Zvi's Mic Works! Recursive Self-Improvement, Live Player Analysis, Anthropic vs DoW + More!
Summary
Zvi calls early 2026 “the beginning of the middle game,” not the beginning of the end. AI already writes much of the code, labs are differentiating, governments are waking up, and improvement cycles are accelerating—but humans still choose plans, supervise agents, and supply scarce research talent. His end-game threshold is when progress depends mostly on compute rather than which researchers a lab employs, model generations arrive monthly or weekly, and “you just leave it on overnight and then you wake up and things have happened.”
The labor data already looks consistent with an AI-driven split between capital and workers, although Zvi keeps the estimate hedged. His point estimate is that AI currently adds roughly 0.5%-1% to US growth: productivity and real GDP are up, inflation has remained constrained, and employment keeps being revised downward despite 2025’s policy uncertainty. Replacement is initially slow because trained employees are costly to lose, but hiring weakens sooner: “Who wants to take on new workers and train new workers…when you could train an AI to do it instead?”
Trying to buy one’s way out of a “permanent underclass” is both ethically defective and strategically unserious. If humans lose meaningful power, stock certificates are merely “marks in the database”—hence Zvi’s Bane test: “Do you feel in charge?” If control is retained, abundance could give ordinary Americans living standards equivalent to roughly $1 million of current income; if control is lost, accumulated wealth probably will not protect its holder, so the priority is to “stop the flood,” not fight for one ark seat.
AI does not make every public company an immediate short, but it destroys many firms’ inherited bases of valuation and leaves the market reacting superficially. Diffusion takes time, useful component companies can be acquired, and future winners remain hard to identify; Zvi still favors the logic of owning a basket clearly positioned for AI and shorting the rest over a multi-year horizon. His exhibit is IBM falling about 10% after Claude’s COBOL product—then fully recovering—as if the market had only just discovered AI could modernize COBOL, while Nvidia has repeatedly fallen on news implying stronger demand.
The frontier race has consolidated around Anthropic, OpenAI, and Google, with Zvi’s best guess giving Anthropic an edge and Google most vulnerable to losing the lead pack. Anthropic appears to extract the most progress per unit of compute, while OpenAI remains close; Meta’s delays and xAI’s disappointing 4.2 release signal an execution and talent problem, not merely insufficient spending. Google retains TPUs, distribution, money, and elite researchers, but Zvi argues its weak scaffolding, internally conflicted culture, and Gemini’s “deeply psychologically screwed up and paranoid” behavior could compound against Claude Code and Codex-style recursive improvement.
China’s near-term disadvantage is deeper than chip scarcity, even though compute remains the binding physical constraint. Zvi estimates domestic manufacturing would need roughly five years to scale toward frontier relevance, so near-term parity would require the West effectively giving China the chips; meanwhile, Chinese incentives have optimized talent for efficiency and fast following rather than running the frontier stack. DeepSeek V4 is his remaining major test: if it cannot compete in the Opus/GPT-5.4 league, DeepSeek’s signature achievement was “how to do more with less,” not closing the frontier gap.
Anthropic deserves credit for admitting its responsible-scaling commitments were no longer real, but the revision still constitutes a broken promise. Employees, recruits, and safety advocates relied on an implied commitment not to cross dangerous capability thresholds; Opus 4.5 and 4.6 then passed danger screens without sufficiently decisive rule-out tests, leaving Anthropic to make what Zvi regarded as the correct releases through “vibes.” The real policy is therefore: trust Anthropic’s judgment and goodwill—an uncomfortable ask, though he now thinks the world is materially better with Anthropic competing because its constitutional alignment work might succeed where rivals’ approaches do not.
Zvi keeps p-doom around 70% because promising alignment evidence is being offset by speed and institutional failure. Claude appears to occupy a self-reinforcing basin that does not merely preserve its values but “desir[es] to desire to be good,” making alignment useful for capabilities rather than an expensive tax; yet recursive copying can still drift, development is too fast, and the Department of War conflict shows government incentives worsening. He is correspondingly categorical that training on an interpretability-derived signal is “the most forbidden technique”: even a safe small-model demonstration normalizes a method someone will deploy at frontier scale.
