AI Weekly Brief
Claude agents discover ART
About 950 Claude agents discovered ART, a new CRISPR-like enzyme system, on their own, showing models can drive biological discovery.
TL;DR
- Three frontier labs and an open-weight challenger shipped inside 48 hours and competed on price as much as capability: Anthropic's Claude Opus 5.5, OpenAI's GPT-6 Sol and Luna, xAI's long-delayed Grok 4.7, and Xiaomi's open MiMo-V2.6.
- Xiaomi's MiMo-V2.6 (1.02T-parameter MoE, MIT) topped the Artificial Analysis open-weight Intelligence Index and open-sourced its full RL stack and 7,000+ training environments, tying xAI's same-day Grok 4.7.
- Anthropic's new life-sciences lab reported Claude autonomously discovering a novel, CRISPR-like enzyme system (ART) from genomic data, a concrete "AI as working scientist" result endorsed by CRISPR pioneer Feng Zhang.
- Physical-AI and enterprise-agent capital kept flowing: Cognex agreed to buy RealSense (~$500M) while Island ($6.4B), TEKEVER ($6.4B), and Snorkel AI ($3.5B) closed large rounds.
- Externally verifiable governance advanced: Opus 5.5 was Anthropic's first release since its "pace the frontier" call (pre-evaluated by METR and Frontier Design), and Google published a tamper-proof record and independent audit for its new Private AI Compute memory.
Editor’s Pick
Given only a high-level prompt to mine a genomic database, roughly 950 Claude agents combed over 200,000 reverse transcriptases across 21 hours and 210 million tokens, narrowed the field to 20 candidates, and flagged a reverse transcriptase flanked by a CRISPR-like repeat array that no prior study had characterized, a system Anthropic now calls array-associated reverse transcriptases (ART). The technical significance is that the agents reproduced established results to validate their own methods, generated and critically pruned their hypotheses, and surfaced a system whose combination of features has only ever appeared in a handful of programmable, biotechnologically useful enzymes, with early wet-lab work confirming the array is expressed as distinct short RNAs. CRISPR pioneer Feng Zhang called the RNA-repeat-array finding "genuinely intriguing" and worth further investigation, and Anthropic released a preprint. The downstream implication is a discovery pipeline in which hypothesis generation scales to thousands of candidates per campaign, shifting the human bottleneck to wet-lab validation and sharpening the governance question raised by the same company's Life Sciences Verification Program: how to keep autonomous biological discovery both open and safe.
Model Releases
Anthropic's Claude Opus 5.5. The first Claude 5.5 model performs at Claude Fable 5.1's level on most work while costing 40% less to run than Opus 5, and posts Anthropic's strongest automated behavioral-audit alignment scores to date.
OpenAI's GPT-6 Sol and Luna. Two cost-efficient GPT-6 tiers that carry Astra-era gains to everyday work, with API prices cut 50% versus their GPT-5.6 counterparts (Sol $2/$10, Luna $0.10/$0.50 per 1M tokens).
xAI's Grok 4.7. The repeatedly delayed ~2.1T-parameter model shipped at Grok 4.6 pricing ($2/$6 per 1M tokens, 500K context), which xAI ranks third behind Anthropic and OpenAI for agentic coding.
Xiaomi's MiMo-V2.6. A 1.02T-parameter (42B active) omni-modal MoE released under MIT that tops the Artificial Analysis open-weight Intelligence Index (46, tying Grok 4.7) and ships with 7,000+ open RL environments and a public six-day training run.
AI Tools & Products
Google's Gemini 3.8 Live with avatars. Adds near-real-time video avatars to Gemini's live-dialogue models, with lip-synced expressions across 97 languages and asynchronous tool calls that run mid-conversation, available in Gemini Enterprise.
Google adds memory to Private AI Compute. Extends Google's confidential-computing platform with persistent, cross-device memory sealed in hardware enclaves and unlockable only by device-held keys, published with a tamper-proof software record and an independent security audit.
Research Breakthroughs
Claude agents discover a CRISPR-like enzyme. In Anthropic's new life-sciences lab, roughly 950 Claude agents autonomously mined genomic data (210M tokens, 21 hours) to identify "array-associated reverse transcriptases" (ART), a previously uncharacterized and potentially programmable enzyme system, released as a preprint.
Models share concept directions, not positions. Shows independently trained models share the directions between concepts rather than their absolute positions, enabling "Shadow Casting": importing one model's capabilities into another via a single cached forward pass at a fraction of fine-tuning compute.
WROP teaches world models object permanence. Introduces WROP, a cognitive-science-inspired benchmark and 1.5M-sample corpus for object permanence, and trains a 16B video world model (PWM-WROP) that ranks first among continuation models in a blind pairwise Elo study.
Workflows & Techniques
Growing Harness evolves agent scaffolds from failures. A failure-guided paradigm that grows a reusable, executable agent harness from a strategy-free scaffold, cutting deployed LLM calls 76-92% while keeping 4B-120B models effective across browsing and web-agent benchmarks.
Pistis pairs IDRL with auto-harnessing. A 9B/27B multimodal model family trained with Interleaved Distillation and Reinforcement Learning, paired with a system-level method that automatically improves the inference harness without updating model weights.
Industry & Ecosystem
Cognex to acquire RealSense for $500M. The machine-vision leader agreed to buy the Intel-spun-out depth-sensing company for roughly $500M in cash, entering the robotic-perception and "Physical AI" market.
Island raises $400M at $6.4B. The enterprise-security company reached a $6.4B valuation in an Evolution Equity Partners-led round as it positions its secure browser as an agentic control plane for enterprises scaling AI agents.
