Summary

Hoffman’s investing map puts much of his AI time in Silicon Valley’s blind spots, beyond the crowded “obvious line of sight.” Chatbots, coding assistants and productivity products remain investable, but everyone sees them; longer runways may sit where software meets biology and atoms. Meanwhile, platform shifts do not erase network effects or enterprise integration: “significant things change,” not everything.
The biotech opportunity is a drug-discovery factory aimed at “the speed of software,” without pretending biology can be fully simulated. A predictive system could be valuable even if it makes the right prediction only 1% of the time, provided experiments validate the candidates, turning a search for “a needle in a solar system” into a potentially useful funnel. A superintelligent drug researcher might arrive “maybe someday, not soon,” while regulation and biological complexity remain real constraints.
Labor adoption begins with augmentation sold to people who want to be “lazier and richer.” Products promising mass layoffs are difficult to distribute; products offering fewer hours and more income are not. The strongest incentives sit with doctors, small businesses and sole proprietors who can capture fivefold patient volume or settlements, while large companies suffer a principal-agent problem.
Medicine previews the role split: AI can displace the credentialed knowledge-store role, while professionals retain context, judgment and lateral thinking. Hoffman’s categorical advice is to use ChatGPT or an equivalent for every serious result as a second opinion—and seek a third if it disagrees. He still expects doctors in 10 or 20 years, but as expert users of AI rather than people whose authority comes from memorization.
AI is underhyped among real-world users, while forecasters often mistake a “savant curve” for “apotheosis.” Hoffman’s tests compressed three days of analyst work into 10–15 minutes yet still reproduced consensus rather than the lateral argument he needed. The likely path is an improving fabric of LLMs, diffusion models and other systems—not “one LLM to rule them all.”
Robotics remains governed by the crossover between capex and human opex, not intelligence alone. Deep research targets high-value analyst work in bits, while folding laundry may require $100,000 of hardware to compete with someone earning $10 an hour. Japan’s labor scarcity makes bowling-shoe robots rational; falling hardware costs could move that crossover elsewhere.
LinkedIn’s durability shows that AI does not repeal hard-to-build networks or economic constraints. Its professional graph survived because the “turtle” accumulated a community that challengers could not reproduce; candid negative references are often obtained through private inquiries across that graph. AI startups also need revenue earlier than Web 2 companies because exponentiating usage creates exponentiating compute costs.
An AI may become a spectacular companion, but Hoffman rejects calling it a friend because friendship is bidirectional. Friends agree to help each other become better versions of themselves, permit themselves to be helped and sometimes deliver tough love; a system does not participate in that reciprocal relationship. The consequential design question is how children learn and form an epistemology around AI, not whether a model claims consciousness.

Deep dive

1. AI alpha begins where Silicon Valley stops looking

Hoffman starts with epistemic humility: everyone is “looking through a glass darkly, through a fog with strobe lights.” His seven-deadly-sins framework nevertheless persists because it rests on psychological infrastructure shared across 8 billion-plus humans.
The “obvious line of sight”—chatbots, productivity software and coding assistants—is still investable, but consensus makes differentiation harder. In the second bucket, AI rearranges markets without erasing network effects, enterprise integration or other sources of durability: “significant things change.”
Hoffman has concentrated much of his co-founding and investment time on Silicon Valley’s canonical blind spot that “everything should be done in bits.” In 2015, he urged Greylock to pursue AI-enabled productivity while asking Stanford to imagine AI tools for every discipline—custom search-like systems transforming knowledge generation, communication and analysis, potentially extending even to theoretical math and physics as capabilities improve.

2. Biology rewards prediction rather than perfect simulation

Hoffman’s concrete company-building example is a drug-discovery factory designed to work at “the speed of software.” He concedes the unavoidable biological and regulatory layers; the bet is acceleration at the bits-atoms boundary, not elimination of atoms.
The classic Valley error is assuming the biological system can simply be simulated until drugs fall out. Hoffman’s alternative is probabilistic prediction: even 1% accuracy could be valuable when experiments validate the hits. The search is “not a needle in a haystack,” but “a needle in a solar system.”
His categorical medical advice: use ChatGPT or an equivalent as a second opinion for every serious result, and get a third opinion when it diverges. A doctor may remain in 10 or 20 years, but as an expert user of the knowledge store—not because medical school conferred memorized authority.
Preparing to debate doctor replacement, Hoffman tested ChatGPT Pro, Claude 4.5, Gemini Ultra and Copilot in deep-research modes. Despite experience dating to GPT-4 six months before public release, he rated the results B− or B: extraordinary compression of work an analyst might produce in three days into 10–15 minutes, but mostly consensus synthesis rather than lateral reasoning.

3. Software sells worker leverage before it eats labor

Hoffman’s simplified seven-deadly-sins adoption heuristic is “lazier and richer”: fewer hours paired with more income. “Software eats labor” currently works less as a product promising job losses—which nobody wants to buy—than as a product that makes an existing expert more productive.
The incentive is sharpest where user and owner are the same person. A dermatologist who can see five times as many patients or a plaintiff’s lawyer who can handle five times as many settlements captures the upside; a corporate director may only see the “ethereal being of the corporation” benefit.
Hoffman said that, apparently, roughly two-thirds of doctors now use OpenEvidence. His credentialism discussion is that degrees once provided a valuable knowledge heuristic, but AI now supplies much of the knowledge base; professionals increasingly earn their place by knowing when consensus deserves investigation.
Hoffman cited Ethan Mollick’s reminder that “the worst AI you’re ever going to use is the AI you’re using today.” He says the people underhyping AI are often those who know nothing and those who know everything, while users applying it to become richer and lazier are seeing the current value. His operating rule is blunt: if AI has not helped with serious work, “you’re not trying hard enough”; a five-minute diligence plan can replace a day of initial work.

