Zheng Dong Wang | Dialectic Summary

1. Key Themes

The "Compute Theory of Everything": Scaling Laws as the Master Narrative

Zheng Dong Wang argues that the entire arc of AI progress reduces to a single compounding trend — compute and scaling — and that once you accept this framing, the trajectory becomes almost inevitable. The difficulty isn't understanding the math; it's emotionally accepting what the math implies.

"The curve is exponential. So if it happens, it doesn't matter when, and then... what does it matter if it's two or 20 years away? The only change that's mattered to my AI timelines is that I used to think it wasn't going to happen in my lifetime. And now I think it is." — Zheng Dong Wang [00:46:23]

Feeling the Wave: The Gap Between Knowing and Believing

A central theme of the episode is why even technically sophisticated people fail to viscerally feel the magnitude of AI progress. Wang argues this is a fundamental feature of human psychology — we normalize change so rapidly that even transformative progress becomes invisible at the margin.

"It actually took me five years of full-time being an AI researcher for me to become this AGI pill, so to say. I felt like it was so hard for me to feel the wave crash over me." — Zheng Dong Wang [00:35:10]

"Like I remember the first time I got into Waymo and I was videoing the Waymo and I'm like, this is the craziest thing that's ever happened to me. I got in the Waymo 60 seconds later, I was on my phone and I had forgotten." — Jackson Dahl [00:36:35]

The "Model Does the Eval" Frame: A Better Lens Than AGI

Rather than debating whether we've reached AGI — a definitionally unstable target — Wang proposes a more tractable frame: everything in AI research reduces to specifying an evaluation and then optimizing it. Progress is continuous, not binary. AGI is not a threshold but a horizon that recedes as we approach it.

"You might even say that the only time AI researchers are doing AI research is when they choose the evaluation. The rest of the time, they're just optimizing a number." — Jackson Dahl quoting Wang [00:13:20]

"The first awesome conclusion of the model does the eval is that we will achieve every evaluation we can state... the second awesome conclusion is that we will fall short on every capability we struggle to state." — Jackson Dahl quoting Wang [00:18:36]

Recursive Self-Improvement: Already Underway, But the Eval Problem Remains

Wang thinks RSI (recursive self-improvement) is closer than most appreciate and will be "easier than people think" — but that achieving it will be less transformative than imagined, because the hard unsolved problem is what to optimize for, not how to optimize.

"RSI is even easier than people make it out to be... the artifact that comes out of it, this AGI that does very well on all of these evals, will also be easier than people think. But that, once we have that, it won't be what it's cracked up to be." — Zheng Dong Wang [00:24:56]

"Jack Clark wrote this post that he thinks based on public information that 60% chance of achieving RSI — so achieving this totally end-to-end frontier research, totally automated — by the end of 2028." — Zheng Dong Wang [00:25:26]

The Myth-Making Deficit: AI's Communication Problem

Wang argues that the AI research community has powerful internal myths (curing cancer, touching the fabric of reality, being on the Manhattan Project) but has failed to develop myths that resonate broadly with the public, policymakers, and workers whose lives will be disrupted.

"If research engineers continue to see myth-making as a chore, second class, a lesser use of their time to quote unquote real technical work, they will keep working in a world under myth they keep complaining is inferior." — Jackson Dahl quoting Wang [00:54:03]

"Fantasy is a natural human activity. It certainly does not destroy or even insult reason... On the contrary, the keener and clearer is the reason the better fantasy will make." — Jackson Dahl quoting Tolkien [00:54:31]

Decentralization as the Best Risk Mitigation

Wang is not a doomer, and his non-doomerism is grounded in a structural argument: the risks he most fears are malicious humans, not unaligned superintelligence. His preferred mitigation is not centralized control but more competition, more diversity of actors, and more decentralization.

"A lot of the risks are very, like humans are very involved in them, like malicious humans using the AI for bad purposes... I think, I think it is consistent and, um, the kind of solution that a vague definition of doomer would support, I feel like would be counterproductive." — Zheng Dong Wang [00:07:27] [01:07:27]

"Competition is generally good... one company's idea of what is moral and good and how much to defer to the user versus how much input the model should have — all of this, I think, competition is generally good." — Zheng Dong Wang [01:04:21]

Evidence Across Time: The Only Way to Truly Feel the Wave

Wang identifies a very specific cognitive mechanism by which people can genuinely internalize AI progress: not being told about a result, but personally setting a test in a domain you know deeply, predicting the outcome, watching AI confound that prediction — and then repeating this over years.

