Post #3940920

2026-07-19 16:37 UTC

@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

Replies (3)

@mikalai@privacysafe.social 2026-07-19 19:41

@troed@swecyb.com We stand on shoulders of giants. In particular, we pickup and follow good patterns in coding. Hence, I'd expect what you say: - your code wad never in training data, - you are good developer, and use known coding pattern, - an averaged Large Language Condensate has some shards of patterns you use. It really checks out. Just don't mistake it for machine understanding. The non-coder just run with it as understanding, in legal, insurance claims, i.e. horrific @david_chisnall@infosec.exchange @hfalcke@mastodon.social

Open ##3940914
@mikalai@privacysafe.social 2026-07-19 19:46

@troed@swecyb.com When humans add agents, processes of cross guessing to remove really bad guesses, when humans inject parsers under LLM skin, when humans add these, how can we call it emergent. Emergent is when it shows up itself. You may say I am nitpicking. But this is a fundamental difference that also speaks to potential, and funding/investing. Emergent was promised, but hadn't emerged, yet. AGI goes into the same bin. @david_chisnall@infosec.exchange @hfalcke@mastodon.social

Open ##3940979
@mikalai@privacysafe.social 2026-07-19 19:52

@troed@swecyb.com If best description is "thinking", how come it can't already write an OS for non-programmer? This isn't the best description. Tools never seen - it doesn't see the tool. Else you could point me to part doing seeing/understanding. It isn't there. @david_chisnall@infosec.exchange @hfalcke@mastodon.social

Open ##3941065