Our AI is already smart enough; we just need to orchestrate it

February 15, 2025

I wrote recently about the difference between Functional and Technical AGI.

I argued that Technical AGI is the real deal (AI that can generalize it's knowledge and learning), and Functional AGI is an emulation that's good enough to replace a human knowledge worker.

But there's a follow-on point that I want to make here as well.

I don't think we need any additional model advances to get to Functional AGI.

I think our best AIs in early 2025 are already smart enough to do the job. They just haven't been properly orchestrated into unified systems that can emulate general intelligence. Yet.

People keep waiting for the big one—some new AI model that will conclusively give us AGI, ASI, or whatever acronym we are currently chasing.

But I've always thought of AGI as a system, not a model. A set of capabilities, not a single piece of tech.

So rather than there being one race for AGI, I see it as two.

1. The race to make the models/ecosystems smarter.
2. The race to build the orchestration that emulates human-like general intelligence.

I'm simply arguing that #2 does not require #1 anymore. A combination of o3, deepmind, and hell—even gpt-4—if built properly into a cohesive system of coordinated agents—will be able to replicate or exceed human performance on a wide range of knowledge work tasks.

Orchestrated agent systems, turned into products, and built on 2024-level AI, can get us to Functional AGI. And I'm predicting that happens by late 2025 or sometime in 2026 (and no later than 2028).

Improvements in the fluid intelligence of the models themselves will continue, of course, but who knows how quickly. And each time that happens, reliance on the orchestration will decrease.

Summary ​

1. We can get to Functional AGI by models themselves getting there.
2. We can also get there by 2024-level AI being properly orchestrated into complex products that emulate general intelligence.
3. These tracks of improvement will continue in parallel, but the orchestration track is enough to get us there even if model progress were suddenly frozen tomorrow (like due to a terrorist attack that stopped development through regulation).
4. Either way, I expect Functional AGI by 2028, and more likely in 2026 or 2027.