"I had to see so much evidence, like just before my evidence across time... I tried this now. I saw it be not that good two years past. Oh my gosh. So evidence across time, like repeated instances, and tests that I set for myself. So it's not like you're telling me what this test is... I know this problem, I know it would be impressive. You have finger feel on that problem." — Zheng Dong Wang [00:35:38]

The Inevitability-Agency Paradox

Wang holds simultaneously that macro AI progress looks like a law when you zoom out, and that at the micro level it is deeply contingent — every individual researcher is rolling dice. This isn't a contradiction but a scale-dependent truth, and it preserves the meaning of individual effort even in a deterministic-seeming trajectory.

"When you look at the trend of like 2% GDP growth for, you know, since we invented invention or something like that, and if you zoom in really closely and you look really closely, it's like all these people trying very hard and doing very different things every time... And then only when you zoom out, it looks like we call it a law." — Zheng Dong Wang [01:11:10]

2. Contrarian Perspectives

AGI Will Be Easier to Build Than Expected — and Less Impressive When We Get There

The mainstream view oscillates between "AGI is decades away" and "AGI will be transformative beyond imagination." Wang holds a third, less common view: we will achieve RSI/AGI sooner than skeptics think, but the result will be anticlimactic because the hard problem — what should we optimize for? — will remain completely unsolved.

"RSI is even easier than people make it out to be... the artifact that comes out of it, this AGI that does very well on all of these evals, will also be easier than people think. But that, once we have that, it won't be what it's cracked up to be." — Zheng Dong Wang [00:24:56]

Language Was Supposed to Be the Hardest Thing — It Turned Out to Be the Easiest

Early AI researchers, including OpenAI's founding team, were uncertain whether language would be the first or last capability to fall. The prevailing assumption was that physical or game-playing skills would come first. The actual result — that language is strikingly statistically predictable — is both surprising and, Wang implies, slightly deflating about human exceptionalism.

"Maybe, like early research agenda, I think OpenAI published it, they had all these bets early on in the company, including RL, including robotics and things like that. And it wasn't so clear that language would be like the first solve. Maybe it would be the last solve... it's surprising or could be a little depressing just how predictable all of this is — that we could do so much of language... so much of what we do with language is actually like extremely statistically predictable." — Zheng Dong Wang [00:29:40]

Intelligence Itself May Not Be a Real Threshold — Just Search and Data at Scale

Wang suggests that what we call "intelligence" may simply be a label we've placed on the output of large-scale statistical search, and that there is no meaningful discontinuity between "smart" and "not smart" — only compute. This has radical implications for how we think about human specialness.

"There's no threshold of intelligence. And it's just, again, big enough search, big enough data... maybe at the end of the day, we find that all of the things that we humans do that we think are very complicated, it's actually like an extremely low dimensional statistical thing in the entire universe." — Zheng Dong Wang, paraphrased in dialogue [00:31:28]

The AI Industry Has French Revolutionary Tendencies — and Has Earned the Backlash It's Getting

The AI industry's self-conception as bold, inevitable, and on the right side of history is precisely the kind of hubris that historically produces violent resistance. Wang argues this is not just an image problem but a substantive flaw in how the field engages with the rest of society.

"The AI industry has all kinds of French revolutionary tendencies and our own self-conception: we're bold, inevitable, and on the right side of history. At our worst, we're ignoring our inheritance to remake the world from abstract, rational personal principles, dismissive of accumulated experience and impatient that no one else is keeping up. I think we're meeting a resistance to that impulse that is earned." — Jackson Dahl quoting Wang [01:03:24]

Doomerism's Preferred Solution Is Counterproductive Even If the Risks Are Real

Wang doesn't dismiss AI risks; he reframes them. He thinks centralized control — which is what most doomer policy prescriptions implicitly require — is itself the bigger long-run risk. The decentralized, competitive, pluralistic alternative is messier but superior precisely because the risks are mostly about malicious human actors, not rogue superintelligence.

"I put a much greater weight on these, like, more human-being-involved kind of risks and that the best solution to them is, you know, very decentralized, very like a lot of good people, in a very decentralized way thinking about... yeah. You can't consistently hold that like this AI thing is a scam and also like, oh please stop disrupting my field." — Zheng Dong Wang [01:08:49]

3. Companies Identified

Notion

Description: Collaborative workspace and productivity software, rebuilt from the ground up for the AI era. Why mentioned: Episode presenting partner; highlighted for its vision of "malleable software" — tools that users can shape to fit their needs — and for extending this to AI agents and teams.

"Notion's ultimate goal is stated by founder and CEO Ivan Zhao, is that it helps us think together in the long run as we have more and more capability." — Jackson Dahl [00:04:21]

Waymo

Description: Autonomous vehicle company operating robotaxi services. Why mentioned: Used as a concrete example of how humans normalize transformative technology almost instantly, undermining the "macro moment" theory of how people come to feel the wave.

