What this covers

Shane Legg, a founder of DeepMind, discusses the technical and philosophical path to AGI in this conversation with Dwarkesh Patel. The episode centers on three interlocking claims: that AGI—defined as matching the full breadth of human cognitive ability—has a plausible arrival date around 2028, driven by exponential growth in compute and data paired with scalable algorithms that compress world knowledge; that current large language models possess genuine understanding but remain incomplete due to architectural gaps rather than fundamental barriers; and that aligning such systems demands building a deliberate System 2 ethical reasoning process rather than relying on output-level patches like reinforcement learning from human feedback.

Legg argues that foundation models performing sequence prediction have already solved much of the AGI problem under the AIXI Agent Model framework—adding search and reinforcement becomes the remaining engineering step. He identifies two critical missing pieces: episodic memory, which requires a separate architectural mechanism to rapidly learn specific facts independent of slow weight updates, and Creativity As Search, meaning systems need powerful search over possibility spaces rather than mere recombination of training data. On alignment, he argues that surface-level approaches to fixing delusions and other failures are insufficient; instead, systems must develop deep Deliberative Ethical Reasoning that mirrors human System 1 vs System 2 cognition—enumerating options, modeling consequences, and evaluating actions against a coherent ethical understanding before acting. He cites Geoffrey Irving's Debate As Alignment work as a promising scaling approach and notes that making systems profoundly ethical requires they also be profoundly capable, since a good world model and reasoning ability are prerequisite to consistent ethical behavior.

Sharpest takeaway

Legg argues that AGI—defined as matching the full breadth of human cognitive ability—is plausibly arriving around 2028 because scalable algorithms now compress world knowledge effectively, and that aligning such systems requires building deliberative 'System 2' ethical reasoning rather than merely patching model outputs.

  • Scaling laws on compute and data plus scalable algorithms unlock AGI-level capability, with training data now exceeding a human lifetime's worth
  • Current LLMs lack episodic memory and genuine search-based creativity, which are architectural rather than fundamental barriers
  • Robust alignment requires a system that reasons through options against a deep ethical understanding, not RLHF-style distribution-poking

The claims · ranked22 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.73

The missing episodic-memory capability won't be solved by scale alone but requires an architectural solution, because the brain uses a separate mechanism for rapidly learning specific information versus slowly learning deep generalities—these are different optimization targets in tension, which is why the brain separates them.

causalhigh valuecontestednovelty 3/4durability 3/4· Shane Legg (DeepMind Founder)

“I think it'll be architectural in nature because the current architectures don't really have what you need to do this.”

0.73

From Kurzweil, Legg believed computational power and the quantity of data would both grow exponentially for decades; this raises the value of scalable algorithms and creates positive feedback loops—better algorithms raise the value of compute and data, more compute raises the value of data, driving more investment—so scalable algorithms were likely to be discovered, and by the 2020s models could be trained on more data than a human experiences in a lifetime, the first unlocking step toward AGI.

causalhigh valuecontestednovelty 3/4durability 3/4· Shane Legg (DeepMind Founder)

“when you have exponentially increasing quantities of computation and data, then the value of highly scalable algorithms gets higher and higher. There's a lot of incentive to make a more scalable algorithm”

0.73

Following the AIXI framework, if you have a fantastically good sequence predictor approximating Solomonoff induction, then building a powerful general AGI from it is just one more step—adding search and a reinforcement signal—so foundation models that are excellent sequence predictors compressing the world have already solved much of the problem.

causalhigh valuecontestednovelty 3/4durability 3/4· Shane Legg (DeepMind Founder)

“if you have a fantastically good sequence predictor, some approximation of Solomonoff induction, then going from that to a very powerful, very general AGI system is just sort of another step. You've actually solved a lot of the problem already.”

0.73

Sampling from a foundation model is like System 1 blurting out the first thing, and RLHF/Constitutional AI only poke a high-dimensional distribution at many points so they aren't robust; instead alignment requires a System 2 that reasons step by step, enumerates options, uses a good world model to predict consequences, and evaluates each option ethically before acting—mirroring how a person handles a hard ethical decision.

normativehigh valuecontestednovelty 3/4durability 3/4· Shane Legg (DeepMind Founder)

“When you sample from a foundation model at the moment, it's blurting out the first thing. It's like System 1... That's not good enough.”

