CROSSPOST: GAUTAM MUKUNDA: The Part That Looks Like Waste
The Indispensable Newsletter #64, by Gautam Mukunda :: What look like “frictions” are often the safeguards that we really really need to save us when our typical abstraction layers catastrophically leak and fail: examples from corporate governance, companies trying to utliize MAMLMs, knowledge workers dealing with this cognition technology from the other side, and assessing the net benefits of rapid Schumpeterian innovation…
You must read this because you need to know what Gautam Mukunda says here, with respect to a surprisingly common failure mode:
leaders who strip out constraints because they slow things down,
but those “wastes” were actually error-correction.
Concentrated control at firms like SpaceX removes governance checks and is likely to turn Elon Musk into something we have not yet seen: his worst self at full throttle.
Firms trying to get ahead of cognitive technology revolution burn money as they discard their legacy workflow capabilities that are their insurance against uncertain timing and limits. Vibe-coding humans on autopilot risk hollowing out their own skills unless they deliberately nurture and preserve “stick‑and‑rudder” abilities for the rare but critical 1% of situations in which the MAMLM-driven abstraction layer at which they can work most rapidly fails. And the socially inept princelings of AI appear incapable of making the case that “AI” will be an engine of human liberation rather than drowning people in seas of AI-slop while undermining their labor-market bargaining power, with the likely consequence of this last being a powerful Polanyist backlash against the mantra “the Market giveth, the Market taketh away: blessed by the name of the Market”.
The four pieces this is drawn from are very much worth reading as well:
<https://www.bloomberg.com/opinion/articles/2026-06-04/uber-s-ai-doubts-are-just-normal-growing-pains>
CROSSPOST: GAUTAM MUKUNDA: The Part That Looks Like Waste
<https://gautammukunda.substack.com/p/the-part-that-looks-like-waste> <https://gautammukunda.substack.com/>
Dear Friends,
Four columns, four domains, one mistake: stripping away the safeguard precisely because it slows you down.
The most valuable component in many systems is the part that looks like waste. The board meeting that slows the CEO down. The skill the machine made archaic. The capability you half-abandoned because a new technology seemed to make it obsolete. The neighbor who can block your permit. Each one is friction. Each one is also the thing that catches you when you’re wrong.
I wrote four columns over the last two weeks, in domains that don’t obviously belong together — corporate governance, AI strategy, energy infrastructure, and the future of work. Underneath, they are all about the same temptation: to strip away the safeguard because it slows you down, and to discover too late that slowing you down was the entire point.
Thanks for reading The Indispensable Newsletter by Gautam Mukunda! Subscribe for free to receive new posts and support my work.
The cleanest version is the one that started the run. SpaceX’s IPO structure gives Elon Musk 85% voting control, three titles, a board he appoints, and immunity from being fired — and it was made possible by the institutions that are supposed to push back. Texas rewrote its corporate law to lure the listing; Nasdaq bent its index rules to force money into the stock. Governance is error-correction. It exists to test a leader’s decisions before the checks clear — the way someone might have flagged that spending billions so your Metaverse avatars could finally have legs was not, in fact, a triumph. Remove it and you don’t get a freer genius. You get the worst version of whatever leader you started with, running at full power. The diagnosis these CEOs share — that Wall Street is too short-term — is correct. Their cure, making themselves monarchs, is worse than the disease.
The same move shows up as strategy. Uber’s president told a podcast the company can’t yet connect its AI coding spend to a return, and the skeptics seized on it as proof the bubble is popping — even as Anthropic raised $65 billion at a $965 billion valuation, a number you don’t pay for a technology going nowhere. Both things are true. The economics are unsustainable today, and the technology is still transformative. Aviation and biotech each incinerated capital for decades before they paid off. The companies that survived the wait, like Vertex in pharma, didn’t bet everything on the new approach and discard the old one. They kept their existing capabilities as a hedge against a technology whose timing they couldn’t predict. The old skill looked redundant. It was insurance.
