The point isn't that progress is not happening but that it's slowing down. You get more of the same, smaller memory footprint, faster responses, less hallucinations, etc. Significant progress would be another deep seek kinda of breakthrough, near 0% hallucination rate, performing like current models with less than half of their dataset, epistemological self awareness (i.e. I am not sure of the correctness of the answer I just gave you, ability to override assumptions from the training dataset, etc).

We are just getting cars of different shapes and colours, with built-in speakers and radio. Not exactly progress

Reasoning models like o1 are on a whole new level compared to previous models, for example they are incredible at math, something that previous models struggled with a lot, this seems pretty huge to me as the performance of previous models was flattening, it's a kinda a new paradigm.

> another deep seek kinda of breakthrough

That was only six month ago. I don't think this is an argument that things are slowing down (yet).

I didn't say that progress stopped only that it is slowing down (ie they become less frequent). Deep seek happening 6 months ago doesn't counter what I said.

>but that it's slowing down

Progress isn't a smooth curve but more step like.

Also, the last 10% of getting AI right is 90% of the work, but it doesn't seem to us humans. I don't think you understand the gigantic impact that last 10% is going to make on the world and how fast it will change things once we accomplish it.

Personally, I hope it takes us a while. We're not ready for this as a society and planet.

This assumes that LLMs are only useful if they are AGI. I don't think we do - what we have today is already sufficient to unlock an enormous amount of value, we just haven't done so yet.

Thus, I think we can compare them to electricity - a sophisticated technology with a ton of potential, which will take years to fully exploit, even if there are no more fundamental breakthroughs. But also not the solution to every single problem.

Arguably, LLMs - or whatever systems succeed them - are only useful if they are not AGI. Given the evidence already collected about how willing humans are to make these systems "agentive", we pretty well have to worry about the possibility of an AGI using us instead. Even if there's some other logical barrier to recursive self-improvement ("hard takeoff") scenarios.

The consequences of a new thing being invented are not entirely dependent on progress of innovation in the thing itself. It takes quite a long time for people to build _on top of_ a new technology. LLMs could not appreciably improve at all, and we've still barely scratched the surface of applying what we have.

Your car example is a perfect one -- society was _completely reordered_ around the car, even though the fundamental technology behind the car didn't change from the early 20th century until the invention of the electric car.

Surely, finding new applications to existing tech can't be considered progress in the development of that tech

That's also one of those things that would probably confuse LLMs as readily as it confuses North Americans (for much the same reason - training).

Spring 2024 for me was from the 1st of September to the 30th of November.

Only about 10% of earths population live in the southern hemisphere. It’s pretty fair to assume northern.

Maybe, but if you're trying to appear folksy by using region-dependent seasonal identifiers rather than, you know, dates AND you're happy that instantly means 10% of your readers will be confounded by that, then I guess go for it.

I'll note that about 40% of the earth's population lives in the tropics, where seasons also don't break down the same way they do for the author - typically that demographic will recognise two identifiable seasons through the year.

What's the confusion here? The author is from Sweden, also neither I nor the author are North Americans.

While one half of the planet is having spring, the other half is having autumn.

Yeah that's something you generally learn when you are a kid.

We should catch up next autumn for a quiet ale to talk about the ambiguity of that date format.

o1 was not a major advancement