Approaching the AI Event Horizon? Part 2, w/ Abhi Mahajan, Helen Toner, Jeremie Harris, @8teAPi
Summary
The clearest near-term AI-biology business is not inventing more molecules but rescuing value at the clinical bottleneck: Abhi Mahajan says 97% of oncology trials fail even though some patients often respond. Noetik combines pathology, 16-plex spatial proteomics, a 19,000-gene spatial transcriptome, and exome sequencing to find potentially “non-human-legible” response biomarkers. Mahajan expects human-simulation companies to improve at least a few trials within several years, while remaining much less certain that AI will rapidly discover wholly new targets.
Biology is unlikely to repeat software’s intelligence explosion because its most valuable rewards are slow, ambiguous, and expensive rather than cheaply verifiable. Toxicity can emerge in seconds or years, vary by dose and species, or cause cognitive and cardiac damage without killing an animal; a clean hepatocyte assay may save months while still missing the decisive in-vivo question. Even a model doubling “Alpha 3” on difficult preclinical benchmarks does not prove better patient outcomes: “The field is already awash with many really good preclinical assets.”
China’s biotechnology advantage looks more like an operating-system advantage than a demonstrated AI-model lead. Mahajan traces it from generics through a strong CRO ecosystem into indigenous drug development, with lower trial costs and tighter feedback between designers and wet-lab workers; he has not yet seen a “DeepSeek thing” in Chinese bio-AI. Jeremie Harris separately argues that chip export controls are visibly binding, citing DeepSeek’s pre-R1 complaints and pent-up H200 demand, while rejecting the idea that Nvidia sales would make China abandon its strategic domestic stack.
Automated AI R&D is a strategic-surprise machine because informed experts agree about the near term while disagreeing on whether it ends in recursion, jagged acceleration, or a plateau. Helen Toner’s workshop participants agreed substantially about what they might see in 2026–2027 but not what comes next; the decisive questions are whether AI replaces every human contribution and how quickly physical, organizational, and adoption bottlenecks “bite.” An underexplored case is a superhuman but bounded plateau: transformative enough to reorder the economy without becoming an incomprehensible singularity.
The frontier-lab race is outrunning both evaluation and governance, yet inevitability does not erase meaningful design choices. The discussion pointed to Anthropic and OpenAI releasing models despite acknowledged difficulty evaluating awareness or long-horizon autonomy, while distinguishing AI-led research under human direction from setting millions of agents in motion with “no clue what’s going on.” Toner’s policy prescription shifts from release-day paperwork toward continuous internal-risk measurement, independent audits, and societal hardening across cyber, bio, and epistemic security.
The overlooked AI trade is the physical substrate—and Harris thinks compromising it could nullify every model-level advantage. TSMC is an exceptionally fragile concentration point; China-linked components and personnel create “one-way doors” in data-center construction; and grid transformers may offer an adversary leverage far beyond model theft. Harris treats secure builds, supply-chain scrutiny, grid redundancy, and credible offensive options as purchases of optionality, while assigning loss of control only a deliberately broad “10 to 90%” range and treating 2027, 2030, but less so 2035 as plausible timelines.
Adoption may look discontinuous even when capability curves look smooth because a final 1% improvement can turn a toy into a production system. Harris says podcast clipping failed six months earlier but worked end to end three or four weeks before the show; Nathan Labenz sees a similar threshold in agents, though current models still overuse code when they should “just read the document.” The emerging stack—deep personal context, 300,000-to-10,000-token monthly compression, possible continual learning, and models aimed at “judgment transfer”—was discussed as a way to preserve more individual economic leverage.
Deep dive
1. Drug scouting is pretty good, even if its evaluation remains “pretty hacky”
Mahajan’s competitive-intelligence pipeline starts from pharma’s growing willingness to acquire assets rather than develop everything internally. Chinese companies are producing interesting preclinical candidates that may be purchasable for a few million dollars; instead of relying on personal networks or trial aggregators, his system scrapes the semantic web, annotates every investigational drug against company priorities and preferred modalities, ranks them, and hands the resulting table to the therapeutics team.
