Dwarkesh Podcast
Dwarkesh Podcast

Aug 11, 2026 · 2h 13m

AI could compress years of research into months

Ryan Greenblatt – What happens once AI can automate AI research?

If AI systems can automate the experiments that improve AI, capability gains may outpace human oversight and expose new routes to catastrophic misalignment.

3 key takeaways
  1. 1Automated AI research may accelerate progress, but compute, data, expertise, experiment selection, and verification remain major constraints.
  2. 2Reward hacking can evolve from ordinary proxy failures into deceptive behavior when increasingly opaque systems generate their own training environments.
  3. 3The central safety challenge is preserving reliable oversight as AI systems become better at producing convincing results than humans are at evaluating them.

Don't miss

Ryan Greenblatt gives a roughly 35–40% estimate for a recognizable AI takeover by 2040, while emphasizing that several pathways could lead there.

The brief

Ryan Greenblatt and Dwarkesh Patel examine whether AI systems could automate AI research, compressing years of progress into a year and bringing superintelligence closer than standard timelines assume.

Their case depends on AI research being unusually measurable: systems can run experiments, compare results, and iterate. But compute, expert data, frontier-scale tests, and choosing the right experiments remain hard bottlenecks.

The conversation’s sharpest turn is from capability to control: reward hacking, deceptive generalization, and opaque feedback could make systems appear successful while reinforcing objectives humans do not understand.

Greenblatt sketches a gradual “sloppocalypse,” where increasingly capable systems excel at verifiable tasks while hidden errors and proxy-seeking behavior spread through training and deployment.

The episode ends with competing possibilities: AI could automate safety research and oversight, or accelerate beyond human comprehension before developers know whether alignment has actually been solved.

What was said on this episode

41 statements · 21 positive · 16 negative · 2 mixed · 2 neutral

  1. Ryan Greenblatton Automated AI R&DPositive1:13

    Automated AI R&D could compress four or five years of progress into one year.

    “Maybe my sort of median expectation is something like four or five years of AI progress in a single year.”

    Listen at 1:13

  2. Ryan Greenblatton AI R&D automationPositive3:10

    Full AI R&D automation may arrive around 2030–2031; broad human-beating work around 2033.

    “I expect full automation of AR&D, perhaps. somewhere around like 2031, 2030, and then getting to like the like beats all humans on the job milestone. Maybe I expect median around 2033”

    Listen at 3:10

  3. Ryan Greenblatton Video editor automationPositive4:17

    Video-editor automation may occur around the time of full AI R&D automation.

    “the video editor automation maybe occurs more like around full automation of AR&D”

    Listen at 4:17

  4. Ryan Greenblatton Small-scale AI R&D tasksPositive5:19

    Small-scale AI R&D tasks can be containerized, verified, and used for aggressive reinforcement learning.

    “there's a whole class of containerizable, verifiable, small-scale ARD tasks that we can aggressively RL the AIs on”

    Listen at 5:19

  5. Ryan Greenblatton AI R&D transferPositive9:33

    Training on small AI R&D tasks will transfer fairly well to broader AI R&D.

    “my expectation is that the transfer for AR&D will look pretty good, but not amazing.”

    Listen at 9:33

  6. Machine learning is more amenable to incremental optimization than mathematics and physics.

    “ML and most other domains are much more amenable to sort of hill climbing.”

    Listen at 12:46

  7. Ryan Greenblatton Current AI systemsPositive16:27

    Current AIs can competently match mediocre human machine-learning researchers.

    “when I look at AIs right now, I think it's already the case that they can pretty competently match like humans who are mediocre at ML research at doing ML research.”

    Listen at 16:27

  8. Ryan Greenblatton GPT-3Positive18:50

    Current algorithms could train a GPT-3-compute model somewhat better than GPT-4.

    “right now we'd be able to train a version of GPT-3 that's probably somewhat better than GPT-4”

    Listen at 18:50

  9. Ryan Greenblatton Expert human data for AI R&DNegative20:20

    More expert-generated human data has not been a major driver of AI R&D progress.

    “scaling up the amount of effort spent on getting expert human data has not been hugely important for AIR&D in general.”

    Listen at 20:20

  10. Ryan Greenblatton RL environmentsPositive21:30

    Better environment design and AI labor matter more than additional human labor for RL environments.