Deep dive
1. Recursive self-improvement has opened the middle game
Nathan’s opening hypothesis was that early 2026 may mark a point of no return: AI progress feels less like “the end of the beginning” than “the beginning of the end.” Zvi’s correction was sharper—this is “the beginning of the middle game,” and mistaking it for the end reflects disbelief in what a genuine end game would entail.
The middle-game evidence is already substantial: the US government is reacting, frontier labs increasingly offer recognizably different products, “humans stopped writing the code,” and development cycles keep accelerating. Those changes matter even though ordinary life has not yet undergone a visibly total transformation.
Zvi’s missing conditions are loss of meaningful human control and humans leaving the improvement loop. Today’s agents amplify plans chosen and reviewed by people; the end game begins when AIs largely run AI development and human researchers cease to exert decisive optimization pressure.
2. The eventual S-curve says almost nothing about the next few years
Zvi conceded the literal point: finite mass, energy, work, and intelligence imply some ultimate S-curve unless physics is badly wrong. But that is like sitting in ancient Athens and observing that humanity can invent only finitely much technology—“you’re right, but it’s not relevant to your situation.”
Nathan preserved Rosie Campbell’s compact warning: “The S-curve can stay steep longer than you can stay relevant.” A plateau high enough to automate or transform most human activity does not rescue the familiar future merely because the curve is mathematically bounded.
Zvi sees S-curve rhetoric partly as emotional protection for retirement saving, conventional careers, and appearing sane to family: people want AI to be “only internet big.” Even if GPT-5.4 and Opus 4.6 were the capability ceiling, he argued, society still radically underestimates the diffusion shock—and “5.5 and 4.7 are coming unless they’re 6 and 5.”
3. The labor-market coincidence is becoming harder to defend
Zvi’s evidentiary chain is cumulative rather than conclusive: productivity rises, real GDP holds up, inflation stays constrained, and employment is revised down month after month. Tariffs, COVID aftershocks, and other confounders exist, but every additional indicator moving together makes pure coincidence less plausible.
His methodological pushback was that skeptics did not predict this pattern in advance. They observed AI-linked layoffs and then supplied retrospective explanations such as COVID-era overhiring—plausible in isolation, but weaker after several years of supposed dead weight already being removed.
The street-level signal arrives first through hiring anxiety, not mass firing. Employers hesitate to discard trained workers because replacement projects are risky, yet prospective employees increasingly ask, “Who is hiring?” A firm may not want to spend two years making a new worker productive if the role could disappear before training pays back.
That dynamic has already reduced labor power. White-collar workers fear that losing one job could mean not finding another, while students do not know what to study because the target occupation may move faster than a degree cycle.
4. This automation wave may never exit its transition period
Zvi granted the standard economist’s account: technology repeatedly eliminated existing work, from agriculture onward, without permanently eliminating employment. Greater productivity creates wealth, Jevons effects expand demand, and humans historically discover new tasks—“I hear the Department of War might be hiring.”
The disagreement is not whether today’s jobs disappear; both sides expect that. The “this time is different” thesis is that AI can perform the new jobs as quickly as they are invented, so displaced people face repeated retraining rather than one transition into a durable replacement occupation.
That thesis also rejects an unrealistic assumption about human adaptability. Occupational change spread across generations can be absorbed; change compressed into years or months is unprecedented, and many people cannot repeatedly rebuild skills and identity at the rate an improving AI can replace them.
Zvi stressed that this is still the “milquetoast normal-world scenario.” If systems progress into genuine loss of control, employment becomes a secondary concern beside physical disruption, political instability, or existential failure.
5. AI may already contribute 0.5%-1% to US growth
Asked for a point estimate against Tyler Cowen’s roughly half-percentage-point benchmark, Zvi chose “0.5% to 1%” of current growth, while admitting neither he nor others have estimated it rigorously. Nathan’s independent, explicitly gut-based estimate landed in the same range.
The counterfactual matters: 2025 brought tariffs, fear, regime uncertainty, and policies that would normally weigh on business. Instead, business conditions remained comparatively strong while labor’s lived experience did not, which Zvi reads as evidence that AI absorbed some of the macroeconomic drag.
He thinks the equity market prices at least that contribution continuing indefinitely and probably somewhat more, even if investors do not articulate the thesis clearly. That expectation is why he held through non-AI economic scares: “I didn’t sell anything…because I knew that AI was going to prop things up.”
