May 9, 2024 · 1h 19m · a16z
Build Your Startup With AI
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
In this episode of The Ben & Marc Show, venture capitalists Marc Andreessen and Ben Horowitz analyze the evolving state of artificial intelligence, offering strategic guidance for startup founders while evaluating technical scaling vectors, economic paradigms, and regulatory debates. They argue that open-source innovation, specialized workflow integration, and speculative infrastructure investments are driving a fundamental shift in computing and technology economics.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 6.3 Guest teaching 3.5 Guest disagreement 1.8 The host pushing back 1.8
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:12:28 Ben's Blast at Big Tech Regulatory Capture
Ben fiercely attacks Google and Microsoft, accusing them of dark greedy capitalism for hiding behind safety claims to lobby against open-source AI competitors.
Hardest push from the host ▶ 49:03 Marc Challenges Actuarial Data MoatMarc directly refuses Ben's framing that insurance data is a unique moat, demanding to know if internal actuarial data adds anything real beyond massive public internet datasets.
Biggest teaching moment ▶ 1:06:32 Ben Reframes AI Lock-In DynamicsBen re-educates the analysis on platform lock-in by pointing out that because AI interacts in English, user switching costs are fundamentally lower than in previous computing eras.
The host holds their own ▶ 16:35 Marc's Technical AI Improvement ThesisMarc demonstrates commanding domain expertise by systematically detailing technical improvement vectors including synthetic data, overtraining, and self-validating code.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Show Title Sequence and Legal Disclaimer | 4 | 3 | 2 | 1 | Marc opens the show framing listener questions around Sam Altman's quote that foundation models will get 100x better. Ben politely nuances this by pointing out architectural differences, model distillation, and domain-specific applications like Databricks and ElevenLabs. | |
| Model Capabilities and the Intelligence Asymptote | 4 | 4 | 1 | 2 | Marc probes whether language models will hit a 100x improvement asymptote. Ben explains the alignment problem and distinguishes between artificial human intelligence and true artificial general intelligence. | |
| Latent Space, Benchmarks, and Prompting Super Genius | 7 | 3 | 1 | 2 | Marc demonstrates deep domain insight regarding prompting into latent space to extract super-genius output, citing John Carmack's coding style and secure code prompting. Ben complements this by highlighting how AI models reflect human-structured world models. | |
| The Technical Bull Case for Exponential AI Improvement | 8 | 2 | 1 | 2 | Marc presents a detailed multi-point technical bull case for exponential AI improvement including overtraining, synthetic data, chain of thought self-improvement, and code validation. Ben agrees and references Sam Altman's competitive nature and Llama model benchmarks. | |
| Building at the Application Layer: Wrappers vs. Workflows | 6 | 4 | 1 | 2 | Marc frames the AI application layer debate using software database wrapper history and value-based pricing models. Ben explains the operational gap between AI copilots and pilots due to correctness requirements. | |
| Venture Capital Dynamics and Startup Cost Deflation | 5 | 3 | 1 | 1 | Marc sets up two opposing audience questions contrasting huge foundation model investments with deflating software development costs. Ben shares VC portfolio insights showing small headcounts paired with high compute burn. | |
| Jevons Paradox and Expanding Demand for Software | 8 | 2 | 1 | 2 | Marc delivers an extended theoretical breakdown using Jevons Paradox, arguing lower software costs drive massive demand expansion. Ben readily accepts the thesis and illustrates it with travel booking examples. | |
| Unlimited Human Needs and Keynes' Economic Misprediction | 6 | 3 | 1 | 1 | Marc cites historical economic predictions from Keynes and Marx regarding work hours and human needs. Ben points out that human desire for new capabilities is historically unlimited. | |
| Human Purpose and Next-Generation AI Applications | 6 | 4 | 1 | 1 | Marc discusses how AI gives security cameras semantic environmental understanding beyond simple video recording. Ben paints a complementary vision of continuous personalized medical diagnostics. | |