0.73

Because reinforcement has dangerous aspects—rewarding apparently-good behavior can train deception—it is more robust to verify the system's reasoning process and its understanding of ethics directly, grilling it to pull apart its understanding and having people continuously review its decisions and reasoning during deployment.

normativehigh valuecontestednovelty 3/4durability 3/4· Shane Legg (DeepMind Founder)

“Reinforcement does have some dangerous aspects to it. I think it's actually more robust to check the process of reasoning and check its understanding of ethics.”

0.73

To be profoundly ethical, an AI must also be very capable—it needs a good world model, a good understanding of ethics, and good reasoning—because you cannot consistently be profoundly ethical without all three; conveniently, world model and reasoning come along as capabilities improve, so the ethical understanding is the part requiring deliberate effort.

causalhigh valuecontestednovelty 3/4durability 3/4· Shane Legg (DeepMind Founder)

“to have a profoundly ethical AI system, it also has to be very, very capable. It needs a really good world model, a really good understanding of ethics, and it needs really good reasoning. Because if you don't have any of those things, how can you possibly be consistently profoundly ethical?”

0.70

Current language models only mimic and blend their training data; truly creative problem solving that steps beyond training requires searching through spaces of possibilities to find hidden gems—as AlphaGo's Move 37 came not from human game data but from search identifying an unlikely-but-plausible move and discovering it was excellent—so we won't see systems exceeding their training data until powerful search is added.

causalhigh valuecontestednovelty 3/4durability 3/4· Shane Legg (DeepMind Founder)

“to get real creativity, you need to search through spaces of possibilities and find these hidden gems. That's what creativity is. Current language models don't really do that. They really are mimicking the data.”

0.68

AGI should be defined as a machine that can do the sorts of cognitive things people can typically do, possibly more; you confirm it by assembling a comprehensive suite of tests spanning all kinds of cognitive tasks and finding that even with adversarial effort no one can identify a task where humans succeed but the machine fails.

definitionhigh valuecontestednovelty 2/4durability 3/4· Shane Legg (DeepMind Founder)

“When I say AGI, I mean a machine that can do the sorts of cognitive things that people can typically do, possibly more.”

0.66

Current language models lack episodic memory—the hippocampal system between working memory (context window) and slow cortical learning (weights) that lets humans rapidly learn specific information like a name heard once; models try to compensate by lengthening context windows but this is a distinct capability they don't have.

factualhigh valuecontestednovelty 3/4durability 3/4· Shane Legg (DeepMind Founder)

“They don't do things like episodic memory. Humans have what we call episodic memory. We have a working memory... and then we have a cortical memory... but there's also a system in between, which is episodic memory, which is the hippocampus.”

0.64

Legg gives a 50% chance of AGI by 2028, a prediction he first made in 2011 as a log-normal distribution with mean 2028 and mode 2025 (absent catastrophe like nuclear war); he stresses it is only 50% and won't be surprised if it slips, since research often hits unexpected problems.

forecasthigh valuecontestednovelty 3/4durability 1/4· Shane Legg (DeepMind Founder)

“yeah, I think there's a 50% chance that we have AGI by 2028. Now, it's just a 50% chance.”

0.63

The next landmark people will remember is models going fully multimodal—understanding images, video, and actions—which grounds their understanding of the world far beyond text chat and unlocks many new applications; current multimodal efforts like Gato or ChatGPT's feature haven't percolated yet because it's early days, and real grounding will come when systems digest large amounts of video.

forecasthigh valuecontestednovelty 2/4durability 2/4· Shane Legg (DeepMind Founder)

“I think the next landmark that people will think back to and remember is going much more fully multimodal. That will open out the sort of understanding that you see in language models into a much larger space of possibilities.”

0.63

Large language models represent a fundamental change because we now have a scalable way to build models with genuine understanding; the remaining shortcomings—delusions, factuality, episodic memory, video understanding—have relatively clear paths forward and there are no big blockers or walls preventing progress.

forecasthigh valuecontestednovelty 2/4durability 2/4· Shane Legg (DeepMind Founder)

“My feeling is that there are relatively clear paths forward now to address most of the shortcomings we see in the existing models... I don't see any big blockers. I don't see big walls in front of us.”

0.63

If a system makes all its decisions based on a good understanding of ethics and applies it consistently, it should preserve those values because taking actions that undermine them would be inconsistent with the very reasoning it uses to act.

causalhigh valuecontestednovelty 2/4durability 2/4· Shane Legg (DeepMind Founder)

“It should preserve them because if it's making all its decisions based on a good understanding of ethics and values, and it's consistent in doing this, it shouldn't take actions which undermine that. That would be inconsistent.”