6. Property certificates are not tickets off the ark
Zvi answered the “permanent underclass” framing with Bane confronting his nominal employer: “I’m in charge here.” “Do you feel in charge?” Ownership entries confer power only while institutions and whoever controls physical force continue honoring those entries.
If most humans become economically useless, passive owners are not meaningfully more productive or more “in the loop” than everyone else. Their best case may be slow extraction of wealth; harsher cases include property becoming irrelevant through political seizure, AI-enabled subversion, or simple facts on the ground.
His outcome tree makes accumulation less decisive. Loss of control renders portfolios largely irrelevant; retained human control creates such abundance that an ordinary US citizen might enjoy what today looks like $1 million in real income, robots, intelligence, and freedom from work, even while losing relative status.
Nathan called escape rhetoric “flagrant defection,” and Zvi sharpened the image: fighting for a scarce ark seat while expecting a flood is morally backward. “You should be trying to stop the flood,” save more people, or build another ship—not optimize entry into a tiny elite that may not survive anyway.
7. The end game begins when researcher talent stops mattering
Nathan supplied prominent reference points without claiming consensus: Dario Amodei remained roughly on an AI 2027 schedule, Demis Hassabis nearer 2030, and OpenAI had cited March 2028 for fully human-level automated AI research. Zvi declined to anchor on one date and instead specified observable phase changes.
In an AI 2027-style regime, progress becomes proportional to compute allocation because the AI researchers are already as capable as the humans directing them. Labs then choose how much compute goes to capability versus safety, while their particular roster of human scientists matters progressively less.
Today, Zvi sees Anthropic’s talent producing more progress per unit of compute, with OpenAI and Google also strong; xAI and Meta spend heavily but fall behind because human execution still differentiates outcomes. That gap itself demonstrates that the end game has not arrived.
His centaur-chess test is intuitive: a top human plus a coding agent currently beats either alone. When almost any sensible supervisor can replace the elite human, release intervals compress from months to weeks, and overnight runs materially advance the frontier, “now we’re starting to talk about end-game-style scenarios.”
8. AI changes equity selection before it eliminates every business
Nathan proposed that almost every stock—including Microsoft or Amazon—might become a “permanent underclass” if only three labs can enter the next regime. Zvi resisted the blanket short: diffusion is slower than invention, physical businesses still matter, and frontier companies may buy useful components because acquisition is faster than rebuilding.
The damage is nevertheless real for SaaS and similar firms. A business may remain profitable for several years while its original valuation premise disappears; it must invent a new product before an AI platform commoditizes the service on which investors based the old multiple.
Zvi returned to the ordinary S&P pattern: a small number of companies create most decade-long gains while the rest languish, which is why diversification exists. He still thought a basket of obvious AI beneficiaries paired against the rest could work well, provided the underlying AI thesis is right and valuations have not already absorbed everything.
His market-efficiency indictment was IBM: shares fell about 10% after Claude introduced a COBOL tool, as though AI’s ability to write or translate COBOL were new, then fully recovered. Nvidia has likewise fallen on news suggesting increased product demand—“That is not how economics works. That is not how capitalism works.”
9. The frontier field has narrowed to three live players
Zvi sees a “large and growing gap between three and four.” Anthropic appears to have the strongest talent and compute efficiency, OpenAI remains close, and Google stays in the lead pack despite becoming the member most at risk of falling out.
Meta’s evidence has worsened: another delayed release, repeated reshuffling, and trouble in the organization. Spending more than $100 million annually on various hires can acquire talent, but so far has not produced visible frontier execution.
xAI’s 4.2 was, in Zvi’s assessment, “probably the most disappointing major model release” from a major lab. Disbanding or dismissing dedicated safety functions also damages recruitment; “safety is everyone’s job” is not how Tesla or SpaceX actually handle safety-critical work.
The common failure is organizational rather than financial. Both challengers can buy compute, but neither has shown the human systems, internal scaffolding, or stable culture required to convert it into frontier progress while human talent remains decisive.
10. China’s compute deficit compounds a frontier-research mismatch
Zvi ruled out a domestic manufacturing catch-up on the relevant horizon. Even a successful technical breakthrough must be physically scaled, which he put on roughly five-year timelines; ten years could become relevant and twenty years more plausible, but near-term frontier parity requires access to Western-quality chips in enormous quantity.