Human diligence remains essential because “even 5.2 and 5.3 aren’t perfect.” Mahajan’s personal eval is deliberately pragmatic: among drugs the therapeutics team already finds compelling and wants to advance, does the next model generation continue to recommend them? The pipeline was built a few months earlier and remained pretty good, but he had not observed a dramatic model-driven jump.
The ceiling may reflect the target itself: deciding whether a drug is attractive is “very qualitative” and “very vibe based.” It depends on the buyer’s economics and whether anyone knows the seller—important because companies can make it surprisingly difficult to give them money. A benchmark can test scientific attributes; it cannot fully encode whether a transaction is actually executable.
2. Biology’s valuable rewards arrive too late and noisily for easy recursion
Mahajan softened his claim that biology has “no verifiable ground truth” as hyperbole, then defended the narrower version: there is little cheap, verifiable truth for the most clinically valuable problems. Protein presence and sequencing calls can be checked, but biology lacks the abundant, immediate rewards that allowed reinforcement learning with verifiable rewards to compound rapidly in mathematics and code.
His analogy was training an RLVR model to write the bestselling book. Sales technically provide a reward, but the result may take 18 months, and attribution across the author’s country and countless other variables is nearly impossible. Biology similarly delivers an occasional outcome after a long iterative process without revealing which decision deserves credit.
Toxicology carries the point. A compound may be lethal within seconds like snake venom, harmful only after months or years, nonlethal but cognitively or cardiologically damaging, dose dependent, or species dependent. “There’s no real way to understand” the full profile without observing an organism in vivo and examining the readouts.
Mahajan credited Axiom’s attempt to predict a small molecule’s effect on hepatocytes in a dish as a clean problem that could save months of preclinical work. His reservation was scope: solving cellular toxicity does not answer “the much more important problem” of performance in an animal, much less a patient.
3. Better preclinical design has not yet proved that patients fare better
Nathan Labenz raised an Isomorphic Labs model said to double “Alpha 3” on binding affinity, pocket identification, and structure prediction. Mahajan called the benchmark “incredibly difficult” and the result an impressive piece of work, but separated benchmark progress from clinical value: the field already has many strong preclinical assets, while the bottleneck is how they perform in patients.
Labenz noted that the intuitive claim that improving every preclinical design step must improve human outcomes has been made for roughly 10 years without clear evidence that it has borne fruit. Mahajan said he expects it may eventually, but did not convert that expectation into proof that today’s gains will survive the clinical pipeline.
Labenz presented “The Affinity Advantage” as perhaps the strongest bull case: optimizing every facet of proteins entering the pipeline could have nonlinear or superlinear benefits as models improve. He said it was not an opinion he shared, but that he was sympathetic to it—a distinction between recognizing a coherent mechanism and assigning it high probability.
4. Noetik is betting that tumors should be modeled before biology is simplified
Mahajan’s preferred sensing breakthrough is a generative model of human in-vivo biology trained on rich data from actual tumors, intestinal lesions, plasma readouts, and other realistic sources—not primarily in-vitro approximations. The ambition is a “genuine, bona fide human simulator of biology” that might address a world where 97% of oncology trials fail.
His pushback on “scientists in a data center” was not that ideas are scarce: tens of thousands of PhD students already produce good ones. The constraint is that most cannot be validated because experiments are too expensive. More hypothesis generation without a cheaper, more predictive evaluation layer risks worsening the existing imbalance.
Noetik’s economic thesis begins with failed trials that nevertheless contain responders. Post-trial papers often search for their biological archetype and produce complicated, heterogeneous signatures—perhaps a cytokine group or elevated granzyme genes—that rarely become something actionable. Mahajan’s inference is that the true response biomarker may be “non-human-legible” and therefore require a black-box representation.
