Note "(unlike Llama 2 license)" means there aren't restrictions in who can use the model. See "On Question Definitions and New Questions" section below.
Spirit of this market is seeing if any "big tech" company will release their LLM foundation model in a similar way that Meta ultimately released their Llama 2 model. This market will focus on "big tech" but that definition is pretty fuzzy. At creation of this market, I'm starting with the most discussed companies developing an LLM that is comparable to something like GPT-3 or GPT-4, specifically OpenAI (GPT-n), Google (Gemini), Meta, Apple, and ByteDance. More than one question can be resolved.
Note that I won't be voting on any of the questions to remain relatively impartial to any question resolutions.
Note on "Open Source" Definition
Something to note is that I use "open source" in quotes because there's a fair argument that although Llama 2 is often referred to as "open source" it shouldn't be such as limiting it's license so companies with more than 700 million monthly users have to ask for special permission (see articles like this from IEEE Spectrum).
On Question Definitions and New Questions
I'll try my best to keep questions reasonably defined to include at least a couple degrees of "open source". There are plenty of potential edge cases but to keep this market getting too complicated, I'll interpret "open source" in two ways:
weights & training code available with some restrictions (similar to the current Llama 2 license)
weights & training code available without restriction to who can use it (unlike the current Llama 2 license)
The idea would be that the "open source" model has available weights and is able to be trained/fine-tuned. If a model falls short of those conditions, I won't define it as "open source" but open to redefining if it makes sense for this market.
I'm open to adding more questions (companies or categories) but I'll be pretty conservative on adding new questions. I'll also keep it to established "big tech" but who knows what can change over the year.
If there ends up being a significant vagueness that makes it difficult to resolve a question as "YES" or "NO", I'll default to "N/A" but will try to resolve as close to the spirit of the market at creation.
Related questions
Google just released a family of open source models this morning. Which other big tech companies will follow suit?
Now that Google released Gemma, their family of open models 'inspired' by their Gemini architecture, I wonder if and how any other companies will respond!
Is 2024 the year of 'open source' models? We're just a little over 50 days into 2024 so a lot can happen over the next 300 days!
Google released open-source Gemma
Terms of use permit responsible commercial usage and distribution for all organizations, regardless of size
@shankypanky Thanks for sharing the link!
On a personal note, I think this is pretty exciting to see Google open this model up for everyone! But of course, I'm a bit biased here ; )
It'll be interesting if this pushes other companies to consider opening up their models
@VictorsOtherVector Surprising early morning news here but that's what I've come to expect in this era tbh 😅 Definitely feels like a big moment and I'm curious to see how the others on the list here move forward. And what comes as people start using the Gemma model.
Nice market by the way!
@shankypanky totally! And thank you!
I've been super excited about this release to see how other companies respond! Even Meta might be considering opening their models more depending how Google's release is received. I honestly don't think anyone really saw this coming (the market seems to support that with relatively few trades)
To be up front, I heard early on that Gemma was leaning hard into being open. I knew that there was going to be an early announcement (~5am my time) so figured some lucky soul would get a jump on me. But also figured it wouldn't be a huge flood of traders since the market isn't particularly active or on people's minds
@VictorsOtherVector I'm on the West Coast too and admittedly I'm curious but it's not necessarily my wheelhouse so I was hoping I chose the right Google option here. It gave me an excuse to spend a little time exploring the current open source offerings.
Funny that my Twitter TL was full of people making snarky comments about Gemini image creation yesterday and now this. Let's see what the random masses have to say today lol
@shankypanky yeah you picked the right one! I know having essentially two options per model can be confusing, but I thought it was important to make that distinction since the org's size restriction to Llama2 is a big critique.
Oh yeah, it was fun to see what people end up talking about Gemini. But it adds to the fun when you also know there are more announcements to come :D But in my opinion, I'd be a bit surprised if the 'mass of people' talk about this. In my opinion, Gemma is meant for the people deeper into LLMs who care there is an open model to work with.
I think most people only are aware of a product (like ChatGPT or Gemini/Bard) or something more 'shareable'. Like just last week, I think the Gemini Pro 1.5 with the long context was arguably more ground breaking & in the long-term more important than Sora. (In the AI research community, I believe most would agree that AI video would get to near 'realistic' depictions eventually.) But Sora 'make pretty pictures' and dominated the general conversation.
And to be fair, I believe Sora is quite an achievement but I think the 1M-10M context with the claimed capabilities is more unexpected by those working on AI/ML in the field. However, I might be showing my bias, at least partially, in working for Google lol
@VictorsOtherVector I can absolutely agree on these points - I didn't mean the average person like my mom or my neighbour (tbf my neighbours are probably in tech though lol) but even my Twitter TL masses in tech/adjacent roles. That said, when people spend their whole day shitposting about not being able to get Gemini to create images they don't actually want or need it's less about legitimate interest and more about trends. And lack of understanding never stopped a swarm of people from having an opinion to tweet 😅
Your specificity on the answers makes total sense. Without being particular there are too many grey areas and debates on resolutions. Even with your knowledge of one roadmap there are still outliers here for sure.
@shankypanky that makes sense— less 'average person' & more the general conversation on Twitter
But even then I would still be surprised if a lot is talked about in regards to Gemma. I think at least in part the conversation naturally tends towards 'meme-able' content. So it'll be interesting to see how Gemma gets talked about, if much at all by this 'group'
(Of course, I'm seeing a certain filtered lens of the conversation of these things. I tend to follow people who are working in the space & usually decently knowledgeable/aware of the field. So I tend to see the conversation about how these AI systems are doing compared to benchmarks, flubs, etc.)
----------
And there's totally blind spots! This definitely happened to me when running the Chatbot Arena market about Gemini's performance. This field is moving so quickly I wouldn't be surprised if something unexpected comes up!
Reels announcement from Zuckerberg: https://www.instagram.com/reel/C2QARHJR1sZ
tl;dw
Zuckerberg confirms Llama 3 confirmed being trained now and will be "open source"