AI Weekly Brief
Olympiad gold without frontier scale
SU-01, a 30B-A3B MoE, reaches olympiad gold through its training recipe, showing expert-level reasoning does not need frontier scale.
TL;DR
- OpenAI launched the Deployment Company ($4B+, 19 PE and consulting partners, plus the Tomoro acquisition) as a direct counter to the Anthropic, Blackstone, and Goldman services firm, opening an explicit race to embed forward-deployed AI engineers inside operating companies.
- OpenAI's Daybreak cybersecurity platform (GPT-5.5-Cyber plus Codex Security, eight security-industry partners) answered Anthropic's Project Glasswing, as the IMF warned that AI-enabled cyberattacks now pose a credible systemic risk to global financial stability.
- Anthropic passed OpenAI in verified business adoption for the first time (34.4% vs. 32.3% on Ramp data) and extended its enterprise reach with a $200M Gates Foundation partnership, SAP Business AI Platform embedding, and an expanded PwC alliance.
- SU-01, a 30B-A3B MoE trained with a curriculum SFT, two-stage RL, and test-time scaling recipe, reached gold-medal level on IMO 2025 and USAMO 2026, matching the top human total among 340 USAMO competitors.
- Attractor Models delivered a Pareto improvement over standard Transformers at constant training memory, with a 770M model matching a 1.3B Transformer trained on twice the tokens.
Editor’s Pick
The result matters less as another benchmark record than for what it reveals about the capability boundary. At roughly a tenth of the parameter count of the largest known frontier models, SU-01 scores 35 points on IMO 2025 and clears the USAMO 2026 gold line by 10 points, matching the best of 340 human competitors, using a three-stage pipeline: reverse-perplexity curriculum SFT to instill proof search, two levels of RL (outcome-verifiable, then proof-quality), and test-time verification-and-refinement loops.
The upstream implication is that the bottleneck for olympiad-level reasoning was training dynamics, not scale. The reverse-perplexity curriculum prioritizes the trajectories the model finds hardest, instilling proof-search and self-checking behavior before RL amplifies it. Moving from outcome-correctness RL to proof-quality RL, with a generative reward model scoring complete proofs and experience replay preserving rare successes, directly attacks the sparse-reward problem on hard problems. These are engineering decisions, not scale decisions.
The downstream implication is competitive: if gold-medal formal reasoning is achievable from a 30B-A3B MoE with a documented recipe, it becomes reproducible across labs within one to two training cycles. The paper's specializable-generalist framing, a broadly capable backbone specialized toward expert proof reasoning while preserving scientific transfer, applies to every domain where expert judgment can be operationalized as a verifiable reward: legal reasoning, software verification, clinical diagnosis, and financial modeling. The IMO result is the existence proof; the replication cascade is the strategic consequence.
Model Releases
IBM's Granite Embedding Multilingual R2. A sub-100M-parameter Apache-2.0 multilingual embedding model with a 32K context window that leads models under 100M parameters on standard multilingual retrieval benchmarks, aimed at cost-sensitive enterprise RAG.
AI Tools & Products
OpenAI launches the $4B Deployment Company. The majority-OpenAI-owned forward-deployed engineering firm is capitalized by 19 partners including Goldman Sachs, SoftBank, McKinsey, and Capgemini, and is acquiring Tomoro (~150 FDEs serving clients such as Tesco and Virgin Atlantic), mirroring Anthropic's services firm from a week earlier.
OpenAI's Daybreak cybersecurity platform. Daybreak combines GPT-5.5-Cyber and Codex Security into a continuous vulnerability detection and remediation pipeline across three access tiers, with partners including Cisco, CrowdStrike, Palo Alto Networks, Cloudflare, and Zscaler, positioned against Anthropic's Project Glasswing.
SAP embeds Claude in its Business AI Platform. Claude becomes the primary reasoning and agent engine of SAP's new platform, connecting to S/4HANA, SuccessFactors, and Ariba via MCP, with initial agents for quarter-end close, supplier rerouting, and employee leave management.
OpenAI brings Codex to mobile with Hooks. Codex, now past 4M weekly users, reached the ChatGPT iOS and Android apps in preview for steering agent tasks remotely, and added project-level Hooks that intercept prompts and responses for PII screening or compliance logging, Remote SSH, and HIPAA compliance.
ThurrottSiliconANGLEPreviouslyApr 25NextMay 22
Anthropic's Claude for Small Business. A packaged bundle of connectors and ready-to-run workflows that embeds Claude in tools small businesses already use, with lower setup overhead than enterprise plans, forming the downmarket tier below Anthropic's vertical products.
Google DeepMind's AI Pointer. A research-stage, Gemini-powered mouse pointer infers intent from what the user points at, shifting context disambiguation from user to system, with first integrations in Chrome's point-to-ask and Googlebook's Magic Pointer.
Research Breakthroughs
SU-01 reaches olympiad gold at 30B-A3B. A 30B-A3B MoE trained with reverse-perplexity curriculum SFT, coarse RLVR then proof-level RL, and test-time verification reaches gold on IMO 2025, USAMO 2026, and IPhO 2024–25, scoring 35/42 and matching the top human among 340 USAMO competitors.
