I recently had a discussion about whether we are at the 300baud modem level of AI models, or at the first home broadband stage.
This is basically what is asking here, although partially a bigger question too about not just quantity but capability.
I got a 1200baud modem at home as my first Internet connection, connected to an Apple IIe. Just a little past the analog receiver in a cradle phase1.
I used it to connect to Compuserve and local Bowen Island Users Group (BUG) bulletin board system. This would have been around 1985 or so.
10~ish years later, in 1996 when I moved into a big shared house just off campus while going to the University of Victoria (UVIC), we got onto the beta of an always on broadband connection with a Rogers Cable modem. I think like 5Mbps down, and 1Mbps up? And of course LAN parties, blue 10Mbps Ethernet cables stapled to wood paneling2.
At the same time, looking at the Wikipedia modem page, looks like 56k modems came out in 1998.
So I was living in the broadband future in 1996, even while the alternate technology of modems was still being rolled out.
For developers I think the "first broadband moment" happened with AI models for coding purposes in December 2025. We're now 6 months on from that, with another wave of models being released.
And there's also the analogy of AI model as internal combustion engine3: the Model T worked with a basic engine, which was a huge invention, and then there were massive advances aside from the engine - better tires, chassis, streamlining, brakes, headlights, radio etc etc (many of which we see as basic now).
We can see many people rapidly advancing harnesses and other things around the AI model, while the AI models also increase in capabilities.
And of course jaggedness: the uneven success rate of AI models being able to complete tasks on the frontier that appear similarly hard to humans. Read Helen Toner's Taking Jaggedness Seriously for more on this.
The combination of good success in one domain (coding) and variable access to high performing models, and learning of using surrounding tooling is going to continue to make it difficult to collectively have this discussion.
I have purposely avoided using models because I didn't want to get locked into thinking about what a 300baud modem can do. More on this soon.