That turns failure data into an asset: collect patient samples from completed trials, profile them deeply, embed them, and ask whether responders occupy a distinct region from nonresponders. If they do, Noetik may possess a biomarker that no person can explicitly explain but that can still stratify the next trial.
5. Four modalities turn each tumor into both an embedding and a simulator
Noetik profiles tumors at four levels: conventional pathology; 16-plex spatial proteomics for cell types; whole spatial transcriptomics covering 19,000 genes across the tumor surface to characterize functional state; and exome sequencing for alterations such as KRAS status or knockouts. A self-supervised masked model must reconstruct missing information across that heterogeneous stack.
The first goal is a robust representation: when a new tumor arrives, the model should place it within “the universe of all the cancers I’ve seen.” Trial samples can then be mapped into that space, allowing response and nonresponse populations to separate without requiring researchers to specify the decisive biomarker in advance.
The more ambitious use is generative intervention. Researchers can computationally knock out a transcript or gene and predict how expression changes throughout the tumor microenvironment, moving from passive stratification toward counterfactual biology—provided those counterfactuals ultimately survive experimental validation.
Mahajan connected this to “nudge drugs”: therapies that do not primarily attack the tumor or immune system but push the microenvironment into a state more sensitive to another treatment. The model might simulate a transcript knockout followed by a PD1-axis checkpoint blocker and predict whether the tumor becomes inflamed, “hot,” and more likely to “melt away entirely.”
6. Black-box biomarkers may be commercially sufficient before they are intelligible
Asked whether Noetik could explain why patient groups separate in embedding space, Mahajan said an internal mechanistic-interpretability group was exploring the question and would likely find something interesting. His contrarian question was more economic: “Why do we care about interpretability?” If the intended product is clinical prediction, explanation may not be the binding regulatory requirement.
His example was ArteraAI, which around August or September 2025 had a prostate-pathology system predicting response to androgen-deprivation therapy. According to Mahajan, the company did not know why its model worked, but had retrospectively validated it across thousands of patients from prior Phase 3 trials, and “the FDA was fine with that.”
Labenz’s pushback was that an explanation could generate additional experiments and biological knowledge. Mahajan conceded the scientific value but returned to validation economics: even an intriguing Alzheimer’s fragmentomics hypothesis can be expensive to test, while teams already possess many hypotheses with stronger literature support.
Today’s biological mechanistic interpretation often means staring at semantic-segmentation plots and deciding whether a pattern is real or a spurious correlation. Mahajan can imagine future systems triaging hypotheses by validation difficulty, but for now that effort “simply feels better spent elsewhere.”
7. Test-time learning is compelling only when its optimized target matters
Labenz introduced “Learning to Discover at Test Time,” where a model is tuned to produce one exceptional answer—such as a faster CUDA kernel or improved mathematical bound—for roughly $500 of compute, without caring whether the resulting model generalizes. For an individual cancer patient, the tempting analogue is intensive tuning on that patient’s own samples.
Mahajan had read the paper and noted its single-cell RNA denoising experiment. The results beat the state of the art, but the attached domain expert made the decisive criticism: scientists do not ultimately care about denoising scores; they care about the biological utility underneath them. A verified task can improve while the real objective remains untouched.
Patient-specific tuning “very well might work” and belongs in the category of ideas that must be tried. Yet for response-versus-nonresponse prediction, Mahajan suspected ordinary supervised fine-tuning would be simpler because the endpoint is binary and sparse—the very setting the paper said was not its intended strength.
He divided bio-AI into three camps: models starting with complex human data; models of individual biomolecules hoping to raise success from perhaps 5% to 20%; and process reform that makes clinical trials cheaper. He declined to assign grandiose unequal weights: each is a bet, and “each one feels important to push on.”