"I remember the first time I got into Waymo and I was videoing the Waymo and I'm like, this is the craziest thing that's ever happened to me. I got in the Waymo 60 seconds later, I was on my phone and I had forgotten." — Jackson Dahl [00:36:35]

OpenAI

Description: AI research and deployment company, creator of GPT series and ChatGPT. Why mentioned: Referenced for its early research agenda showing uncertainty about which modality (language, RL, robotics) would be solved first; also referenced regarding GPT-5 and government access restrictions.

"I think OpenAI published it — they had all these bets early on in the company, including RL, including robotics and things like that. And it wasn't so clear that language would be like the first solve." — Zheng Dong Wang [00:29:40]

4. People Identified

Demis Hassabis

Description: Co-founder and CEO of Google DeepMind, Nobel laureate. Why mentioned: Cited for his Nobel lecture framing — that classical algorithms given sufficient scale can solve vast categories of problems — and for keeping the field focused on the genuine enormity of what remains unknown in the universe.

"Demis does a really good job of like keeping us focused on these stakes of like, even that is like so small or just like doesn't pale in comparison to the true nature of reality, which is like, we have no idea what's going on in this whole universe." — Zheng Dong Wang [00:32:53]

Jack Clark

Description: Co-founder of Anthropic, former OpenAI policy director. Why mentioned: Cited for a specific published probability estimate on recursive self-improvement, which Wang treats as a credible public data point.

"Jack Clark wrote this post that he thinks based on public information that 60% chance of achieving RSI — so achieving this totally end-to-end frontier research, totally automated — by the end of 2028." — Zheng Dong Wang [00:25:26]

Andrej Karpathy

Description: AI researcher, former Tesla AI director and OpenAI founding member. Why mentioned: His "auto research" project cited as an early, basic version of automated AI research — an early instance of the RSI spectrum.

"Andrej Karpathy's auto research is like a very basic version of this." — Zheng Dong Wang [00:26:53]

David Ha

Description: AI researcher, former head of research at Stability AI, previously at Google Brain. Why mentioned: Praised for a tweet — "AI research is just applied philosophy" — that Wang found deeply resonant as a framing of the field's true scope.

"One time David Ha tweeted, AI research is just applied philosophy. And I just really, really liked that tweet. It's a shame he doesn't tweet anymore." — Zheng Dong Wang [00:16:18]

Leopold (Aschenbrenner)

Description: Former OpenAI researcher, author of the "Situational Awareness" essay series. Why mentioned: His long-form PDF on superintelligence is cited, specifically the section on "superintelligence harness" and the acknowledgment that the future training signal will be "something kind of like RLHF, but TBD."

"I was just rereading the Leopold PDF and he gets to the superintelligence harness part and he's kind of just like, I think we're going to come up with something kind of like RLHF, but like TBD." — Jackson Dahl [00:24:03]

Scott Alexander

Description: Blogger (Astral Codex Ten), rationalist writer. Why mentioned: Cited for a post imagining future historians treating everyone even marginally connected to the creation of powerful AI as celebrities — used to illustrate that the personal and local is historically significant and always will be.

"Scott Alexander, I think, has written a very good post that makes you feel this particular wave crashing over you of like, imagine in the future, if there are like people in other galaxies and they think they read all the history of what was the creation of powerful AI like. And they're like every single character that was even marginally related is like a celebrity." — Zheng Dong Wang [01:19:22]

Laura Deming

Description: Longevity venture investor, founding partner of The Longevity Fund. Why mentioned: Her essay "The Rage of Research" is recommended on Wang's blog as capturing the obsessive, compelled spirit of great research.

"On your blog, you recommend Laura Deming's The Rage of Research... It captures that spirit that just like, I have to." — Jackson Dahl [00:08:07]

Jeffrey Litt

Description: Described as a guest on Dialectic (episode 21), subsequently joined Notion. Why mentioned: Cited for articulating the concept of "malleable software" — software that should feel like a craftsperson's workshop, shapeable to the user's needs.

"One of my favorite articulations of Notion comes from an idea from Jeffrey Litt, who is episode 21 on the podcast. And shortly after I interviewed him actually joined Notion — and that is malleable software." — Jackson Dahl [00:03:29]

Ivan Zhao

Description: Founder and CEO of Notion. Why mentioned: Cited for Notion's stated long-run mission of helping humans think together as collective capability increases.

"Notion's ultimate goal is stated by founder and CEO Ivan Zhao, is that it helps us think together in the long run as we have more and more capability." — Jackson Dahl [00:04:21]

Jasmine Sun

Description: Writer/researcher; mentioned in passing during the conversation. Why mentioned: Referenced for a vivid line about skyscrapers appearing "like mushrooms" in rapidly developing Chinese cities — used to illustrate how sustained, compounding change can eventually be felt.

"Jasmine Sun was at this line, somebody sent in about like the skyscrapers popping up like mushrooms." — Jackson Dahl [00:37:29]

Rebecca Lowe

Description: Philosopher, blogger. Why mentioned: Cited for a recent blog post expanding the definition of "work" beyond tasks and jobs to encompass purpose, meaning, and contribution to something larger — directly relevant to what automation means for human identity.

