I agree with your premise, nice write up. I do have a question to dig into your thinking:
Yes, GPT-5 isn't much of a leap from o3 (slightly better on benchmarks and cheaper?). However, if the only model you had experience with was the original GPT-4, the one people were using to rewrite the Declaration of Independence in the style of Snoop Dogg in early 2023, wouldn't GPT-5-Thinking seem like a fairly shocking upgrade in capability?
It's only been 2.5 years since GPT-4 was released and we've covered a lot of capability ground in that time. The incremental releases we've seen since them has led to a 'frog in boiling water' effect, which OpenAI has said is their intent: many small incremental upgrades will give society a better change at adapting.
All that said, I do agree that GPT-5's overall capabilities make me more pessimistic that we will see somethin akin to AGI in the next ~2-3 years.
To my mind there have been three significant milestones in LLM development:
1. GPT-3, which absolutely blew my mind the first time I saw it
2. GPT-4, where they scaled up the model an order of magnitude
3. o1 and other reasoning models
GPT-5 is certainly impressive compared to GPT-4, but that's mainly because it incorporates reasoning.
I'm sure we'll see another big jump in capabilities at some point but I'm not sure anyone even knows yet how we'll do it. Compare that to, say, Moore's Law, which really was exponential progress. You can bet that every year the chip designers knew exactly how they were going to achieve the performance improvements planned for next year's version.
I agree with your premise, nice write up. I do have a question to dig into your thinking:
Yes, GPT-5 isn't much of a leap from o3 (slightly better on benchmarks and cheaper?). However, if the only model you had experience with was the original GPT-4, the one people were using to rewrite the Declaration of Independence in the style of Snoop Dogg in early 2023, wouldn't GPT-5-Thinking seem like a fairly shocking upgrade in capability?
It's only been 2.5 years since GPT-4 was released and we've covered a lot of capability ground in that time. The incremental releases we've seen since them has led to a 'frog in boiling water' effect, which OpenAI has said is their intent: many small incremental upgrades will give society a better change at adapting.
All that said, I do agree that GPT-5's overall capabilities make me more pessimistic that we will see somethin akin to AGI in the next ~2-3 years.
To my mind there have been three significant milestones in LLM development:
1. GPT-3, which absolutely blew my mind the first time I saw it
2. GPT-4, where they scaled up the model an order of magnitude
3. o1 and other reasoning models
GPT-5 is certainly impressive compared to GPT-4, but that's mainly because it incorporates reasoning.
I'm sure we'll see another big jump in capabilities at some point but I'm not sure anyone even knows yet how we'll do it. Compare that to, say, Moore's Law, which really was exponential progress. You can bet that every year the chip designers knew exactly how they were going to achieve the performance improvements planned for next year's version.