← Back to context

Comment by postalcoder

5 months ago

An AI engagement farmer on twitter claimed to create a llama 3.1 fine tine, trained on "reflection" (ie internal thinking) prompting that outperformed the likes of Llama 405B and even the closed source models on benchmarks.

The guy says that the model is so good because it was tuned on data generated by Glaive AI. He tells everyone he uses Glaive AI and that everyone else should use it too.

Releases the model on HF, is an absolute poopstorm. People cannot recreate the stated benchmarks, the guy who released the model literally said "they uploaded it wrong". Pretty much turns to dog-ate-my-homework type excuses that don't make sense either. Turns out people find it's just llama 3.0 with some lora applied.

Then some others do some digging to find out that Glaive AI is a company that Matt Schumer invested in, which he did not disclose on Twitter.

He does a holding pattern on Twitter, saying something to the effect of "the weight got scrambled!" and says that they're going to give access to a hosted endpoint and then figure out the weight issue later.

People try out this hosted model and find out it's actually just proxying requests through to anthropic's sonnet 3.5 api, with some filtering for words like "Claude".

After he was found out, they switch the proxy over to gpt 4o.

The endgame of this guy was probably 1. to promote his company and 2. to raise funding for another company. Both failed spectacularly, this guy is a scammer to the nth degree.

Edit: uncensored "Glaive AI".

This is accurate, but you don't need to censor GlaiveAI. They helped create the model. They're complicit in the scam.

  • I took out Glaive so as not to give them free publicity – all I did was mess up the formatting of my comment.

    And yes, you're correct. Glaive employee(s) contributed to the model uploaded on HF.