Chaos

🧠 + 🖥️ = 👁️ | Tweeting here when my ADHD meds have stopped working | Officially descended into madness

[Chaos @FoldMani]

I would like to kindly ask anyone or any journal that presents the results BEFORE the methods to just

[Chaos @FoldMani]

@ChombaBupe Just do it the Hinton way: redefine consciousness so that it includes current AI in it, QED.

[Chaos @FoldMani]

@vineettiruvadi I've always engineered my decisions around the idea that my "functional time" or "cognitive battery" is my most valuable resource. And I think this is becoming more and more true as technology allows us to embed more complex "cognition" into * somewhat * reliable systems.

[Chaos @FoldMani]

@vineettiruvadi But it's a helluva good model to shovel money in the mouth of big publisha! Nature eating good.

[Chaos @FoldMani]

@ChrisWLynn I don't think your results give you the support to make that generalization jump, regardless of any other criticism about it. And if I was to guess, any phrasing that hints at this generalization might be a motivator for the "lower degree of accuracy" criticism that you received.

[Chaos @FoldMani]

@KordingLab It's in Nature, what did you expect? They're more like pop science, regardless of what academia's toxic relationship with it seems to suggest. Just let her go man, she's not good for you.

[Bernardo Sabatini @blsabatini]

1/n Are biological neurons linear-nonlinear computers, like perceptrons, or is their output governed by non-linear interactions between inputs? If the activity of a neuron is well fit by linear models that sum inputs, does that mean that the neural computation actually is linear?

[Chaos @FoldMani]

@QuantumTumbler Rule of thumb when learning on your own: If you haven't fully validated the math (if theoretical) or went through things manually/visually/interactively as closely as feasible (if empirical) before moving forward, then you're playing assumption roulette, and the house always wins

[Chaos @FoldMani]

@vineettiruvadi And with the incoming AI slopfest that will arrive when SWE tools reach academia more directly, I think this is going to get more and more true, and even at that point it'll be the stronger statement with "[...] that you haven't QC yourself" at the end.

[Chaos @FoldMani]

@Steve_Yegge It seems like you're delegating the one thing that you really should be keeping yourself responsible for: defining the trajectory through decision making. If you want thought dynamics ask & give rules for it. Otherwise you're just asking for something they don't have: an opinion.

[Chaos @FoldMani]

It's 4am and I've already worked more than I've slept. Insomnia is a helluvadrug. 0/10 highly don't recommend.

[Chaos @FoldMani]

Everything is gradient-free optimization (unless it's explicitly gradient-based optimization).

[Chaos @FoldMani]

@Zakrea_312 @vikhyatk Lmk when we find the free lunch

[Chaos @FoldMani]

@sourabhbgp "Can it?" is 100% still a relevant question, it decides on the boundary between "that's something they will be instructed to do" and "that's something the system design will ensure". That's what distinguishes this approach from "just throw an LLM at it".

[Chaos @FoldMani]

@vineettiruvadi

[Chaos @FoldMani]

@henrytdowling @vineettiruvadi Now the question isn't "how can LLMs do research?", it's "which parts of my research workload can I sequester and pre-think through sufficiently thoroughly to built an LLM-driven system around them?". It's basically just 1 hierarchy up from plain scripting really, but stochastic.

[Chaos @FoldMani]

@henrytdowling @vineettiruvadi I think a "cognitive enhancement" approach with carefully engineered guardrails and support structures optimized for epistemic transparency and end-to-end traceability is the way to go: Identify a pattern -> build it into LLM process -> trigger it when needed -> traceback at fail

[Chaos @FoldMani]

@vineettiruvadi LLMs are kinda ass at most things that aren't gated by direct verification with actionable and informative error signals. I literally wrapped my code "intents" around dedicated tests that fail if the code stops following the "intents" to protect its "essence" from LLM slop edits.

[Chaos @FoldMani]

@vineettiruvadi with or without head voice?

[Chaos @FoldMani]

@vineettiruvadi I'm curious, what's the usecase for that? Also, you can probably extract the transcript get the conversation text and output it to a file with a script (if using CC anyway), then have that version-controlled, and just use multiple sessions (file is VC so sessions can just be new)

[Chaos @FoldMani]

@BlancheMinerva I hate the idea that citations are just "oh I need to defer authority for that thing that I just have come to believe is true but didn't bother to keep mental note of my epistemic trace for its origin in my belief systems", like wtf, how do you even define knowledge for yourself?