"Recently Rebecca Lowe wrote this blog post — she's a philosopher — about some broader definition of work where this idea of like replacing jobs, but what even is a job?" — Zheng Dong Wang [01:20:42]

Peter Thiel

Description: Co-founder of PayPal and Palantir, venture investor. Why mentioned: Cited for a framing about Elon Musk — that a world with a billion robots and massive national debt are not mutually exclusive outcomes — used to illustrate that technological abundance and financial crisis can coexist.

"Teal's point is that Elon's talking about we're gonna have billion robots and we're going to have all of these national debt problems, and Teal is critiquing those two things aren't mutually exclusive." — Jackson Dahl [00:44:30]

Elon Musk

Description: CEO of Tesla and SpaceX, owner of X. Why mentioned: Referenced in the context of the Thiel critique about robots and national debt coexisting.

Referenced in Jackson Dahl's discussion of the Thiel quote [00:44:30]

5. Operating Insights

Choose Better, Not More: Selection Quality Beats Work Volume in Fast-Moving Fields

Wang draws a sharp conclusion from observing who has done best in AI research over the past five to seven years: it is not the people who did the most, but those who chose best. In a field with infinite things to try, the meta-skill is curation and prioritization — knowing what to skip is more valuable than stamina.

"It's chosen better, just because of the amount of different things you could try all the time. Even if you're not doing research and you're just scrolling through Twitter, or looking at all these papers, you defer a lot of curation to people telling you what to read." — Zheng Dong Wang [01:10:34] [00:10:34]

Set Personal Tests in Your Domain of Expertise — Then Watch AI Confound Them Over Time

Wang identifies a specific, replicable practice for internalizing technological change that is more reliable than reading about it or being told about it: pick a hard problem you know intimately, predict what AI will do with it, observe the result, and repeat over years. This is the mechanism by which he went from skeptic to convinced, and it transfers to any domain.

"You need to pick your own test. It should be a problem you know well, when you've worked on for years and one you're supposed to be an expert in. You need to predict the results year after year, then you need to watch AI confound those predictions anyway." — Jackson Dahl quoting Wang [00:33:47]

Being in the Conversation Is a Prerequisite — Even If the Conversation Is Wrong

For anyone trying to enter or influence a fast-moving field, Wang argues that participation in the current discourse — even a discourse that is memetically flawed or directionally wrong — is a necessary first step before you can introduce new ideas. You cannot shortcut to influence by staying outside the conversation.

"Just getting yourself into the conversation, whatever that might be, even if it's now just being on Twitter instead of publishing journal articles or things like that, is an important first step." — Zheng Dong Wang [00:11:28]

6. Overlooked Insights

The Eval-Specification Problem Is the Real Bottleneck — and It Will Absorb More Human Labor, Not Less

Wang makes a point that passes quickly but has large implications for anyone thinking about where value accrues in an AI-abundant world: as AI automates task execution, the scarce and valuable activity shifts entirely to specifying what to optimize for. This means the profession of "writing good evals" — or more broadly, translating human values and goals into precise, machine-actionable specifications — is set to become one of the most economically important activities in existence. Wang notes that AI agents will eventually join humans in this task, but for now it is a growing human bottleneck.

"A bunch of humans are trying very hard at this right now. More and more humans will join the fields of like specifying things just like, for all of history, we've been trying to specify things. Soon there will be more AI agents joining in this task of like looking at all the things and trying to specify them better." — Zheng Dong Wang [00:23:10]

The investment implication is direct: companies and tools that help humans write, test, and iterate on evals — or that make high-quality specification more accessible to non-technical domain experts — are positioned at exactly the bottleneck Wang is describing.

The "Personality Hire" May Become the Last Defensible Human Job Category

In a brief, almost throwaway aside, Wang introduces a concept that cuts to the heart of what work survives automation. He describes a class of roles he calls "personality hires" (explicitly non-derogatory) — people whose value to an organization is fundamentally relational and emotional: they raise colleagues' spirits, create belonging, and make people feel like they live more fulfilled lives. Wang's point is that this category of contribution is structurally very hard to specify as an eval, which means it is structurally hard to automate. In a world where nearly everything else can be optimized by a machine, the human who makes the room feel better may be irreplaceable precisely because their value cannot be quantified.

"If you're like a personality hire, non-derogatory, then it's really hard to specify, but you'd be doing a good job if you are charismatic and your colleagues enjoy working with you and you raise the spirits of everyone and everyone just feels like they live a more fulfilled life. I think this is also hard to specify and maybe your performance review at the end of the day justifiably is just like, how happy did you make everyone? How much does everybody like you?" — Zheng Dong Wang [00:21:49]