Breakfast At Ralf's

i hate seeing people drink the openai/chatgpt koolaid 😭😭😭 genuinely feels like watching someone get seduced by scientology or qanon or something. like girl help it's not intelligent it's Big Autocomplete it's crunching numbers it's not understanding things i fuckign promise you. like ohhh my god the marketing hype fuckign GOT you

An excellent question, @realcleverscience!

The answer is that traditional AI is already useful, in that there exist successful applications of AI doing important work RIGHT NOW like rapidly diagnosing cancer, sequencing DNA, folding proteins, and docking/landing spacecraft.

The difference is that those applications are expert systems. They're very good at specific tasks, but suck at everything else. You're not going to ask a cancer-identifying AI to do your taxes, for example.

The generative LLM AI we're bitching about today will never become reliable, free from hallucination & lies. AI companies keep pushing larger LLM data sets, but that just seems to make them worse rather than better. The current hype around AI is a grift, a way for investors to make a pile of money before the bubble pops.

Remember 3D television? About 15 years ago everyone was absolutely convinced 3D was the future. Until suddenly it wasn't.

The current AI situation is kinda like that.

No, but this is actually crazy in the most dystopian sense.

Hi, we'd like to steal your voice and clone it for resale to other companies to say whatever they want. Just imagine, your own voice selling erectile dysfunction drugs and possibly promoting hateful political candidates.

No, you won't get any money for this!

And by the way, the TOS you agreed to in 2017 and that we modified in 2023 (once this technology became available) allows us to do this anyway without your knowledge or consent. Retroactively! Using your existing catalog of youtube content!

Have a great day!

Oh for fucks sake

idk if elon will kill his new chatbot grok when he finds out its woke, but he'll definitely kill it once he discovers it's funnier than him

#i dont think grok is an ai#i think grok is a few dozen people elon has in a shed somewhere pretending to be an ai

I've been saying this for a while. Musk cares way more about being seen as a genius than being liked. Any given halfway-coherent ai has a 50/50 chance of being exposed as a mechanical turk these days and they're just the ones that have been found out so far; Grok seems far too cohesive to be artificial compared to other current ais. It's entirely possible that Grok is actually not great and there's selection bias at play here (I'm not on twitter so I only see the grok stuff that gets picked up and reposted here, if it's generally boring and incoherent then I wouldn't know) but everything I've seen suggests that someone's turkin' it.

As a former mturk grunt, I see our fingerprints all over Grok.

Likely it's a combination.

Musk's xAI is operating a huge server farm in Memphis, Tennessee and that wouldn't be happening if Grok was 100% mechanical turk.

What's probably happening is that easily handled responses are generated via LLM, while certain phrases trigger an override and get shunted into a queue to be serviced by humans.

Yes that's what I mean. It would be unfeasible for any very large llm that's expected to provide instant responses to be 100% individual human responses. I'm given to understand that the usual way they mturk an llm is to manually write or edit canned responses to popular trigger terms, give high profile cases to human operators to either write fully or check and edit before release, or shunt stuff over to a human operators if the llm fucks up spectacularly and pretend it's still the llm. Background fuzz can be ignored and people pretending to be a robot can make the whole thing sound insightful if you're clever with your intervention. But I don't work in the industry and don't know the details, a turker can probably explain if I'm wrong.

That makes sense.

Unless I am mistaken, Grok responds on twitter as if it's a user replying to another user. Which means there is a built-in delay where the user isn't wondering why Grok isn't responding instantly. It's the perfect setup for a turk solution.

idk if elon will kill his new chatbot grok when he finds out its woke, but he'll definitely kill it once he discovers it's funnier than him

#i dont think grok is an ai#i think grok is a few dozen people elon has in a shed somewhere pretending to be an ai

I've been saying this for a while. Musk cares way more about being seen as a genius than being liked. Any given halfway-coherent ai has a 50/50 chance of being exposed as a mechanical turk these days and they're just the ones that have been found out so far; Grok seems far too cohesive to be artificial compared to other current ais. It's entirely possible that Grok is actually not great and there's selection bias at play here (I'm not on twitter so I only see the grok stuff that gets picked up and reposted here, if it's generally boring and incoherent then I wouldn't know) but everything I've seen suggests that someone's turkin' it.

As a former mturk grunt, I see our fingerprints all over Grok.

Likely it's a combination.

Musk's xAI is operating a huge server farm in Memphis, Tennessee and that wouldn't be happening if Grok was 100% mechanical turk.

What's probably happening is that easily handled responses are generated via LLM, while certain phrases trigger an override and get shunted into a queue to be serviced by humans.

I’m publishing a book that makes fun of AI, but is disguised as a self help book for business people who want to use AI. I commissioned my friend Isabell Struble to illustrate a cover that looks like AI art, and she fucking nailed it!! Look how good this is!

"what did students do before chatgpt?" well one time i forgot i had a history essay due at my 10am class the morning of so over the course of my 30 minute bus ride to school i awkwardly used by backpack as a desk, sped wrote the essay, and got an A on it.

six months later i re-read the essay prior to the final exam, went 'ohhhh yeah i remember this', got a question on that topic, and aced it.

point being that actually doing the work is how you learn the material and internalize it. ChatGPT can give you a short cut but it won't build you the the muscles.

I think of an art teacher I had once who said he used to be terrible at drawing hands until he bought a book with 100 pictures of hands shown from various angles and drew each hand in the book 10 times. Magically, after drawing 1,000 pictures of hands, he wasn't bad at drawing hands anymore.

The reason OP could bullshit a decent essay in a 30 minute bus ride is because OP knew how to write and craft a well-reasoned analysis thanks to a childhood full of writing and critical thinking practice.

Feel like I'm transforming into one of those reddit assholes, but this genuinely is just a "skill issue" and you really do just have to "git gud."

"No time to refill"

Good news! I just checked, and as of June 2025 the Titanic swimming pool is still full.

Thank God for watertight compartments.

https://www.cbsnews.com/news/sextortion-generative-ai-scam-elijah-heacock-take-it-down-act/?utm_source=firefox-newtab-en-us

Fucking awful. Imagine being 16 and getting threatened with your AI generated nudes sent to all the stolen contacts in your phone.