Grappling with this hard right now. Anyone who is still of the "these things are stupid and will never replace me" mindset needs to sober up real quick. AGI level agentic systems are coming, and fast. A solid 90% of what we thought of as software engineering for the last 30 years will be completely automated by them in the next couple years. The only solution I see so far is to be the one building them.
As someone who's personally tried ( with lots of effort) to build agentic assistants/systems 3+ times over the course of the last few years I haven't seen any huge improvements in the quality of output. I think you greatly underestimate the plateau these models are running into.
Grok and o1 are great examples of how these plateaus also wont be overcome with more capital and compute.
Agentic systems might become great search/research tools to speed up the time it takes to gather (human created) info from the web, but I don't see them creating anything impressive or novel on their own without a completely different architecture.
>As someone who's personally tried ( with lots of effort) to build agentic assistants/systems 3+ times over the course of the last few years I haven't seen any huge improvements in the quality of output. I think you greatly underestimate the plateau these models are running into.
As someone who's personally tried with great success to build agentic systems over the last 6 months, you need to be aware of how fast these things are improving. The latest Claude Sonnet makes GPT-3.5 look like a research toy. Things are trivial now in the code gen space that were impossible just earlier this year. Anyone not paying attention is missing the boat.
>As someone who's personally tried with great success to build agentic systems over the last 6 months.
Like what? You're the only person ive seen claim they've built agentic systems with great success. I dont regard improved chat-bot outputs as success, im talking about agentic systems that can roll their own auth from scratch, or gather data from the web independently and build even a mediocore prediction model with that data. Or code anything halfway decently in something other than Python.
