China's DeepSeek applying trial-and-error learning to its AI 'reasoning' Chinese AI company DeepSeek has shown it can improve the reasoning of its LLM DeepSeek-R1 through trial-and-error based reinforcement learning, and even be made to explain its reasoning on math and coding problems, even though explanations might sometimes be unintelligible. The release of DeepSeek-R1 in January 2025 inspired a …
It does offer the possibility to make connections that are out there, but that we haven't yet seen.
For example, if we didn't have any idea about the link between poor dental hygiene and heart disease, if it saw large scale studies that showed a correlation between missing teeth and heart attacks (i.e. the type of large studies where many factors are shown so this particular correlation was buried in the data) and saw sequencing of arterial plaque and also of dental plaque it might see the correlation that wasn't seen by humans in such a large dataset, and also see that arterial deposits contain the same bacterial DNA as dental plaque.
Then it could infer a correlation between heart disease and oral health that had not been noticed by humans due to missing it in a large dataset and the fact that people studying dental plaque and people studying arterial deposits were (in the past at least) never the same people so that was missed as well.
So it is capable of "reasoning" in circumstances like the above. In this case we already know about the correlation, but it is a certainty there is other stuff out there. Maybe a major factor in autism is buried in the global dataset and will just need the right training data combined with the right set of queries to unearth it. It doesn't matter too much if it hallucinates sometimes, because people are going to have to investigate such suggestions regardless so even if it takes you down rabbit holes so long as it is sometimes right it is worth it.
The problem is that if AI was responsible for such a "breakthrough" a lot of people would not understand WHY it was able to do it when humans had not, and falsely believe it is far more capable than it actually is.
There are plenty of more effective ways to do this kind of narrow domain inference that have been in use for decades. In fact there are entire journals dedicated to specific classes of them.
Prodding an LLM with a stick and hoping it turns into something useful seems expensive and unlikely to work at this stage, they've already plateaued and we don't have a single working use-case for them.
"As the child navigates their avatar through the game world, they learn through trial and error that some actions (such as collecting gold coins) earn points, whereas others (such as running into enemies) set their score back to zero,"
No, the child learns by understanding context and visual cues. Shiny flashing gold thing => good, ugly scary monster => bad. There may well be hidden features that you stumble across by accident but I'd hardly describe that as trial and error.
"This contrasts with previous prompting-based approaches, which were more akin to expecting a child to learn to master a video game by having them read the instructions, or supervised-learning approaches, which can be likened to expecting the child to master a game by watching a sibling play it hundreds of times,"
Err, that's exactly how children, indeed adults too, learn, by watching and copying. A few of us even RTFM...
I suppose some people adopt trial and error but the cost of the errors is usually too great to keep that up for very long. Cutting down tree with chainsaw, learned by trial and error, not usually successful. Driving articulated lorry through city centre, learned by trial and error, poor outcome.
Its not a very good analogy for the real world, but it might be great for LLM's. I don't really care.
Yeah, I guess the authors of that "Nature" (so-called) paper never played Pac-Man, as that neither requires the use of the word "avatar", nor any sort of RTFM!
And I have to agree that there's a rotund buttock level of mumbo jumbo bollocks in that paper, with phrases like "superior, non-human-like reasoning pathways", or "incentivize the emergence of new reasoning capabilities in LLMs", and the notion that "human-defined reasoning patterns may limit model exploration". I guess they're suggesting LLMs should be programmed to follow extra-terrestrial alien approaches instead, or plants ... which sounds like the latest fashion trend on the catwalks of girthy models of languagerie these days, with the goal of producing outputs "that a human would be unlikely to rationally design" (per Niko McCarty, linked in the "functional viruses" TFA).
But I'm most bothered by the claim that "LLMs can exhibit emergent behaviours, including reasoning abilities, when scaled to a sufficient size". AI scaling "laws" (again, so-called) of this type are an obvious wishful thinking sham that really shouldn't need debunking (2016) over and over again (2025), imho.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
