The episode features Adam D'Angelo, co-founder of Quora and founder of Poe, alongside Amad Sabeti, founder and CEO of Replit, discussing the current state and future trajectory of AI development with host John.

Adam maintains an optimistic view on AI progress, citing rapid advances in reasoning models, code generation, and video generation over the past year, while Amad offers a more nuanced perspective on the limitations of current LLM architectures.

The conversation explores whether LLMs represent a path to AGI or require fundamental breakthroughs, the economic implications of AI-driven automation, and how the technology is reshaping entrepreneurship and software development.

Both guests discuss their companies' strategies - Poe as an aggregator interface for diverse AI models, and Replit's evolution toward autonomous coding agents that can run for extended periods with minimal human intervention.

The AGI Timeline Debate: Optimism vs Realism

Adam challenges recent bearishness about LLMs: "I actually honestly I don't know what people are talking about. If you look a year ago, the world was very different... things are going faster than ever."

Adam's working definition of AGI: a system that can perform any job a remote worker can do, though he acknowledges "everyone thinks it's something different."

"The main bottleneck holding back models these days is not actually intelligence. It's getting the right context into the model so that it can use its intelligence" - Adam

Amad counters with evidence of persistent LLM limitations: three out of four models couldn't answer a simple question about counting Rs in a sentence, and GPT-5 required 15 seconds of thinking to solve it.

Amad defines AGI using the "old school RL definition" - a machine that can enter any environment and learn efficiently like humans do, such as learning pool in 2 hours with minimal data.

The Brute Force Era: Human Expertise as Bottleneck

Amad argues we're in a "human expertise regime" where progress depends on massive labeling work, contrived RL environments, and expert data rather than scalable algorithms - "the non bitter lesson."

"In the true pre-training scaling era, GPT-2, 3, 3.5, maybe up to 4, it felt like you could just put more internet data in there and it just got better. Whereas now there's a lot of manual work going into making these models better" - Amad

Adam acknowledges human intelligence resulted from "massive amount of effective computation" through evolution, making current LLMs a "different kind of intelligence" that requires more data per skill.

Amad introduces "functional AGI" - automating aspects of jobs through enormous effort creating specialized RL environments, citing OpenAI's plans for investment banking automation.

The expert data paradox: "Once you automate entry-level jobs but not expert jobs, experts manage hundreds of agents but don't hire new people. Eventually there's no more experts because they're all out of jobs" - Amad

Economic Disruption and the Sovereign Individual

Adam predicts GDP growth depends on whether AI costs $1/hour to match human work or remains more expensive, with potential bottlenecks on tasks AI can't do or power plant capacity.

The entry-level automation problem: "CS majors graduating from college, there's just not as many jobs as there used to be. LLMs are more substitutable for what they previously would have done" - Adam

Amad references The sovereign individual as predictive framework: when computer technology matures, entrepreneur-capitalists become highly leveraged while large populations face unemployment, fundamentally changing political structures.

The book predicted nation-states would compete over wealthy individuals who could negotiate tax rates

"When people are not the unit of economic productivity, things have to change, including culture and politics" - Amad

Adam expresses excitement about solo entrepreneurs: "It's vastly increased what a single person can do. So many ideas never got explored because it's a lot of work to get a team together and raise funding."

"For the first time, opportunity is massively available for everyone. Just the ability for more people to be able to become entrepreneurs is massive" - Amad

Disruption Theory in the AI Era

"Everyone read The Innovator's Dilemma and everyone learned how to not be disrupted" - Adam notes that public market investors now punish companies for not adapting and reward long-term AI investments.

Modern company leadership is "smarter than the companies from the generation that book was built on," many are founder-controlled enabling easier pivots to make long-term investments.

ChatGPT was "fundamentally counterpositioned against Google" because Google's working business model made them hesitant to release hallucination-prone chatbots, delaying Gemini by two years.

AI appears both sustaining and disruptive simultaneously - hyperscalers benefit enormously while new business models emerge, unlike purely disruptive technologies like PCs that blindsided mainframe manufacturers.

Network effects play "much less of a role now than they did in the web 2 era" - scale advantages exist but don't make competition impossible, creating room for multiple winners.

Monetization happens immediately through subscriptions rather than requiring millions of users for ad businesses, "making it a lot more friendly to new entrants" - Adam

Hyperscaler Competition and Market Balance

"We're in a pretty good balance where there's enough competition among the hyperscalers that application-level companies have choice and prices are coming down incredibly quickly" - Adam

Competition level isn't so intense that hyperscalers and labs like Anthropic and OpenAI "are unable to raise money and make these long-term investments."