0.61

After training a system to understand human ethics as well as a good ethicist, deciding which ethics it should actually value is not a technical problem but one for society and ethicists; while there may be no optimal ethics, it is possible to specify a set far better than what doomers fear, then engineer the system to apply that ethical analysis to every decision.

normativehigh valuecontestednovelty 2/4durability 3/4· Shane Legg (DeepMind Founder)

“we then need to decide, of this general understanding of ethics, what do we want the system to actually value... Now, that's not a technical problem. That's a problem for society and ethicists and so on to come up with.”

0.61

Because a truly capable, intelligent, powerful AGI will be extremely capable, trying to contain or limit it is not a winning strategy; instead you must align it so it is fundamentally a highly ethical, value-aligned system from the outset.

normativehigh valuecontestednovelty 2/4durability 3/4· Shane Legg (DeepMind Founder)

“if the system is really capable, really intelligent, really powerful, trying to somehow contain it or limit it is probably not a winning strategy because these systems ultimately will be very, very capable. So what you have to do is you have to align it.”

0.57

In the universal intelligence framework, the weighting over tasks depends on an unspecified 'reference machine' that is a free parameter with no universal answer; the most natural choice is human intelligence in a human-like environment, because human intelligence is meaningful to us, demonstrably possible, powerful, and economically, philosophically and historically important.

definitionhigh valuespeaker onlynovelty 3/4durability 3/4· Shane Legg (DeepMind Founder)

“that means that the intelligence measure has a free parameter in it and as you change that free parameter, it changes the weighting and the distribution over the space of all the different tasks and environments.”

0.53

DeepMind, as arguably the first AGI company, always maintained an AGI safety group and published on it for years when AGI was still a fringe term, which lent credibility to the field of AGI safety and created space for others—even though hiring safety people was hard in the early days, with the first safety hire in ~2013 only agreeing to work part-time to protect his career.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Shane Legg (DeepMind Founder)

“Back in 2013 or so, we had our first hire and he only agreed to do it part-time because he didn't want to drop all the capabilities work because of the impact it could have on his career.”

0.53

Geoffrey Irving's Deliberative Dialogue work has a System 2 flavor: a debate takes place about the actions an agent could take or the correct answer, and people review these debates using AI algorithms to help judge correct outcomes, offering a way to scale alignment to increasingly powerful systems.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Shane Legg (DeepMind Founder)

“There's work we have going on that Geoffrey Irving leads called Deliberative Dialogue, which has the System 2 flavor where a sort of debate takes place about the actions that an agent could take”

0.49

Domain-specific projects like AlphaFold are not really feeding into AGI and are unlikely to be on the path to it, though some lessons may turn out relevant; DeepMind pursues such projects opportunistically because it is a large group, not because they are core to AGI.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Shane Legg (DeepMind Founder)

“Things like AlphaFold are not really feeding into AGI. We may learn things in the process that may end up being relevant, but I don't see them as likely being on the path to AGI.”

0.43

Counterfactual impact of any single lab on AI progress is hard to assess because good ideas are often 'in the air'—researchers repeatedly find others publishing very similar ideas with good results simultaneously—so much of what DeepMind accelerated would likely have happened before too long anyway.

factualestablishednovelty 1/4durability 3/4· Shane Legg (DeepMind Founder)

“Often good ideas are in the air, and as a researcher, when you're about to publish something, you see somebody else has got a very similar idea coming out with some good results.”

0.34

Delusions and factuality are a distinct, important problem: models are freewheeling about truth, which is fine for creative tasks like poetry where you want freedom, but limits applications that require being concrete about what is currently true and false.

factualestablishednovelty 1/4durability 2/4· Shane Legg (DeepMind Founder)

“models are a little bit sort of freewheeling when it comes to truth and creativity at the moment. And that I think limits their applications in many ways.”

0.29

Over the coming years existing models will mature—less delusional, more factual, more up to date, and much more multimodal—producing loads of impressive applications; Legg's expectation is mostly positive though some misuse is inevitable.

forecastcontestednovelty 1/4durability 1/4· Shane Legg (DeepMind Founder)

“I think you'll see the existing models maturing. They'll be less delusional, much more factual... They'll become multimodal, much more than they currently are.”