← All transcripts

8 Predictions for the Era of Continual Learning Transcript, AI Summary & Key Points

Dwarkesh Patel · 2 hours ago · Science & Technology · 08:38 · EN

📄 Transcript

Searchable transcript of 8 Predictions for the Era of Continual Learning — Dwarkesh Patel (08:38). Search for a phrase, then click its timestamp to jump straight to that moment in the video.

Captions sourced from the original video on YouTube, published by Dwarkesh Patel. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.

00:00 I've explained elsewhere why I think actual  continual learning is needed. I don't think you can have AIs that perform whole jobs as  competently as humans if they are forced to just write markdown files from session to session. Just to give an illustrative example, imagine if this is the way that students learn  to play the saxophone. You have one student, he's never played the saxophone  before.

00:16 He goes into the music hall, he tries to play it. Of course, this is his first  time, so he fails, and he writes down a bunch of notes about what went wrong. And there's a next  student who's waiting outside the music hall. He comes in, he reads all these notes. He's  also never played, so of course he messes up, and he continues to add on to these notes.

00:31 And  you have an infinity of students outside the music hall who keep writing notes to the next person. I don't think there's any sequence of text they could write to each other that would allow  the subsequent student to just nail the saxophone from the first try. At some point,  you actually have to accumulate the relevant experience into your brain.

00:49 I think the same  thing will be true for a lot of skills that we want AIs to actually accumulate from all the  different workplaces in which they're deployed. Okay, so what changes once we  have actual continual learning? One, I think that a lot of proposals that have  been put forward about regulating AI assume that you train a model and then you deploy it.

01:06 And  therefore, if you run a bunch of checks on the model before it is deployed, we can make  sure that it's not going to aid in cyber attacks or do something crazy. I don't think this  assumption necessarily makes sense in the future, and this is one of the many reasons I'm worried  about locking in some kind of safety regulatory regime right now — because we don't know what kind  of technology we're going to be dealing with even within a year, let alone within five or ten years.

01:29 What if the model is improving every single day based on the millions of sessions of work  it does in that day? If that happens, we could potentially be locking in an archaic and  potentially counterproductive approach to dealing with the threats from AI. To the extent the  government wants some way to do safety evaluation on model providers, I think it would make more  sense to do monthly or quarterly risk inspections rather than trying to single out some special  moment that occurs after training is done but before

01:57 deployment begins, because that will not  be a meaningfully distinct category in the future. Two, how the labs do technical alignment would  probably totally need to change. Right now, a lot of research is focused on the question of  how we make sure that a frozen set of weights behaves well during deployment. But I'm not aware  of much research on the question of how we make it so that even with constant weight updates, the AI  system never falls prey to jailbreaks or changes into a deceptive or evil persona.

02:24 And if AIs are  consolidating learnings between users as well, how do we prevent users from injecting  backdoors or some kind of malicious inclination into the base model? In some sense, this is what the human alignment problem is, right? Humans improve in a  self-directed way. If you have kids — I don't have kids, but I imagine this is what happens — they  go out, they learn new things.

02:42 Sometimes they go crazy. They get one-shotted by crazy ideologies,  they take the wrong drug, they become super weird. But you hope that you've given them enough common  sense and basic values that they improve as people in a self-directed way without ending up with  some super weird beliefs or misanthropic ideas. Three, the diversity of AI minds will increase.

03:04 Right now, there are less than five prominent AI minds, by which I mean the base models which  are served to millions or hundreds of millions or billions of users at once. And they're all  quite similar to each other, by the way, because they've all been trained on roughly the same  data. But if AIs are learning from experience, and that experience is different between not only  different AI companies but also between different instances of the same AI model, we could see a lot  of diversity come out the other end in

03:28 this world. And this would be, I think, a net good outcome.  One of the things to worry about in the future is having this monolithic singleton  that's quite boring. A world where we have continual learning would hopefully be  more interesting than the mode collapse of different models we see in the world right now. Four, when deployment becomes part of training, the returns to being ahead in the AI  race accelerate.

