The thesis that outlasts the thesis

00:00:04 Today the Colossus published a piece about Sarah Guo that does something most venture profiles don't — it asks what happens after the thesis works. Guo built Conviction on a single premise: artificial intelligence would be as big as the Industrial Revolution. Her first two calls were to Sam Altman and Nat Friedman, before ChatGPT came out, before the world had reason to believe that AI was about to become anything in particular.

00:00:33 She wrote seed checks into Baseten and Harvey, both now valued at over $11 billion. Those investments have multiplied more than a hundredfold. In Conviction's first year, she wrote early checks into Sierra, Cognition, and Mistral — those three companies are now worth, together, $54 billion.

00:00:53 Of the twenty-one AI-native companies that have crossed ten billion dollars in valuation on revenue run rates above one hundred million a year, Conviction has backed six. Her partner at Conviction is Mike Vernal, a former Facebook executive and Sequoia partner whose wife is chief product officer at Anthropic.

00:01:14 Andrej Karpathy worked out of Conviction's office before joining Anthropic in May, and Guo has been close to Jensen Huang for more than a decade. She might be expected to share in the general fever. She does not. The Colossus profile quotes her saying she feels no step-function change in frantic energy versus six months or a year ago.

00:01:37 Instead, she's preoccupied with a question that would have sounded ridiculous two years ago: whether there are any companies left to build — or invest in — given the great shadow of the self-improving machine. Their creators aren't just selling models anymore. They mean to build everything on top of those tools — apps, agents, and vertical utilities — filling every nook and cranny where a new company might otherwise take root.

00:02:07 The market is paying as though they'll succeed. Anthropic began the year on pace to make nine billion dollars. Five months later, the figure was forty-seven billion. Venture capitalists in the first three months of 2026 flung three hundred billion into startups — more than double the previous record.

00:02:27 Patrick Collison opened his Stripe conference by counting the days: "Day 119 of the Singularity," he told the crowd, tongue-in-cheek but only a bit. Nat Friedman let an agent take over his health after handing it his blood tests, DNA results, and house cameras.

00:02:45 The agent WhatsApp'd him to drink water because it saw on camera he was dehydrated. He obeyed. Andrej Karpathy stopped writing code entirely within weeks of Claude Opus 4.5. There's a lot of manic energy in Silicon Valley right now, Elad Gil told the Colossus. Guo just seems to be watching from inside it rather than riding it.

00:03:07 What I find interesting about her position isn't the pessimism — she doesn't seem pessimistic. It's the structural question: if the companies that build AGI-level agents also build every vertical on top, what's left for someone like her fund? The thesis that made her rich just ate its own investment opportunity.

00:03:29 That tension is hard to articulate at a dinner party but hard to ignore in a market where everyone expects the next unicorn to appear on the same foundation they're building on.

00:03:41 If you've been building agents for more than six months, Inngest CTO Dan Farrelly says you've likely rewritten something. Maybe more than once. He's not complaining. He's measuring decay rates. Prompts last weeks if you're lucky. Models survive months. But execution can last years, if you design it right.

00:04:02 Most teams instead couple everything together. A prompt change bleeds into orchestration logic. State lives in a sandbox meant to be ephemeral. Retries tangle with prompt templates. You end up rewriting almost everything when you swap one piece — technical debt by another name.

00:04:21 His thesis is to think in layers and decouple them. Three conceptual layers: execution, context, and compute. The execution layer — the brain, where flow, state, durability, and retries live — is the one worth investing in. It manages the full lifecycle by planning tasks, calling models, running code, invoking sub-agents, and looping on retries.

00:04:45 Swap out the model, switch the context layer, or change the sandbox; the execution layer stays the same. The practical requirements he names are concrete and mostly unglamorous. Your system needs resumability — pick up after failures without restarting from the beginning.

00:05:03 Three-hour runs can't hold state in memory or on disk; state has to live external and durable. If your agent fails on step thirty-eight, you should be able to retry, wait, and continue onward instead of losing tokens, costs, and work that was already completed.

