The Memo - 30/Oct/2026

1. Key Themes

Theme: AI is moving from assistant to autonomous scientific discoverer

AI labs are reporting concrete biomedical results, not just demos

Anthropic reported early drug-development results across multiple partners, with measurable outperformance of standard approaches.

"Claude Code designed a CAR T-cell component ~70% shorter than standard. In lab tests it produced ~10x more receptors and killed tumor cells more efficiently." "Claude generated ~35,000 antibody designs in a month for a novel food allergy approach. Three of those outperformed the standard medication."

AI agents are now executing physical experiments end to end

An agent chose the therapy, the target, and the coordinates, then operated the surgical hardware itself.

"An AI agent using Claude Sonnet 5 autonomously performed gene therapy brain surgery on a mouse, choosing the drug, the brain region, and the coordinates, then driving the stereotaxic arms itself." "It worked out exactly where to inject using three research papers, then injected a tiny dose while watching through the camera for bleeding, checking the camera 32 times during the operation."

Frontier math and theory results are being produced by internal models and formalized in Lean

OpenAI claims it overturned a ~30-year-old widely believed hypothesis and advanced Millennium Prize-adjacent problems.

"The Exponential Time Hypothesis was proposed ~30 years ago, was widely believed true, and has been used in hundreds of papers to derive hardness results." "As a subproblem, this includes a proof of the existence of pure Yang–Mills theory, part of the Clay Millennium Prize problem." "OpenAI says it does not seek academic credit for the discoveries, and it is funding a conference on the Exponential Time Hypothesis result."

Theme: AI is becoming a geopolitical, export-controlled strategic asset

China is building a licensing regime that could eventually gate open-weight releases

The draft rules set parameter-count, compute, and ranking thresholds and require pre-export registration.

"China announced plans to restrict or ban exports of 40 technologies spanning AI, chips, and solar power as its tech race with the US intensifies. 23 AI technologies are on the list... with exports of models having parameter counts of 2 trillion or above potentially restricted." "China's draft would require licences to export the weights, code, and know-how of any model over 2T parameters, 10²⁶ FLOPs, or in the global top 10 (open or closed), ban exports of models trained on 'core data' outright, and already requires pre-export registration, which gives Beijing a way to ban future open-weight releases without banning them today."

Technology access is fragmenting along national lines

"'The world is getting more fragmented in trade and investment, but also in access to technologies,' noted Gary Ng, senior economist at Natixis, as both superpowers increasingly treat AI as a strategic asset."

Theme: The author frames Oct/2026 to Mar/2027 as a compressed transition window

A tracked checklist of ASI indicators underpins the thesis

"I still think October 2026 through to March 2027 will be viewed as the six-month envelope that delivered a complete evolution for humanity." "While none of the following indicators made the evening news, all are tracked on 'Alan's ASI indicators'."

The header also shows the author's own scoreboard: "AGI: 99%" and "ASI: 2/50".

Theme: Embodied AI and robotics endurance is advancing

Humanoid robots are demonstrating sustained physical operation

(The body is truncated here, so only the headline is available.)

"Figure 03 robot covers 560km (348mi) in nine-day treadmill endurance test (20/Oct/2026)"

2. Contrarian Perspectives

AI backlash reflects human insecurity, not AI failure

Fields Medalist Richard Borcherds argues the anger at AI is misattributed. The author calls it "a spectacular and confronting quote."

"'I think what people are getting upset about is they're slowly realizing that they are dumb and somehow blaming this on AI, even though it's not really AI's fault.'"

Scientific norms are being bypassed by AI labs, and experts are objecting

The newsletter highlights the friction between AI lab releases and working mathematicians. Terence Tao criticized OpenAI's release of about 719 AI-generated maths papers.

"'there was no consultation with people who are already working on these problems… it was extremely rude and discourteous.'"

Counterpoint from the article: OpenAI "does not seek academic credit" and is funding a conference, with the original proposers of ETH helping to shepherd the disproof.

The biggest AI milestones are underreported

The author argues that major ASI indicators are not making mainstream news, suggesting a gap between real progress and public awareness.

"While none of the following indicators made the evening news, all are tracked on 'Alan's ASI indicators'."

3. Companies Identified

Anthropic

• Description: AI lab behind Claude.
• Why mentioned: Reported drug-discovery results with six organizations and acquired Mithril Medicine.
• Quotes: "Anthropic reported early results from drug development partnerships with six organizations, plus its own Claude-run programs for rare childhood muscular dystrophy." / "Mithril Medicine: Anthropic acquired this one-person company."

OpenAI

• Description: AI lab.
• Why mentioned: Announced three internal-model discoveries formalized in Lean, and published about 719 AI-generated maths papers.
• Quotes: "OpenAI announced three new discoveries by its internal models, all formalized in Lean."

Generate Biomedicines

• Description: Biotech partner of Anthropic.
• Why mentioned: Claude Code designed a shorter CAR T-cell component with improved performance.
• Quotes: "Claude Code designed a CAR T-cell component ~70% shorter than standard."

Divergence Labs

• Description: Biotech partner of Anthropic.
• Why mentioned: Claude generated antibody designs for a food allergy approach.
• Quotes: "Three of those outperformed the standard medication."

