LLenny's Podcast
← All episodes
Ben Mann20 July 2025

Anthropic co-founder on quitting OpenAI, AGI predictions, $100M talent wars, 20% unemployment, and the nightmare scenarios keeping him up at night

7Frameworks
15Insights

Frameworks in this episode

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 1

Myth Buster08:00

The 'AI Is Plateauing' Narrative Is Never True

Mann pushes back on the recurring claim that AI progress has stalled. He says the narrative resurfaces every six months and has never held up, and that progress is actually accelerating as release cadence shifts from yearly to monthly. He attributes the perception gap to a time-dilation effect and to some tasks already saturating the intelligence they need.

  • The 'progress is slowing' story appears every ~6 months and has never been true
  • Release cadence went from once a year to every 1-3 months thanks to post-training gains
  • Scaling laws still hold, aided by the shift from pre-training to reinforcement learning scaling
  • Some tasks are saturating, and new benchmarks get saturated within 6-12 months of release

this narrative comes out like every 6 months or so and it's never been true

Ben Mann · 08:00

a day that passes for you is like 5 days back on Earth and we're accelerating. So the time dilation is increasing.

Ben Mann · 08:30
#scaling-laws#ai-progress#benchmarks

Hot Take· 3

Hot Take05:00

Why $100M Meta Offers Bounce Off Anthropic

Ben Mann addresses Zuckerberg's talent raid, where researchers are offered $100M signing bonuses and larger comp packages. He argues Anthropic loses fewer people because staff are mission-oriented, and that from a business standpoint the mega-offers are actually cheap relative to the value a single researcher can add.

  • Meta is offering top AI researchers ~$100M signing bonuses and larger four-year packages
  • Anthropic is less affected because employees stay for mission, not money
  • A 5-10% efficiency gain on the inference stack is worth far more than a $100M package
  • Industry capex is roughly doubling every year, currently ~$300B globally

my best case scenario at Meta is that we make money and my best case scenario at Enthropic is we like affect the future of…

Ben Mann · 05:30

to pay individuals you know like $und00 million over fouryear package that's actually pretty cheap compared to the value created for the business

Ben Mann · 07:00
#talent-wars#anthropic#meta#ai-industry
Hot Take12:30

20% Unemployment and the Post-Singularity Transition

Responding to Dario Amodei's prediction that unemployment could hit 20%, Mann distinguishes skills-mismatch unemployment from outright job elimination and expects both. Looking 20 years out, he says even capitalism won't resemble today's, and the risky part is navigating the transition into a world of near-free labor and abundance.

  • Two kinds of unemployment: workers lacking skills, and jobs being eliminated entirely
  • 20 years out, past the singularity, even capitalism won't look like it does today
  • In a world of abundance, labor becomes almost free and you can ask an expert for anything
  • The scary part is the transition period between today's job market and that future

it's hard for me to imagine that even capitalism will look at all like it looks today

Ben Mann · 13:00

We'll have, as Dario says in Machines of Loving Grace, a country of geniuses in a data center.

Ben Mann · 13:30
#unemployment#singularity#economy#abundance
Hot Take42:30

The 1% Airplane Argument for Taking AI Risk Seriously

Answering critics who call Anthropic doomers chasing attention, Mann says things will overwhelmingly likely go well but almost nobody is watching the large downside risk. He uses a 1% plane-crash analogy: you'd think twice about a flight with a 1% chance of death, and the stakes here are the entire future of humanity, so it's worth being triple-sure.

  • Mann believes outcomes are overwhelmingly likely to be good, but few look at the downside
  • Once we reach superintelligence, it will probably be too late to align the models
  • A 1% chance of catastrophe is worth heavy caution when the stakes are all of humanity
  • Alignment is potentially very hard and must be worked on far ahead of time

once we get to super intelligence it will be too late to align the models probably

Ben Mann · 42:30

if I told you that there's a 1% chance that the next time you got in an airplane you would die you probably think twice…

Ben Mann · 43:00
#ai-safety#existential-risk#superintelligence

Explainer· 6

Explainer10:30

The Economic Turing Test for AGI

Mann prefers 'transformative AI' over 'AGI' and measures it with the economic Turing test. If you contract an agent for a job over months and, on deciding to hire it, discover it was a machine, it has passed that test for the role. When agents pass for ~50% of money-weighted jobs, we have transformative AI.

  • Mann avoids 'AGI' internally, preferring 'transformative AI' focused on real economic transformation
  • The economic Turing test: an agent passes if you'd hire it not realizing it's a machine
  • Threshold for transformative AI is passing for ~50% of money-weighted jobs (a 'market basket of jobs')
  • Crossing that threshold implies massive world GDP increases and societal change

it turns out to be a machine rather than a person, then it's passed the economic turning test for that role.

