Betting Before the Answer Arrives — Notes on Sarah Guo and What the 250 People Building AI Believe

Having weighed what is right, I act; and being right, I see nothing to regret.

—— Wang Anshi, Reply to Sima Guang (Northern Song, 1070; my own translation)

What this episode is about

In the September 1, 2026 episode of Invest Like the Best, Patrick sits down with Sarah Guo, founder of the early-stage AI firm Conviction. She joined Greylock at 23 and left four years ago to build her own thing. The spine of the conversation is a simple question: when something moves too fast to backtest, what do you bet on?

She opens by quoting an investor friend, and I laughed when I heard it: “I keep saying I want to press the brakes as hard as I can, but I’m not doing it — I’m going 90 miles an hour.” That’s most people’s actual situation right now. The real question comes next: are you afraid of missing the opportunity, or are you about to repeat the same mistake every boom-bust cycle in technology has made?

The episode covers how she picks companies, what her team argues about, what the roughly 250 people pushing the frontier believe right now, where compute and energy are stuck, and the calls she got wrong. What stayed with me wasn’t really the AI part — it was how fussy she is about the word judgment. Here’s what I kept, plus some of my own thinking.

The main points

1. Her competitive universe is about 250 people. Patrick mentions that her partner keeps a mental list of roughly 250 people — researchers and founders mixed together — doing the most interesting things at the frontier. One of the firm’s goals is simply to be close to all of them. What I liked was how honest she is about her own contribution: asked how many of her portfolio companies wouldn’t have been funded without her in that round, she said 5%, maybe 10% max. These people are resourceful; they’d find other investors. What she might actually change is whether they get the next $100 million of compute.

2. Frontier researchers are feeling oddly powerless. She describes a belief that’s new in the last twelve months: because AI can now improve AI research itself, a lot of researchers think exponential intelligence is one or two years away. She immediately undercuts it — Karpathy has been saying “two years away” for about ten years, and, to be fair, he thinks it again. What she cares about is the side effect. When a lab had 200 people, you felt you could move the needle. When the framing becomes “we need $750 billion of compute and thousands of people,” people stop feeling ownership. Either “what I do doesn’t matter, the model will do it,” or “only compute scale matters.” Both are disempowering.

3. She picks application areas by working forward from what the models are good at. In late 2022, if you believed next-token prediction over language worked, then what is law? Law is structured language: you read enormous numbers of documents, you generate text, and there’s a mountain of precedent to retrieve from. She knows plenty of people would call that nonsense — you’re supposed to work backwards from the customer problem. Her answer is that you do both. What appealed to her about Harvey was the leap in ambition, from answering a question about a California landlord-tenant agreement to doing 85% of the work on an Activision Blizzard M&A, plus a technical reason to believe it could happen.

4. The real danger isn’t being wrong — it’s substituting pedigree for judgment. This is the part that made me sit up. She debated a very good investor friend about a company: what is this thing going to be that would be big? His explanation boiled down to do you know the quality of this person? She’d known the person for eight years and agreed he was excellent — but the technical theory of the business didn’t make sense to her. Her conclusion: a lot of capital is making large-scale research bets without any intuition or opinion about them, proxying judgment to pedigree, to who else invested, to who referred them. She says it plainly: having no point of view on the business beyond the pedigree of the person is dangerous.

5. When raising money, she refused to tell people what they wanted to hear. A private-equity friend told her that LPs need a highly differentiated story, that every part of the funnel should spell out exactly what you’ll do. Her reaction: honestly, I don’t know yet, and I need to raise money so I can go experiment and find out. So she never made those slides. She gave people a two-pager on her background and investing history, claiming she was good at identifying extraordinary people, being useful to them, and genuinely supporting them. Some LPs didn’t like it. She says she’ll always be grateful to the ones who went to their investment committee and wrote, essentially, she’s going to execute like hell — we’ll find out.

6. On open models: the cat is out of the bag. Two layers to her view. The factual one: over three years, open-source models from China, the US and Europe have all become genuinely competitive, and they’re in use everywhere. The policy one: restricting open models inside the United States restricts law-abiding American businesses — the people with adversarial uses aren’t affected by your restrictions. What she wants instead is rigorous safety testing, including on the backdoor behaviors everyone worries about in Chinese models: go find out, rather than speculating about it. And my favorite line of her argument: the economy is enormous, and every individual has use cases no frontier-lab researcher will ever imagine — you can’t imagine the diversity of reality.

7. The compute and energy bottleneck isn’t a technology problem. She talked to the infrastructure lead at one of the hyperscalers, who told her nothing will move the needle at sufficient scale before 2030. She doesn’t think America lacks the technical or entrepreneurial capability. She thinks it’s a regulatory and persuasion problem: to build data centers in New York, you have to convince New Yorkers they want them; to make nuclear competitive as baseload, you have to convince people it’s safe and allow enough construction for the cost curve to bend. The physical supply chain is harder still — tacit knowledge, labor, raw materials, none of which move at software speed. Her verdict: the only way through is through.