USC's Attractor Models. A Transformer backbone proposes an output that a small weight-tied module refines to a fixed point at constant training memory; a 770M model matches a 1.3B Transformer trained on 2x tokens, and a 27M model solves 91.4% of Sudoku-Extreme where frontier LLMs score 0%.
University of Washington's Dolores agent. Dolores builds task-specific reasoning scaffolds just-in-time from human meta-reasoning traces, improving 24.8% on average over the strongest scaffold baseline across four hard benchmarks, with an 8B agent beating all 32B baselines in over half of settings.
Agents discover new neural architectures. AIRA-Compose's 11-agent ensemble searched Attention, MLP, and Mamba primitives on a 24-hour budget and found 14 novel architectures; at 1B scale AIRAformer-C scales 54% faster than Llama 3.2 and AIRAhybrid-D gains 3.8% in downstream accuracy.
Industry & Ecosystem
Anthropic passes OpenAI in business adoption. Ramp expense data shows 34.4% of businesses paying for Anthropic versus 32.3% for OpenAI, after Anthropic's share grew 26 points in a year from a 9% base; OpenRouter data last ranked OpenAI ahead in December 2025.
Anthropic and Gates Foundation commit $200M. A four-year commitment of grants, Claude credits, and engineering support for global health, education, and economic mobility, spanning vaccine screening, malaria and tuberculosis forecasting, and K-12 tools in sub-Saharan Africa and India, with benchmarks and datasets released as public goods.
PwC expands its Anthropic alliance. PwC is rolling Claude Code and Claude Cowork out to 30,000 professionals, expanding to its global workforce, after production deployments cut delivery times by up to 70%, and launched a Claude-anchored Office of the CFO unit.
EU seeks cyber model access from OpenAI, Anthropic. The European Commission confirmed talks on supervised access to GPT-5.5-Cyber, with Claude Mythos talks at an earlier stage, after the IMF warned that AI-powered cyberattacks pose a credible systemic risk to global financial stability.
Trends & Predictions
Frontier labs build services capacity to deploy into enterprises
Enterprise adoptionStrengthenedContinues fromMay 8
OpenAI's $4B DeployCo, backed by 19 partners including Goldman Sachs, SoftBank, McKinsey, and Capgemini and buying Tomoro's ~150 FDEs, mirrors Anthropic's services firm, while Ramp data put Anthropic ahead in business adoption (34.4% vs 32.3%) as SAP and PwC deepened Claude rollouts.
- Prediction (Nov 2026): At least one more frontier lab, most likely Cohere or Mistral, announces a structurally comparable forward-deployment vehicle as implementation capacity overtakes model capability as the binding enterprise constraint.
- Resolves: Yes if by November 15, 2026 a frontier lab other than OpenAI or Anthropic (such as Cohere, Mistral, Google DeepMind, or xAI) announces a separately capitalized forward-deployed engineering or services company with outside investors, per its press release.
Based on OpenAI launches the $4B Deployment CompanyAnthropic passes OpenAI in business adoptionSAP embeds Claude in its Business AI PlatformPwC expands its Anthropic alliance
OpenAI built Daybreak on GPT-5.5-Cyber and Codex Security with partners including Cisco, CrowdStrike, Palo Alto Networks, Cloudflare, and Zscaler, and the European Commission opened supervised-access talks with OpenAI and Anthropic after the IMF called AI cyberattacks a systemic financial risk.
- Prediction (Nov 2026): At least one G7 government proposes or enacts legislation requiring frontier AI cybersecurity tools to be licensed or registered with national-security oversight bodies, analogous to dual-use export controls.
- Resolves: Yes if by November 15, 2026 a G7 government introduces or enacts legislation or binding regulation requiring frontier AI cybersecurity models or tools to be licensed by or registered with a national-security body, as shown in its legislative record or official gazette.
Based on OpenAI's Daybreak cybersecurity platformEU seeks cyber model access from OpenAI, Anthropic
AI starts producing verified scientific and mathematical results
AI in scientific discoveryStrengthenedContinues fromMay 8
SU-01, a 30B-A3B MoE trained with curriculum SFT, staged RL, and test-time verification, reached gold on IMO 2025, USAMO 2026, and IPhO 2024–25, scoring 35/42 and matching the top human among 340 USAMO competitors.
- Prediction (Nov 2026): A compact model of 50B parameters or fewer reaches IMO and USAMO gold in a direct pass without test-time refinement, using increasingly standardized training recipes.
- Resolves: Yes if by November 15, 2026 a paper or model card reports a model with 50B total parameters or fewer reaching gold-medal scores on both IMO and USAMO problems in a single pass without test-time verification or refinement, as graded on MathArena or by the competition rubric.
USC's Attractor Models let a 770M model match a 1.3B Transformer trained on 2x tokens and solve 91.4% of Sudoku-Extreme where frontier LLMs score 0%, while AIRA-Compose agents found 14 architectures, including AIRAformer-C, which scales 54% faster than Llama 3.2.
- Prediction (Nov 2026): At least one frontier lab incorporates an implicit refinement mechanism (fixed-point, attractor, or convergence-based) into a production model architecture.
- Resolves: Yes if by November 15, 2026 OpenAI, Google DeepMind, Anthropic, Meta, xAI, DeepSeek, Alibaba Qwen, or Mistral documents a fixed-point, attractor, or other convergence-based refinement module in a production model's technical report or model card.
Based on USC's Attractor ModelsAgents discover new neural architectures