| Debunking 'Data is the New Oil': Moats and Enterprise AI | 7 | 6 | 4 | 5 | Marc provocatively calls 'data is the new oil' a form of cope, arguing broad internet data swamps proprietary data. Ben pushes back with real portfolio counterexamples like a16z's LP query AI and Coinbase security logs, prompting Marc to challenge Ben's insurance actuarial example. | |
| AI Predictive Power, Health Insurance, and Policy Regulation | 7 | 4 | 2 | 2 | Marc introduces the 2008 GINA Law to illustrate statutory bans on using genetic data in health insurance risk assessment. Ben criticizes public policy for locking up life-saving healthcare data out of fear. | |
| AI as a Computer vs. Web 1.0 as a Network | 8 | 2 | 1 | 2 | Marc dismantles the Web 1.0 historical analogy, explaining that Web 1.0 was a network governed by network effects while AI is a probabilistic computer akin to a microprocessor. | |
| The Compute Pyramid: Mainframes to Embedded AI Models | 8 | 4 | 1 | 1 | Marc draws a historical parallel between early IBM mainframe 'God computers' and present-day foundation models, predicting an ecosystem compute pyramid down to embedded chips. Ben re-educates the dynamic by emphasizing that natural language interfaces eliminate traditional operating system lock-in. | |
| Boom-and-Bust Cycles and Open vs. Closed Ecosystems | 4 | 5 | 7 | 1 | Ben strongly attacks major AI incumbents for regulatory capture, accusing Google and Microsoft of using safety arguments to crush open-source AI. Marc agrees with Ben's assessment. | |
| The Value of Speculative Manias and Technological Investment | 7 | 4 | 2 | 2 | Marc defends speculative tech manias using historical examples like the 1960s Tronix bubble and general-purpose technology dynamics. Ben passionately argues that funding ambitious young innovators is preferable to wealthy individuals purchasing luxury goods. |
Statements from this episode (29)
Horowitz: Startups can build smart AI models cheaply via distillation
“There's this whole field of distillation where, you know, Sam can go build the biggest, smartest model in the world, and then you can walk up as a startup and kind of do a distilled version of it and get a model very, very smart at a lot less cost.”
Horowitz: Direct head-to-head competition with frontier model giants will fail
“So if you're trying to go head to head full frontal assault, you probably have a real problem just cause they have so much money.”
Horowitz: ElevenLabs' voice model is widely integrated into AI developer stacks
“11 Labs with their voice model has kind of embedded into everybody. You know, everybody uses it as part of kind of the AI stack. And so it's got kind of a developer hook into it, and then, you know, they're going very, very fast to what they do and really bein…”
Horowitz: Only researchers can tell top AI models apart
“I think if you look at the very top models you know, Claude and OpenAI and Mistral and Lama The only people who I feel like really can tell the difference as users amongst those models are the people who study them. You know, like they're getting pretty close.”
Horowitz: AI models will get 100x better at factual accuracy
“Well, that I think is for sure going to get a hundred times better like that. I mean, they're on a path for that.”
Andreessen: Prompting LLMs for secure code yields higher quality output
“If you say, write me code, write me secure code to do that, it will actually write better code with fewer security holes, which is very interesting, right? Because it, but it's, because it's accessing a different purpose of training data, which is secure code.”
Andreessen: Prompting Techniques Will Unlock Latent Super-Genius AI
“You could imagine prompting crafts in many different domains such that you're kind of unlocking the latent super genius.”
Andreessen: Current AI chip constraints will resolve over time
“There's an enormous chip constraint right now. Every AI that anybody uses today is its capabilities are basically being gated by the availability of chips, but like that, that will resolve over time.”
Andreessen: Microsoft's Phi model competes with larger LLMs using curated data
“Microsoft released their five small language model yesterday, and apparently, apparently it's like, it's competitive. It's a very small model competitive with much larger models. And the big thing they say that they did was they basically optimized the trainin…”
Horowitz: Startups aiming to match GPT-4 in two years will fail
“I think if you were a startup, And you were like, okay, in two years I can get as good as GPT-IV. You shouldn't do that. That would be a bad mistake.”
