I also feel continual learning is overrated as a barrier, but for slightly different reasons. Firstly, it’s worth reviewing why Dwarkesh believes that continual learning is necessary for economically transformative impact[1]:
The LLM baseline at many tasks might be higher than an average human’s. But there’s no way to give a model high level feedback. You’re stuck with the abilities you get out of the box. You can keep messing around with the system prompt. In practice this just doesn’t produce anything even close to the kind of learning and improvement that human employees experience.
One of my main criticisms is that he neglects the ability of LLMs to be recursively specialised. For example, a legal AI company could fine-tune a frontier model specifically for handling legal tasks and then further specialise the model to handle law in different countries. Major lawfirms could have a model further fine-tuned to their needs and these models might be further specialised to the needs of individual departments or even roles. This is only happening to a limited degree now because general models are advancing so fast that it’d hardly be worth it, but if progress stalls, we’ll start seeing more layers of specialisation, especially since a significant proportion of the work involved in fine-tuning could be automated.
Whilst Dwarkesh doesn’t comment on recursive fine-tuning specifically, he does comment on RL fine tuning more generally: “But it’s just not a deliberate, adaptive process the way human learning is. My editors have gotten extremely good. And they wouldn’t have gotten that way if we had to build bespoke RL environments for different subtasks involved in their work. They’ve just noticed a lot of small things themselves and thought hard about what resonates with the audience, what kind of content excites me, and how they can improve their day to day workflows.”
I agree that it’d be much harder for the economics to work out for a podcaster like Dwarkesh vs. a major firm. However, AI wouldn’t have to automate every niche role in order to have an economically transformative impact. It would just have to automate a sufficient number of major sectors. But I’m pretty sure AI models would be transformative in terms of podcasts too. Not all podcasts are as unique, or have the same quality standards, as Dwarkesh. I expect someone will fine-tune a model specifically for editing podcasts that will be good enough for most professional podcasters, just like how many copy writers have been replaced by LLMs that are noticably worse.
I’m not claiming that this will solve the problem by itself nor does it have to. Recursive fine-tuning just have to teach the model enough about the context that long-context windows and RAG can handle the rest. And, as you observe, there we will develop better schemes for ensuring that the right information makes it into the context.
I’m also not claiming that more than 25% of white collar employment would disappear because there’s the potential for many of the displaced workers to shift into new jobs, at least in the short term. But I suspect more than 25% of current jobs could be displaced and that current capabilities could be more economically transformative than the internet.
Dwarkesh writes: “Sometimes people say that even if all AI progress totally stopped, the systems of today would still be far more economically transformative than the internet. I disagree. He writes: “If AI progress totally stalls today, I think <25% of white collar employment goes away.”
I also feel continual learning is overrated as a barrier, but for slightly different reasons. Firstly, it’s worth reviewing why Dwarkesh believes that continual learning is necessary for economically transformative impact[1]:
One of my main criticisms is that he neglects the ability of LLMs to be recursively specialised. For example, a legal AI company could fine-tune a frontier model specifically for handling legal tasks and then further specialise the model to handle law in different countries. Major lawfirms could have a model further fine-tuned to their needs and these models might be further specialised to the needs of individual departments or even roles. This is only happening to a limited degree now because general models are advancing so fast that it’d hardly be worth it, but if progress stalls, we’ll start seeing more layers of specialisation, especially since a significant proportion of the work involved in fine-tuning could be automated.
Whilst Dwarkesh doesn’t comment on recursive fine-tuning specifically, he does comment on RL fine tuning more generally: “But it’s just not a deliberate, adaptive process the way human learning is. My editors have gotten extremely good. And they wouldn’t have gotten that way if we had to build bespoke RL environments for different subtasks involved in their work. They’ve just noticed a lot of small things themselves and thought hard about what resonates with the audience, what kind of content excites me, and how they can improve their day to day workflows.”
I agree that it’d be much harder for the economics to work out for a podcaster like Dwarkesh vs. a major firm. However, AI wouldn’t have to automate every niche role in order to have an economically transformative impact. It would just have to automate a sufficient number of major sectors. But I’m pretty sure AI models would be transformative in terms of podcasts too. Not all podcasts are as unique, or have the same quality standards, as Dwarkesh. I expect someone will fine-tune a model specifically for editing podcasts that will be good enough for most professional podcasters, just like how many copy writers have been replaced by LLMs that are noticably worse.
I’m not claiming that this will solve the problem by itself nor does it have to. Recursive fine-tuning just have to teach the model enough about the context that long-context windows and RAG can handle the rest. And, as you observe, there we will develop better schemes for ensuring that the right information makes it into the context.
I’m also not claiming that more than 25% of white collar employment would disappear because there’s the potential for many of the displaced workers to shift into new jobs, at least in the short term. But I suspect more than 25% of current jobs could be displaced and that current capabilities could be more economically transformative than the internet.
Dwarkesh writes: “Sometimes people say that even if all AI progress totally stopped, the systems of today would still be far more economically transformative than the internet. I disagree. He writes: “If AI progress totally stalls today, I think <25% of white collar employment goes away.”