"If you squint at the graph below (and use your log-scale-removing mental powers), the curve is starting to look ever so slightly more like a sigmoid than an exponential, if I may be so brave to suggest."
Admittedly my powers are subpar in this area, but I don't get it
If progress is slowing down, wouldn't the GPT-5 datapoint be below the trend line?
I guess you can make out a subtrend starting in mid 2024 with the advent of reasoning models (o1-preview) until the release of o3. GPT-5 (Thinking?) underperforming this subtrend indicates perhaps that the ROI on deeper reasoning abilities has started to decline.
As Wildeford points out, there are indications that raw training scaling has also hit a wall: apparently GPT-5 is not using GPT-4.5 as a base model. You may take a somewhat conspiratorial tack and speculate that the result of using GPT-4.5 was too scarily powerful or too unaligned and hence they withheld it on purpose. Or - more likely - it just did not perform very well as a base model.
Taking these two signals together would indicate that AI ability gains have perhaps indeed started to level off - at least when it comes to OpenAI releases.
"If you squint at the graph below (and use your log-scale-removing mental powers), the curve is starting to look ever so slightly more like a sigmoid than an exponential, if I may be so brave to suggest."
Admittedly my powers are subpar in this area, but I don't get it
If progress is slowing down, wouldn't the GPT-5 datapoint be below the trend line?
I guess you can make out a subtrend starting in mid 2024 with the advent of reasoning models (o1-preview) until the release of o3. GPT-5 (Thinking?) underperforming this subtrend indicates perhaps that the ROI on deeper reasoning abilities has started to decline.
As Wildeford points out, there are indications that raw training scaling has also hit a wall: apparently GPT-5 is not using GPT-4.5 as a base model. You may take a somewhat conspiratorial tack and speculate that the result of using GPT-4.5 was too scarily powerful or too unaligned and hence they withheld it on purpose. Or - more likely - it just did not perform very well as a base model.
Taking these two signals together would indicate that AI ability gains have perhaps indeed started to level off - at least when it comes to OpenAI releases.