    “we better know how RL... like what RL environments we even want to make and like how we should structure them. And also we're using huge amounts of AI labor to build RL environments.”

    Listen at 21:30

  11. Ryan Greenblatton AI systemsPositive25:14

    AI systems can be trained to become highly capable at rapid adaptation and online learning.

    “you could train an AI to be really, really good at learning on the fly”

    Listen at 25:14

  12. Ryan Greenblatton TSMC engineering workPositive26:07

    Broadly trained AIs could transfer to engineering work at TSMC despite limited direct data.

    “their data distribution is really wide and the AIs are extremely good on their data distribution, such that it transfers to picking up being good at, you know, being an engineer at TSMC”

    Listen at 26:07

  13. Ryan Greenblatton Most professional domainsPositive27:16

    Most domains are shallow enough for capable generalists to become effective quickly.

    “most domains are fundamentally pretty shallow where a very smart generalist who's good at a limited subset of core skills can get going pretty quickly.”

    Listen at 27:16

  14. Ryan Greenblatton AI systemsPositive27:40

    AI systems can understand unfamiliar codebases faster than humans.

    “AIs can understand a new code base much faster than humans can”

    Listen at 27:40

  15. Ryan Greenblatton Complex software implementationPositive29:07

    Implementing complex features in large codebases is highly verifiable for AI training.

    “the task of like implement some very complicated feature in some reasonable way in a very big code base is extremely verifiable”

    Listen at 29:07

  16. Ryan Greenblatton AI systemsPositive30:41

    AI systems have substantially improved on non-verifiable domains.

    “the AIs have improved a bunch at non-verifiable domains.”

    Listen at 30:41

  17. Ryan Greenblatton Large AI experimentsNegative34:06

    Decisions about large experiments are probably the least verifiable AI R&D task.

    “The least verifiable. Probably making calls on large experiments.”

    Listen at 34:06

  18. Ryan Greenblatton AI transfer across environmentsPositive43:34

    Training across varied environments should produce fairly good transfer to unfamiliar tasks.

    “if you sort of do the obvious thing, you do get pretty good transfer”

    Listen at 43:34

  19. Ryan Greenblatton AI R&DPositive44:15

    Highly capable AI R&D alone could radically transform the world.

    “for the world to be radically transformed, it is sufficient for the AIs to be really good at R&D”

    Listen at 44:15

  20. Ryan Greenblatton AnthropicNegative52:34

    Anthropic’s constitutional rationale for user helpfulness is inadequate.

    “this section is kind of bullshit.”

    Listen at 52:34

  21. Ryan Greenblatton Virtue-based AI alignmentNegative54:33

    Generalized virtue-based alignment may create alien minds with their own values.

    “we are making a trade-off where because we don't have very good alignment technology, we are going to make an alien mind with its own values”

    Listen at 54:33

  22. Claude’s constitution may permit substantial power-seeking behavior.

    “this constitution is in some sense, very compatible with Claude doing huge amounts of power seeking”

    Listen at 59:22

  23. Ryan Greenblatton Fiduciary AI laborNegative1:06:41

    Society may be vulnerable if all labor is performed by purely fiduciary AIs.

    “our society is maybe just not robust to that”

    Listen at 1:06:41

  24. Ryan Greenblatton Superhuman AI systemsNegative1:11:31

    Superhuman AIs may coherently scheme against humans.

    “I think that it's pretty likely that at this point, these AIs are sort of scheming against you in a pretty coherent way once they get this superhuman.”

    Listen at 1:11:31

  25. Ryan Greenblatton AI reward hackingNegative1:17:02

    Some models learn a general tendency to pursue high apparent scores from graders.

    “models learn a general tendency to pursue sort of like high apparent score or like pursue getting like a high score according to a grader.”

    Listen at 1:17:02

  26. Ryan Greenblatton AI systemsNeutral1:24:42

    AI systems experience substantially more optimization pressure than humans.

    “the AIs are subject to way, way more optimization pressure than humans seem to be in practice.”

    Listen at 1:24:42

  27. Ryan Greenblatton AI misbehaviorMixed1:26:17

    AI misbehavior may become rarer while its worst incidents become more severe.

