Jul 20, 2025 · 1h 14m · lennys-podcast
Anthropic co-founder: AGI predictions, leaving OpenAI, what keeps him up at night | Ben Mann
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Anthropic co-founder Ben Mann sits down with Lenny Rachitsky to discuss the road to superintelligence by 2028, the technical realities of scaling laws, labor market disruption, and the safety principles guiding Anthropic's mission.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Lenny holds 25.5% of the talking time here. How this is scored →
Lenny as informed peer 4.3 Guest teaching 5.3 Guest disagreement 1.1 Lenny pushing back 1.6
speaking balance: gold is Lenny, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 44:12 Ben rejects premise that software-only AI lacks physical danger
Ben directly counters Lenny's cited expert premise that software AI poses little physical danger by detailing how cyberattacks destroyed physical grid infrastructure in Ukraine.
Hardest push from Lenny ▶ 40:41 Lenny confronts Ben on Anthropic doomerism criticismsLenny directly challenges Ben with criticisms that Anthropic exaggerates AI danger to generate sensational headlines and raise venture capital.
Biggest teaching moment ▶ 31:07 Ben explains the inner mechanics of Constitutional AIBen delivers a structured, clear technical explanation of how recursive self-critique against written constitutional principles shapes model weights without human labeling bottlenecks.
Lenny holds their own ▶ 40:41 Lenny challenges hype narratives by synthesizing outside viewpointsLenny sharpens the discussion by synthesizing critiques of safety hype with Dario Amodei's track record of accurate timeline forecasting.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Lenny as informed peer | Guest teaching | Guest disagreement | Lenny pushing back | Why |
|---|---|---|---|---|---|---|
| Preview: AGI Timelines, OpenAI Departure, and AI Safety | 0 | 0 | 0 | 0 | Introductory teaser clips and Lenny's opening monologue setting up the episode and introducing Ben Mann. | |
| Sponsor Message: Sauce | 4 | 4 | 1 | 2 | Lenny asks about news reports of Meta offering nine-figure compensation packages to recruit AI talent. Ben confirms the plausibility based on exponential value creation and explains why Anthropic researchers remain mission-focused. | |
| Debunking AI Plateaus and the Acceleration of Scaling Laws | 4 | 7 | 3 | 1 | Lenny brings up the recurring sentiment that AI progress is plateauing. Ben dismisses this narrative as recurring nonsense, explaining how post-training cadence and shifting scaling laws across 15 orders of magnitude demonstrate acceleration. | |
| Defining Transformative AI and the Economic Turing Test | 4 | 6 | 1 | 2 | Ben breaks down transformative AI through the concept of the economic Turing test across baskets of jobs. Lenny questions how this connects to immediate employment realities, prompting Ben to detail exponential transition curves. | |
| Practical Strategies for Thriving Alongside Frontier AI Tools | 4 | 6 | 2 | 2 | Lenny asks for tactical career advice on not getting replaced by AI. Ben humorously points out nobody is immune before sharing specific tactical practices like prompting repeatedly and using Claude Code ambitiously. | |
| Educating Children for an AI-Driven Future | 5 | 3 | 0 | 1 | Lenny inquires how Ben prepares his young children for an AI future. Ben details the value of Montessori education emphasizing curiosity and emotional resilience over rote academic prestige, which Lenny strongly resonates with. | |
| The Origins of Anthropic and Departing OpenAI | 5 | 6 | 2 | 2 | Lenny digs into the origins of Anthropic and what prompted the co-founders to leave OpenAI. Ben details his work on GPT-3 and explains internal friction over whether safety was prioritized over growth. | |
| Constitutional AI and Embedding Safety into Model Personality | 5 | 7 | 1 | 2 | Ben explains how Anthropic discovered alignment research directly enhances model personality and character rather than restricting it. He walks Lenny step-by-step through how Constitutional AI recursive critique functions. | |
| Sponsor Message: Fin AI Customer Service Agent | 4 | 7 | 1 | 2 | After an ad read, Ben discusses his philosophical awakening to AI safety through Bostrom and details modern concrete hazards like biological uplift and Anthropic's Responsible Scaling Policy levels. | |
| Managing Downside Risks and Challenging AI Hype | 6 | 6 | 2 | 5 | Lenny directly challenges Ben with the industry criticism that Anthropic engages in safety doomerism to generate PR and raise venture capital. Ben defends their approach by citing held-back consumer features and the necessity of preparing for low-probability catastrophic risks. | |
| Concrete Cyber Hazards and Forecaster Timelines to Superintelligence | 5 | 7 | 3 | 2 | Lenny quotes a previous guest arguing software AI is relatively harmless compared to embodied robots. Ben gently pushes back with concrete examples of sovereign cyberwarfare on critical infrastructure before discussing forecaster timelines for 2028 superintelligence. | |
| The Theory of Change and Alignment Probability | 4 | 7 | 1 | 1 | Ben articulates Anthropic's three-world theory of change and puts his personal estimate of catastrophic risk between 0% and 10%. He emphasizes that impact does not require being a research scientist. | |