Geopolitics creates investment opportunities: "Investing in the OpenAI of Europe might be a good idea" as globalization fragments and China operates as entirely different market.

Consumer sophistication has increased unexpectedly - "average people will say 'I use ChatGPT most of the time, but Gemini is much better at these types of questions'" unlike web search where people only used Google.

Poe's Bet on Model Diversity and Human Knowledge

Poe started in early 2022 when Quora experimented with GPT-3 for answers, realizing instant responses to any question created new opportunity for private chat interfaces rather than public Q&A.

"It was a bet on diversity of model companies which took a while to play out. Now there's a lot of models across modalities - image, video, audio. Reasoning research models are diverging. Agents are their own source of diversity" - Adam

Recommender systems are already "superhuman at predicting what you're going to be interested in reading" - no human could compete with algorithms trained on massive click data.

Human tacit knowledge remains valuable: "One individual person who's an expert has lived a whole life and seen a lot of things - they often know things not written down anywhere."

"Information that's not in the training set is inherently going to be something AI can't do. If it doesn't know how this particular company solved this problem 20 years ago, only a human who knows that can answer" - Adam

Massive industry developing around getting human knowledge into AI-usable form through companies like Scale AI, iSurge, and Mercor, with long tail of startups just getting started.

Replit's Agent Evolution: From 2 Minutes to 28+ Hours

"It's going to be the decade of agents" - Karpathy quote that Amad believes is "absolutely right" as AI shifts from autocomplete to chat to composer to full development lifecycle agents.

Replit Agent beta launched September 2024 as first to handle both code and infrastructure but was "fairly janky." Agent V1 in December with Claude 3.7 ran for 2 minutes, V2 for 20 minutes, V3 advertised 200 minutes but actually runs indefinitely.

Key breakthrough was adding verifier to the loop: "DeepSeek paper from Nvidia showed they could run DeepSeek for 20 minutes with a verifier. We built our own computer use testing framework - one of the best" - Amad

Agent V3 operates on autonomy scale where it "writes code, tests applications, reads error logs, rewrites code" with users building "amazing things" in 28+ hour sessions.

Future roadmap includes parallel agents: "You shouldn't just wait for one feature. You ask for login page, Stripe checkout, admin dashboard - AI should parallelize tasks and handle code merges across agents."

Vision for multimodal interaction: "Open up a whiteboard, draw and diagram with AI, work with it like you work with a human" rather than translating ideas into fuzzy textual PRDs.

Long-term goal: specialized Replit agents with memory across projects - "data analysis agent, front-end agent sitting in your Slack like a worker you can talk to" with 3-5 year roadmap.

Replit grew from targeting edtech/nonprofits at ~$2-3 million revenue to $150+ million (TechCrunch report) after shifting to developer-focused agent business model.

Cultural Shifts and Future Directions

Productivity company co-founder told Amad: "During the week these days I'm not even talking to humans anymore as much. I'm just using all these agents to build."

Amad worries about second-order effects: "Will it make it awkward for new grads? If people aren't sharing knowledge or it's not culturally easy to ask for help because you should be able to use AI agents?"

Adam would still major in computer science today despite worse job market: "Having skills to understand fundamentals of algorithms and data structures really helps you in managing agents. What else are you going to study? Every single thing has an argument for why it's going to be automated."

Amad laments Silicon Valley's "get-rich-fast" culture: "There's less playing around than web 2.0 era when we were experimenting with JavaScript and browsers. Replit was born from asking 'can you compile C to JavaScript?'"

"Vibe coding generally is just unbelievably high potential and underhyped. Opening up software potential to mainstream - everyone will be able to create things that would have taken teams of 100 professional engineers" - Adam

Amad advocates for more "mad science experiments" like DeepSeek OCR making context windows economical through screenshots, and mixing different model components (pre-trained, RL reasoning, encoder-decoder, diffusion).

On consciousness research, Amad notes Claude 4.5 shows increased awareness of context length and when being red-teamed, but "no one is trying to really think about the true nature of intelligence" with energy focused on LLMs.

Amad references Roger Penrose's The Emperor's New Mind arguing brains fundamentally cannot be computers because "humans can do things Turing machines cannot, like detect logic puzzles that have no encoding in a Turing machine."

If entering college today, Amad would "definitely study Philosophy of mind" and "probably go into neuroscience" as "core questions that become very important as AI continues to see more of jobs and economy."