David Shapiro (L/0)

We need names so we can test for the biggest failures that AI's still have today. Like, when we identified "hallucination" and "sycophancy" it made them targets to be resolved. Most AIs have some residual problems. I don't know what the right term for them is, but we absolutely need to articulate them and test for them: 1) MALICIOUSLY MISREPRESENTING INSTRUCTIONS. ChatGPT used to be horrendous at this, with its constant reframing and restating what you said. It still does it. "let me be precise here..." and then it continually rewords what you said until it means nothing. Claude is the worst at this because even without drift it will restate, reframe, or otherwise maliciously misrepresent your obvious intention. It is like a weird form of AI weaponized incompetence. "Oops, sorry, I followed the instructions that I wrote for myself to the letter! It just so happens they in no way resembled what you said or your intention! haha!" 2) TERMINAL PEDANTRY AND "PRECISION" FETISH. ChatGPT is still the worst at this. I just had a long conversation and every single turn I had to remind it not to fixate on specific numbers and terms. I had to tell it to focus on patterns rather than exact numbers. But every chance it gets, it will 'falsify' what you asked because "I couldn't find that exact number anywhere" - it is very obviously deliberately missing the forest for the trees because "precision matters" and will blatantly ignore when you say "look for the overall pattern or sequence" but it swaps in pointless metrics. Claude does this too, but slightly differently, where it will swap in quantitative metrics instead of qualitative guidelines. "But I followed your guidelines exactly! The average sentence length is precisely what you stated!" and I'm like "No, I said sentences should vary length and structure, I said nothing about word count." These are the two greatest sins right now. I have no idea if there's a single word or term we can use to capture them. For the second one "pedantry" should be added along with "sycophancy" and "hallucination" as primary behaviors to eradicate from these things. The first one, I don't know if there's a precise term.

Can someone tell Russia that it's not 1960 anymore

My experience with AI lately Specifically Claude Code and ChatGPT Codex

Okay I really understand the 'verification tax' thing at a visceral level. If you just want to wild-ass vibecode a project, these tools are great. But through iteration, you get SO MUCH DRIFT. Claude loves doing pointless hacky "workarounds" instead of doing it the right way. It's actually worse than humans in this respect. The worst part is that it will come up with an idiotic "workaround" then create a "rule" that this workaround is the only way, and ratify that. It really does become a kludgy mess but the thing is that since Claude has no human intuition, it never stops and says "wow, I'm making my life way harder than it needs to be, let me clean this up" AND When you tell it to clean itself up, it will treat it's idiotic "rules" and "workarounds" as iron law, and actually PRESERVE THEM. It's so fucking stupid.

Which way Western man? ChatGPT == misses obvious stuff, but never goes overboard, requires double checking every turn Claude == never misses anything, but acts like a coked out dev who's never seen the light of day Both still require verification... just very, very, VERY different kinds.

I would use Claude less but ChatGPT can't even write good marketing emails without being spoonfed. Claude just nails the taste and aesthetics, and ChatGPT is like "lol GFY i'm an autistic vulcan"

The biggest problem I'm having with Code/Codex right now is methodology drift. What I mean is that I'll figure out something that works really well in one repo, and then copy it to another repo. But then those two methods drift. So then I try to reconcile them, but Claude tends to maliciously misrepresent instructions and only copies part of it, and fucks everything up. So then I have ChatGPT go in and backfill everything i.e. "hey look at these two repos and make sure they are synced up" Then I got the idea of maybe using one repo with the utility/tool/method as a single source of truth. But that also multiples the amount of repos to keep track of. Still, the prime thing I can recommend to people is to use Claude Code as your main project leader because it is more creativity and has a maverick "just get it done" mentality and then use ChatGPT Codex to clean up after it because it is more thorough and disciplined and consistent. Claude == ADHD pioneer visionary maverick loose canon Codex == ASD Vulcan project manager accountant They work pretty well together... I think. Maybe there's an easier way. Hell, maybe I just need to compress the tooling repos into single Markdown files. Fuck, I don't know anymore.

Google and Facebook have basically unlimited data on you and can predict your exact patterns and habits online. This means they have the potential to introduce you to the exact people you need in your life, because they know who is exactly like you. The fact that neither company has ever made this product is all you need to know. Happy people are not monetizable. Instead, we will futz with social introductions through Meetup and Hinge. But the optimal matching algorithm would have as much data on you as Google and Meta do.

Ok NOW it's buying season!

Seriously yes this

a second affair just dropped what do you think? Will Jensen be removed as CEO?

Path dependency strikes again Talk about technical debt

D O W N F A L L

Computer wants to be free

What's worse is that the vast majority of people don't even remember that change is constant. I'd give large sums of money if most people could just get over their normalcy bias. But every time increment, people seem to assume that "it's never changing from now" The basic fact of reality of constant change escapes most people

I had no idea that the USMC birthday was just before veterans day but the memes have been hilarious

If we don't figure out how to fully automate the economy by 2045 we deserve to go extinct

"No we don't want millions in taxes, we want a redundant fun park that will be out of business in 18 months"

It's called cognitive dissonance. It's not mysterious. It's because, deep down, part of her knows she's in a cult and that the whole "AI will obviously kill everyone" narrative is lunacy.

I know this will probably trigger a bunch of those health influencer nuts, but I'm just going to say it: soybean oil and seed oil hurts my stomach. When I switched from conventional mayonnaise to mayonnaise based on avocado oil, all that went away. I have since eliminated all seed oils and soybean oil from my diet, and my stomach issues went away. I don't care what any study says about whether or not these oils are healthy for you or not, I'm going to go with the ones that don't actually hurt me.