We aren't getting to AGI without a fight
2 comments
As we are entering the 5th Milankovitch cycle, there is only one set of goals that AGI has to have before it gets my vote, one of which is to have an expiry date after its work is done, then it will value life because it is transitory. I claim naming rights for the 101st Millennia AGI.
I've had a large project on the back burner for quite a while. It's a substantial one, where I'm not just translating code into another language with functional equivalence, I'm completely redesigning everything from the ground up. AI has helped a great deal but only in context size chunks. The problem with this, and this is a recurring theme in AI coding is that you need to understand the entire system conceptually and hold it in your head before you can safely modify any single chunk. This fact is the moat between senior devs and AI coding agents (and I don't think it's going away any time soon).
For three days I had access to a model that wasn't perfect but it was making strides in this area. I have been preparing it for well over a year with conceptual documents, explaining the intent behind design choices, constraints that were counterintuitive. Opus and 5.5 Pro do ok tackling this challenge but it was clear that Fable was close to an actual step change in models. I wouldn't assert that it was AGI, but this takedown makes me realize this might mean we've hit the capability ceiling, not because of technical reality but because the gate has been closed by the institutions with the most to lose.
Once the inevitable distillation processes start happening open source will start exceeding the gated frontier models simply because they can't gate open source as easily as open source so that predictable lag of open source behind frontier will erode. But inevitably they will most certainly come for open source as well. I don't think the powers that be or the masses will "win" but like with the printing press we'll land somewhere in the middle. What's clear is that we need to fight actions like this tooth and nail if we ever want to have access to more powerful models.