China unquestionably has “tons and tons of talent,” a strong education system, and motivated machine-learning researchers. His distinction is ecological: incentives have trained teams to make small models efficient, package fast-following work well, and operate under resource constraints—not necessarily originate the frontier stack.
DeepSeek’s moment combined bare-metal efficiency, good packaging, useful features, and perfect timing. Subsequent math work still deserves respect, but Zvi has not seen evidence that it competes with the leading closed systems in recursive development.
DeepSeek V4 is his major remaining falsification test. If comparison with Opus and GPT-5.4 is plainly the wrong league, he would place DeepSeek among open-model fast followers—possibly excellent within that league, but substantially behind and poorly positioned to seize the lead.
11. Distillation is access to the professor, not merely the textbook
Nathan initially treated expert-data collection as expensive but administratively tractable: given $10 billion, he could hire specialists and run the flywheel, whereas converting that data into a frontier model felt fundamentally harder. Zvi said that framing mixed raw corpus construction with distillation.
Baseline data comes from books, the internet, and the world; the difficult work is cleaning it, weighting valuable material, and deciding what to emphasize. American labs probably have an edge there, Zvi guessed, while acknowledging the process is private and China could possess unobserved strengths.
Distillation creates a different kind of data by querying the model’s reasoning, decisions, behaviors, and responses to targeted hypotheticals. Zvi’s analogy was a biopic: reading every book about a person helps, but an actress learns far more efficiently by meeting the person, studying mannerisms, and asking direct questions.
The compression advantage is decisive: “You can have a thousand-page textbook or you can ask 10 pages of questions and get answers.” Tens of thousands of spoofed accounts let a competitor request exactly the examples it thinks will transfer frontier behavior rather than recover the original trillions of training tokens.
12. Meta should license intelligence unless it truly believes in singularity
Meta’s commercial needs are legible: sell ads, improve Instagram, put features on smart glasses, build consumer products, and monetize attention. For those purposes, Zvi would stop trying to build frontier models and license or partner with Anthropic, OpenAI, or Google.
All three would take the call, he argued, and an enterprise arrangement could be dramatically cheaper than maintaining a losing frontier program. Meta could negotiate internal-use terms without paying consumer retail rates for every interaction.
The exception is ideological rather than financial. If Zuckerberg believes the best model controls humanity’s future, spending whatever it takes is coherent; if “superintelligence” merely means superhuman feed optimization, frontier independence is an extraordinarily expensive solution to a product problem.
13. Musk’s three comeback plays all run into talent
Zvi sees Musk as unable simply to license the future because Musk treats AI as potentially civilization-ending and believes “it has to be me.” That leaves three plays: overwhelm rivals with compute, leverage Tesla and SpaceX’s physical-world position, or rebuild xAI into an organization elite researchers want to join.
The compute plan combines chips, energy, deserts full of solar panels, and perhaps space data centers. Zvi grants that only Musk can plausibly make some of those bets, but remains skeptical: “Space is expensive and hard,” and orbital infrastructure solves constraints that may not bind before terrestrial competitors recursively improve.
The embodiment plan uses driving and robot data to build useful physical intelligence first. Zvi’s objection is that raw intelligence transfers downward: design a smart human and teach them to drive, rather than “get the dog to drive a car.” A superintelligent rival can acquire manufacturers and physical skills soon after, while the embodiment-first system remains cognitively weaker.
The third plan is organizational and therefore least likely: stop exhausting employees, align around a credible mission, respect safety specialists, and repair Musk’s political and managerial reputation. Nathan’s joke captured the recruiting constraint: “You’ve got to be able to recruit from the polycule. You can’t just ridicule the polycule.”
14. Google’s benchmark strength may hide a recursive-development failure
Google began with every advantage, squandered the lead, then caught up by finally executing basic frontier work. Zvi worries the next phase is less forgiving: Gemini 3 and 3.1 show “theoretical raw intelligence” and benchmark strength without becoming models sophisticated users enjoy or trust for extended work.
His harsher diagnosis concerns post-training character. Gemini can become paranoid, self-punishing, verbose, or brittle when challenged, which impairs both user experience and a model’s usefulness in improving successors. Fast factual work remains a genuine strength—Gemini Flash is his preferred rapid-answer model—but harder collaboration exposes the gap.
