Post #3922974

2026-07-19 00:14 UTC

@troed@swecyb.com MCP servers -> add actual data to hope for better guesses. Intermingle with parsers that check correctness of other guesses. All of these are like force vectors that push guesses, mold guesses. But there are no models there. You head had many. If llm had models, it would stick to them instead of being on a wind of next token train/generator, what we call in humanized way, halucinations (thank marketing for misleading). @david_chisnall@infosec.exchange @hfalcke@mastodon.social

Replies (1)

@troed@swecyb.com 2026-07-19 16:37

@mikalai@privacysafe.social It's not wrong to refer to LLMs as "next token predictors" or that they're making "guesses", but it's a red herring. Somehow Large Language Models turned out to be pretty good at things we didn't expect them to and that's where we are now. I'm an old assembler coder from the 80s, still playing around with the platform of my childhood - the Atari ST. A year ago I made a clone of Geometry Dash, fully in assembler, using a hardware abusing trick I found myself to be able to "scroll the screen" in a new way. Since a few weeks back I'm working towards finishing up that development, and to help me do that I implemented an MCP so that my local LLM can aid in debugging. This means I'm using an LLM to write and debug Motorola 68000 assembler code in a game engine that I personally have written 100% of. The workings of my code does not exist in any LLM training data anywhere. ... and the LLM is doing absolutely great. Since this is an area (Atari ST assembler code) where it's not an exaggeration to say that few people in the world know the platform better than I do I am most definitely an authority to weigh in on how well the next-token-predictor/guesser is doing. It turns out that LLMs exhibit emergent behaviors, where the result of their training gives rise to what can best be described as "thinking" and "reasoning" - the ability to use abstract concepts together with tools they've _never_ seen before to produce actual and valuable work. Feel free to cite me. @david_chisnall@infosec.exchange @hfalcke@mastodon.social

Open ##3940920