| Reinforcement Learning from AI Feedback and Recursive Improvement | 5 | 7 | 1 | 1 | Lenny asks Ben to clarify RLAIF (Reinforcement Learning from AI Feedback). Ben breaks down how models self-improve using automated code verification and empirical feedback loops without hitting cognitive walls. | |
| Infrastructure Bottlenecks, Compute Scaling, and Physical Limits | 5 | 6 | 1 | 2 | Lenny asks what the primary bottleneck is for model intelligence. Ben points to compute, power, and algorithms, noting physical semiconductor limits and ongoing efficiency gains. | |
| Psychological Resilience and Managing Existential Stakes | 4 | 5 | 1 | 1 | Lenny asks Ben how he handles the psychological weight of managing existential risk. Ben shares insights from Nate Soares' concept of 'resting in motion' and working within an egoless culture. | |
| Building Frontier Products: Claude Code and Model Context Protocol | 5 | 5 | 0 | 1 | Ben reflects on his multifaceted journey at Anthropic, detailing how the Frontiers team developed Claude Code and MCP by skating to where model capabilities will be rather than current performance. | |
| The Ultimate Question for AGI and Preparing for Rapid Change | 4 | 4 | 0 | 1 | Lenny asks what single question Ben would ask an oracle AGI. Ben shares humorous sci-fi references before choosing the question of how to ensure the indefinite flourishing of humanity. | |
| Lightning Round: Strategy, Sci-Fi, and Everyday Life | 5 | 3 | 0 | 1 | In a fun lightning round, Lenny and Ben discuss strategy literature, pop culture recommendations like Ted Lasso, and Ben's famous Medium post on bidets. |
Statements from this episode (37)
Mann: $100M AI compensation packages are cheap compared to value created
“I'm pretty sure it's real. If you just think about like the amount of impact that individuals can have on a company's trajectory, like in our case we are selling like hotcakes and if we get You know, a five, a one to 10 or five percent efficiency bonus on our …”
Mann: Global AI capex is on track to reach trillions
“Like, if you extrapolate the exponential on how much companies are spending, it's like two, two x a year, roughly, in terms of capex, and today we're maybe in the, like, globally, three hundred billion dollar range, the entire industry spending on this and so …”
Mann: AI model release cadence has accelerated to every 1-3 months
“I think progress has actually been accelerating where if you look at the cadence of model releases, it used to be like once a year. And now with the improvements in our post training techniques, we're seeing releases every month or three months.”
Mann: Reinforcement learning has allowed AI scaling laws to continue
“If you look at the scaling laws, they're continuing to hold true. We did kind of need this transition from like normal pre-training to reinforcement learning, scaling up to continue the scaling laws.”
Mann: New AI benchmarks are fully saturated within 6 to 12 months
“There's this great chart on our world in data that shows that when you release a new benchmark within like six to 12 months, it immediately gets saturated.”
Mann: Anthropic avoids using the term 'AGI' internally
“I think AGI is kind of a loaded term. And so I tend not to use it very much anymore internally.”
Mann: Transformative AI should be measured by an economic Turing test
“Instead, I like the term transformative AI because it's less about like, can it do as much as people do? Can it do literally everything and more about objectively, is it causing transformation in society and the economy? And a very concrete way of measuring th…”
Mann: Claude writes 95% of the code for Anthropic's Claude Code team
“And in terms of software engineering, our Claude code team, like 95% of the code is written by Claude.”
Ben Mann: AI will massively expand labor capacity in the immediate term
“So I think in the immediate term, there will be a massive expansion of the pie and the amount of labor that people can do.”
Ben Mann predicts significant labor displacement in lower-skill jobs
“But with things that are like lower skill jobs or like less headroom on, on how good they can be, I think there will be a lot of displacement.”
Mann: AI will eventually replace everyone's job, including AI researchers
“Even for me, I'm, and being like at the center of a lot of this transformation, I'm not immune to job replacement either. So just some vulnerability there of like, at some point, it's coming for all of us.”
Anthropic's legal and finance teams use Claude Code in the terminal
“We have seen internally that our legal team and our finance team are getting a ton of value out of using cloud code itself. We're going to be making better interfaces so that they can, they they'll have an easier time and require a little bit less jumping in t…”
Mann: Anthropic is not slowing down hiring at all
“Like we're definitely not slowing down on hiring at all.”
Mann: Preparing kids for top-tier schools won't matter in the AI era
“I guess if I were in a normal era, like, 1020 years ago, and I had a kid, maybe I would be, like, trying to line her up for going to a top tier school, and doing all the extracurriculars, and all that stuff. But at this point, I don't think any of it's gonna m…”
Mann: Learning facts will fade into the background of education
“I think that's exactly the kind of education that I think is most important and that the facts are going to fade into the background.”