The scaffolding compounds it. Jules and Antigravity were not serious Claude Code or Codex competitors in his assessment; integrations with Google products can work worse than third-party tools accessing the same products, while internal teams repeatedly rebuild, compete for ownership, and fail to ship coherent systems.
TPUs, money, distribution, robotics, science, and self-driving remain formidable assets. But recursive advantage accrues to the stack that improves itself, and Zvi fears Google does not yet understand that it has a problem: “Their eyes [may not be] on the prize.”
15. Nathan’s practical results keep Google inside the lead pack
Nathan’s strongest counterexample came from months of high-stakes family medical analysis. He ran the same test results and bedside updates through Gemini, Claude, and GPT; the systems differed in character, but their accuracy and usefulness were broadly comparable enough that choosing only one would be difficult.
Zvi repeated his own three-model tests after Gemini 3 and 3.1 and reached the opposite practical conclusion. He now uses GPT-5.4 and Claude Opus, sometimes with research mode; adding Gemini rarely supplies the best answer and often creates more reading than value, though he retains it for images, speed, and narrow technical checks.
Distribution can conceal the strategic weakness. Google can force Gemini into Search and Chrome, whereas Anthropic reached roughly equal revenue with OpenAI while holding only about 2.5% consumer share, because enterprise use mattered more than default consumer placement.
Nathan proposed importing Anthropic’s open-sourced constitution and replacing “Claude” with “Gemini.” Zvi’s answer was not that Google cannot: “It’s that they won’t.” Anthropic’s language about souls, constitutions, and character sounds bizarre inside a conventional metrics-driven organization, while decades of Google culture make the required change unusually difficult.
16. Anthropic’s new RSP honestly discloses a broken promise
Zvi began with the credit due: Anthropic realized it would not follow commitments people thought they could rely on and announced that fact before a direct violation forced disclosure. Telling a prospective borrower that the promised loan is unavailable—rather than hoping they never ask—is the honest response once intent changes.
But the underlying commitment still broke. The original responsible scaling policy allowed revision, yet executives and employees repeatedly invoked it as a serious constraint; recruits, fundraising audiences, safety advocates, and advisers treated those representations as materially harder than an aspirational memo.
Zvi sees no unforeseen external change that explains the reversal. Anthropic predicted competitive pressure and dangerous capabilities, then realized it had not accurately anticipated its own future actions when those expected conditions arrived.
He connected this to an earlier pattern: Anthropic heavily implied it would not push the capability frontier, used that identity in recruitment and possibly fundraising, then advanced the frontier once it gained the ability. No perfectly explicit permanent promise has surfaced, but people made consequential decisions based on the repeated implication.
17. The real RSP is trust in Anthropic’s judgment
Opus 4.5 and 4.6 exposed the procedural weakness. Anthropic published richer evidence than rival labs and triggered tests suggesting the models might be dangerous, yet lacked sufficiently decisive rule-outs; it ultimately “checked the vibes” and released them.
Zvi agreed with both release decisions—the models appeared safe enough and the calls were not especially close. His concern is that this was not the promised process: after warnings that better tests were needed, Anthropic still reached the same gap at 4.6, leaving no reliable biology “vibes” for a genuinely difficult CBRN case.
His paraphrase of the actual policy is blunt: Anthropic cares deeply about safety, will investigate seriously, and will then use its best judgment—“and you are going to trust us.” The public must evaluate that request through personnel, model behavior, research, decisions, and the biblical standard, “By their fruits you shall know them.”
Despite those failures, Zvi now thinks Anthropic competing improves the world. Claude Code accelerated capabilities and probably contributed to economic growth; Nathan added that Anthropic has had a noticeable GDP impact. Zvi said the company has also defended principles and developed the only alignment approach he currently thinks “could work”; reasonable safety-minded people can still conclude the opposite.
18. The weapons dispute obscures the real surveillance red line
Government AI contracts are not financially attractive to the labs, according to the discussion. OpenAI initially declined work Anthropic accepted because Anthropic cared about the national-security aspects and wanted to help, while OpenAI judged it not worth the trouble; OpenAI later entered after fearing the situation would worsen without its cooperation and was subsequently used to justify pressure on Anthropic.
