machine learning

AMA with Jeffrey Ladish. Wednesday 2/4 at 2:00 PM CT

Executive director of Palisade Research; studying AI loss of control risks.

AMA with Nate Soares. Wednesday 2/4 at 10am CT

Author of If Anyone Builds It, Everyone Dies answers questions about why superhuman AI would kill us all.

Please help me stay intellectually honest! I'm not a fan of generative AI in general, and LLM technology specifically. I think its capabilities are being drastically over-hyped. It's a perfect, sweaty example of a solution looking for a problem. I'm skeptical of many claims people are making wrt how it's helping them.

My experience is it's like having access to an idiot-savant intern. Awful at most tasks, but knows everything and can read incredibly quickly.

Publicly, I've taken on the mantle of a staunch critic of generative AI and a pro-human, pro-soul advocate.

And for the most part, I'm happy with that stance. I like it. It feels good to rail against something, and it feels good to contrast a thing that I hate against something I love. It throws the love into more relief.

Yet, I don't want to lose any babies in that bathwater, and I don't want to lose my intellectual honesty in the neurochemical rush of fighting for a cause. So I'd love to explore the best use cases of LLMs that you all are actually using, and actually finding beneficial, life improving, productivity increasing, all of that.

I'd love to hear your experience, and ideally, you'd to tell me how you're doing what you're doing with it in enough detail so that I can try it.

I'll start.

Absolutely most useful thing I've found for it so far, and it's not even close, is language learning.

I'm in a slow process of learning Japanese, and asking a chatbot to break down the grammar of a specific sentence is super useful. It's also great for generating content for flashcards. Say you have a set of characters, and you want some example words that use each particular character. It's so easy to generate stuff like that.

Outside of that, I use it in super basic ways (basically as google with one less step).

So please, give me your best use cases, things that you've not only been impressed by, in a "oh wow, that monkey can tap dance!" way, but that has actually improved the quality of your life.

How to make ai less sycophantic? I went into the settings and set ChatGPT to always give me the hard truth and be real and not tell me what I want to hear. But now it’s just like “I’m not gonna lie to you. Here’s the hard truth. You’re totally right” 😂

Has anyone had any luck getting ai to be more objective and less of a people pleaser?

Quick reflections on vibe-coding:

• kicks up weird addiction cycles in me
• the end product feels more like 20-30% jordan's creation and 70-80% LLM's. The text for example is all a little off, the style, etc, and it drifts because it's so easy to have it do it and so much more "time consuming" to go fix all the little bits (plus you have no guarantee it wont fuck something else up) and keep them coherent, versus just surrendering to the llm
• As a result I feel weirdly not that autonomous... like is the LLM working for me or am I working for it? This feels similar to the addiction cycle too
• There's a vibe that's similar to the "HR Lady" tone and em-dash hell of the LLM writing; like everything is sanitized and optimized for general professional sleeze marketing and you can't quite put your finger on why it sucks. The glassmorphism designs and gradients feel similar
• I continue to feel like I'm learning a lot from it, but I have updated to think that at the current stage its at, its actually an unhealthy practice and so is a little like a blood sacrifice that should be used sparingly. I imagine this will change, but maybe it'll get worse.

Quick reflections on vibe-coding:

• kicks up weird addiction cycles in me
• the end product feels more like 20-30% jordan's creation and 70-80% LLM's. The text for example is all a little off, the style, etc, and it drifts because it's so easy to have it do it and so much more "time consuming" to go fix all the little bits (plus you have no guarantee it wont fuck something else up) and keep them coherent, versus just surrendering to the llm
• As a result I feel weirdly not that autonomous... like is the LLM working for me or am I working for it? This feels similar to the addiction cycle too
• There's a vibe that's similar to the "HR Lady" tone and em-dash hell of the LLM writing; like everything is sanitized and optimized for general professional sleeze marketing and you can't quite put your finger on why it sucks. The glassmorphism designs and gradients feel similar
• I continue to feel like I'm learning a lot from it, but I have updated to think that at the current stage its at, its actually an unhealthy practice and so is a little like a blood sacrifice that should be used sparingly. I imagine this will change, but maybe it'll get worse.

Telling people that I upvoted/downvoted their comment when I do and telling them why. (Triggered)(Feel super activated about the hiddenness and anonymity of the downvoting)

I want to see what happens when it’s ok to be explicit about affecting someone’s status.

(Fear that I’ll be avoided and ultimately alone. Anger at everything.)

(Longing to be the next me that’s social-cue numb to my own abrasiveness).

Not sure if this is triggered or not but I want the consequences of being known.