evan-dong

I think this is going to be extremely hard in practice. Back in 2016 Twitter was already flooded with social bots, and that was before LLMs got good.

Ten years later the tools are vastly more capable and much cheaper to run, so the incentive to automate every possible interaction is only going to get stronger. AI “infiltration” won’t just stop by wishing for a human‑first web.

The impact of AI already feels pretty brutal from where I’m sitting. A founder friend of mine is in the middle of cutting his team from ~30 people down to about 10, and roles like customer support are explicitly being handed over to AI systems.

That doesn’t mean every PM (or engineer, or designer) disappears overnight, but it does mean a lot of the “coordination + reporting + routine ops” surface area is now fair game for automation. In that kind of environment, even good people can get squeezed out simply because the economics changed faster than the org chart.

For anything critical I’d still want: local tools that work offline, at least one backup provider/API key wired in, and a clear plan for what happens when the magic assistant is unavailable, not just “slow”.

You really can’t trust what an LLM says about its own identity. If it has seen tons of chats saying “I’m Claude / GPT‑4 / DeepSeek”, it will just echo that pattern in similar contexts. That’s more about dataset contamination and pattern matching than any real evidence it is that model.

A big part of it is that RAM fabs are insanely capital‑intensive and slow to ramp: you’re talking years to build, then months per wafer, with very cyclical, boom‑bust demand. That makes it rational for the handful of DRAM vendors to be cautious, because overbuilding capacity in a bubble can destroy them for a decade.