5 ms·
Maybe it's just the nature of being early adopters. Other fields will get their turn once a baseline of best practices is established that the consultants can
by grim_io 1y ago
Maybe it's just the nature of being early adopters.
Other fields will get their turn once a baseline of best practices is established that the consultants can sell training for.
In the meantime, memes aside, I'm not too worried about being completely automated away.
These models are extremely unreliable when unsupervised.
It doesn't feel like that will change fundamentally with just incrementally better training.
- ACCount37 1y agoDoes it have to? Stack enough "it's 5% better" on top of each other and the exponent will crush you.
- cjs_ac 1y agoAre LLMs stackable? If they keep misunderstanding each other, it'll look more like successive applications of JPEG compression.
- ACCount37 1y agoBy all accounts, yes. "Model collapse" is a popular idea among the people who know nothing about AI, but it doesn't seem to be happening in real world. Dataset quality estimation shows no data quality drop over time, despite the estimates of "AI contamination" trickling up over time. Some data quality estimates show weak inverse effects (dataset quality is rising over time a little?), which is a mindfuck. The performance of frontier AI systems also keeps improving, which is entirely expected. So does price-performance. One of the most "automation-relevant" performance metrics is "ability to complete long tasks", and that shows vaguely exponential growth.
- grim_io 1y agoThe jpeg compression argument is still valid. It's lossy compression at the core.
- ACCount37 1y agoI don't think it is. Sure, you can view an LLM as a lossy compression of its dataset. But people who make the comparison are either trying to imply a fundamental deficiency, a performance ceiling, or trying to link it to information theory. And frankly, I don't see a lot of those "hardcore information theory in application to modern ML" discussions around. The "fundamental deficiency/performance ceiling" argument I don't buy at all. We already know that LLMs use high level abstractions to process data - very much unlike traditional compression algorithms. And we already know how to use tricks like RL to teach a model tricks that its dataset doesn't - which is where an awful lot of recent performance improvements is coming from.
- grim_io 1y agoSure, you can upscale a badly compressed jpeg using ai into something better looking. Often the results will be great. Sometimes the hallucinated details will not match the expectations. I think this applies fundamentally to all of the LLM applications.
- muldvarp 1y agoAnd if you get that "sometimes" down to "rarely" and then "very rarely" you can replace a lot of expensive and inflexible humans with cheap and infinitely flexible computers. That's pretty much what we're experiencing currently. Two years ago code generation by LLMs was usually horrible. Now it's generally pretty good.
- grim_io 1y agoI think you are selling yourself short if you believe you can be replaced by a next token predictor :)
- ACCount37 1y agoI think humans who think they can't be replaced by a next token predictor think too highly of themselves. LLMs show it plain and clear: there's no magic in human intelligence. Abstract thinking is nothing but fancy computation. It can be implemented in math and executed on a GPU.
- Aloisius 1y agoGiven the number of academic papers about it, model collapse is a popular idea among the people who know a lot about AI as well. Model collapse is something demonstrated when models are recursively trained largely or entirely on their own output. Given most training data is still generated or edited by humans or synthetic, I'm not entirely certain why one would expect to see evidence of model collapse happening right now, but to dismiss it as something that can't happen in the real world seems a bit premature.
- ACCount37 1y agoWe've found in what conditions does model collapse happen slower or fails to happen altogether. Basically all of them are met in real world datasets. I do not expect that to change.
- anthem2025 1y agoPretty crazy, and all you have to do is assume exponential performance growth for as long as it takes.
- OtherShrezzing 1y agoAI training costs are increasing around 3x annually across each of the last 8 years to achieve its performance improvements. Last year, spending across all labs was $150bn. Keeping the 3x trend means that, to keep pace with current advances, costs should rise to $450bn in 2025, $900bn in 2026, $2.7tn in 2027, $8.1tn in 2028, $25tn in 2028, and $75tn in 2029 and $225tn in 2030. For reference, the GDP of the world is around $125tn. I think the labs will be crushed by the exponent on their costs faster white-collar work will be crushed by the 5% improvement exponent.
- johnnienaked 1y agoYour math is a bit less than it should be because you doubled instead of trebled 2026
- pkaye 1y agoThe current trained models are already pretty good enough for many things.
- Collisteru 1y agoBe careful you're not confusing the costs of training an LLM and the spending from each firm. Much of that spending is on expanding access to older LLMs, building new infrastructure, and other costs.
- OtherShrezzing 1y agoThat’s a fair criticism of my method, however model training costs are a significant cost centre for the labs. Modelling from there instead of from total expenditure only adds 2-3 years before model training costs are larger than the entire global economy.
- muldvarp 1y ago> These models are extremely unreliable when unsupervised. > It doesn't feel like that will change fundamentally with just incrementally better training. I could list several things that I thought wouldn't get better with more training and then got better with more training. I don't have any hope left that LLMs will hit a wall soon. Also, LLMs don't need to be better programmers than you are, they only need to be good enough.
- grim_io 1y agoNo matter how much better they get, I don't see any actual sign of intelligence, do you? There is a lot of handwaving around the definition of intelligence in this context, of course. My definition would be actual on the job learning and reliability i don't need to second guess every time. I might be wrong, but those 2 requirements seem not compatible with current approach/hardware limitations.
- muldvarp 1y agoIntelligence doesn't matter. To quote "Superintelligence: Paths, Dangers, Strategies": > There is an important sense, however, in which chess-playing AI turned out to be a lesser triumph than many imagined it would be. It was once supposed, perhaps not unreasonably, that in order for a computer to play chess at grandmaster level, it would have to be endowed with a high degree of general intelligence. The same thing might happen with LLMs and software engineering: LLMs will not be considered "intelligent" and software engineering will no longer be thought of as something requiring "actual intelligence". Yes, current models can't replace software engineers. But they are getting better at it with every release. And they don't need to be as good as actual software engineers to replace them.
- grim_io 1y agoThere is a reason chess was "solved" so fast. The game maps very nicely onto computers in general. A grandmaster chess playing ai is not better at driving a car than my calculator from the 90s.