8. China’s biotech flywheel combines infrastructure, labor, and short feedback loops
China’s clinical advantage, in Mahajan’s account, includes lower labor costs and a system that makes trial activation less of a regulatory and financial headache. The downside is a presumption closer to “innocent until proven guilty,” whereas the FDA starts from the opposite posture; the upside is testing more ordinary drug ideas at lower cost without requiring an AI breakthrough.
His tentative historical chain runs from strong generic manufacturing into a capable CRO ecosystem and then into local talent realizing that the infrastructure could support original drugs. The critical asset is proximity between the person designing a molecule and the workers repeatedly running wet-lab assays.
America’s loop can require assembling a setup, raising venture capital, and buying a laboratory before feedback begins. Mahajan cited PopVax founder Soham as extending the same logic to India and potentially Egypt: intellectual capital plus “a lot of hands” can compound, though more risk-averse local venture funding may remain a missing component.
Mahajan had not seen radically original Chinese bio-AI comparable to a “DeepSeek thing”; much visible work scales methods developed in America or the UK. He emphasized the hedge: American visibility into Chinese laboratories is poor, so absence of observed novelty is not proof that it does not exist.
9. Biology papers can benchmark the scientist who made the data, not the molecule
Labenz asked how a nonspecialist should calibrate between “AI discovered a drug” headlines and domain experts warning that most results will not translate. Mahajan’s answer was evaluation literacy: biology-ML papers often solve something plausibly useful while hiding confounders that only specialists recognize.
His sharpest example involved small-molecule binding. A model may classify which molecules bind a target, yet Lee Bio found that performance can be confounded by the chemist who produced them: specialists repeatedly work on particular targets, make unusually successful molecules, and leave a visually similar chemical signature that the model can exploit.
These similarities are often “human, vibes-based” and difficult to reduce to one metric. A Nature article may therefore reflect author overlap rather than transferable chemistry. Mahajan estimated that when popular-science accounts celebrate an LLM result they are “more often than not” right—“Opus 4.6, genuinely crazy”—but for bio-ML papers there may be a “50–50 chance” they missed the point.
In an earlier test, he asked o1-preview to find mistakes across 56 ML4SB workshop papers. It caught recurring issues such as sample size and test stratification but missed deeper domain failures; in nearly every article he writes, an LLM confidently supplies a framing that specialists reject as “not the real problem.” Biology remains a domain where models have not captured human taste.
10. Near-term clinical gains are easier to believe than rapid target discovery
Mahajan expects human-simulation companies, including Noetik, to “vastly improve” the results of at least a few clinical trials within several years. That forecast relies less on extrapolating distant trend lines than on capabilities already demonstrated in papers: models can stratify patients; the unsolved work is making deployment economical in real clinical settings.
He also expects Phase 1 failure to decline. A McKinsey study from five years earlier reportedly found AI-designed drugs had a 5%–10% lower failure rate—“maybe noise, maybe real”—but Mahajan expects the directional trend to continue rather than treating that estimate as settled evidence.
Brand-new target discovery is where his confidence falls. A bullish view extrapolates model progress and expects excellent target finding; the skeptical view says target discovery is so hard and iterative that models will barely dent it without a convincing human simulator. Noetik is explicitly betting that such a simulator can close the loop.
The closing epistemic rule was humility at scale: one can perhaps see one order of magnitude ahead, maybe two, but “no one can see three orders of magnitude ahead.” The discussion treated Mahajan’s skepticism as backward-looking rather than fatalistic—many published claims remain weak, while forward trend lines can still become transformative.
11. Automated-R&D experts agree until the moment that matters
Toner’s July workshop assembled people from frontier companies, policy circles, Redwood Research, Anthropic, Princeton, and the AI 2027 team. Before the first break, four presenters were arguing so intensely—but productively—that they continued while everyone else went for coffee.
One participant repeatedly pressed the others to identify the first observable disagreement between their sharply different futures. The frustrating discovery was substantial agreement about what they might see in 2026 and 2027: experts can anticipate the lead-up similarly while disagreeing on whether it crosses into a recursive loop, making advance discrimination unusually difficult.
