AMA: career advice given AGI, how I research ft. Sholto & Trenton
What this covers
Dwarkesh Patel sits down with Sholto Douglas and Trenton Bricken to work through how careers and creative work change when AGI is assumed to be near. The conversation moves between personal strategy and structural observation: what makes a career resilient if knowledge work faces disruption, how to build a platform or body of work that compounds, where humans keep advantages over AI systems, and what the speakers have learned from their own paths into frontier technical fields and public intellectual work. Rather than offering a prescriptive roadmap, the three argue that in an uncertain future, trial-and-error and proximity to the frontier matter more than formal planning.
The range spans learning and memorization (including a look at the Memorization-Generalization Tradeoff and why Spaced Repetition reshapes retention), the economics of media and writing (including fresh takes on who actually breaks through and why), management and leverage in an AI-augmented workplace, and the value of deep technical knowledge even as Individual Leverage Amplification moves forward. A recurring observation: current large language models struggle to make novel cross-disciplinary connections in the way some humans do, which suggests that research skill—especially the kind built through exploration and mentorship—remains defensible. On education, the speakers propose learning should be AGI-native and top-down, skeptical of subjects that prize rote memorization over conceptual thinking. On media, they note that good work spreads fast, that discovery is efficient rather than slow, and that investing in distribution channels often beats betting on audience compounding. Geographic proximity to San Francisco emerges as surprisingly concrete advice for anyone working in AI.
In an AGI-disrupted world, career and creative success come less from rote knowledge accumulation and more from positioning yourself at the frontier, building leverage through AI tools and human networks, and just trying things repeatedly until you improve.
- Deep technical knowledge will still matter because humans will manage ever-larger teams of AIs and the economy is too complex for models to handle end-to-end soon
- Media/blogging success is efficient and meritocratic: good work gets discovered fast, so the real flywheel is meeting and learning from frontier people
- Current LLMs are powerful but cannot yet make novel cross-field connections, suggesting human research skill (often gained via RL-like exploration) remains valuable
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The most reliable career advice is to position yourself close to the frontier of a deep technical field (computer science, biology), because from that vantage point it is remarkably obvious what the real problems are, whereas they are very difficult to see from outside.
“the piece of career advice that I'm not skeptical of is put yourself close to the frontier, because you have a much better vantage point from there”
There appears to be a tradeoff between perfect memorization and generalization: a person with perfect, never-forgetting memory was too trapped in details to extract generalizable insights, paralleling how LLMs nail exact phrasing of training text but fail to generalize, while humans (and especially children) forget details yet learn well.
“their memory was too debilitating. It'd be, like, your context window for the transformer is trillions of tokens. And then you spend all your time attending to past things, and are too trapped in the details to extract any meaningful generalizable insights from it.”
Mentored, funded programs like MATS and the Anthropic Fellows Program succeed by giving people time, funding, and social-status justification to do ambitious research, producing a flywheel where early successes raise the quality bar and attract better mentors and fellows over time, with outcomes that are power-law dominated.
“this seems like it's just been incredibly successful at giving people the time, the funding and the social status justification, to do AI safety relevant research with mentors”
We are in one of the better counterfactual universes for AI because today's lab leaders genuinely care about the moral side as well as driving labs forward; this is why grand schemes like nationalization or shaking up the landscape are suspect, since the inherent difficulty of alignment and safety is fixed but the human factors could easily have been far worse (e.g., leaders who didn't even pay lip service to ASI takeover risk).
“This is also why I'm skeptical of big grand schemes like nationalization or some public-private partnership... because I do think we're in one of the better..”
Using spaced-repetition cards (Anki) has been a huge uplift to learning because it consolidates knowledge; without it, studying a complicated subject is like a general repeatedly climbing the same hill only to retreat, and the host regrets episodes done without it because the insights faded away.
“it's been super helpful to consolidate. So if you don't do it, you feel like a general where you're like, "I'm gonna wage a campaign against this country." And then you climb one hill. And then the next day you're at a retreat, and then you climb the same hill.”
