AI stock prices could collapse and AI intelligence is over-hyped
Nikos Katsikanis - 13 October 2025
I have spent enough nights pairing with agents to know they are accelerators, not oracles. The market keeps telling me AI will replace whole teams any day now, but I keep seeing the cracks up close.
Agents amplify what I already know
When I ask an agent to extend my Codex and GPT pairing workflow, it mirrors whatever context I provide. If I understand the module, the agent helps me draft tests and fill in the glue code quickly. If I do not grasp the subsystem, the suggestions wobble. I still need to set the direction, review every important change, and correct the spots where the agent hallucinated business rules. The quality of my prompt is limited by the quality of my own understanding.
I treat code as the single source of truth. The more clarity I bring to the table, the more the agent amplifies my strengths. It is not going to invent a product vision for me, and it definitely will not debug a problem that I cannot describe.
Teams are worth more than rushed layoffs
I keep hearing from founders who think they can trim their payroll once an agent lands in the IDE. That move ignores how long it takes a developer to internalise a codebase. When a teammate leaves, I lose historical context, the unwritten conventions, and the reflexes built from late night incidents. Training an agent without that human review just compounds the risk.
The smarter play is to keep the core team and teach them how to steer the tooling. Once developers are comfortable pairing with agents, they review diffs faster, they batch more work per session, and they catch regressions early. I have seen this first hand on client work where the agent drafts migrations and the developer signs off. That balance is healthier than firing staff and hoping an AI will learn the system solo.
Markets and AI models can still deflate
Model quality could slide back to whatever we can run locally on a chunky laptop if AI companies go bust. Right now that means something around GPT-3 level if you are willing to wire up an Ultra spec EPC rig in your home office.
Cloud models will not vanish, but their price points can climb fast. When budgets get tight, companies will pick fewer API calls and shift more workloads to self hosted checkpoints.
Local models will always be around to do boring CRUD stuff
Even if the market cools, GPT-3 class models still automate the tedious onboarding tasks that usually slow me down. They write boilerplate, scaffold CRUD screens, and sanity check my docstrings. That frees me up to focus on the gnarly parts that demand a stronger hosted AI model.