Citation :
A Job is not just a bundle of predefined skills and tasks
Came across this substack post from podcaster Dwarkesh Patel and it cleanly summarized something I think a lot of AI bears have been saying the past few years. The tldr is that a job is not just a set of skills, and even the jobs you think are easy require open-ended reasoning, learning, and adaptation that no AI is capable of and will not become capable of just because you create a billion learning environments for reinforcement learning .
Citation :
I was at a dinner with an AI researcher and a biologist. The biologist said she had long timelines. We asked what she thought AI would struggle with. She said her work has recently involved looking at slides and decide if a dot is actually a macrophage or just looks like one. The AI researcher says, “Image classification is a textbook deep learning problem—we could easily train for that.”
I thought this was a very interesting exchange, because it revealed a key crux between me and the people who expect transformative economic impacts in the next few years. Human workers are valuable precisely because we don’t need to build schleppy training loops for every small part of their job. It’s not net-productive to build a custom training pipeline to identify what macrophages look like given the way this particular lab prepares slides, then another for the next lab-specific micro-task, and so on.
What you actually need is an AI that can learn from semantic feedback or from self directed experience, and then generalize, the way a human does.Every day, you have to do a hundred things that require judgment, situational awareness, and skills & context learned on the job. These tasks differ not just across different people, but from one day to the next even for the same person. It is not possible to automate even a single job by just baking in some predefined set of skills, let alone all the jobs.
|
Patel also makes a great point about shifting goalposts, although I don't think he really understands the implications (what I'll explain below)
Citation :
AI bulls will often criticize AI bears for repeatedly moving the goal posts. This is often fair. AI has made a ton of progress in the last decade, and it’s easy to forget that.
But some amount of goal post shifting is justified. If you showed me Gemini 3 in 2020, I would have been certain that it could automate half of knowledge work. We keep solving what we thought were the sufficient bottlenecks to AGI (general understanding, few shot learning, reasoning), and yet we still don’t have AGI (defined as, say, being able to completely automate 95% of knowledge work jobs). What is the rational response?
It’s totally reasonable to look at this and say, “Oh actually there’s more to intelligence and labor than I previously realized. And while we’re really close to (and in many ways have surpassed) what I would have defined as AGI in the past, the fact that model companies are not making trillions is revenue clearly reveals that my previous definition of AGI was too narrow.”
https://substack.com/home/post/p-180546460
|
Despite understanding that the goalposts aren't meaningful, Patel is still, in his words, bullish on agi in the long-run. I guess if you define the long run as anytime between now and the heat death of the universe, bullishness may be justified. But long-term bullishness is usually like 25-50 years timeline, and I don't think that is justified.
The problem I would argue is two-fold. First, there's only really been one actual method for cognitive automation that has worked: programming rules and heuristics into a model. That was what expert systems was in the 1980s, and I would argue, what deep learning essentially still is. The difference is that with deep learning you are using an immense amount of compute and data to identify some of the rules (or patterns) in the data that can be applied to slightly different contexts. But both expert systems and deep learning are brittle. They fail when they encounter any problem which cannot be solved by the rules which they have already been programmed with or that the learned during training. Here is how one AI researcher put it
Citation :
When we see frontier models improving at various benchmarks we should think not just of increased scale and clever ML research ideas but billions of dollars spent paying PhDs, MDs, and other experts to write questions and provide example answers and reasoning targeting these precise capabilities ... In a way, this is like a large-scale reprise of the expert systems era, where instead of paying experts to directly program their thinking as code, they provide numerous examples of their reasoning and process formalized and tracked, and then we distill this into models through behavioural cloning.
|
https://www.beren.io/2025-08-02-Mos [...] -Progress/
With expert systems, you are trying to come up with all the rules which may be applicable future deployment of the system. With reinforcement learning, you are trying to brute force simulate all possible futures and bake those pathways into the models weights. Both systems, to reiterate, are incapable of out-of-distribution generalization or of continual learning. The only difference between now and the 1980s and we have a lot more compute and data.