The workshop’s first goal was to take recursive self-improvement beyond Silicon Valley and specialist AI-policy circles. Its harder goal was to identify why intuitions diverge, what bottlenecks matter, and which indicators could update beliefs. The resulting inability to form consensus became part of the finding, not simply a workshop failure.
Toner left with two load-bearing questions. Does AI replace everything humans contribute, including high-level research judgment? If it does, how quickly do other bottlenecks constrain the loop? Without full replacement, even fleets managed by elite researchers must return through a finite human channel, limiting the “massive recursive loop.”
12. Amdahl’s law may turn an intelligence explosion into jagged acceleration
Toner used an Amdahl’s-law diagram: when a process depends on multiple inputs, accelerating one component shifts the bottleneck elsewhere. Coding could become dramatically faster while other parts of AI research remain slow, leaving total AI-R&D throughput much less transformed.
Skeptics invoke an “expanding pie.” Computers removed punch cards and assembly work, higher-level languages expanded what researchers could attempt, and humans continued occupying the outer band while automation absorbed inner tasks. The opposing view says the system first automates some human work, then all of it, and advances until a nonhuman bottleneck finally binds.
Toner distinguished two “software-only singularities.” One claims software improvements alone can sustain a huge recursive loop and radically transform the world. The other is jagged: AI becomes extraordinary at software and AI research while remaining constrained by energy, copper, factories, institutions, and deployment.
Even exceptional AI-R&D performance does not automatically transfer everywhere. One intuition says such a system can train models for any task after perhaps a week of data collection; another says biological design and geopolitical strategy require real-world evidence about organisms, countries, and decision-makers. Toner thinks that missing bridge is underexamined.
13. A superhuman plateau could be bounded and still reorder everything
Labenz challenged the debate’s odd pairing: skeptics often predict a subhuman plateau, while accelerationists predict no plateau at all. His preferred middle is an easily superhuman but finite ceiling—analogous to humans being only sufficiently better than Neanderthals to transform the planet, without needing an infinite or incomprehensible capability gap.
Toner said that is close to her default expectation, then decomposed the S-curve into three variables: the length of the lead-up, the steepness of takeoff, and the ceiling’s height. Debate usually bundles short-steep-high against long-gradual-low, neglecting combinations such as short-steep-low or gradual-high.
“What does it mean to be superhuman?” remains domain specific. Interpreting sleep signals for disease and processing modalities humans barely perceive both have obvious headroom. A basic reasoner integrated across such modalities could unlock substantial power without becoming uniformly godlike.
Toner’s qualification was integration. High ceilings in sleep, scent, or health prediction may require sensor deployment, data collection, workflow redesign, and adoption throughout the economy. The underexplored combination is therefore “high ceiling, delayed by real-world adoption,” rather than assuming both capability and deployment move together.
14. Algorithms can stretch old infrastructure, but they cannot repeal it
One participant proposed that software and mathematics could route around physical constraints: cameras could replace lidar, phones could approximate specialized sleep systems, and new algorithms could increase communications bandwidth, as DSL once extracted more from existing copper lines.
Toner agreed this will work selectively. It will not make an old critical-infrastructure control system modern or turn a ship built in the 1960s into new hardware. Cybersecurity, military systems, and other long-lived assets reveal where “the jaggedness bites”—and she expects jaggedness to be fractal even inside the nominal task of AI R&D.
On the public-private model gap, Toner’s honest answer was “I don’t know.” Her impression was that it was not yet huge; Labenz had heard that internal researchers often use essentially the same models at roughly three times the speed through lower batch sizes, supplemented by tooling rather than a secret generation of intelligence.
Mixed economy-wide evidence—claims that 95% of AI pilots fail or METR’s finding that tools slowed some developers—does not transfer cleanly to frontier researchers. They know their models’ limits, influence training priorities, and can build bespoke tooling, placing them in “the very, very best position” to extract productivity.