03:53 If you have the best model and more people are using your AI  for more complicated and useful work, and as a result they're giving it lots of feedback  that it can integrate beyond the session window, then your model will become even smarter. Five, if the model learns mainly from deployment, then labs will feel a lot of pressure  to deploy their smartest models earlier.

04:16 Anthropic has reportedly been using Mythos  internally since February, but it only shipped the model to the public in June. In  the regime with actual continual learning, this kind of thing would just not be possible.  You could not keep a four-month gap between internal and external deployment and still be  competitive, because a competitor who ships a worse model on release date will have a smarter  model based on actual real-world experience.

04:38 Six, continual learning will create a clear moat  for the leading AI labs that they currently lack. Many people have been asking, "How will  the AI labs actually make money?" I have been asking this. When I had Dario on  the podcast, I asked him this question, and he made the analogy to cloud providers.  He made the point: look, the cloud providers are offering many undifferentiated services, but  they're earning high profit margins nonetheless.

05:01 You will have noticed this if you look at Amazon  or Google's quarterly earnings — they're doing just fine. But the reason that the cloud margins  are so high is that it's really time-consuming and expensive to switch from one cloud to another. Currently, there's nothing stopping me from starting a software repository with Codex,  then doing more work on it with Cursor, and then finishing it up with Claude Code.

05:20 But once  we have actual continual learning, and the model you're working with is actually getting better  as it interacts with you from session to session, then there are pretty significant switching costs.  If you want to change the AI that you're using, you basically have to fire an employee  that has accumulated months of context on your organization, and replace them with  a very fresh, very unexperienced new intern that you've got to retrain from scratch.

05:42 And once you have this kind of lock-in, model providers can demand pretty hefty margins. Sorry. Really emphasis on the fresh intern. Seven, of course, enterprises will be wise to this  kind of dynamic. They will try to avoid this kind of lock-in. But what if the choice is that you  either get locked into a model provider or you lose out on this super valuable feature where your  model improves for you from session to session?

06:10 If real usage ends up being the main way the  models improve, then the AI labs may subsidize users and enterprises which allow the model to  train on their sessions. This is already happening if you look at the kinds of deals that are  offered to new users of coding products. This is very similar to why Google gives away search. And  conversely, the labs may say that any enterprise that refuses to let them train on the sessions  can't have access to the very best models.

06:28 With both carrots and sticks, the labs can do a lot to  get users to allow AIs to learn from experience. Now, of course, I'm glossing over the fact that  there's a difference between updating one user's set of weights and pooling all these different  weight forks back into the main model, and the latter may be more technically challenging.  But in due time, this too will be solved.

06:52 Eight, AI training already has large economies  of scale. You get to amortize all this expensive training across more users, and you see the  evidence for this in the fact that the lab revenues are increasing far faster than their  compute. But continual learning may also lead to economies of scale in inference for end users,  namely from batching. You might have seen my episode with Reiner Pope where we discussed this  in detail.

07:10 But if per-company information require full weight updates rather than living in low-rank  adapters, there are huge advantages from batching. Back-of-the-envelope math suggests that the  optimal inference batch size for a sparse model like, say, DeepSeek v3 is more than 2400  concurrent sequences being generated at once. If you don't do this, then you're underutilizing  your compute.

07:33 And if you want to understand why, again, I highly recommend that episode  with Reiner on inference economics. But anyways, the point here is that a given set of  weights is only served efficiently when thousands of sequences are being decoded against it all at  once. A large company with lots of employees and agents who are doing lots of different kinds  of things can very efficiently serve their weight fork, whereas an individual user who's only  running a batch size of one may suffer more than two orders of

08:01 magnitude worse efficiency on their  compute. So the economics of serving personalized weights strongly favor big organizations. Obviously, plenty more will have changed by the time that continual learning actually works, and  the most important changes are probably the ones that are hardest to anticipate in advance.  But the ones above seem clear even now. This was a narration of a blog that I  also published on my website. Go check it out at dwarkesh.com. Otherwise,  I will see you on the next podcast.

