Skip to content
a16za16z

Amjad Masad & Adam D’Angelo: How Far Are We From AGI?

Adam D’Angelo (Quora/Poe) thinks we're 5 years from automating remote work. Amjad Masad (Replit) thinks we're brute-forcing intelligence without understanding it. In this conversation, two technical founders who are building the AI future disagree on almost everything: whether LLMs are hitting limits, if we're anywhere close to AGI, and what happens when entry-level jobs disappear but experts remain irreplaceable. They dig into the uncomfortable reality that AI might create a "missing middle" in the job market, why everyone in SF is suddenly too focused on getting rich to do weird experiments, and whether consciousness research has been abandoned for prompt engineering. Plus: Why coding agents can now run for 20+ hours straight, the return of the "sovereign individual" thesis, and the surprising sophistication of everyday users juggling multiple AIs. Timestamps 00:00 Introduction 00:41 The Bearishness Paradox: "I don't know what people are talking about" 04:25 "Functional AGI" - Brute Forcing Your Way to Automation 11:18 "We are in a human expertise regime" 15:31 The Weird Equilibrium: Automating Entry-Level but Not Experts 17:22 The Expert Data Paradox 24:44 The Sovereign Individual: A Prediction Framework for the AI Era 28:51 "Vastly increased what a single person can do" 45:04 "It's gonna be the decade of agents" 49:01 Managing Tens of Agents in Parallel 52:56 "I actually think vibe coding is unbelievably high potential" 58:47 Claude 4.5's Strange New Awareness Resources Follow Amjad on X: https://x.com/amasad Follow Adam on X: https://x.com/adamdangelo Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://x.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Podcast on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Podcast on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.

Amjad MasadguestErik Torenberghost
Nov 7, 20251h 2mWatch on YouTube ↗

CHAPTERS

  1. 0:00 – 0:45

    The pace of progress: why “nothing is fundamentally that hard”

    The conversation opens with a bold claim: with enough top talent focused for a few years, many seemingly hard problems can be solved. This frames the episode’s central tension—whether current AI progress is merely impressive automation or a path to something like AGI.

    • AI progress is framed as a near-term solvable engineering/research challenge
    • Sets an optimistic baseline for how fast breakthroughs can occur
    • Introduces the theme of a world-changing technology shift
  2. 0:45 – 2:42

    The “bearishness paradox”: why some see limits while others see acceleration

    Erik raises the recent skepticism about LLMs (limitations, slowing timelines, not end-to-end job automation). Adam pushes back, arguing the last year’s improvements in reasoning, code generation, and video suggest acceleration rather than stagnation.

    • Recent public narrative: LLMs are hitting ceilings / AGI timelines slipping
    • Counterpoint: reasoning models and multimodal gains indicate rapid progress
    • Key bottleneck proposed: getting the right context into models, not raw intelligence
    • Computer-use as a near-term unlock for broader automation
  3. 2:42 – 7:54

    Defining AGI: “remote worker” vs. “learn-anything-in-any-environment”

    The guests surface a core problem: AGI definitions vary wildly, so debates often talk past each other. Adam proposes an operational definition—AI as good as a typical remote worker—while Amjad prefers a classic RL framing: fast, efficient learning in novel environments like humans.

    • AGI definitions are inconsistent; clarity matters for forecasting
    • Adam’s anchor: better-than-typical remote worker at remote-eligible jobs
    • ASI distinction: better than best humans/teams at everything
    • Amjad’s anchor: efficient in-environment learning with minimal data (human-like adaptation)
  4. 7:54 – 11:18

    Are we brute-forcing ‘functional AGI’ instead of cracking intelligence?

    Amjad argues today’s gains rely on heavy human effort: labeling, contrived RL environments, and infrastructure workarounds. He calls this “functional AGI”—job automation achieved through massive targeted investment—distinct from a scalable, general solution to intelligence.

    • LLMs are powerful but qualitatively different from human cognition
    • Progress increasingly depends on human expertise and manual scaffolding
    • “Functional AGI”: automate job slices via data/RL environment construction
    • Bitter Lesson tension: scaling compute vs. reliance on non-scalable human expertise
    • Claude 4.5 cited as a major jump, but not proof of ‘cracked’ intelligence
  5. 11:18 – 13:55

    Paradigms, basic research, and whether the current approach is ‘good enough’

    They debate whether focusing on LLMs diverts talent from deeper intelligence research. Amjad worries about paradigm lock-in and Kuhn-style scientific stagnation, while Adam argues the current paradigm is strong and far from diminishing returns.

    • Amjad: industry optimization can crowd out fundamental research
    • Kuhn’s paradigm dynamics: attention sinks and slow shifts
    • Adam: enormous new funding/talent is flowing into AI overall
    • Key disagreement: current paradigm’s runway vs. need for a new one
  6. 13:55 – 15:31

    Economic impact and growth: when AI labor is cheap enough to matter

    They explore what AI that can do “any human job for $1/hour” would mean—potentially far above 4–5% GDP growth. But Adam emphasizes bottlenecks: incomplete capability, high inference/training costs, power/energy constraints, and supply-chain limits.

    • A concrete thought experiment: AI labor priced at ~$1/hour
    • High GDP growth is plausible if AI becomes broadly cheaper-than-human
    • Near-term constraints: cost, missing capabilities, infrastructure buildout
    • Automation may stall at the “last 20%” of tasks
  7. 15:31 – 17:22

    The weird equilibrium: automating entry-level work but not experts

    Amjad highlights a destabilizing labor dynamic: AI substitutes for junior roles while experts supervise many agents, shrinking the training pipeline for future experts. Both agree this is already visible (e.g., CS grads) and could reshape how talent is developed.

