JULY 12, 2026 — WHAT AI CAN DO, WHAT LABS ADMIT, WHO CONTROLS THE CHIPS
Today's three stories sit on three different layers of the AI stack simultaneously: capability (Sol Ultra and 50-year math), governance (FLI safety grades and moving goalposts), and infrastructure (RISC-V chips and the Nvidia alternative).
- GPT-5.6 Sol Ultra — Claimed Proof of 50-Year Math Conjecture — 64 subagents, under one hour, July 10. OpenAI published both the proof PDF and the full 700-word prompt. Mathematician Thomas Bloom: 'very nice,' 'elementary,' missing citations. Not peer reviewed. The conjecture has attracted flawed proofs before. The prompt technique is the immediately useful part. Full analysis →
- FLI AI Safety Index — Anthropic C+, xAI Failing, All Labs Accused of Moving Goalposts — Anthropic ranked first but only C+. OpenAI and Google DeepMind each C. xAI fell from 4th to 7th with a failing grade. All four major labs accused of weakening commitments to pause development near danger thresholds. Also: Anthropic launched Claude Corps — paid 12-month nonprofit fellowship. Full story →
- Qualcomm in Early Talks to Acquire Tenstorrent for $8-10B — Jim Keller's RISC-V AI chip company. Talks unconfirmed. Would give Qualcomm its first competitive data center AI hardware position. The $8-10B is a bet on Keller's engineering team, not current revenue. Full story →
Story 1 — Sol Ultra and the 50-Year Math Conjecture: What to Make of It
OpenAI's Ethan Knight announced on July 10 that GPT-5.6 Sol Ultra — made generally available just the day before — had produced a proof of the Cycle Double Cover Conjecture, a graph theory problem that has sat unresolved since it was independently posed by George Szekeres in 1973 and Paul Seymour in 1979. The company published both the proof and the full prompt that generated it as PDFs on its CDN, attributing the mathematics entirely to the model.
The honest assessment has two parts. Mathematician Thomas Bloom called the proof 'very nice' and 'elementary,' noting it could have been discovered in the 1980s, and criticised it for lacking citations for foundational prior work. The conjecture has attracted multiple previous 'proofs' over the years that were later found to have gaps or were withdrawn. The proof has not undergone peer review. Treat the mathematical claim as pending until expert review completes over the coming weeks.
The prompt technique is useful now regardless of whether the math holds up. The 700-word prompt instructed Sol Ultra to deploy up to 64 concurrent subagents, managing them 'aggressively and dynamically,' with early rounds designed to maintain diversity across different mathematical formulations, algebraic angles, and structural inductions, and independent review mechanisms built in. HN commenters estimate the run cost between $275 and $13,000 — accessible to any well-funded team. The dynamic multi-agent orchestration architecture is immediately applicable to any complex structured problem, not just mathematics. Read the full Sol Ultra math proof analysis →
Story 2 — FLI Safety Grades: The Best Lab Gets a C+
Anthropic ranked first in the FLI's AI Safety Index but received only a C+ overall. OpenAI and Google DeepMind each received a C. Meta improved to fourth from sixth, while xAI fell to seventh from fourth. xAI, DeepSeek, and Mistral received failing overall grades. The most significant findings from the index: the reviewers said Anthropic, OpenAI, Google DeepMind, and Meta have weakened or eliminated earlier commitments to pause development if their systems approached specified danger thresholds — described as 'moving the goalposts.'
xAI's fall from 4th to 7th is the most significant movement. No system card for Grok 4.5, no published red-teaming framework comparable to Anthropic's RSP or OpenAI's Preparedness Framework, a $530M legal reserve disclosed in the IPO prospectus, and the adult content controversy all contribute. For enterprise procurement teams: a failing FLI grade is not a procurement barrier by itself, but it is a governance documentation flag. Meanwhile, Anthropic launched Claude Corps, a paid 12-month fellowship placing early-career professionals in nonprofit organisations to build AI capability — open to applicants aged 18 or older with less than two years of work experience, authorised to work in the US. Read the full FLI Safety Index story →
Story 3 — Qualcomm and Tenstorrent: The Nvidia Alternative Takes Shape
Qualcomm is in early talks to acquire Tenstorrent for between $8 and $10 billion. Tenstorrent designs AI chips using the open RISC-V standard and features chip veteran Jim Keller's engineering expertise. The acquisition would give Qualcomm real seats at the AI hardware table currently dominated by Nvidia and AMD. While talks are early and unconfirmed, the reported valuation reflects how valuable AI chip engineering talent has become in 2026.
The strategic picture: Qualcomm dominates edge AI through Snapdragon but has no competitive data center AI accelerator. Tenstorrent's RISC-V architecture avoids Arm and x86 licensing fees and is designed explicitly to compete with Nvidia's H-series GPUs. Jim Keller previously led AMD's Zen CPU revival, Apple's A4/A5 chips, and Tesla's FSD chip — the most credible chip architect attached to an Nvidia alternative in the current market. An $8-10 billion valuation for a pre-revenue-at-scale company is a bet on that engineering team and architecture, not on Tenstorrent's current cash flows. The talks, if they proceed to completion, would be the most significant AI chip M&A since Intel's $5.4 billion Habana Labs acquisition in 2019. Read the full Tenstorrent story →
The Week in Review — July 6-12, 2026
| Date | Event | Significance |
|---|
| Jul 6 | xAI rebrands as SpaceXAI, SPCX joins Nasdaq-100 | Confirms AI is now SpaceX's primary identity by TAM ($26.5T of $28.5T) |
| Jul 7 | LongCat-2.0 revealed as Owl Alpha, SWE-Together benchmark | Chinese open-source beats GPT-5.5 on SWE-bench Pro; Claude leads multi-turn |
| Jul 8 | JADEPUFFER confirmed, SK Hynix IPO, Fable 5 credits begin | First autonomous AI ransomware; $29.4B AI memory IPO |
| Jul 9 | GPT-5.6 Sol/Terra/Luna + Grok 4.5 public launch announced | Most competitive single-day AI launch since GPT-4; regulatory precedent set |
| Jul 10 | SK Hynix +13% Nasdaq debut, ChatGPT Work, Apple sues OpenAI | AI infrastructure IPO window confirmed open; co-operation era formally over |
| Jul 10 | Sol Ultra claims Cycle Double Cover proof (64 subagents) | First AI-claimed proof of 50-year open problem; peer review pending |
| Jul 12 | FLI Safety Index, Claude Corps, Qualcomm/Tenstorrent talks | Governance accountability, nonprofit AI expansion, RISC-V alternative emerging |
What to Watch Next Week
Cycle Double Cover peer review: Mathematicians are reading the proof now. The first credible expert assessment — either confirming the proof is correct or identifying a gap — will land within 1-2 weeks. This is the most watched mathematical question in AI right now.
China July 15 deadline: Three days from now. ByteDance Doubao and Alibaba Qwen have confirmed compliance with the anthropomorphic AI rules. Watch for third-party app removals and enforcement actions in Chinese app stores during the week of July 13.
Gemini 3.5 Pro GA: Still the only major unrestricted frontier model in the queue. Google targeted July 2026 general availability. No date confirmed. If it ships next week, it will reset the science benchmark table — Gemini 3.1 Pro already leads on GPQA Diamond at 94.3%.
Muse Spark 1.1 independent benchmarks: Meta's first paid model has no independent SWE-bench Pro score yet. Third-party evaluations in the next 1-2 weeks will determine whether $4.25/M output delivers quality that justifies switching from Claude Sonnet 5 or GPT-5.6 Terra.