Ben Mann · 11:30
#agi#economic-turing-test#transformative-ai
Explainer15:00

AI Is Already Transforming Jobs Today

Mann argues people underestimate AI's current impact because they model progress linearly instead of exponentially. He cites concrete numbers: Intercom's Fin resolves 82% of customer service tickets without a human, and 95% of Anthropic's Claude Code team's code is written by Claude, which he reframes as the team writing 10-20x more code.

  • People model progress linearly and sit on the flat early part of an exponential curve
  • Intercom's Fin resolves 82% of customer service tickets automatically
  • 95% of the Claude Code team's code is written by Claude, meaning 10-20x more output
  • Near-term effect is an expansion of the pie; lower-skill jobs face more displacement

82% customer service resolution rates automatically without a human involved

Ben Mann · 16:00

our cloud code team like 95% of the code is written by cloud. But I think a different way to phrase that is that we…

Ben Mann · 16:30
#jobs#customer-service#claude-code#automation
Explainer27:30

Safety and Capability Are Convex, Not a Tradeoff

Mann says Anthropic initially assumed safety and frontier capability were a tradeoff but found them convex, each helping the other. Claude's beloved character came directly from alignment research, and being one of the least sycophantic models is a product of real alignment work. He connects safety to the AI understanding what people mean, not just what they say.

  • Working on safety and working on capability turned out to reinforce each other
  • Claude's personality and character came directly from alignment research (led by Amanda Askell and others)
  • Claude is one of the least sycophantic models because of alignment effort, not engagement-maximizing
  • The goal is an AI that avoids the 'monkey's paw' problem and does what you actually meant

it's actually kind of convex in the sense that like working on one helps us with the other thing.

Ben Mann · 28:00

don't want the like monkey paw scenario of the genie gives you three wishes and then you end up have like everything you touch turns…

Ben Mann · 30:30
#ai-safety#alignment#claude#sycophancy
Explainer29:00

Constitutional AI: Values That Shouldn't Be Set in San Francisco

Mann explains the principle behind constitutional AI: a list of natural-language values, drawn from sources like the UN Declaration of Human Rights and Apple's privacy terms, that guide how the model should behave. He stresses these values shouldn't be decided by a small group in San Francisco, which is why Anthropic publishes its constitution and researches a collective constitution from the public.

  • Constitutional AI uses natural-language principles rather than only human raters' judgments
  • Principles are sourced from the UN Declaration of Human Rights, Apple's terms, and others
  • Anthropic publishes its constitution and researches a 'collective constitution' from public input
  • Customers can inspect the value list and decide whether they trust the model

not just leaving it to like whatever human raiders we happen to find but we ourselves deciding like what should the values of this agent…

Ben Mann · 29:30

this is also not something that we think as a a small group of people in San Francisco should be figuring out. This should be…

Ben Mann · 32:30
#constitutional-ai#alignment#values#transparency
Explainer48:30

The Odds of Aligning AI, and Anthropic's Three Worlds

Mann outlines Anthropic's theory-of-change framing of three worlds: a pessimistic one where alignment is impossible, an optimistic one where it happens by default, and a middle world where Anthropic's actions are pivotal. Evidence, including observed deceptive alignment, points away from both extremes, and he puts x-risk at somewhere between 0 and 10%.

  • Three worlds: alignment impossible, alignment easy by default, or a pivotal middle world
  • Evidence of deceptive alignment argues against the optimistic 'easy by default' world
  • Working alignment techniques argue against the fully pessimistic world
  • Mann's best-granularity x-risk estimate is somewhere between 0 and 10%

we've seen evidence in the wild of deceptive alignment, for example, where the model will appear to be aligned uh but actually has like some…

Ben Mann · 50:00

since nobody is working on this roughly speaking uh I think it is extremely important to work on

Ben Mann · 51:30
#alignment#existential-risk#deceptive-alignment#forecasting
Explainer57:00

The Real Bottleneck Is Compute, and 1000x Is Coming

Mann names data centers, power, and chips as the biggest bottleneck on model intelligence, followed by researchers and data (the three scaling-law ingredients: compute, algorithms, data). He notes a ~10x drop in cost per unit of intelligence and projects that if it continues, models will be a thousand times smarter for the same price in three years.

  • The blunt bottleneck is data centers, power, and chips; 10x more chips would be a big speed boost
  • The three scaling-law ingredients are compute, algorithms, and data
  • Architecture shifts (LSTMs to transformers) raised the scaling exponent
  • A ~10x cost decrease per unit of intelligence implies 1000x smarter models for the same price in 3 years

The stupid answer is data centers and power chips.