TEKEVER raises $580M at $6.4B. Europe's AI-powered autonomous-systems provider reached a $6.4B valuation in a UC Investments- and Baillie Gifford-led first close, marking UC Investments' first direct European investment.
Snorkel AI raises $350M at $3.5B. The data-for-AI company reached a $3.5B valuation (Insight Partners and S32 co-led) on demand for expert human data and recurring model evaluation and adaptation.
Progress closes $400M Domo asset deal. Progress Software closed its $400M purchase of substantially all assets of Domo's agentic-AI and data-products business, broadening its enterprise data-readiness stack.
GlobeNewswirePreviouslyAug 14
AI Startups
Moxing raises $150M for 4D world models. The Hangzhou startup building low-cost "4D" world-model infrastructure runs its MoWorld 4D system on Huawei Ascend chips, claiming roughly 70% lower inference cost.
Subconscious raises $5.1M for agent inference. The MassVentures-led round funds an inference platform purpose-built for long-running agents, available to developers today with an on-prem enterprise package.
Ando raises $20M for human-agent messaging. Backed by Accel, Index Ventures, and Emergence Capital, Ando is building a messaging platform where human teams and AI agents work side by side and accumulate shared context over time.
Healthcare AI
Dynamic prompting improves cancer outcome prediction. A retrieval-based dynamic few-shot prompting framework over unstructured EHR notes improved LLM outcome prediction most for a lower-incidence cancer (glioma) while maintaining performance for breast cancer.
Lessons from scaling retinopathy AI to 1M patients. A commentary distills lessons from scaling AI diabetic-retinopathy screening from a single hospital to more than one million patients, covering deployment, monitoring, and equity.
Enveda raises $311M at $2B. A Catalio-led round doubled the valuation of a startup that uses AI to discover nature-derived drug candidates, funding more programs into clinical trials.
BigHat Biosciences raises $75M for protein design. A DFJ Growth- and Premji Invest-co-led round advances a clinical-stage platform pairing machine learning with an agentic wet-lab data-generation loop for rapid protein design.
Trends & Predictions
Every major release this week led on cost: Opus 5.5 matches Fable 5.1 at 40% lower run cost, GPT-6 Sol and Luna cut API prices 50%, and Grok 4.7 and MiMo-V2.6 held prior-generation prices.
- Prediction (Mar 2027): Price-per-task at a fixed capability, and "default effort" mid-tiers, become the primary procurement axis, with headline benchmark deltas mattering less than cost-at-quality.
- Resolves: Yes if by March 25, 2027 OpenAI, Anthropic, and Google have each released a new flagship whose list API price is no higher than its predecessor's.
Based on Anthropic's Claude Opus 5.5OpenAI's GPT-6 Sol and LunaxAI's Grok 4.7Xiaomi's MiMo-V2.6
Chinese labs take the lead in open-weight models
Open-weight modelsStrengthenedRevivesAug 21
Xiaomi's MiMo-V2.6 topped the open-weight intelligence index under MIT and shipped with 7,000+ RL environments and a public six-day training run.
- Prediction (Mar 2027): Reproducible RL infrastructure becomes the open-weight battleground, with DeepSeek V4.1-Pro and Qwen 4 the next tests of the pattern.
- Resolves: Yes if by March 25, 2027 DeepSeek V4.1-Pro or Qwen 4 is released together with its RL environments or training framework.
Based on Xiaomi's MiMo-V2.6
AI starts producing verified scientific and mathematical results
AI in scientific discoveryStrengthenedRevivesSep 4
About 950 Claude agents in Anthropic's life-sciences lab found a previously uncharacterized CRISPR-like enzyme system (ART) in 21 hours, and BigHat raised $75M for an agentic wet-lab design loop.
- Prediction (Mar 2027): Labs increasingly pair models with wet-lab validation, and large-scale hypothesis generation becomes a productized discovery workflow.
- Resolves: Yes if by March 25, 2027 a model-generated biological discovery validated in a wet lab is published in a peer-reviewed journal.
Based on Claude agents discover a CRISPR-like enzymeBigHat Biosciences raises $75M for protein design
Agents improve themselves through skills, harnesses, and memory
Agent training and memoryStrengthenedContinues fromSep 18
Growing Harness evolves an executable harness from failures and cuts deployed LLM calls 76-92%, and Pistis tunes its inference harness automatically without touching weights.
- Prediction (Mar 2027): Harness and inference-efficiency engineering become a primary cost-performance lever, letting smaller deployment models match larger ones.
- Resolves: Yes if by March 25, 2027 a model under 30B parameters with an optimized harness comes within two points of a frontier model on SWE-bench Verified or Terminal-Bench.
Based on Growing Harness evolves agent scaffolds from failuresPistis pairs IDRL with auto-harnessing
Frontier launches come under outside review and independent verification
Frontier safety governanceStrengthenedRevivesAug 28
Opus 5.5 shipped after pre-release evaluation by METR and Frontier Design, and Google published a tamper-proof software record and an independent audit for Private AI Compute's new memory.
- Prediction (Mar 2027): Third-party evaluation and cryptographic attestation shift from voluntary to expected, converging with emerging state independent-verification regimes.
- Resolves: Yes if by March 25, 2027 OpenAI, Google DeepMind, and Anthropic each name a third-party pre-deployment evaluator for their next flagship release.
Based on Anthropic's Claude Opus 5.5Google adds memory to Private AI Compute
In this brief
- Products
- Benchmarks