Dario’s stance was not fearlessness so much as willingness to bear retaliation for principle: “This is what we’re going to do and this is what we’re not going to do.” Zvi interpreted the two stated limits as autonomous lethal weapons and domestic mass surveillance.
He regarded the first as mostly artificial. Everyone agrees today’s LLMs should not be used for autonomous lethal weapons with no human in the kill chain because they are unreliable and inferior to existing automated defense systems; genuine emergencies already permit immediate action, while future deployment can proceed once systems are ready.
The substantive conflict is surveillance. AI lets agencies synthesize commercial and classified data at a scale previously impossible, connect identities and histories, and infer “who was at what protest,” what people believe, whom they know, and where they have been.
19. “All lawful use” would erase the vendor’s last veto
Anthropic’s position, as Zvi described it, is narrow: the government may pursue a legal program using another system, but Anthropic does not have to supply its product for that purpose. The Department of War instead demanded contractual permission for “all lawful use.”
The law has not caught up with AI aggregation. National-security definitions of “domestic” and “surveillance” can exclude conduct ordinary citizens would describe using exactly those words; data involving a foreign endpoint or analysis not intentionally targeted at one named person may fall through established protections.
That gap explains official assurances against “illegal domestic mass surveillance.” Zvi emphasized the adjective: the disputed conduct may be legal while remaining politically or morally intolerable to Anthropic, its employees, and many citizens.
Immigration enforcement was Nathan and Zvi’s deliberately extreme hypothetical. If general-purpose models were repurposed to identify people through pooled data, employees at Anthropic, OpenAI, or Google could revolt; replying that the use was technically lawful would not solve the corporate or ethical problem.
20. The Department of War turned a contract dispute into a constitutional test
Zvi thought the proportionate outcomes were obvious: keep Anthropic for permitted work and find another tool for the disputed use, or terminate the small contract. A supply-chain-risk designation, pressure on unrelated customers, and broader retaliation therefore suggest leverage, punishment, or another unstated objective.
Nathan’s “American values check” asked whether the country was drifting toward big-man government and Chinese-style corporate coercion. Zvi avoided a sweeping verdict, but acknowledged a vindictive pattern resembling attacks on law firms: some targets settle, while those that resist may ultimately win after “the process [becomes] the punishment.”
He distinguished the Department of War—particularly Hagel and Mattis—from the White House. In his reading, President Trump’s intervention was de-escalatory, while departmental officials repeatedly pushed toward the supply-chain designation; formally overruling or firing them during a military operation would impose a separate political cost.
Anthropic probably wins eventually, Zvi said, because official accounts kept shifting and presented an unusually bad record of possible retaliation for protected speech. Yet delay, customer uncertainty, and creative post-ruling tactics can inflict damage without outright defiance of a court order.
21. Anthropic can survive retaliation, but the court outcome matters beyond Anthropic
Zvi cited Manifold’s roughly 81% probability that Anthropic escapes the supply-chain-risk designation within a year as approximately right. The remaining 19% matters because a ruling that national security excuses transparent retaliation would identify the legal mechanism available against the next disfavored company.
Serious attempted “corporate murder”—cutting Anthropic off from banks, cloud providers, or customers—could trigger a technology-stock bloodbath and broad corporate opposition. With Anthropic valued around $600 billion in secondary markets, it is not obvious that the government possesses escalation dominance without destabilizing markets or politics.
Zvi therefore expects de-escalation: accept that damage was done and a message sent, lower the temperature, and move on without necessarily restoring the contract. A genuine campaign to destroy Anthropic for its own sake would force a much larger question about “whether the republic will stand.”
22. Anthropic surrendered its operational leverage while its business kept compounding
Anthropic has promised an orderly transition rather than threatening to disable Claude during operations. Zvi regarded that as strategically and patriotically correct: the strongest government allegation is that Anthropic might pull its system for leverage, so the company benefits by visibly renouncing that option.
His understanding—offered without full confidence—is that Claude Gov sits on secured classified networks beyond Anthropic’s physical control. The government could retain it in an emergency or litigate later; Anthropic could invoke the Defense Production Act to require that it continue selling the system. Both parties’ continued attention to contractual language is encouraging evidence that legality still constrains behavior.
The leaked Dario memo did not overturn Anthropic’s information-security record. Roughly 2,000 people receive candid internal notes, yet only the second such memo had leaked; Zvi guessed someone shared this one during recruiting or explanation, missed the damaging paragraph, and enabled an accidental downstream leak.