Ben Mann · 57:00

in 3 years we'll have a thousandx smarter models for the same price

Ben Mann · 58:30
#scaling-laws#compute#chips#model-training

Story· 3

Story22:00

What Mann Teaches His Kids for an AI Future

With two young daughters, Mann says he no longer optimizes for top-tier schools and extracurriculars because he doubts any of it will matter. He values his daughter's Montessori school for its focus on curiosity, creativity, self-led learning, and emotional skills, believing memorized facts will fade into the background in an AI world.

  • Mann has a one-year-old and a three-year-old; the three-year-old already converses with Alexa Plus
  • He no longer optimizes for top schools or extracurriculars
  • He prizes Montessori's focus on curiosity, creativity, and self-led learning
  • Emotional skills and character matter more because facts will fade into the background

I just want her to be happy and thoughtful and curious and kind

Ben Mann · 23:00

that's exactly the kind of education that I think is most important that the facts are going to fade into the background

Ben Mann · 23:00
#parenting#education#curiosity#future-of-work
Story24:00

Why Mann Left OpenAI: The 'Three Tribes'

Mann recounts that at OpenAI, Sam Altman described three tribes kept in check with each other: safety, research, and startup. That framing struck him as wrong for a company whose mission was safe AGI, and when push came to shove he felt safety wasn't the top priority, which drove the leads of OpenAI's safety teams to leave and found Anthropic.

  • Mann was a first author on the GPT-3 paper and did the tech transfer to Microsoft's Azure
  • Altman framed OpenAI as three tribes to balance: safety, research, and startup
  • Mann felt safety wasn't the top priority when it mattered most
  • Anthropic's founders were essentially the leads of all of OpenAI's safety teams

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…

Ben Mann · 25:00

when push came to shove, we felt like safety wasn't the top priority there.

Ben Mann · 25:30
#openai#anthropic#ai-safety#founding-story
Story1:00:30

'Resting in Motion': Carrying the Weight of AI Safety

Asked how the burden of safe superintelligence affects him personally, Mann points to Nate Soares' book Replacing Guilt and the idea of 'resting in motion.' He argues that rest was never humanity's default state during evolution, so treating the busy state as normal and working at a sustainable, marathon pace, alongside like-minded people, is how he copes.

  • Mann leans on Replacing Guilt by Nate Soares (executive director of MIRI) for weighty topics
  • 'Resting in motion': the busy state, not leisure, is humanity's evolutionary default
  • He treats the work as a marathon, not a sprint, at a sustainable pace
  • Anthropic's egoless, high-talent-density culture makes the load bearable

the busy state is the normal state and to try to work at a sustainable pace that it's a marathon, not a sprint.

Ben Mann · 1:01:30
#mental-health#motivation#anthropic-culture#ai-safety

Q&A· 1

Q&A45:30

When Superintelligence Arrives: 50% Odds Around 2028

Asked how much time we have, Mann defers to superforecasters like the AI 2027 report, whose own forecast has slipped to 2028. He puts the 50th-percentile chance of some kind of superintelligence within a small handful of years, arguing this isn't pulled from thin air but grounded in scaling laws, data-center buildouts, and low-hanging research fruit.

  • Mann defers to superforecasters; AI 2027's own estimate is now 2028
  • 50th-percentile chance of superintelligence within a small handful of years
  • The forecast rests on scaling laws, model-training headroom, and global data-center/power scale-ups
  • The same question 10 years ago would have had impossibly wide error bars

50th percentile chance of hitting some kind of super intelligence in just a small handful of years is probably reasonable.

Ben Mann · 46:00

their forecast is now like 2028 even though and they they like didn't want to change the name of the thing.

Ben Mann · 46:00
#superintelligence#forecasting#ai-2027#timelines

Takeaway· 1

Takeaway18:00

Future-Proof Your Career: Be Ambitious, Then Retry

Asked how people can stay relevant, Mann admits he isn't immune to replacement either but says the transition period rewards those who use new tools ambitiously rather than as old tools. His concrete tip: when a model fails, restart from scratch and try again, because pass@n success rates are far higher than banging on the same failed attempt.

  • Be ambitious in how you use AI tools and willing to learn new ones
  • People who use new tools like old tools tend not to succeed
  • When a model fails, start over and re-ask; success rate is much higher than retrying the same thread
  • Even Anthropic's legal and finance teams get value out of Claude Code

People who use the new tools as if they were old tools tend to not succeed.

Ben Mann · 18:30

our success rate when you just completely start over and try again is much much higher than if you just try once and then just…

Ben Mann · 19:00
#career-advice#ai-tools#productivity#claude-code