🏅 Top traders

GPT-4, DALL-E 3. Any others?

@IsaacKing Below people are advocating for Midjourney.

@Joshua Wait that came out in 2022.

Why was this at 98%?

@Joshua Maybe Grok? Obviously it filled up twitter for a bit on release, but I don't see much of it since. And the Wikipedia page isn't very long...

I hear about Bard and Gemini on Manifold but not really anywhere else that I can recall. I never hear anything about Claude even on Manifold.

@Joshua midjourney upgrades models. It's an interface in a way, like chatGPT. The underlying model swaps out with each version.

V6 came out in 2023.

@IsaacKing suno.ai, everyone on twitter is going mad about it, karpathy tweeting etc

@IsaacKing I just saw this market, but I think Gemini is very clearly high-profile and was just unveiled this month despite longstanding rumors. Google has been thought of as the #1 competitor to OpenAI. You do say "experimenting with them," and only a weaker version of Gemini is publicly available at the moment, but I think that still counts.

(I think other systems should probably qualify too, particularly Claude 2.0, but Gemini seems like the slam dunk.)

GPT-4, Claude, Bing Chat?

@ShadowyZephyr Bing is just GPT-4 though.

@firstuserhere midjourney should quality, clearly being talked about more than ever since v5 came out

@firstuserhere Yeah, I've upgraded my opinion of Midjourney since my last comment. They're targeting it to professional artists to make concept art, and the outputs are extremely high quality. And even outside of professionals, you're going to see little Midjourney-based videos everywhere.

If I was e.g. making a video game, I would get the "look and feel" down with AI art and probably Midjourney before hiring an artist... It definitely works for concept art.

@Mira didn't know about their targeting prof artists, cuz i just didn't search for it ig. Mostly seen it used with hobbyists like myself. It's absolutely my fav img gen model out there, even if i pay for it. V high quality, and with v5, it's only gotten better than ever. I think v1 came out a year ago? Don't remember, but the pace of releases is incredible.

Bing

Bard

Gpt4

Llama

Alpaca

Midjourney v5

@Chad I'm not sure if Bing and GPT-4 can be separated from ChatGPT here, and same for Midjourney v5 from Midjourney as a whole. I think Bard counts (but I don't know if its Wikipedia page is "long" or if it's filling up Isaac's Twitter feed), but Llama and Alpaca don't seem as high-profile as the examples in the market description.

@JacyAnthis autoGPT? That's definitely being talked about in the public a lot, even when you compare to things like "Bard" (source: google trends)

@firstuserhere (despite that, I took a quick loss because at 94% it seemed a bit overpriced)

@firstuserhere So, sold to 85

I'll certainly count Bing. I think GPT-4 should probably be folded into Bing.

Midjourney... maybe. I did see enough discussion about it that I went and joined their Discord server to test it out, but it didn't seem to inspire nearly as much chatter as other launches.

AutoGPT has gotten a lot of press, but does that really count as a "new system"?

Llama and Alpaca I've only heard of from people trying to make this market resolve YES, so they definitely aren't high profile enough.

Bard I've only heard about in the context of "wow it sucks", and even that seems to have died off pretty quickly.

@IsaacKing LLaMA is absolutely being experimented with. Many graduate students, etc. are working on collecting data(such as from GPT-4) to finetune it, optimizing it, distilling it, training it further, trying to replicate its dataset with public domain copyright as best they can, etc. It's got Facebook behind it, the "leak" got reported on, I see people turning it into Discord bots and running it locally and porting it to weird architectures.

Every llama derivative should be folded into the same thing, just like Bing/GPT-4. But should be counted.

Bard.... is getting the wrong kind of attention. I would say hold your market on that one

Claude doesn't count. It's on a waitlist, I don't have access, they seem to be targeting enterprise and not THAT successful.

Midjourney looked okay. Maybe it'll be more high-profile later this year, when someone makes a movie with it or something.

I bought M10 on YES and would score it 2/3 currently from options I've seen other people say in this market, with the option to 3/3 when Google's Magi system is released, maybe Elon's TruthGPT or whatever, the Stability.AI language model they're working on, the UK's BritGPT. Lots of options for a 3rd one.

Segment Anything!

Don't know if you've tried the demo yet, @IsaacKing but you can try here.

Not to mention HuggingHPT and AutoGPT

We already have it, right? GPT-4, Claude and LLaMA. Also Alpaca. And maybe Gen-[1/2].

And also perhaps Midjourney 5.

@qumeric definitely midjourney v5. It's so good

@qumeric I can't recall having seen any discussion of Claude,, Alpaca, or GEN-[1/2], and while I've seen the name "LLaMA", I don't know what it is.

@IsaacKing okay so LLaMA is a set of models with different sizes. LLaMA <7B, 33B, 65B> are three main sizes for the same model.

Important thing here is that these were trained using only publically available data. The biggest 2 models on 1.4 trillion tokens, the 7B one on 1 trillion tokens.

The reason for the massive popularity they achieved is that most companies will keep such a massive R&D project to themselves and only provide access via API (like OpenAI). But a Meta user(?) Accidentally posted the key to accessing all the LLaMA weights while submitting a pull request on GitHub. People used this key to access and download the LLaMa weights for each of the three models, and this gave them access to proprietary research done using only publically trained data to get LLM as good or better than GPT-3

@firstuserhere since the weights are out in the open for anyone to download and customize, we've got personalized applications that are achieved by tuning the weights themselves for a specific task(transfer learning), rather than adding the task on top of the fixed weights (fine-tuning).

This means people can create apps as good as >GPT-3 but the model is specifically good at the task they further train it at, instead of being a general langauge model.

That's where stuff like "the LLaMA is on the loose!" Comes from

Everyone is rushing to customize LLaMA weights for their specific application.

People have been optimizing the code for Llama models and now we've got them running on smartphones, laptops, with incredibly fast response times. Running locally! Entire models. Not api, but entire models fitting inside a small device, and being as good as GPT-3 for 7B and better for the other 2.