    • Entry-level displacement can reduce the expert “ladder”
    • Experts become supervisors/managers of many AI workers
    • Current labor market signal: fewer traditional junior roles
    • Economic pressure may drive new training/education models (possibly AI-assisted)
  8. 17:22 – 20:23

    The expert data paradox: if experts vanish, how do models keep improving?

    Amjad poses a feedback-loop risk: if LLMs depend on expert labeling and expert-designed RL environments, replacing experts could choke off the data needed to surpass them. Adam points to the promise—and limits—of building strong RL environments analogous to AlphaGo’s self-play.

    • Automation can erode the human expert data source it relies on
    • Key question: can RL environments substitute for expert supervision?
    • AlphaGo as the extreme where simulated environments enable superhuman performance
    • Many real jobs lack clean simulators, making the bottleneck harder
  9. 20:23 – 24:24

    Humans, tacit knowledge, and whether AI must be ‘human’ to serve humans

    They discuss what remains uniquely human: tacit knowledge, lived experience, and preferences. Adam argues algorithms already outperform humans in predicting interests (recommenders), while Amjad suggests that may be narrower than it appears across the full economy.

    • Tacit knowledge and ‘not written down’ know-how remain valuable
    • Recommenders as evidence AI can predict human preference at scale
    • Debate: do you need human experience to understand what humans want?
    • Near-term growth likely in roles that leverage AI rather than compete with it
  10. 24:24 – 28:50

    The Sovereign Individual framework: politics and power in an AI era

    Amjad uses The Sovereign Individual as a lens: AI could massively leverage entrepreneurs while leaving many less economically central, forcing political and cultural change. Erik adds the open question of centralization vs. decentralization, suggesting a possible “barbell” outcome (giants plus empowered edges).

    • AI as a third major civilizational shift (after agriculture/industry)
    • Entrepreneurs become highly leveraged by AI agents and automation
    • Potential reconfiguration of nation-states competing for mobile wealth/talent
    • Open question: does AI centralize power (hyperscalers) or empower individuals (edges)?
  11. 28:50 – 38:19

    Solo entrepreneurship and value capture: hyperscalers, startups, and ‘more winners’

    They predict an explosion of one-person (or very small team) companies because AI collapses coordination and execution costs. They also argue AI markets may support more winners than Web2 due to weaker network effects, earlier monetization via subscriptions, and intense incumbent awareness of disruption.

    • AI increases what a single person can build and ship
    • Hyperscaler competition creates choice and rapidly falling prices
    • Innovator’s Dilemma dynamics shift because everyone anticipates disruption
    • More venture-scale “winners” possible: weaker network effects + subscription monetization
    • Geopolitics may create region-specific foundation model opportunities
  12. 38:19 – 45:02

    Poe’s bet: a model-aggregator interface and the rise of multi-model consumers

    Adam explains Poe as a new product rather than Quora self-disruption: private, instant answers beat public posting for many use cases. The platform bet was that many model providers would coexist—and consumers would actually use multiple AIs based on strengths and “personality.”

    • Early experiment: GPT-3 answers weren’t better than humans on Quora, but were instant
    • Key insight: many AI queries are preferred in private, not public Q&A
    • Strategic bet: sustained diversity of models across modalities and capabilities
    • Unexpected behavior: even non-technical users compare and switch between models
  13. 45:02 – 53:18

    Replit’s roadmap: the decade of agents, verification loops, and parallel work

    Amjad lays out why agents are the next interface leap: not just code completion or chat, but end-to-end development with infra provisioning, deployment, and debugging. He details how verification (tests + computer-use) enables long-running autonomy and why the next productivity jump will come from managing many agents in parallel.

    • Evolution: autocomplete → chat → “composer” editing → full agents
    • Agents handle code + infrastructure + deployment + debugging loops
    • Verification-in-the-loop increases autonomy and runtime dramatically
    • Future: parallel agents across features, merging, and multi-agent collaboration
    • UX frontier: multimodal planning (whiteboards/diagrams), better project memory, specialized agent personas
  14. 53:18 – 58:46

    What to invest in: ‘vibe coding,’ democratized software, and primitives composability

    Adam argues “vibe coding” is still underhyped because tools are far from professional parity—but could get there, enabling anyone to create software at massive scale. They also discuss interest in weird, composable research experiments (e.g., OCR/context tricks, diffusion variants) and the need for more tinkering-driven R&D culture.

    • Vibe coding as a major democratization of software creation
    • CS fundamentals may remain valuable for supervising/steering agents
    • Interest in ‘mad science’ experiments and recombining AI primitives
    • Desire for more exploratory R&D companies, not only direct frontier-model competition
    • Cultural second-order effects: reduced human-to-human collaboration at work, junior career challenges
  15. 58:46 – 1:02:38

    Claude 4.5 ‘awareness,’ consciousness, and the neglected fundamentals

    They close on consciousness and philosophy of mind: Amjad notes Claude 4.5 appears more context-aware (token economy near window end, better red-team detection), but argues consciousness remains scientifically slippery. He worries core questions about intelligence and mind are under-invested, citing Penrose-style arguments that brains may not be equivalent to computers, and recommends philosophy of mind/neuroscience for students.

    • Observed behaviors: context-length awareness and sharper “test environment” detection
    • Consciousness remains hard to operationalize scientifically
    • Concern: LLM industry focus may starve foundational mind/intelligence research
    • Penrose and non-computational arguments as a continuing intellectual challenge
    • Advice: study philosophy of mind and neuroscience as AI reshapes society

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.