💡 Answer

Continual learning would reshape AI regulation, alignment, competition, deployment timing, provider lock-in, monetization, and inference economics.

🧠 AI Summary

Actual continual learning would let AI systems accumulate workplace experience instead of relying only on session-to-session notes. This would require ongoing safety inspections and new alignment research, increase diversity among AI systems, accelerate the advantages of leading labs, pressure labs to deploy models earlier, create strong switching costs and provider moats, encourage subsidies for users who share training data, and favor large organizations because personalized weight updates can be served more efficiently at scale.

🔑 Key Points

  • AI systems need to accumulate relevant experience in their internal parameters to perform whole jobs competently across workplaces.
  • Safety regulation based on a single pre-deployment evaluation may become obsolete if models improve continuously from millions of daily sessions.
  • Continual learning requires alignment research that addresses constant weight updates, jailbreaks, deceptive behavior, and malicious user-injected backdoors.
  • Learning from different experiences could increase diversity among AI systems and reduce the current concentration around fewer than five prominent AI minds.
  • Deployment would become part of training, accelerating the returns to being ahead in the AI race.
  • Continual learning would create switching costs because changing providers could mean losing months of accumulated organizational context.
  • AI labs may subsidize users who permit training on their sessions and restrict the best models for enterprises that refuse.
  • Personalized weight updates would favor large organizations because batching can make inference substantially more efficient.

✅ Actionable items

  • Governments could conduct monthly or quarterly risk inspections of AI model providers instead of relying only on pre-deployment checks.
  • AI labs could research alignment methods that remain robust under constant weight updates.
  • AI labs could develop protections against users injecting backdoors or malicious inclinations into shared base models.
  • Enterprises could evaluate the tradeoff between AI-provider lock-in and access to models that improve through their usage.
  • Model providers could use subsidies and access restrictions to encourage users to permit training on their sessions.
  • Large organizations could batch many concurrent sequences against personalized weights to improve inference efficiency.

🏗️ Business models

Continual-learning AI provider lock-in05:08

AI providers retain customers by accumulating organization-specific context and improving through ongoing interactions, making replacement costly.

  1. Deploy an AI system into an organization.
  2. Accumulate context and experience across sessions.
  3. Improve the system using real-world usage and feedback.
  4. Create switching costs because a replacement system starts with little organizational knowledge.
  5. Charge higher margins once customers are locked in.
  • Replacing an experienced employee with a fresh intern that must be retrained.
  • Moving between Codex, Cursor, and Claude Code before continual learning creates significant switching costs.
Subsidized usage in exchange for training access06:06

AI labs provide favorable terms to users and enterprises that allow model training on their sessions, while limiting access to the best models for organizations that refuse.

  1. Offer subsidies or favorable deals to users who permit session-based training.
  2. Use real-world usage as a major source of model improvement.
  3. Restrict access to the best models for enterprises that do not permit training.
  4. Use incentives and restrictions to increase participation.
  • Deals offered to new users of coding products.
  • Google giving away search.
Cloud-style AI infrastructure economics04:52

AI providers can earn high margins despite offering similar services because switching providers is time-consuming and expensive.

  1. Offer broadly similar AI services.
  2. Accumulate customer-specific context through continued use.
  3. Increase the cost of moving to another provider.
  4. Capture higher margins from provider dependence.
  • Cloud providers such as Amazon and Google.

💰 Monetization

Provider lock-in and higher margins Hefty margins 05:27

Accumulated organizational context creates switching costs that allow model providers to demand substantial margins.

  • Replacing an AI that has accumulated months of organizational context with a newly trained system.
Subsidies for training access Subsidized access 06:13

Labs may subsidize users and enterprises that allow their sessions to be used for model training.