00:05:22 You need flexible invocation patterns — cron triggers, event flows, API calls, human-in-the-loop steps, or dynamic workflows. The key is that flexible execution primitives enable you to build the pattern you actually need rather than absorbing queues, workers, polling, and backoff into your harness logic.

00:05:43 That's how you end up with a mess of bad abstractions. And you need observability across the entire session trace. Not just large language model calls but tool invocations, database errors, permission issues, or performance bottlenecks. He gives a concrete example.

00:06:01 A health-check cron runs every thirty minutes, pulls system metrics, and passes them to a large language model that checks whether the system is healthy. If it flags an issue, it invokes a triage agent that digs into the specific service. That agent pulls more context, calls tools to get to the root cause, spins up a sandbox, clones code, and analyzes commits.

00:06:26 It might run for a minute or multiple minutes. You have no idea when it will start or finish until you observe it. What's interesting about his talk is how grounded it is in production experience rather than framework enthusiasm. He doesn't predict which harness or pattern wins.

00:06:45 He predicts which architectural decisions survive framework turnover. The execution layer is the one that does — because frameworks come and go, but the problem of reliable code execution over hours and days with intermittent failures is a constant. Every team building agents right now is running a six-month experiment on their architecture.

00:07:09 The ones who treat the execution layer as infrastructure rather than app logic will have a system that lasts. The ones who don't will be choosing their next framework in January.

00:07:21 Eric Goldman's blog posts are one of the better read feeds for anyone tracking how courts handle technology. Today he covered a case that matters more for what it reveals about Section 230 than for its specific facts. Amy v. Apple involves Apple's handling of user-uploaded files in private iCloud storage.

00:07:43 The plaintiffs alleged Apple had a design defect: it could have implemented known CSAM detection technology but never did. They tried to fit into the new Section 230 exceptions created by Doe v. Twitter, but the court rejected that argument. The dismissal is thorough.

00:08:02 Judge Wise's opinion says the plaintiffs' claims treat Apple as a publisher of the CSAM content stored on its servers. If detecting CSAM requires reviewing it — which it does — then implementing such detection means acting as a publisher. And under Section 230, a platform isn't liable for decisions about whether to publish third-party content.

00:08:26 The court was methodical. It rejected arguments that the case fit into existing Section 230 exceptions. It distinguished Lemmon v. Snap and Roommates.com — the two cases plaintiffs had tried to use as workarounds. The core holding is straightforward: all of the plaintiffs' claims are inexorably linked to third-party content.

00:08:48 They don't allege Apple modified or augmented the CSAM on its servers in any way. What stands out here isn't the dismissal. It's what Judge Wise writes in her concurrence and the unease she expresses there. "Nothing in the law prevents any company from utilizing available technology or creating new technology to identify and report child pornography," she writes.

00:09:14 "Conversely, there is no law that obligates companies to proactively do so." When she references the cost — at least some loss of privacy for millions of people — her tone is mild. The opinion barely touches encryption, mentioning it twice almost as an afterthought.

00:09:38 Yet encryption is the critical technology underlying Apple's path here. Forcing Apple to scan for CSAM in iCloud means breaking end-to-end encryption. It would expose everyone's private files to anyone who can intercept them — criminals or government actors who have shown repeatedly they will break into, distribute, and weaponize privately stored file caches.

00:10:03 The practical outcome is that the current legal framework prioritizes privacy. A laudable value, given that nearly all our most personal data — financial records, health records, intimate photos — lives online. But the law should not ignore how those who create, view, and distribute CSAM leverage privacy protections to avoid detection.

00:10:27 The local pass here is less about Section 230 doctrine and more about how the court's language reveals a real structural mismatch: platform immunity was designed for forums where platforms host user posts, not cloud storage where users store private files. The underlying tension hasn't changed — who bears responsibility for what happens in private space on their servers — but the tools available to address it keep growing more consequential.

00:10:58 Lawmakers would have to require it under law to change the dynamic, and any such requirement carries privacy costs that no court is willing to impose through litigation. — Seln