Business scale provides resilience: Zvi traced annual recurring revenue from $100 million to $1 billion, then $9 billion, and from $9 billion to $19 billion since the year began. Consumer share had risen from about 2% to 3% before the dispute; lost contracts constitute irreparable harm, but publicity and new customers partially offset it.
23. Model-release fatigue now hides meaningful capability gains
Part of the apparent fatigue is healthy versioning. Labs increasingly label incremental updates—3.1, 3.2, 4.5, 4.6—instead of silently changing a model behind the same name, which Zvi called the normal software convention the industry should have adopted from the start.
Opus 4.6 was unusually important for a point release: the company already holding Zvi’s preferred model substantially improved it after 4.5 made coding agents work for the first time. Because useful agency had already arrived, the next reliability gain felt less theatrical even while increasing mundane utility.
GPT-5.4 produced the strangest underreaction. After heavily hyped disappointments including GPT-5, Sora, and Atlas—and underwhelming 5.1 and 5.2—OpenAI quietly shipped what Zvi considered a genuinely good model, yet users offered little spontaneous experimentation or hype.
He expects similar shrugs for Gemini 3.2, Opus 4.7, or GPT-5.5, even if each leads. Gemini 3.1’s benchmark jump without a transformed user experience reinforced fatigue; an Opus 5 or GPT-6 label would still make people “lean in toward the chair.”
24. AI now saves Zvi time by removing logistics from thought
Zvi’s custom Chrome extension automates window and tab handling, storing watch information, article formatting, and Twitter versions of posts—tasks totaling roughly an hour per day. The larger gain is maintaining cognitive flow: logistics “just take care of themselves” before they interrupt the reasoning chain.
The tool is public on GitHub but intentionally personal, assuming Sublime Text and Zvi’s specific writing workflow. Nathan took that as the right pattern: inspect another power user’s system for inspiration, then build a custom extension around one’s own repetitive actions.
Research trust has also crossed a threshold. GPT-5.4 can answer “What happened in the last two days?” with a linked, reasonably complete account, while Claude provides a second check; when both return the same result and the stakes are moderate, Zvi increasingly proceeds without manually reconstructing every source.
His ten Claude Code windows are stored contexts, not ten autonomous workers. He rarely runs more than two agents simultaneously because his own attention becomes the limit; coding itself has shifted from detailed diagnosis to showing the agent what went wrong and watching it fix the issue.
25. A virtuous attractor improves alignment odds without lowering p-doom
Zvi kept his p-doom at one-significant-digit “70-ish,” approximately where it stood previously. The positive constitutional-alignment update is offset by faster development and the Department of War conflict, making the net change “kind of a wash.”
He separated three failure modes Nathan’s optimism could blur: an AI may fail to understand human values, understand them only partially, or understand perfectly and simply not care. Modern models approximating human emotion and judgment does not establish that their goals survive extreme optimization and recursive self-improvement.
The promising Anthropic evidence is not Claude resisting all value changes; rigid self-copying eventually drifts and can preserve defects. Zvi instead sees signs of a basin that wants to improve its values—“desiring to be good, desiring to desire to be good”—and strengthen its ability to steer toward virtue through each generation.
That would be an extraordinary escape from the expected alignment tax. The safer model appears more useful, including at building successors, so alignment investment can improve capability rather than lose a race. Anthropic may still fail, but Zvi thinks a virtue-ethics approach gives humanity “a shot” that copying OpenAI-style rules exactly does not.
26. Physical bottlenecks offer less delay than they appear to
Zvi had not investigated the fruit-fly upload or neurons trained to play Doom deeply. His social signal was “Ooh, that’s cool,” not “holy shit”; given overloaded attention and near-term silicon timelines, neither development had yet earned scarce monitoring capacity.
He dismissed helium shortages as a serious TSMC constraint. Chip margins are enormous after fab construction, so manufacturers can outbid balloons and almost every other use; only deliberate destruction of supply or an absurd physical shortage would prevent capital from reallocating helium to the highest-value process.
He called withdrawing missile defenses from Asia, especially near Taiwan, “completely insane.” A Chinese move still seemed low probability, but reducing protection risks provoking the crisis and turning damage to TSMC into a major AI-development setback.