So when AI bulls claim that they are going to solve limitations such as continual learning or self-motivation or out-of-distribution generalization or world modeling in the next 5-10 years, that is a statement of faith rather than anything that can be derived from so-called scaling laws. And, I would suggest, if the ai companies really believed that, they wouldn't be talking about the need for trillions of dollars worth of GPUs. An actual AGI would be cheap.
The second problem, following from what I just said, is that no one in the AI field actually knows what intelligent is or what it entails. In fairness, I don't either, but I'm not trying to sell you anything. The long history of, "if AI can do this, then it must be generally intelligent" should be ample proof of that, going to back to the days when AI researchers believed that a program which could play chess at a human level would have to be generally intelligent.
Take one example of "not having a clue." A few weeks ago on the Patel podcast Andrej Karpathy, the former head of self-driving at Tesla, proposed that we could achieve or improve generalization among these models by implementing what he called sparse memory. His reasoning: human have bad memory and generalize well, while AI has great memory and generalizes poorly. Therefore, we should shank the AI's memory to make it better at generalization.
But the relationship between poor memory and generalization may be coincidental rather than causal. Evolution is not goal-directed. Evolution is 100 quadrillion organisms with an average of a million cells each with each of those capable of mutating at any moment and this has been going on for over 3 billion years. It results in the production of almost infinite diversity, but it is not an optimizing algorithm. Humans might have mutated much greater memory or much worse memory and still have the same level of generalization, but the memory we have is just what happened to have mutated in the past and it didn't discourage procreation and therefore it passed on. But certainly evolution didn't select specifically for our type of intelligence because there are billions of other species which are less intelligent yet manage to survive (as a species), some for millions of years. Nature has created an infinite variety and levels of intelligence through random mutation.
But even if we look at the specific configuration of human intelligence through a lens of optimization, there are much better explanations for the combination of great generalization and poor memory than direct causality. Human brains are ravenous. They make up 2% of body mass yet consume 20-25% of our calories. Chimpanzee brains, by contrast, only consume 8% of their calories. Higher intelligence confers survival advantages, but in the hunter gatherer world where they often went long periods without foods, the brains high energy demand could be a liability. A brain that can remember the migration patterns of prey animals probably has a good balance of intelligence to energy consumption. A brain that can remember any minute detail of what a person was doing on any random day 15 years earlier probably has a bad balance of intelligence to energy consumption.
The point is, looking at human intelligence as a way to model artificial intelligence is not so easy given we don't even really understand human intelligence, and the lessons we try to draw are often wrong. Another example, an ai researcher compared the problems of catastrophic forgetting, the case where trying to finetune a trained model results in the model forgetting some of the skills it learned during training, to how humans have a hard time learning a new language when they get older. Problem with this analogy is that an older person learning a new language is not going to forget the langue he currently speaks. The field of AI research is full of bad, misleading anthropomorphisms.
A more concrete example, nano banana pro has a hard time making 6 finger hands. It can, but it is extremely prompt sensitive. I asked nano banana to "generate an image of a hand with six fingers" and it drew a 5 finger hand. I asked it to "generate an image of a six-fingered hand" and again it drew a 5 finger hand. I then asked it to "generate an image of a hand that has 6 fingers" and it succeeded, but one of the fingers was splitting off from another finger. So then I asked it to "generate an image of a hand that has 6 normal fingers" and again, it drew a 5 finger hand. They've clearly done a lot to make sure the model can draw normal, 5 finger hands, but now the model struggles to draw 6 finger hands. A human who improves his ability to draw a 5 finger hand isn't going to forget how to draw a 6 finger hand.
This is getting too long, but just one more thing to address: the idea that AI doesn't have to work like human intelligence in the same way that a plane doesn't work like a bird. Here's the problem with that analogy. A plane can't do all the things that a bird can do. A plane can't fly in a forest or among houses and building. It can't take-off without a very long, clear runway nor can it land without these conditions. It was designed to do a very specific thing (carry heavy cargo fast through clear space) under very specific conditions. That is pretty much all AI is today. In other words, we already have the plane version of AI. What researchers are trying to build is the bird version of it.
|
Il va y avoir un embargo total pour obtenir la fin de l'armement nucléaire nord coréen. Les réserves nous font tenir 3 ans avant de manger l'écorce des arbres.
)