Gary Marcus Slams Zuckerberg's Data Labeling Push For AGI

No, Mark Zuckerberg, you cannot get to AGI through data labeling alone.

Turning a once-top notch AI research division at $META into a data-labeling sweatshop has to be one of dumbest blunders in corporate history.

@GaryMarcus AI is existential for $META

Zuckerberg’s is panicking and throwing anything at the wall and hoping it sticks

This may end up as one of THE epic founder/CEO fails in modern corporate history…

There is one way out, but the $META of old is gone forever…

@GaryMarcus But Gary, the guy that they hired to lead the AI Research division is a famous labeller, seems to be the only thing that he knows.

@sampeppard @GaryMarcus He achieved nothing exceptional that’s why Mark wanted to change the leadership in AI division. He might have picked the wrong person but he’s definitely not happy with LeCun’s performance because Meta is way too far behind in AI race.

@sampeppard @GaryMarcus Just give me a shipped AI product from Meta while he was there for 10 years.

@GaryMarcus Dumber than the "Metaverse" ?

I'd say yes when you compare the money burnt.

@jonanon18 @GaryMarcus What he was working towards was far more ambitious than just data labeling, meta was in a much better place under him. He’s an actual scientist unlike Wang.Meta has since burned through so much cash trying to up compute and fallen behind.

@sampeppard @GaryMarcus Yeah but what did LeCun achieve while he’s at Meta?

@jonanon18 @GaryMarcus He laid the foundation for the Joint embedding predictive architecture and open source llama models. That’s at least something. Can’t think of anything the latter did.

@GaryMarcus We all knew Zuck has lost it when he got Wang for the superintelligence lab.

@GaryMarcus What about capturing users most intimate moments in their homes, sending all corresponding video and audio to another country for labeling, and then having contractors annotate that data. Gotta be worth something, right?

@GaryMarcus It’s crazy how people on Twitter (X) think they are smarter than the guys that run a trillion dollars company. All they do is speak negatively when share price is down without any clue about that underlying business fundamentals.

@GaryMarcus we need to get away from human language-based models entirely, for one thing

the models need to "think" in as close to the pure "human algorithm stack" as possible and this won't be in english

@GaryMarcus conversion into Scale AI 100% complete

@GaryMarcus ⌚️ "us"👀 ____ ⏳ 🧠 @neuralink 👓 @Meta 💄 @amazon 💻 cc: 👁️ @JeffBezos 💠 @elonmusk 👁️ bcc: 🐞 @kgriffasi 🍷 @finkd 🫎 (zuck)

@GaryMarcus Not surprised the labeller is selling his books.

@GaryMarcus Honestly, I really think, at least in the LLM space, and as much as I hate it, Anthropic's methods for Claude (constitutional, emotive woo woo) makes a better model. Don't think any LLM is going to be "AGI".

@GaryMarcus labeling your way to AGI is like building a spaceship by stacking bricks

META research used to run laps around everyone, now its just a big catch-up squad

@GaryMarcus "Elect a clown, expect a circus", Hire a data-labeler,...

@GaryMarcus It is quite instructive to see how various vendors approach AI. Google: bulletproof. Anthropic: focused and deep first, then expand sideways. OpenAI: depth and breadh but still more competent than what you give them credit for.

Then the wannabes: Meta, xAI.