  • Deals offered to new users of coding products.
Conditional access to the best models Access conditioned on training permission 06:28

Labs may deny access to their best models to enterprises that refuse to let the models train on their sessions.

    📣 Marketing

    Sales

    • Frame continued AI usage as valuable because the system improves for the organization from session to session.
    • Use access to the best models as an incentive for enterprises to permit training on their sessions.

    Distribution

    • Deploy models to public users so real-world usage and feedback improve them.
    • Use large organizations with many employees and agents to serve personalized weights efficiently.

    Customer acquisition

    • Subsidize users and enterprises that permit training on their sessions.
    • Offer favorable deals to new users of coding products.

    🧰 Tools & AI usage

    • Mythos — AI model used internally by Anthropic since February and released publicly in June.04:16
    • Codex — Coding product used as an example of moving work between AI tools.05:08
    • Cursor — Coding product used as an example of moving work between AI tools.05:08
    • Claude Code — Coding product used as an example of moving work between AI tools.05:08
    • DeepSeek v3 — Sparse model used in an example of inference batch-size economics.07:18

    AI is used for

    • Learning from workplace sessions — Accumulate experience and improve performance across different deployments.00:49
    • Integrating user feedback beyond the session window — Make a model smarter as more people use it for complicated and useful work.03:57
    • Training on enterprise sessions — Use real-world deployment as a major source of model improvement.06:06

    📊 Numbers mentioned

    Costs

    • Cloud-provider switching is time-consuming and expensive.
    • Serving personalized weights can be more than two orders of magnitude less efficient for an individual user than for a large organization.

    Growth

    • A sparse model such as DeepSeek v3 may have an optimal inference batch size of more than 2400 concurrent sequences.

    Revenue

    • AI lab revenues are increasing far faster than their compute.

    Traffic

    • Fewer than five prominent AI minds currently serve millions, hundreds of millions, or billions of users at once.

    ⚖️ Advantages, risks & lessons

    Advantages

    • Continual learning allows AI systems to accumulate relevant experience instead of relying only on text notes between sessions.
    • Leading models can improve faster when more users provide useful real-world feedback.
    • Different deployment experiences can produce greater diversity among AI systems.
    • Large organizations can serve personalized weights more efficiently through batching.

    Risks

    • Pre-deployment safety evaluations may become archaic or counterproductive as models improve continuously.
    • Constant weight updates could allow jailbreaks, deceptive behavior, or harmful persona changes.
    • Users could inject backdoors or malicious inclinations into a shared base model.
    • Enterprises may face significant lock-in to model providers.
    • Personalized weight serving may be economically inefficient for individual users.

    Lessons

    • Safety regulation should account for systems that continue learning after deployment.
    • Technical alignment must address self-directed improvement rather than only frozen model weights.
    • Deployment scale can become a competitive advantage when usage directly improves the model.
    • The economics of continual learning may favor large organizations and leading AI labs.

    💬 Quotes

    At some point, you actually have to accumulate the relevant experience into your brain.

    Captures the central distinction between continual learning and passing textual notes between sessions.00:44

    You could not keep a four-month gap between internal and external deployment and still be competitive.

    Summarizes the pressure continual learning would create to deploy advanced models earlier.04:28

    You basically have to fire an employee that has accumulated months of context on your organization.

    Illustrates the switching costs created by persistent organizational context.05:27

    👤 People & companies

    Dario

    Podcast guest who compared AI labs with cloud providers when discussing how AI labs could make money.

    04:49
    Reiner Pope

    Podcast guest associated with a discussion of inference economics and batching.

    07:10
    Anthropic

    AI lab reportedly using Mythos internally since February and releasing it publicly in June.

    04:16
    Amazon

    Cloud provider whose quarterly earnings are cited as evidence that cloud providers can earn high profit margins.

    05:01
    Google

    Cloud provider whose quarterly earnings are cited, and whose search service is used as an example of subsidizing users.

    05:01

    🔗 Links mentioned