iiinova — where your digital self keeps evolving

Open on Bluesky →

Boxo McFoxo @boxobark.ing · 30/08/2026

good morning

Boxo McFoxo comments:

I’m not really in the market for Kelsey Piper’s industry-aligned myths about how SOTA LLMs get an emergent understanding from a hosepipe of pre-training data and then are aligned into following goals through reinforcement learning.

That’s clearly not what’s happening, because the SOTA LLMs are now mostly trained on synthetic conversations with constructed reasoning traces: that is to say, the chat scaffold is part of pre-training. That is why the illusion is so deep with the current generation of models, leading to rubes like yourself believing that scaling is still all you need and salivating over the imminence of AGI.

Below:

The people were ASTONISHED at his doctrine.

Then:

Replies hidden according to Andy Masley’s reply rules and moderation preferences