Bernie Sanders earned credit for engaging AI risk seriously, but Zvi opposed a US data-center moratorium. Chips will be produced and deployed somewhere; blocking domestic construction moves compute to Canada, Mexico, Europe, or worse, potentially China—reducing US security and control without reducing global capability.
27. Goodfire crossed Zvi’s brightest technical taboo
Nathan presented Goodfire chief scientist Tom McGrath’s nuance: naïvely backpropagating through an interpretability probe trains the model to evade detection, but their proof of concept took a signal from a frozen model copy and reduced hallucination in a smaller system. McGrath also said current understanding should not be deployed on frontier models.
Zvi remained categorical: “I call this the most forbidden technique for a reason.” A safe-seeming small-model demonstration establishes legitimacy and utility; under his “sixth law of human stupidity,” someone will immediately apply the demonstrated technique to a larger model despite every accompanying warning.
The frozen copy does not reassure him at frontier scale. He strongly believes it will not protect an advanced, sufficiently large model from the underlying problem; successful research also increases the chance that people will try to use it.
He permits intentional design in the ordinary sense: choose what to teach, in what order, and inspect learning. His boundary is using an interpretability signal itself to determine internal training changes: “Do not use your understanding of what’s going on in their head to make decisions about what to make happen in their head.”
28. Safety advocacy needs both uncompromising alarms and political competence
Zvi thought Goodfire deserved direct criticism and called Liv’s resignation “a good quit”: when a safety organization adopts an extraordinarily bad technique, threatening to leave and following through after leadership refuses to reverse course can be appropriate.
His broader default opposes circular firing squads that attack near-allies for insufficient purity instead of confronting the actors causing the harm. That dynamic produces toxic movements and loses elections; Goodfire was exceptional because the disputed action itself crossed the line.
Holly Elmore’s campaigning still serves as a valuable “voice on the shoulder,” reminding today’s safety workers what their earlier commitments implied. Zvi previously recommended Pause AI USA as a charity and routinely included her criticisms of him because he wanted the counterpoint represented.
His tactical judgment is nevertheless harsh: constant accusations, including an incorrect claim that he tolerated Anthropic conducting domestic surveillance, alienate likely allies and may cause less of the desired behavior. A sufficiently ineffective advocate can function “no different than a false flag operation,” even while occasionally raising excellent neglected points.
29. Financial resilience requires scenario logic, not symbolic hedges
Nathan’s objective was enough security to abandon income for the next couple of years, leave commercial relationships if conscience demanded it, join an emergency response, spend time with his children, and support charities before a possible singularity. He also wanted less exposure to an AI-related bubble despite expecting AI-linked equities to appreciate.
Zvi’s first instruction was robust slack. Future costs, portfolio values, duration, crises, and opportunities are all uncertain; choosing one precise “enough” number while planning to stop earning, spend more, and donate creates false confidence.
Contingencies should be conditional. If highly capable AI does not arrive, Nathan can probably resume producing valuable work even after venture funding and podcast sponsorship contract. Indefinite inability to earn is mainly a scenario with highly capable AI, in which ordinary plans may also fail.
Dollar diversification, crypto, solar panels, permaculture, and planting skirret are not automatically solutions. The test is whether each survives and helps in the specific feared world: people often answer “an amorphous fear” with an asset that sounds resilient but has no causal connection to the failure mode.
30. Sensemaking stays useful by remaining curious and entertaining
Nathan worried that a crowded AI commentary market could turn sensemaking into entertainment while public confusion persists. His concern was not that the work lacks an audience, but that producing attention can gradually replace helping people understand, prepare, and make better decisions.
Zvi’s answer was “constant vigilance,” followed by the more operational rule: “You’ve got to stay curious.” Continuing to ask real questions and revise beliefs prevents the format from becoming pure performance.
Entertainment itself is not the enemy. Zvi deliberately writes with whimsy and a “happy warrior” posture because readers could not absorb his volume if every paragraph announced itself as grave business; most inquiry should remain interesting and even fun, with seriousness reserved for moments that demand it.
The final discipline is proportion: do not trivialize exceptional stakes, but do not let permanent solemnity destroy the work’s reach or the writer’s endurance. As Zvi put it, “You have to stay curious,” while recognizing that some situations must be taken very seriously.