That is exactly what aviation teaches about the future of knowledge work. Autopilot makes flying safer, but a pilot who lets the computer fly 99% of the time loses the stick-and-rudder skill she needs for the other 1% — the difference between Sullenberger on the Hudson, whose reflexes were intact, and Air France 447, whose crew had been allowed to let theirs rust, falling recoverable the whole way into the Atlantic. AI is about to do this to everyone who does cognitive work. Whether it hollows people out or empowers them won’t be decided by the technology. It will be decided by whether leaders deliberately preserve the perishable human skills the machine made optional — because, unlike aviation, there is no FAA requiring it.
Thanks for reading The Indispensable Newsletter by Gautam Mukunda! Subscribe for free to receive new posts and support my work.
Three of these safeguards have something in common: they’re yours, and the danger is that you let them quietly atrophy. The fourth is different. In the revolt against AI data centers, the safeguard is the public itself — and the public doesn’t atrophy. It votes. Kevin O’Leary’s plan to pave 40,000 acres of Utah drew overwhelming local opposition, and his answer was to call the objectors proxies for the Chinese government. That’s the same contempt that killed GMOs in Europe after Monsanto’s man called wary Britons “sad sacks.” The lesson of Rockefeller, the original hated billionaire, is that the public never actually hated his oil — it got cheaper every year. They hated him. Tech’s leaders are offering a worse bargain: I get richer, you pay more for power, breathe dirtier air, and might lose your job. Data centers still need permits. And that means every data center is a referendum. Tech leaders can outspend the public. But they can’t outvote it.
Notice who keeps appearing. Musk’s governance structure removes every check on him at SpaceX; his data center ran 35 unpermitted gas turbines over a Memphis neighborhood where cancer risk already runs four times the national average. He is the purest case of the instinct all four columns are about — remove the constraint, at every level, and call it freedom.
The safeguards in these stories are somebody’s stick and rudder: the manual control you keep for the day the autopilot flies you into a mountain. Cutting it loose feels like buying speed. It isn’t:
You’re buying the worst version of yourself, at full throttle, with nothing left to grab.
—Gautam
Brad here again: The coëxistence of Anthropic’s giving a 6.7% equity ownership stake in return for $65 billion coëxists uneasily with Andrew Macdonald’s observation that Uber’s progammers’ use of AI in their code is still not connected to the company’s bottom line. The programmers may be building skills that will be useful in the future, but the rubber has not yet met the road—or, rather, consider this metaphor: The programmers are getting practice improving their stick-thwacking form, but the ass has not yet moved, and the programmers are still having to carry the load as well as thwacking the recalcitrant ass.
The two coexist, of course, because the perception is that variance is absolutely enormous and that we are not even in the realm of calculable risks. In such times of great uncertainty, the systems and organizations that survive will be those that prioritize robustness and optionality. They will not be the ones that equate marginal cost and benefit. They will, rather, be those that do a lot of things that look to be ex-ante and turn out to be ex-post wasteful, but one of which winds up generating the capabilities that actually matter in the unknown future. I see three major things going on here:
People paying $65 billion for a 1/15 stake of Anthropic’s equity are almost surely overpaying, but they are enmeshed in a system that makes not having a stake in Anthropic a career-destroying move should Anthropic turn out to be more than marginally profitable over the long run.
Uber and other tech companies telling their programmers to burn tokens on AI training as a way of investing in the future are also almost surely overpaying, but again they are enmeshed in a system that makes not building AI-capabilities now a career-destroying move in the future in which these technologies do turn out to be world-changing.
Consider the Amazons, Googles, Facebooks, and Microsofts spending fortunes building-out as a way of attempting to stay close enough to the frontier to keep Anthropic or Open-AI from breaching their moats and taking over their platform monopoly profits for themselves. They are also almost surely burning their shareholders’ money, but, again, they see an existential disruption threat.
These three episodes of likely massive overspending and overinvestment from the perspective of the bottom line, however, may well be not bad for society as a whole. We do underinvest in activities with powerful positive externalities, and exploring the spaces of new technologies is prime among those activities in which we underinvest. But that is an argument that applies to governments thinking about whether they should be taking steps to curb the current AI bubble, not an argument that those who are currently gung-ho and investing it have their heads screwed on right. Rather, those currently doubling down on the bubble should be thinking about building optionality and robustness instead.