andrew davis ✨

Check out my first ever short!! How'd you like it? How can I improve? Does it need more BOOM or memes? youtube.com/shorts/rEUEXVq…

Everybody is talking about recursive self-improvement (RSI) and meta learning. Here is my old 2020 talk about this [1]. It has aged well. Example: humans still define the starts & ends of trials of many modern meta learners. My RSI systems since 1994 LEARN to (re)define them [2]! [1] Meta Learning Machines in a Single Lifelong Trial (talk for workshops at ICML 2020 and NeurIPS 2021, based on earlier talks since 1994). Abstract: the most widely used machine learning algorithms were designed by humans and thus are hindered by our cognitive biases and limitations. Can we also construct meta learning algorithms that can learn better learning algorithms so that our self-improving AIs have no limits other than those inherited from computability and physics? This question has been a main driver of my research since I wrote a thesis on it in 1987 [2]. Here I summarize our work on meta reinforcement learning with self-modifying policies in a single lifelong trial (since 1994), and mathematically optimal meta-learning through the self-referential Gödel Machine (since 2003). Many additional publications on meta-learning since 1987 can be found in the RSI overview [2]. [2] J. Schmidhuber (AI Blog, 2020-2025). 1/3 century anniversary of first publication on recursive self-improvement (RSI) and meta learning machines that learn to learn (1987). For its cover I drew a robot that bootstraps itself. 1992-: gradient descent-based neural meta learning. 1994-: meta reinforcement learning with self-modifying policies. 1997: meta RL plus artificial curiosity and intrinsic motivation. 2002-: asymptotically optimal meta learning for curriculum learning. 2003-: mathematically optimal Gödel Machine. 2020-: new stuff!

The Curse of Grafting follows unfiltered contact with Deep magic, or reckless, repeated transformation without knowledge... it rewrites flesh as if correcting a mistake. Only the powdered bone of a Divine can halt the body’s failed attempt to become something else.

Vibe coders be like

we were already on track for "the end of culture", with everything being a sequel, remix, or reference... nearly nothing new. AI just cements that.

AI isn't intelligence it's Xerox. It's a copy machine and suffers from the same "lossy" reproduction. Still an incredible and world changing machine, but just a machine.

In this notebook, @LaoTzunami visualizes a state space of possible chart types and transitions among them along defined edges. Transitions between unconnected types are achieved by routing through intermediate types. observablehq.com/d/3c8b082818ef…

@JoelAbenhaim Yeah, AI code gen output. Lots of duplication. Often overly complex, overly verbose, and sometimes wrong logic in very subtle "looks right, but not quite!" ways. Hard to maintain. Hard to read. Often "cheats" during validation steps.

@andrewdavisdev Is this how you feel about the output of AI coding agents? Duplication, premature optimizations, non respect of modularity, narrow focus fixes breaking app invariants, false positive unit testing, etc.

I am not as smart as many of the engineers I follow who are all in on LLMs, but I do have good instincts. And my gut keeps telling me there is something not quite right about the output of these tools.

@WaterMuseum_ i love this game.

Stress testing the inventory

Michael I. Jordan on the new MLST. Four things: > AGI is a PR term. It confuses young people. > Discourse is bipolar, either alarmist or exuberant, this is in his words "so demoralizing" for 20- and 25-year-old researchers. > ML's methods came from statistics and operations research, NOT the AI tradition. > Data markets are Stackelberg games, not optimisation problems. A lot of ML researchers have never computed an equilibrium. Michael I. Jordan is a no-nonsense original gangster of the field and was described by Science magazine, back in 2016 as the most influential living computer scientist.

@gregthompsonjr An LLM as I understand it is a huge collection weights and relationships, yes, there are an infinite number of things encoded in there, but they are completely inert until activated by a prompt. The prompt is all it is. It's next token prediction. But it must have an input.

@andrewdavisdev You’re ignoring the non-deterministic aspects of what’s happening despite the system prompts, which are very significant aspects of getting you an output.

Saw a description of AI: it’s an instruct-able copy machine.

@gregthompsonjr Look at all the claude code code & system prompt. The model does nothing by itself. The model is very very good autocomplete.

@andrewdavisdev It’s “self-directed” in that you’re not telling it exactly how to go about solving a problem. Instead, you’re leaving it to its own reasoning devices to determine what needs to be solved and how, which includes identifying sub-problems.

@gregthompsonjr It's using the prompts and system prompts to do that. It's literally never self directed.

@gregthompsonjr how by any defintion is it "self-directed". given no prompt, it does nothing.

@andrewdavisdev Wouldn’t it be a “self-directed instruct-able copy machine”, with emphasis on “self-directed”?

still not quite over the fact that i watched 15 year olds get sued for millions of dollars for downloading twelve songs and now we all have to accept AI slop because every tech company in the known universe decided that IP laws don't exist now that they're inconvenient for them

The crushing weight of a handful of completely manageable tasks.

Whewwww weee. What a week.

i absolutely can not stomach advertising. its just so bad.