The more these foundational AI companies focus on product development, the more convinced I am that improvements in intelligence of the base models have slowed down to a point where AGI/"superintelligence" isn't happening, and they have to sustain their valuations in alternate ways.
Agreed 1000%. And the rate ate which OpenAI is spitting out new product tests like this I think they know very well they're hitting the limits of what the underlying model can do
We can wait for Gemini 3.0 to see if it's a huge improvement, but my best guess is that if OpenAI couldn't get a meaningful improvement, it's more likely that it's non-trivial to be gotten than they're just incompetent.
I'd think you have to do product development. Unless you get to true SUPERintelligence, even getting to AGI by itself gives you very little. Imagine they got to true AGI, but with an IQ of 100 -- what does that open up for you? I can't think of much. You really need to start doing product development so that you can use the 100 IQ AI to do things like automate some processes or do web queries for you, etc... With today's LLMs we already get most of the value of 100 IQ AGI. Now if you give me 300 IQ super intelligence then that's a game changer.
The explanation behind the acquisition of Statsig is so obvious.
OAI is taking a Meta-old-school approach towards building products. Did it work at Meta? Nah, they ended up acquiring since their internal efforts were not preferred by the users in the market.
In my day-to-day work, after having some time to use it regularly, I haven't noticed GPT-5 being way better than o3 or o4-mini-high. It is marginally better.
Fair but my intent was to hit at something else. It's more about how to delegate a portion of a system over to an agent. In an AGI scenario, there's no world in which I just hand over my computer and browser with all my logins to an agent - just like I'd never do that with a human. It's insane. So whether you're using an AGI or not, there's going to need to be some change to software it's using such that it can be used in a safer, delegated way.
Like, how do you get your human assistant to book you a flight? They have a corporate credit card with limits on it where they can't spend on just anything. I'm guessing it would be the same thing. You'd need to implement safeguards so the agent can't screw up
