[talldayo]

GPT-4, GPT-3 and GPT-2 all coded up small projects, though. If GPT-5 is just another "it gets it right sometimes" dice-roll machine then it won't matter how many modalities it has. It will be DOA, and we'll all be pinning out hopes on GPT-6.

This shtick is running out of steam, and it's becoming increasingly clear that Sam Altman (or anyone at OpenAI for that matter) doesn't have what it takes to keep it going. The honeymoon phase is over, you can't promise capabilities that don't exist without proving that they're possible in the first place. Superintelligence is one of those things.

[falcor84]

> If GPT-5 is just another "it gets it right sometimes" dice-roll machine then it won't matter how many modalities it has. It will be DOA

Every single thing and every single person can be seen as a "dice-roll machine"; when I give even a senior engineer a task to take on this sprint, there's a significant chance they won't complete it that sprint (and it might just have been infeasible at all). At the end, all work boils down to the project management triangle (On Time, On Spec, On Budget - choose two), and the better the AIs get, the more tradeoffs we can choose from and the more and more we'll gradually be using them, regardless of whether we'd able to ever trust them to complete a task perfectly 100% of the time.

[talldayo]

If you want to be reductive, then sure. But even then, "superhuman intelligence" in this case literally means not being a dice-roll. I do not understand how you can pitch AGI without solving the hallucination problem, you might as well be selling people a random-number-generator that can do math with entropy.