Mithril Medicine

• Description: One-person biotech, acquired by Anthropic.
• Why mentioned: Case study of AI-delegated science and operations.
• Quotes: "Its founder delegated much of the science and operations to Claude and took a muscular dystrophy treatment from idea to promising mouse results in under 18 months."

Exin Therapeutics

• Description: Y Combinator-backed therapeutics company co-founded by Dr Gabriel Ocana-Santero.
• Why mentioned: Its CEO ran the autonomous Claude mouse-surgery experiment.
• Quotes: "Dr Gabriel Ocana-Santero (Oxford, co-founder and CEO of Y Combinator-backed Exin Therapeutics)."

Moonshot AI (Kimi K3.1 Pro) (company inferred from model name; the article names only the model)

• Description: Maker of the 2.8T-parameter Kimi model.
• Why mentioned: Released the same week as China's export-curb news, serving as a test case.
• Quotes: "Kimi K3.1 Pro 2.8T was released this week without issue."

Figure (Figure 03)

• Description: Humanoid robotics company.
• Why mentioned: Nine-day treadmill endurance test.
• Quotes: "Figure 03 robot covers 560km (348mi) in nine-day treadmill endurance test."

Natixis

• Description: Financial institution.
• Why mentioned: Its senior economist commented on tech fragmentation.
• Quotes: "'The world is getting more fragmented in trade and investment, but also in access to technologies,' noted Gary Ng, senior economist at Natixis."

Also named in the contents (details paywalled): Meta (Prometheus), Genie 4, CERN, Mistral, DeepSeek, Sanxingdui: Future Past AI.

4. People Identified

Prof Richard Borcherds

• Description: Fields Medalist.
• Why mentioned: Provocative quote on AI backlash.
• Quotes: "'I think what people are getting upset about is they're slowly realizing that they are dumb and somehow blaming this on AI.'"

Terry Tao

• Description: Mathematician.
• Why mentioned: Criticized OpenAI's release of AI-generated papers.
• Quotes: "'it was extremely rude and discourteous.'"

Dr Gabriel Ocana-Santero

• Description: Oxford researcher, co-founder and CEO of Exin Therapeutics.
• Why mentioned: Ran the autonomous AI mouse-surgery experiment.
• Quotes: "gave the agent 19 gene therapy vectors in a spreadsheet, a motorized surgery rig (custom-built on Arduino) with cameras, and one goal: reduce seizures."

Prof Russell Impagliazzo and Prof Ramamohan Paturi (UC San Diego)

• Description: Original proposers of the Exponential Time Hypothesis.
• Why mentioned: Will shepherd the disproof.
• Quotes: "Its original proposers, Prof Russell Impagliazzo and Prof Ramamohan Paturi (UC San Diego), will shepherd the disproof with Prof Ryan Williams (MIT)."

Prof Ryan Williams (MIT)

• Description: Complexity theorist.
• Why mentioned: Co-shepherding the ETH disproof.
• Quotes: "will shepherd the disproof with Prof Ryan Williams (MIT)."

Gary Ng

• Description: Senior economist at Natixis.
• Why mentioned: Commentary on tech fragmentation.
• Quotes: "'The world is getting more fragmented in trade and investment, but also in access to technologies.'"

Dr Alan D. Thompson

• Description: Author of The Memo (LifeArchitect.ai).
• Why mentioned: Source of the ASI indicator framework and the analysis.
• Quotes: "Across over 100 editions, The Memo continues to be the #1 AI advisory, informing 10,000+ full subscribers."

5. Operating Insights

Delegating science and operations to AI enables one-person biotech

Mithril's founder compressed idea-to-mouse-results into under 18 months by leaning on Claude.

"Its founder delegated much of the science and operations to Claude and took a muscular dystrophy treatment from idea to promising mouse results in under 18 months."

Give agents a bounded goal, structured options, and instrumentation

The mouse-surgery setup gave the agent a concrete objective, a constrained menu, and sensors, and let it verify its own work.

"gave the agent 19 gene therapy vectors in a spreadsheet, a motorized surgery rig (custom-built on Arduino) with cameras, and one goal: reduce seizures." "...checking the camera 32 times during the operation."

Generate at volume, then filter, in discovery pipelines

High-throughput AI design followed by empirical selection produced outperformers.

"Claude generated ~35,000 antibody designs in a month... Three of those outperformed the standard medication."

6. Overlooked Insights

Chinese export rules use compute and ranking thresholds, not just parameter count

The draft covers models above 10²⁶ FLOPs or in the global top 10, which could affect any leading Chinese model regardless of size.

"any model over 2T parameters, 10²⁶ FLOPs, or in the global top 10 (open or closed)."

OpenAI's discoveries are being formalized in Lean

Formal verification suggests an attempt to make AI-generated results machine-checkable, which addresses trust and review-burden concerns.

"OpenAI announced three new discoveries by its internal models, all formalized in Lean."

Note: the full edition is paywalled, so sections on Kimi K3.1 Pro, Meta Prometheus, Genie 4, CERN/Mistral, policy items, and DeepSeek pricing could not be summarized.