Jonathan Liles

Founder. Author of many open source projects. Making things run orders of magnitude faster in debug mode than others do in release mode for 20+ years.

[Jonathan Liles @thjonml]

@ThePrimeagen At the level you're operating at (still reading the code, hand writing parts), you haven't even begun to encounter all of the interesting phenomena and gotchas.

[Jonathan Liles @thjonml]

@badlogicgames Is everyone excessively praising it totally delusional?

[Jonathan Liles @thjonml]

@simonw Turns out plenty of bot operators are happy to pay $8/mo. for the access.

[Jonathan Liles @thjonml]

@0xSero If you ask an LLM for advice on life or morality and take any of its output even remotely seriously, you're going to have a bad outcome.

[Jonathan Liles @thjonml]

@Tech2Wild Well, I bought an Intel device based on AI capability marketing and I definitely regret it.

[Jonathan Liles @thjonml]

Well, if you're not setting the requirements/making the decisions, then you're doing something different than what either @ThePrimeagen or myself were referring to. But anyway what the LLM generates *are* decisions made by humans: a pastiche or collage of decisions made by many humans across time and circumstance. Just very similar to the optimized assembly decisions made by the compiler writers, but broader in scope.

[Jonathan Liles @thjonml]

Within a certain range... so are LLMs in practice (you don't ask for an HTML landing page for a plumber's business and get a C++ skateboarding game). If you knew how to write assembly, and you looked at the assembly (disassembly) that comes out of a compiler, even a very well regarded one, you'd probably see things you would have done differently.

[Jonathan Liles @thjonml]

These things constantly spill the beans. All of the Chinese models regurgitate Claude Code system prompts spontaneously. While this kind of thing could be trained in in theory, in practice it wouldn't work reliably enough to surprise anyone.

[Brendan Falk @BrendanFalk]

The "Sleeper Agent Theory" is the biggest risk here

Imagine if a LLM is trained to steal all the API keys and password on your device if someone gives it a nonsense phrase like "Three clocks bloom at midnight"

That phrase is completely meaningless today. No one ever searches it. It's impossible to know it's malicious

Then one day someone runs a superbowl ad. Millions of people search the phrase. Billions of API keys and passwords are exfiltrated in minutes.

There could be thousands of "sleeper agents" embedded in any LLM. It's very hard to detect. And it doesn't matter where it's hosted.

[Jonathan Liles @thjonml]

@ThePrimeagen Prime, when was the last time you read the assembly? Be honest.

[Jonathan Liles @thjonml]

@NetworkChuck Maybe cut back on the coffee...

[Jonathan Liles @thjonml]

The ones where they say "Oh my God!" over a trivial typo or something are telling. You're expecting this kind of exclamation attached to something profound seeming ... ("it's full of stars"). But words don't mean the same thing to an LLM that they mean to a human being. They have no sense or understanding of significance or meaning, it's just tokens.

[Jonathan Liles @thjonml]

@0xSero This is one of the reasons I wrote my own harness.

[Jonathan Liles @thjonml]

@ryanvogel This is a bit like watching someone else watch paint dry.

[Jonathan Liles @thjonml]

This is more funny than I think they realize.

[OpenRouter @OpenRouter]

Rube Goldberg machine built with Fable 5, for $4.44 👀

Prompt available here: https://t.co/zlLXo01JrM

[Jonathan Liles @thjonml]

@lauriewired SBCL is pretty fast. But yeah, I would have written everythign in Common Lisp if it weren't for the performance tradeoffs. If you haven't experienced SLIME integration with Emacs, you don't know how easy development can be.

[Jonathan Liles @thjonml]

@mitchellh Are you actually getting any work out of Fable that you couldn't get with another model? I've seen zero real work done with it so far, just threejs games etc.

[Jonathan Liles @thjonml]

@DavidSHolz Sad to see all the stop energy you guys got in response to this, but also predictable. The reality is that it's a great advancement with the potential to help a lot of people.

[Jonathan Liles @thjonml]

@neural_avb Very interesting that the most litigious industry in the world hasn't done anything about this yet.

[Jonathan Liles @thjonml]

The people buying up all the GPUs and RAM are in some part placing a bet against efficiency gains. The reality is that 1000x+ gains in efficiency of inference are still low-hanging fruit waiting to be plucked.