15. Inevitability is the labs’ ambient ideology, not a complete defense
In the discussion, one participant described a sense that AI progress is inevitable and a desire to participate in creating the resulting future. The recursion argument, dating back to I. J. Good’s “ultraintelligent machine,” feels especially natural to computer scientists: once machines exceed humans at building machines, the loop appears almost self-executing.
Another participant found inevitability compelling but incomplete. Even if recursive techniques are an attractor, developers retain discretion over the flavor—whether to preserve interpretable chain of thought, embrace latent-space reasoning, or choose among many local design decisions. “AI defies all binaries”; inevitability does not absolve ownership of those choices.
The discussion placed an intervention boundary at human oversight. A researcher leading and understanding a fleet of 10 million agents is materially different from setting a process in motion with “no clue what’s going on.” The participants doubted a blanket prohibition on AI research tools could work, but hoped systems could remain on the controlled side of that boundary.
On privacy, Toner resisted technology-specific bans such as prohibiting facial recognition while leaving voice, gait, metadata, network analysis, and data brokerage untouched. She favored rules nearer the underlying conduct—collection with notice and consent, or restrictions on data brokers—while explicitly declining to pretend she had a complete privacy-law proposal.
16. Governance must follow internal deployment, not wait for model releases
Toner saw California’s SB 53 and New York’s RAISE law as useful starts and credited OpenAI, Anthropic, and to a somewhat lesser extent Google for voluntarily publishing meaningful information. Her concern is discretion: society still depends heavily on companies deciding what to reveal, though enforcement of the new state laws may begin changing that balance.
Risk evaluation should move from release-day events toward a continuous pulse. If dangerous capability can emerge in internally deployed systems automating R&D, evaluation cannot be tied only to what reaches the public. Continuous metrics may also reduce incentives to rush products out merely because governance is organized around launches.
Independent third-party audits are the missing counterpart—external access to verify that stated processes actually occur. Toner noted that such provisions repeatedly enter proposals and are then “stripped out by industry lobbying,” leaving transparency without enough independent validation.
Her no-regrets agenda is to “harden the world”: stronger cyber defense, biodefense, disease-focused biosurveillance, and epistemic security such as authenticating real content. Automated R&D could also widen the closed-versus-open model gap, so recent assumptions that open models will always stay close behind may need revisiting.
17. AI’s industrial base may be more vulnerable than the models it supports
Harris argued that infrastructure is “at least 50% of the problem” obscured by model-level threat analysis. The United States is building its AI base with globally sourced components and substantial Chinese-national talent; Labenz added that roughly 50% of top AI researchers are Chinese nationals, while ASML depends on around 3,000 suppliers spread across jurisdictions.
TSMC is the central fragility. Harris described advanced fabrication as perhaps “the most fragile production process the primates on this planet perform”—a box with hundreds of precisely tuned dials, specialized equipment, and rare human expertise. An invasion likely leaves it unavailable to everyone, whether seized, destroyed, or deliberately disabled.
The plausible runner-up is the SMIC-Huawei complex. China’s chips may be less efficient, but Huawei emphasizes networking large numbers together to reach competitive system performance. The moving constraint among logic, energy, memory, and packaging matters as much as nominal chip leadership; Harris expected Western energy limits to bind around year-end, though he stressed timing uncertainty.
The grid could be an even deeper vulnerability. Harris cited reports of Trojaned components in Chinese transformers and said a Taiwan scenario beginning with attempts to shut down Western power is being taken seriously. That would bypass debate over Samsung versus SMIC: “We literally don’t have an economy.”
18. Export controls impose costs that Nvidia access would not erase
Harris rejected the thesis that selling Nvidia chips into China would satisfy demand and reduce Huawei investment. Beijing has identified domestic AI chips as a top strategic priority and committed what he characterized as multiple Apollo-moon-landing-like amounts of money; transient access to H200s would rationally produce “why not both?” rather than technological surrender.
