Yes. I don’t know about the timeline for the higher bar definitions of AGI. For the lower bar definitions, we’re basically already at “good enough” today.
If you’d told me 10 years ago that I’d be able to run a program on my computer which would let me feed in an image along with some CSS and JS files and it would then give me a correction that fixes the bug I indicated visually… I would not have believed you. Here I am in 2026 though, and I have done exactly that several times with local LLMs on my own hardware. That same program can also take a natural language description of characters, motivations, and a vague scenario and write a scene. Not an especially well written scene, most of the time, but good enough to get the characters from the initial conditions to ending conditions via complex intermediate steps. I can also define tools it hasn’t seen before and it can combine them in sequence to solve a problem defined in natural language. Is it perfectly reliable? Hell no. Is it always coherent? Definitely not. The fact that it can do as much as it can is just bonkers though. If we’re getting this far with what I strongly suspect is not the ideal architecture for general intelligence, god only knows what we’ll see when we do hit on the right architecture.
Yes. I don’t know about the timeline for the higher bar definitions of AGI. For the lower bar definitions, we’re basically already at “good enough” today.
If you’d told me 10 years ago that I’d be able to run a program on my computer which would let me feed in an image along with some CSS and JS files and it would then give me a correction that fixes the bug I indicated visually… I would not have believed you. Here I am in 2026 though, and I have done exactly that several times with local LLMs on my own hardware. That same program can also take a natural language description of characters, motivations, and a vague scenario and write a scene. Not an especially well written scene, most of the time, but good enough to get the characters from the initial conditions to ending conditions via complex intermediate steps. I can also define tools it hasn’t seen before and it can combine them in sequence to solve a problem defined in natural language. Is it perfectly reliable? Hell no. Is it always coherent? Definitely not. The fact that it can do as much as it can is just bonkers though. If we’re getting this far with what I strongly suspect is not the ideal architecture for general intelligence, god only knows what we’ll see when we do hit on the right architecture.