> At some point they will want someone new

If we freeze AI models at their current stage of development, then yes. If not, given the rate at which models are improving (compare Claude today to Claude 1 year ago), in 3 to 5 years software development could be fully automated.

Well, if you automate everything, the bad parts will still be there, then you need real engineers to fix them

Yes, however like in modern factories, very few of them.

Including development of AI models?

Possibly. But even if not, how many AI model developers do we need? We will likely need a nonzero number, but it could be very few compared to the total number of software developers we have today.

Called "recursive self-improvement"

we dont' need to "freeze" models if next versions aren't a huge improvement over previous ones. That seems to be case now.

Really? Only last month OpenAI declared the "AGI era" is here.

Do you keep a lot of Tesla and SpaceX stocks in your portfolio?

OpenAI and Anthropic say a lot of shit to keep investments flowing in while they try to find an avenue for profitability.

Are you sure that's how this works?

Let's see, my grandparents were about 160cm when they were 35. My parents were about 170cm when they were 35 and I am about 180cm around 35. You are saying I can expect my grandchildren to be around 200cm at that age and their grandchildren 220cm?

Or we could assume since new data is increasingly AI generated then the next generation of models are inherently born out of incestuous relationships so I would imagine Claude's grandson is going to be a dwarf.

Of course it is possible also that Claude will marry a chinese model (since it's asian odds are it is also smarter) bringing some diversity into the "gene" pool. Or better yet, a Russian supermodel.

Are you sure that AI model progression and the height of your family members are 35 is related? It seems unlikely but I guess you'd know better than I.

That's their point, we can't assume model capabilities will go up linearly forever just because they have increased quickly in the past

"We're running out of data, so LLM progress is hitting a wall" was a prominent late-2024 narrative. Back then many experts including Ilya used the word "plateau" when describing LLM progress. No one is using that word now, and there is no evidence we are about to reach a plateau, even though we ran out of data long ago.

The way Claude writes code comments there will be no shortage of data.

If we didn't have any humans before your grandparents, and these were the only data points we have to extrapolate from, then, yes, this would be a very reasonable prediction.

  > given the rate at which models are improving

Right. https://xkcd.com/605/

  > in 3 to 5 years software development could be fully automated.

Yes, and as we all know, AI has replaced software engineers "in 6 months" several times by now.

Do you have anything to back up this 3-5 years claim or is this another hollow 'full self driving by the end of next year' promise?

Some of that is long pre-existing, where too many companies would go: "Hey, let someone else hire juniors and train them up, we'll them away after that."

Not that the new LLM-as-junior trend doesn't make it worse.