Going further

”The news says AI changes everything. Am I buying the top right now?”

I ask myself some version of this every few weeks. The episode doesn’t answer it directly, but the 90-miles-an-hour image gives a useful way to cut it.

Her situation isn’t ours, but the anxiety has the same root: two fears that are both valid and point in opposite directions. Brake and you miss something genuinely large. Don’t brake and you make the same mistake as every prior cycle. “Be careful” doesn’t resolve that, because it doesn’t tell you which way to move.

What I thought when I heard it: whether the trend is real and whether today’s price is right are two separate questions, and the news only answers the first one. Good news lands and the stock falls — usually the news wasn’t fake, it was already in the price. So I’ve started splitting it into two layers: will this happen in three years, and if it does, how much of it does today’s price already assume? A podcast can teach me the first layer. Nobody is going to hand me the second one.

Her formulation — find the truth, find it wrongly priced, then hold the opinion when others haven’t come around — is those two layers plus a third. And most people fail on the third. Not wrong analysis. Couldn’t hold.

”Everyone says this company is great. Can I just follow them in?”

This is the part I could most directly apply to myself. She’s talking about venture, but substituting pedigree for judgment is worse in public markets, not better.

I’ve done it. I bought a stock where my honest reason was “this management team is excellent and everything they’ve touched has worked.” It fell 60%, and the worst part wasn’t the loss — it was realizing I had nothing to reason with about whether to add or leave, because none of my original reasons were checkable. “He’s great” was still true after the drop, so it told me nothing.

Her line — if I don’t understand what they’re doing, I can’t have an opinion on their judgment — translates roughly to: you rate someone’s judgment because you can walk through it yourself. If you can’t, that’s admiration, not analysis.

I gave myself a dumb little check afterwards: before buying, write one sentence about what would make me admit I was wrong. If I can’t write it, or what I write is “if it falls a lot,” then there’s nothing verifiable in my reasoning, and I’m probably riding someone’s résumé. It doesn’t guarantee anything, but it means that on the way down I have something to do besides refreshing headlines for comfort.

”AI is booming — which layer does the money actually land in?”

Her account of compute independence is more honest than most industry write-ups. She works backwards from a data center full of GPUs: cooling, power, training, inference, and every input those require. The shape of it looks a lot like energy independence, and parts of that chain are a very thin sieve located somewhere not necessarily stable or accessible.

The useful part for a retail investor isn’t that conclusion. It’s what she says next: her team has looked hard at data center builders and solar-and-battery installers — and still hasn’t invested. Those businesses sit between operations, a bit of technology, and real estate, and the dominant variable is financing. She’s fundamentally a technology investor; she wants to understand what the durable product asset is.

That’s a good demonstration: seeing the demand and deciding the business is worth backing with your method are two different acts. The same distinction shows up in their internal debates. Venture-backing semiconductor companies was, she says, a god-awful business for a very long time — and what changed isn’t that demand appeared, it’s that demand became consolidated and at scale, with buyers who themselves want supply chain independence. Demand improving isn’t the same as the business improving. The payoff structure has to change.

My own lesson: spot a structural bottleneck and rush to name a stock, and you usually end up owning the most purchasable name rather than the genuinely constrained layer. One extra question helps — does this layer earn its money from technology, from financing, or from relationships? The answers diverge fast.

Worth following up

• The episode itself: Invest Like the Best EP.489 with Sarah Guo, September 1, 2026.
• No Priors, which she co-hosts with Elad Gil — several of the debates referenced here (which markets suit venture, whether models can make money in biology) get longer treatment there.
• Her factual claims are checkable, and worth checking: nuclear cost curves and SMR build progress, open-model evaluation leaderboards, the single-source links in the semiconductor supply chain.
• If you want to pull on the “what makes you trust a person” thread, pair this with any episode about long-term holding. The same question looks completely different on a different time scale.

The one thing to take away

If only one line survives, I’d keep her warning about having no view on a business beyond the person’s pedigree. In my own words: your confidence in something has to come from reasoning you can walk through yourself, not from how impressive the person you trust is. That isn’t only investing. Following a manager’s direction at work, taking a friend’s advice about a life decision — same structure. Trust is good, but trust can’t substitute for understanding, and trust without understanding has nothing to tell you when things go badly.

Here’s something I’ve tried, and it’s simple. Pick one decision you recently made by following someone else’s judgment — something you bought, a direction you took at work, a school you chose, something you started doing because a person told you to. Take a piece of paper and write one sentence: “If I find out in three months that this was wrong, it will be because ______.”

Just that sentence. No reasons, no plan. If you can write it, you did walk through the reasoning yourself. If you can’t, then what you trusted was the person, not the thing. Either is allowed — but you should know which one you’re standing on.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.