Work with AI directly — without bots, wrappers, aggregators and extra plugins
31 episodes in 12 months · 51 total · first in June 2024
Through a Telegram bot, an aggregator like Perplexity or an assistant built into someone else's program, you do not get the model you pay for but a cheap, cut-down version. Every extra layer in between spoils the result.
With strong models this also applies to plugins and ready-made instructions: more often than not they only limit the model. Open ChatGPT, Claude or Codex and set the task right there, and try a plugin only when you understand what it gives you.
Happening now. The layers in between are dying off: by No. 155 (August 2026) the hype around OpenClaw had died down and work had moved straight into Codex and Claude Code, and in No. 167 (September 2026) — “work directly… without plugins”.
Where to start
Get rid of the Telegram bots and aggregators: open ChatGPT, Claude or Codex itself and work there.
Do not build a product as a wrapper around someone else's model: the platform itself will build that feature in.
Layers in between are justified only in genuinely complex cases.
“Once again, in general I categorically don't support this, various middle layers; I support them specifically in complex, strictly very complex cases.”
When I said it
Quotes
“So through a bot the quality of the answers is completely different. I mean, through a bot you're getting access to, let's say, to cheap infrastructure.”
“That's exactly why, by the way, I never switched to Perplexity. Why would I switch to a third-party platform, an aggregator of, let's say, everything, if inside I can use all sorts of super features.”
“…I, on the contrary, remove these middle layers, and I get more effective, incredibly super cool conclusions, yeah, just super-quality conclusions of a completely different kind.”
“For that you need to work directly with artificial intelligence all the time, without plugins, you need to work with the latest cutting-edge models without plugins…”