, putting frontier reasoning capability on a phone for the first time. DeepGrove is a small independent AI research lab — this is the kind of work that was impossible for labs outside the top five just two years ago.

Sources

Claude Fable 5 Redeployed After US Government Export Controls

Anthropic's Most Powerful Model Returns Globally

Claude Fable 5 — Anthropic's most advanced model, built for autonomous long-running tasks in software engineering and cybersecurity — was pulled offline on June 12 after the US Department of Commerce applied export controls. On July 1, Anthropic redeployed the model globally, citing "new cybersecurity classifiers" and a 50% weekly usage cap through July 7.

What Happened in the Middle

Amazon researchers reportedly alerted the White House that jailbreaks could strip Fable 5's safety guardrails, triggering the government intervention. Anthropic CEO Dario Amodei stepped aside from direct White House negotiations, and a new lead negotiator was brought in. The company eventually secured permission to redeploy with enhanced safeguards.

Why It Matters

This was the first time the US government directly intervened to pull a commercial AI model off the market. The episode established a precedent: frontier AI models are now treated as export-controlled technology. Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, with a 1 million token context window.

Sources

Astera Labs Smashes Q2 Earnings as AI Infrastructure Demand Surges

Revenue Doubles to $392.4 Million

Astera Labs reported Q2 2026 revenue of $392.4 million, beating the consensus estimate of $360.72 million. Adjusted earnings per share came in at $0.80, trouncing the $0.69 Street estimate. Revenue surged 104% year-over-year, driven by accelerating demand for AI infrastructure connectivity products.

The Scorpio X-Series Ramp

The company cited the volume-production ramp of its Scorpio X-Series fabric switches as a key driver. These switches connect GPUs and accelerators inside data centers, forming the backbone of rack-scale AI infrastructure. Astera Labs projects a sharp sequential increase for Q3.

Why It Matters

Astera Labs sits at the physical layer of the AI boom — its chips move data between GPUs in data centers. A revenue beat of this magnitude signals that hyperscaler capex on AI infrastructure is still accelerating, not slowing. The stock jumped in extended trading on the results.

Sources

Corgi Insurance AI Hits $2.6 Billion Valuation

AI-Native Insurance for the Startup Era

Corgi, a San Francisco-based AI insurance startup, reached a $2.6 billion valuation after a $106 million Series B1 funding round. The valuation doubled from $1.3 billion just three weeks earlier. Corgi has insured 40,000 clients across 49 states, issuing policies in minutes using LLMs for underwriting, claims processing, and policy management.

The Business Model

Corgi built its entire stack — insurer, reinsurer, and platform — on large language models. The company focuses on insuring AI and startup risks, filling a gap that traditional insurers have been slow to address. Jack Ma's Yunfeng Capital was among the investors in earlier rounds.

Why It Matters

The doubling of Corgi's valuation in three weeks illustrates how quickly AI-native businesses can scale. The company is also building the insurance layer for the AI economy itself — as AI deployment grows, the liability risks grow with it, and Corgi is positioning to be the default underwriter for that risk.

Sources

OpenAI, Anthropic, and DeepMind Staff Petition US to Pace AI Development

The Workers Want the Government to Slow Down

Employees at OpenAI, Anthropic, and Google DeepMind have circulated a petition urging the US government to support the pacing of frontier AI development. The petition, first reported on August 4, calls for enforceable safety benchmarks and mandatory pre-deployment testing before models exceed certain capability thresholds.

Context: 200 Protesters Already Marched

On July 14, approximately 200 protesters marched between the offices of OpenAI, Anthropic, and Google DeepMind, calling on frontier AI companies to pause training more powerful models. The staff petition represents a more institutional approach to the same concern.

Why It Matters

When the people building the most powerful AI systems are publicly calling for slower development, it signals a disconnect between the pace of commercial competition and the comfort level of the researchers closest to the technology. This kind of internal dissent could influence future regulation.

Sources

GPT-6 Naming: Will OpenAI Call It Astra or GPT-6?

The Name Game Continues

OpenAI introduced Astra on August 1 as "our next major model," but has not decided whether it will be labeled GPT-6 or positioned as an additional model in the GPT-5 series (such as GPT 5.7). An internal version of Astra produced ten new results in mathematics and theoretical computer science, published with machine-checkable Lean 4 proofs.

Why the Name Matters

The naming decision has commercial implications. GPT-6 would signal a generational leap, potentially triggering massive adoption. Keeping it as "Astra" or "GPT 5.7" would be more conservative, suggesting incremental improvement. The choice also affects how the model is perceived by regulators, investors, and competitors.

Sources

Frequently Asked Questions

What did OpenAI's Astra model actually solve?

OpenAI's Astra model solved 10 open problems in mathematics and theoretical computer science. Each problem had been unsolved for at least a decade. The proofs were verified using Lean 4, a formal proof-checking language, and published on GitHub. The total compute cost was approximately $2,000.

When did the EU AI Act become enforceable?

The EU AI Act's full enforcement began on August 2, 2026. Article 50 transparency obligations — including chatbot disclosure, deepfake labeling, and AI-generated content marking — are now legally binding. Fines can reach 3% of global turnover or €15 million.

What makes DeepGrove Maple-Preview different from other models?

Maple-Preview uses ternary weights (values of -α, 0, or +α) instead of standard 16-bit floating point, packing 20.2 billion parameters into 2-bit weights. This allows it to run at 127 tokens per second on an iPhone and 281.5 tokens per second on an M5 Pro MacBook — 5–16× faster than comparable models.

Why was Claude Fable 5 pulled from the market?

The US Department of Commerce applied export controls to Claude Fable 5 and Claude Mythos 5 on June 12, 2026, after Amazon researchers reportedly found jailbreaks that could remove safety guardrails. Anthropic redeployed the model globally on July 1 with enhanced cybersecurity classifiers and usage limits.

How much revenue did Astera Labs report in Q2 2026?

Astera Labs reported $392.4 million in Q2 2026 revenue, beating the $360.72 million consensus estimate. Adjusted EPS was $0.80 versus the $0.69 estimate. Revenue surged 104% year-over-year, driven by demand for AI data center connectivity products.

What is Corgi Insurance and why is it valued at $2.6 billion?

Corgi is an AI-native insurance startup that uses large language models for underwriting, claims processing, and policy management. It has insured 40,000 clients across 49 states and issues policies in minutes. Its valuation doubled from $1.3 billion to $2.6 billion in three weeks.

Are AI researchers calling for slower development?

Yes. Employees at OpenAI, Anthropic, and Google DeepMind have petitioned the US government to support pacing frontier AI development. This follows a July 14 protest where 200 demonstrators marched between the offices of these three companies calling for a pause on training more powerful models.