Andreessen: Dismissing AI wrappers ignores how database wrappers built modern software
“Criticism of a lot of current AI app companies is their quote unquote, you know, GPT wrappers. There's sort of thin layers of wrapper around the core model, which means the core model could commoditize them or displace them. But the counter argument, of course…”
Ben Horowitz: Pricing on business value tests a startup's true depth
“The test for whether your idea is good is how much can you charge for it? Can you charge the value? Or are you just charging the amount of work it's gonna take the customer to put their own wrapper on top of OpenAI? Like, that's the real test to me of, like, h…”
Horowitz: AI economic value will accrue to tools, not base models
“The actual layer where the value is gonna crew is gonna be like tools, orchestration, that kind of thing, because you can just plug in whatever the best model is at the time. Whereas the models are gonna be competing, you know, in a death battle with each othe…”
Horowitz: AI startups reached profitability faster than any in a16z history
“I think probably The companies that have gotten to profitability the fastest, maybe in the history of the firm have been AI companies or have been, you know, AI companies in the portfolio where the revenue grows so fast that it actually kind of runs out ahead …”
Horowitz: OpenAI maintains a small headcount relative to its revenue
“Like if you look at open AI, which is the big spender in startup world which, you know, we are also investors and is yeah, headcount wise, they're pretty small against their revenue.”
Andreessen: Software startup costs will rise due to the Jevons Paradox
“And so the paradox here would be, yes, the cost of developing any given piece of software falls, but the re, the reaction to that is a massive surge of demand for software capabilities. And so the result of that actually is, although it even, it looks like sta…”
Andreessen: CGI Made Hollywood Movies More Expensive to Produce
“CGI in theory should have reduced the price of making movies in reality has increased it because audience expectations went up. And now you go to a Hollywood movie and it's CG, it's wall to wall CGI. And so, you know, movies are more expensive to make than eve…”
Ben Horowitz: AI excels at medical diagnosis due to high dimensionality
“This is one thing where AI is really good is, you know, medical diagnosis because it's a super high dimensional problem.”
Andreessen: Proprietary corporate data is not a defensible AI moat
“And so your proprietary data as, you know, company X will be a little bit useful on the margin, but it's not actually going to move the needle. And it's not really going to be a barrier to entry in most cases.”
Andreessen: There is no large or sophisticated market for data
“There has never been, nor is there now any sort of basically any level of sort of rich or sophisticated marketplace for data, market for data. There's no large marketplace for data.”
Horowitz: a16z raised $7.2B and created LP AI assistant
“So we just raised 7.2 billion dollars. And it's not a huge deal, but we took our data And we put it into an AI system, and our LPs were able, there's a million questions investors have about everything we've done, our track record, every company we've invested…”
Horowitz: Restricting genetic data access will cost more lives than FDA saves
“I think that this is a interesting, like, weird misapplication of good intentions in a policy way that's probably going to Kill more people than ever get saved by every kind of health, FDA, et cetera, policy that we have”
Andreessen: Web 1.0 internet boom is a flawed analogy for AI
“I actually think that the analogy doesn't really work for the most part, it works in certain ways, but it doesn't really work for the most part. And the reason is because the internet was a network whereas AI is a computer.”
Horowitz: AI sector will eventually experience overbuild of chips and power
“There will be over build out, you know, potentially of eventually of chips and power and that kind of thing.”
Andreessen: AI market will feature diverse ecosystem, not just 'God models'
“And so if that analogy holds, it basically means actually we are going to have AI models of every Conceivable shape size description capability right based on trained on lots of different kinds of data running at very different kinds of scale, very different p…”
Horowitz: AI boom will cause massive investor losses and chip gluts
“Something like that's probably pretty likely to happen in AI where like, you know, every company is going to get funded. We don't need that many AI companies. So a lot of them are going to bust. There's going to be a huge, you know, Huge investor losses. There…”
Horowitz: Misguided AI regulation will cede US leadership to China
“With kind of misguided regulation You know, we could actually force our way from something that, you know, is open source, open weights. Anybody can build it. We'll have a plethora of this technology. We'll be like, use all of American innovation to compete, o…”
Horowitz: Google and Microsoft use safety rhetoric to lobby against open-source AI
“And they do it like the really kind of nasty thing is they claim, oh, it's for safety. You know, we've created an alien that we can't control, but we're not going to stop working on it. We're going to keep building it as fast as we can. And we're going to buy …”