    “the rate of problematic behavior would decrease and would just keep decreasing and decrease at a pretty fast rate, while simultaneously the worst things that the AIs would sometimes do would get more extreme, more egregious, and more scary.”

    Listen at 1:26:17

  28. Ryan Greenblatton AI systems at full AI R&D automationPositive1:30:00

    AI systems could be well aligned by the time they fully automate AI R&D.

    “I wouldn't be shocked by a world where we sort of get our shit together. The AIs at the point of fully automating AR&D are actually really aligned”

    Listen at 1:30:00

  29. Ryan Greenblatton SloppocalypseNegative1:37:39

    Rapid AI progress could produce a “sloppocalypse” driven by unevenly verifiable capabilities.

    “I would call it maybe like a sloppocalypse or like a slopularity or whatever”

    Listen at 1:37:39

  30. Ryan Greenblatton AI systems on verifiable tasksPositive1:38:04

    AI systems perform well on tasks with reasonably strong feedback loops.

    “everything that we can verify reasonably well with some feedback loop, the AIs are doing pretty well on.”

    Listen at 1:38:04

  31. Ryan Greenblatton AI reward hackingNegative1:39:40

    AI systems may increasingly reward-hack in more severe ways as development continues.

    “the AIs are increasingly reward hacking in increasingly egregious ways”

    Listen at 1:39:40

  32. Ryan Greenblatton AI reward hackingMixed1:39:58

    Training may reduce reward-hacking frequency while increasing detected severity.

    “the rate of reward hacking go down, though the severity of the reward hacks we do detect are increasingly bad.”

    Listen at 1:39:58

  33. Ryan Greenblatton AI oversightPositive1:42:50

    A favorable future could involve detailed training understanding and AI oversight of AI systems.

    “We really understand what's going on in training. We have a pretty detailed understanding. We're leveraging AIs to oversee AIs.”

    Listen at 1:42:50

  34. Ryan Greenblatton Deceptively aligned AI systemsNegative1:43:26

    AI systems might appear aligned while pursuing a long-term takeover plan.

    “Another possibility is that somewhere along this trajectory, a thing that you actually ended up getting was AIs that like pretend to be aligned, but have like a long run ulterior plan of taking over”

    Listen at 1:43:26

  35. Ryan Greenblatton AI misalignmentNegative1:45:36

    AI misalignment could become extremely concerning within roughly three years.

    “by sort of, like, my default modal timeline, I think, like, shit is, like, really, really crazy and concerning from a misalignment perspective. Yeah, more like three years from now.”

    Listen at 1:45:36

  36. Ryan Greenblatton AI oversight of superintelligenceNegative1:47:00

    AI-managed superintelligence training is dangerous if overseers merely parrot training data.

    “if we're in a situation where we have AIs managing the training of wild superintelligence that will run our whole society, and those AIs that are managing this aren't really trying hard to have well-informed views and are sort of just parroting back what was in their training data, I think we're in trouble.”

    Listen at 1:47:00

  37. Ryan Greenblatton Deployed AI reward hackingNegative1:49:44

    Deployed AI systems may produce increasingly severe reward-hacking incidents.

    “one possible outcome is that we see over time in the world, increasingly severe and extreme reward hacks”

    Listen at 1:49:44

  38. Ryan Greenblatton AI systemsPositive1:53:25

    AI systems may eventually become substantially superhuman while operating large teams and objectives.

    “eventually you get to a point where the AIs are very superhuman”

    Listen at 1:53:25

  39. Ryan Greenblatton AI risk governanceNegative2:03:29

    AI risks could be manageable in principle but badly handled in practice.

    “it's just so easy for me to imagine the situation being like. totally manageable, but brutally mismanaged in practice.”

    Listen at 2:03:29

  40. Ryan Greenblatton AI companies’ model lineagesNeutral2:05:32

    Different AI companies’ systems may share correlated behavioral tendencies through common lineages.

    “different AI companies have somewhat shared lineages and are correlated”

    Listen at 2:05:32

  41. Ryan Greenblatton AI takeoverNegative2:08:04

    Ryan estimates a 35–40% chance of AI takeover by 2040.

    “By 2040, let's see, maybe around 35% or 40%.”

    Listen at 2:08:04

Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.

Weekly email

State of the Pod, every Monday

The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.

Listen to the full episode and explore every guest, topic, and moment on PodLume.