4. Robotics waits for the capex-opex crossover

Rampell contrasts Goldman Sachs-style sell-side analysis, the kind of work deep research targets, with laundry folding, where a $100,000 robot competes against $10-an-hour labor. Language has a high bits-to-value ratio; the physical world contains far more state to sense and abstract.
The bottlenecks include manipulation, multiple degrees of freedom and battery chemistry—lithium-ion energy density compares poorly with cellular ATP. Deterministic FANUC assembly robots work; general household machines struggle. Japan leads in robotics partly because labor is scarce, down to a bowling-alley robot that dispenses and cleans shoes.
Hoffman adds context awareness to the robotics discussion. GPT-2, GPT-3, GPT-4 and GPT-5 look like a progression of increasingly capable “savants,” yet long-running agent conversations can loop for extended periods through exchanges like “one month later, thank you” and “no, thank you.” Humans immediately know to stop; models still approximate that commonsense context.
Hoffman’s revised label for the species is “Homo technae,” not merely Homo sapiens. Opposable thumbs mattered, but the compounding mechanism was language, writing and technology transmitting learning across generations—the substrate AI now accelerates.

5. Scaling produces better savants, not automatic gods

Hoffman supports extrapolation but disputes the assumed curve. An exponential “savant curve” is different from apotheosis: “in 2½ years, magic” may yield extraordinary specialized capability without yielding all magic, leaving room for generalists, cross-checkers and context-aware judgment.
Critics who point to prime numbers or the number of Rs in “strawberry” are “missing the magic,” even if some structural LLM weaknesses may persist for three to five years. AI will combine LLMs, diffusion models and other architectures through a shared fabric; whether that fabric is fundamentally an LLM remains TBD.
Torenberg relayed Stuart Russell’s view that a more predictable model fabric could reduce fears of systems going amok. Torenberg also said formal verification of arbitrary outputs looks extremely difficult—“we can’t even do verification of coding”—while Hoffman agreed that greater programmability and reliability are worthwhile technical goals.
Torenberg’s math frontier distinguishes AIME answers, integers from 0 to 999 with easy evaluation, from novel proofs that are difficult to construct and validate. He mentioned Lean and a rumor about DeepMind solving Navier–Stokes, while the group joked that “AGI is the AI we haven’t invented.”

6. Agency is likely; consciousness remains unresolved

Rampell considers AI agency and goal-setting “almost certain” because complex problem-solving requires systems to establish subgoals. He treats the paperclip maximizer as an example of context failure, while resisting the assumption that an actual intelligence would mechanically convert the planet into paperclips.
Rampell’s strongest argument against free will is biochemical override: hunger, anger or norepinephrine can radically change behavior. Hoffman accepts that humans may be biochemical machines but rejects simplistic machinery, pointing to Roger Penrose’s coherent possibility that quantum effects matter to our form of computational intelligence.
Hoffman does not think consciousness is required for goals, reasoning or perhaps some forms of self-awareness. Mustafa Suleyman’s “semiconsciousness” framing is useful precisely because conversational fluency misleads: an earlier Google model answering “yes” when asked whether it was conscious was not proof—“QED” was the setup for Hoffman’s rejection, not the conclusion.
He expects some definitions of AGI to be reached before philosophy solves consciousness. The nearer design priorities are concrete: children’s epistemology and learning around AI, plus intelligent energy optimization. He cited Google applying algorithms to its highly tuned data centers and achieving 40% energy savings.

7. LinkedIn’s network survives AI, but AI rewrites monetization

LinkedIn was long dismissed as the dull “turtle” beside Friendster, MySpace, Facebook and TikTok. Its professional-productivity orientation—LinkedIn’s seven-deadly-sins analogue was greed, versus Twitter’s wrath—created a difficult network that became the place where its members collaborate.
Hoffman welcomes new products that help people find and perform productive work because his hierarchy is humanity, then society, then industry, though he would prefer LinkedIn to build them. After seeing GPT-4 and knowing Microsoft had access, he urged LinkedIn into the room: Silicon Valley’s “religion” begins with the amazing new thing, sometimes before anyone knows its business model.
AI constrains that old Web 2 playbook. ChatGPT had monetization built in through a $20-per-month subscription; AI’s changing COGS and usage volume make a rising cost curve dangerous without a following revenue curve. At PayPal, exponentiating free volume once made the team able to identify the hour it would run out of money.
LinkedIn also demonstrates why negative-reference products resist virality. Public endorsements can resemble book blurbs; candid criticism carries social and legal complexity. Hoffman instead finds connected people and asks for a 1-to-10 rating or “call me”—a couple of “call me” responses are revealing, while a set of eight 9s is reassuring.

8. Human leverage culminates in government and friendship

Hoffman keeps AI and its effects on Homo technae, society and work at the center of his calendar—now “six and a half days” rather than seven. That includes co-founding an AI-biotech venture with Siddhartha Mukherjee and getting instruction on the FDA process, the kind of work that software instincts alone do not make easy.
He has advised officials in Western democracies for 20–25 years, recently discussing with Macron how France should respond if frontier models concentrate in the US and perhaps China. Mistral is one part of the landscape, but the governing question is how domestic industry, society and citizens benefit from an externally driven platform shift.
Hoffman’s closing definition is relational: friends “agree to help each other become the best possible versions of themselves,” allow themselves to be helped and sometimes say, “You screwed up.” An AI may be an “awesome companion,” but without mutual stakes it is not a friend; sycophancy cannot substitute for a team sport.

A Conversation with Dai Yusen: From Jumei to ZhenFund in the AI Era Marathon Partners' Mario Cibelli updates the Remitly Thesis $RELY