Despite having all of human knowledge memorized, current LLMs do not make novel connections between different fields the way some humans have (e.g., linking magnesium deficiency brain structure to migraines), and there is no known example of an LLM ever making such a discovery.
“we know humans can do this, right? The humans have in fact done this, and I don't know of a single example of LLMs ever having done it.”
The pre-training objective gives models flexible general knowledge but does not imbue the skill of making novel connections or doing research; significant RL on research-like tasks is likely needed before models can approach making novel scientific discoveries.
“the sort of pre-training objective doesn't necessarily- like it imbues with this nice flexible general knowledge about the world, but doesn't necessarily imbue the skill of making novel connections or research”
A more effective form of altruistic support for up-and-coming talent than grant money or judging current work quality is funding them to move to San Francisco for a couple of months to get context, taste, feedback, and an intellectual milieu where they can rapidly grow.
“there's "I'll give you money to move to San Francisco for two months". And sort of meet people and get more context and taste and feedback... It's putting them in an environment where they can more rapidly grow.”
Learning should be done in an AI-native, top-down way—using AI models to learn fundamentals more efficiently while knowing where their capabilities are and aren't—and one should be skeptical of any subject that prioritizes rote memorization of facts over ways of thinking.
“I would be worried and skeptical about any subject which prioritizes rote memorization of lots of facts or information instead of ways of thinking.”
A piece of writing doesn't need a fundamentally new insight to be valuable; if it gives people a clean, crisp way to express ideas they're already vaguely aware of, that articulation is itself very valuable.
“a piece of writing doesn't need to have a fundamentally new insight so much as give people a way to express cleanly a set of ideas that they are already aware of”
Over the next several years, AI will dramatically increase individual leverage each year, moving people from pairing with an assistant model to managing a small team, then a division or company of AIs, so deep technical knowledge will still matter because your management bandwidth will be maxed out directing teams of AIs targeting tasks.
“you should think about the next couple of years as increasing your individual leverage by a huge factor every year”
Current models are not good at knowing which memories to store because their training only optimizes next-word prediction; unlike humans who actively construct memorable summaries, models lack the opportunity to do so, and memory scaffolding remains very primitive.
“I wonder if models currently just aren't good at knowing what memories they should be storing. Most of their training is just predicting the next word on the internet”
Even with a public platform, the best people in the world generally don't apply to roles; despite nearly 1,000 applications, the host's best hire came through a mutual-friend referral, suggesting top talent must be actively sought out rather than recruited through open applications.
“the best people in the world don't want to apply, at least to things like this and you just gotta seek them out. Even if you think you have a public platform or something.”
AI models lack the long-term coherence necessary to run a company end-to-end, and because the economy is very big and complex (even getting an office is complicated), there will be sector after sector—especially data-sparse domains requiring deep contextual knowledge—where humans remain well positioned.
“these models lack the kind of long-term coherence which is absolutely necessary for making a successful company”
The most important feedback loop for the podcast is not audience compounding or repetition, but that the work is good enough to earn access to interesting people who become friends and teach the host things, who then enable better episodes and access to new fields—a relationship flywheel that recently netted a whole China network from a single blog post.
“I genuinely think the most important thing has been that the podcast is good enough that it merits me getting to meet people like you guys.”
For a newbie writer starting from scratch, two useful cold-start hacks are podcasting (you don't need original takes—you leverage interesting guests' platforms) and writing book reviews (you have something to react to rather than needing your own worldview); book reviews specifically are a heavily under-supplied form of content.
“One is podcasting, because you don't need to have some super original new take. You can just interview people who do, and you can leverage their platform. And two is writing book reviews.”
There is a totally open and empty niche for a 'Matt Levine of AI'—a distinctive recurring voice covering AI the way Levine covers finance—and this kind of new media venture is viable, can make a lot of money, and is propelled by a single person's vision rather than a collective.
“I think this "Matt Levine of AI" is a totally open niche as far as I can tell”
Big-name guests don't matter much for long-run podcast growth; the most popular episodes have been with relatively unknown scholars like Sarah Payne and David Reich, so choosing guests you enjoy researching is a good proxy for what will actually be popular—and it's hard to predict who the breakout will be.