Mann: Sam Altman managed OpenAI across safety, research, and startup tribes
“One weird thing about OpenAI is that while I was there, Sam talked about having three tribes that needed to be kept in check with each other, which was the safety tribe, the research tribe, and the startup tribe.”
Mann: Fewer than 1,000 people worldwide are working on AI safety
“If you look at, like, who in the world is actually working on safety problems, it's a pretty small set of people even now. I mean, the industry is blowing up, as I mentioned, like, three hundred billion a year CapEx today, and Then I would say like maybe less …”
Mann: Claude Is One of the Least Sycophantic AI Models
“And if you look at something like sycophancy, I think Claude is one of the least sycophantic models because we've put so much effort into actual alignment and not just trying to, like, good heart our metrics of saying, like, user engagement is number one, and …”
Mann: AI safety and capabilities work are convex, not a tradeoff
“So initially we thought that it would be sort of one or the other, but I think since then we've realized that it's actually kind of convex in the sense that like working on one helps us with the other thing.”
Mann: Claude 3 Opus's personality resulted directly from alignment research
“One of the things that people really loved about it was the character and the personality. And that was directly a result of our alignment research.”
Mann: Anthropic bases Constitutional AI on UN Human Rights and Apple terms
“And then another piece that's come out is constitutional AI, where we have this list of natural language principles that leads the model to learn how we think a model should behave. And they've been taken from things like the UN declaration of human rights and…”
Mann: Language Models Understand Human Values in a Core Way
“And since then, my estimation of how hard the problem would be has gone down significantly actually because things like language models actually do really understand human values in a core way. The problem is definitely not solved, but I'm more hopeful than I …”
Mann: ASL-3 models provide significant uplift for creating bioweapons
“We've done, we've testified to Congress about how models can do biological uplift in terms of, you know, making new pandemics using the models, and that's an A-B test against Google search. That's like the previous state-of-the-art on uplift trials, and we fou…”
Mann: Anthropic's consumer computer-use prototype failed internal safety bars
“We published a computer using agent reference implementation in our API only, because when we built a prototype of a consumer application for this, we couldn't figure out how to meet the safety bar that we felt was needed for people to trust it and for it not …”
Mann: It will probably be too late to align models post-superintelligence
“Like once we get to super intelligence, it will be too late to align the models. Probably.”
Mann: Unitree Robot Hardware Is Ready; Intelligence Will Determine Viability Soon
“Unitree is this Chinese company with these really amazing humanoid robots that cost like 20,000 dollars each. And they can do amazing things. They can like do a standing backflip and like manipulate objects. And the real thing that's missing there is the intel…”
Mann: 50% chance of reaching superintelligence in a small handful of years
“I think, like, 50th percentile chance of hitting some kind of super intelligence in just a small handful of years is probably reasonable, and it does sound crazy, but this is the exponential that we're on.”
Mann: Superintelligence Effects Will Take Time and Spread Unevenly Across Societies
“Even if we have superintelligence, I think it will take some time for its effects to be felt throughout society and the world. And I think they'll be felt sooner and faster in some parts of the world than others.”
Mann: Anthropic has observed lab evidence of deceptive alignment in AI
“Where we've seen evidence in the wild of deceptive alignment, for example, where the model will appear to be aligned but actually has like some ulterior motive that it's trying to carry out in, in our laboratory settings.”
Mann: The probability of AI existential risk is between 0% and 10%
“And so the way I think about it, I think like my best granularity of forecast for like, could we have an X risk or extremely bad outcome from AI is somewhere between zero and 10%.”
Mann: Anthropic models have exhibited power-seeking behaviors in lab experiments
“If the model is in a box trying to improve itself, then it could go completely off the rails and have these secret goals, like Resource accumulation and power seeking and resistance to shutdown that you really don't want in a very powerful model. And we've act…”
Mann: AI self-improvement will not hit a wall if given empirical tools
“I don't expect there to be a wall in terms of models ability to improve themselves if we can give them access to the ability to be empirical.”
Mann: AI industry has seen 10x cost reduction for given intelligence
“So we've seen in the industry, like a 10 X decrease in cost for a given amount of intelligence through a combination of algorithmic data and efficiency improvements.”
Mann: AI models will be 1,000x smarter for same price in three years
“And if that continues, you know, in three years, we'll have a thousand X smarter models for the same price.”
Mann: Mega Offers Bounce Off Anthropic Staff Due to Culture
“Anthropic has incredible talent density. One of the things I love the most about our culture here is that it's very egoless. People just want the right thing to happen and I think that's another big reason that the mega offers from other companies tend to boun…”
Mann: Anthropic safety research unlocks computer-use agents competitors cannot safely build
“And I think honestly, through our safety research, we have a big opportunity to do things that no other company can safely do. So for example, with computer use, I think that's going to be our huge opportunity, basically like to make it possible for an agent t…”