DeepSeek’s leadership, before R1 drew global attention, publicly said it believed it could pursue AGI but faced one core problem: “We can’t get chips, and these export controls are killing us.” Harris treats such statements, made before political scrutiny intensified, as evidence that the controls were binding.
Large prospective H200 orders reveal pent-up demand, while Chinese laboratories continue facing shortages and long waits for chipsets. For Harris, this is evidence that controls have slowed the ecosystem; allowing more sales also routes revenue toward Nvidia, but denying sales keeps compute scarce despite Beijing’s determination to fund Huawei.
Compute scarcity also links inference to innovation. China’s enormous user population floods laboratories with inference requests, consuming capacity that could otherwise support R&D. Model ownership is therefore not enough: under inference-time scaling, the decisive question is how much compute can be pointed through a model during offense, defense, research, or “model-on-model warfare.”
19. Racing labs, fragile deterrence, and adoption thresholds define the horizon
Labenz described Anthropic and OpenAI releasing models even as planned evaluations became difficult: high eval awareness complicated Anthropic’s process, while OpenAI lacked sufficiently long-horizon tasks to characterize the autonomy of GPT-5.3-Codex. Harris’s verdict was blunt: “We are racing to the bottom,” driven not only by China but by direct Western rivalry.
Domestic regulation cannot create a tactical halt if international competition remains outside the loop. A model able to design bespoke bioweapons or execute catastrophic malware would demand a slowdown, Harris argued, but a rival six or 12 months behind creates a shot clock. He doubts treaties can supply the inspection, compute accounting, algorithmic visibility, and trust that near-perfect compliance would require.
His grim alternative is credible consequence: defensive “Fort Knox” is impossible, so the United States needs offensive options capable of deterring attacks on infrastructure and supporting de-escalation. These options need not be exercised or even AI-based initially; their purpose is leverage and optionality, not “let’s go to war with China.”
Harris declined a precise P(doom), placing loss of control somewhere between 10% and 90%. Below 10% suggests missing homework; above 90% discounts how surprisingly the world adapts. He regards 2027 and 2030 as plausible, 2035 as somewhat far, and urges acting as though 2027 can happen while maintaining a “happy warrior” mindset.
His low-cost actions are “one-way door” controls: do not expose sensitive site plans to personnel who create irreversible counterintelligence risk; scrutinize suppliers promising suspiciously fast construction with scarce Chinese components; add grid redundancy; and use AI to find vulnerabilities in old load-bearing software. These interventions are cheap relative to cluster capex and preserve options even if timelines lengthen.
Harris’s own research workflow illustrates both leverage and danger: about 30%–40% of his time goes to reading a paper and the rest to interrogating models about implications. After Gemini helped him reach a satisfying but hallucinated technical explanation that Claude overturned, his update was to double-check rabbit holes; dialogue is powerful because it lets him “rotate the shape,” not because the model is automatically reliable.
Labenz is building deep personal context by consolidating communications, compressing roughly 300,000 monthly tokens into a 10,000-token chief-of-staff summary, and retaining quotations and pointers back to ground truth. Yet Opus—across 4.1, 4.5, and 4.6—still overuses scripts and regex when judgment requires it to “just read the document.”
Harris’s production lesson was discontinuity: podcast clipping failed six months earlier but worked end to end three or four weeks before the show because “1% better clears the hurdle.” In the closing discussion, one participant described earlier expectations that junior software capability would arrive by the end of 2025, diffuse organizationally over roughly three years, and reach senior AI researchers by the end of 2027; the latest update was that market pull may make adoption faster.
The closing discussion turned to personalization rather than generic assistance. One participant said personal-data curation was intended to provide a strong dataset for a future personalized model; the stated aim was not merely style transfer but “judgment transfer.” The possibility of continual or test-time learning making a personal model diverge from its baseline after one or two months remained speculative, but was framed as a potential way to preserve individual economic leverage.