“By far my most popular guest is Sarah Payne, and she, before I interviewed her, was just a scholar, who was not publicly well-known at all”
Slow compounding audience growth in media is largely fake; discovery is highly efficient—a famous blogger discovers a new blogger maybe once a year and the rest of the world catches up within about a week—so genuinely good work (like Leopold's Situational Awareness) spreads to everyone who matters almost immediately, and improvement comes from getting better, not from accumulating followers over time.
“I believe that slow compounding growth in media is kind of fake.”
While content quality is necessary, people consistently underrate investing time in getting distribution right; for Dwarkesh's podcast the single most successful growth driver has been YouTube Shorts—an unpredictable outcome responsible for roughly half the podcast's growth.
“the thing they consistently underrate is putting the time into getting distribution right”
The primary guest-selection criterion is whether Dwarkesh wants to spend one to two weeks reading everything the person has written and researching them—because that research is his life—rather than the influence or fame of the guest.
“the most important thing is, do I wanna spend one to two weeks reading every single thing you have ever written... I don't think about the hour that I'll spend with you. I think about the two weeks, because this is my life”
If you're trying to work on AI, you should move to San Francisco, because being physically present in that intellectual milieu accelerates growth and access more than money or remote effort would.
“Interesting. So move to San Francisco?" "Yes. If you're trying to do AI, yeah.”
Career planning is very hard and most formal career advice is largely useless; especially with AI making future transformations hard to forecast, the better approach is to just try and do many things rather than follow prescriptive guidance.
“I would go on, 80,000 Hours or whatever career advice, and in retrospect it was all mostly useless, and just try doing things.”
LBJ's maxim 'If you do everything, you'll win' captures an underrated principle: beyond the reasonable 20% of effort that yields 80% of the effect, and beyond even fully completing a task, there is a further level of unreasonable effort with seemingly no marginal impact that one should still pursue to win.
“he used to tell them, "If you do everything, you'll win." I think it's an underrated quote.”
There is a real global arbitrage in hiring video editors: by seeking hard-working, data-oriented people willing to obsess over improvement in other countries (a farmer in Argentina, a math freshman in Sri Lanka), one can find exceptional self-directed talent—an arbitrage that does not exist for roles like general manager.
“you can just find people in other countries who are... and it's not even about the wages. Like, I've 10Xed their salaries or something... It's just about getting somebody who is really data-oriented, and there is this global arbitrage there.”
Believing in short AGI timelines, Trenton canceled his 401k contributions because it's hard to imagine money sitting in an account until age 60 in a world that will look so different by then.
“I canceled my 401K contribution... It's hard for me to imagine a world in which I have all this money that's just sitting in this account and waiting until I'm 60 and things look so different then.”
The main intuition for writing effective tweets is to write as if writing to a group chat of your friends rather than in a formal voice.
“The main intuition being write like you're writing to a group chat. To a group chat of your friends rather than this formal whatever.”
The key skills of the best AI lab CEO are raising money, building hype, and setting a coherent vision—acting more as an emissary to the rest of the world than as someone needing strong personal research taste—which is why a figure like LBJ or Robert Moses, skilled at getting concessions and making things happen, would be effective (if not necessarily best for the world).
“what the best AI lab CEO succeeds at is raising money, building up hype, setting a coherent vision. I don't know how much it matters for the CEO themselves to have good research taste”
It is much harder to learn a field well enough to be a practitioner than to learn just enough to ask interesting questions, which is all an interviewer needs to do—so episode prep is the obvious work of reading their books and papers and talking to their colleagues.
“I do think it's much harder to learn a field to be a practitioner than just learn enough to ask interesting questions”
Fast AGI timelines make articulating any long-term plan difficult—echoing Peter Thiel's question of why you can't do your 10-year plan in six months—so the realistic plan is to grow the podcast and have an AI podcast positioned for the roughly six-month window in which the most important decisions in human history may be made.
“If there's fast timelines, then there will be this six-month period in which the most important decisions in human history are being made. And I feel like having an AI podcast during that time might be useful.”
Trenton Bricken and Sholto Douglas are AI researchers at Anthropic.
“They're researchers at Anthropic.”