A message from the founder, Kyle Morris

In my 13 year career in AI—building companies, publishing research, filing patents — I learned to win you need a lot of talented people, capital, compute and the latest research. Today all of these things seem much harder to get, except for a few elite teams.

Here are some examples

• Compute: There are ~20 million GPUs in global data centers — less than 1% we project we need—and most are controlled by few big players. If you need cheap compute you can get free cloud credits, but will find your price marked up 2-5X once locked into a cloud. Your alternative is to spend several weeks migrating software between clouds that name the same concepts different things making it harder to switch. If you want to change the GPU type it can get worse - whole programming abstractions are maintained with an incentive to keep you on one hardware provider.

• Funding: The Stargate announcement has $500 billion committed to one AI company —more than Apollo and the Manhattan Project combined. Fundraising gets a lot harder when you have 100k but investors know teams with 100B and you’re both building the next AI powered social network or search engine, or web browser, and so on. You’ll need to argue that instead you have some combination of a better team, a research breakthrough, and speed.

• Talent: The top AI talent receive millions in compensation and primarily orbit around a few elite AI labs. Given the above two challenges it is even harder to get these people to join you unless you’re already a top level executive leaving a top company.

• Research: Arguably the biggest AI labs of today wouldn’t exist if the Transformer paper wasn’t published openly, yet now modern AI labs publish a small fraction of details they did before - so even if you solve the above 3 problems above you’re likely behind the big AI labs by at least a few months and now you have to out-research them, likely repeating a lot of duplicate experiments first. If you are one of the few teams that can do novel research and publish in the open, there may be fear of losing your secret recipe to a closed competitor, or investors thinking you’re foolish, and to be relegated as a project that isn’t expected to become huge but is for good social cause.

The Silent Majority

The gist is that a lot of investment is being made into a future where AI is more private and access is decided by a few, but go talk with almost anyone and you’ll find few people actually want this - even those at the top claim they don’t.

Our focus on making an open future is split across group chats, open-source repos, and hackathon projects, and a few companies that vocally care about it alongside making profit. Still, this is not even close to the amount of centralized money and talent going into creating private AI research and frontier models within a handful of companies and governments.

Ok so what exactly are we doing?

1. We set a clear name and goal to rally around
   This is what the name Democratize Intelligence stands for.
   The goal is to make it easier for more people to build and access state of the art AI.
   ps. I know some engineers don’t like the word Democratize - but in my defense OpenAI was taken & it’s a lot better in this case to pick something than do nothing, we can iterate and we’re open to other name suggestions!

2. We create a pulse for the community

   We should help you meet your next cofounder, raise money, hire experts.
   We do this in person by hosting dinners at AGIHouse, running hackathons, and the DEMI Summit conference series across tech hubs like San Francisco and New York. Online we’ll experiment with curated breakout groups and an online community chat.
   If a builder wants to make AI more open we want them to feel they have an army on their side.
   As simple as it may sound, a lot of incredible things happen when you bring together smart people who all care about a shared thing. A fun example is that the field of artificial intelligence arguably began like this: as a summer research workshop, in particular the Dartmouth Summer Research Workshop of 1956.

3. We see what builders need most and grow in that direction

   If we do (1) and (2) properly we should have the best understanding of what builders need to make AI more open — be it money, or talent, some a tool, or government policy change. We then expect to be one of the best to help them get this. Traditionally this could look like starting an investment fund, an accelerator, or building a job portal — but AI is not traditional. Even though we may likely do these things I firmly believe waiting it out a bit and seeing what demand emerges from the network of builders will be far more useful in staying focused.

As we grow we’ll have evolving metrics to measure success — one immediate idea is a DEMI index - similar to the global Economic Democracy Index, but focused more on FLOPs and tokens than on adult literacy rates as the global index uses.

Long term a simple heuristic we can use is counterfactual impact ie — in 10 years if we look back at the world and say “You know, AI might have ended up fully centralized and privatized if it weren’t for Tinygrad, or Modular, or ComputeExchange, or SiliconData, or {list your project here}” — I want those projects to win and be really successful because of DEMI and because we did the 3 steps described above.

What we need to build

Here are some known topics — some with existing projects — that make massive difference in the future of AI being open or not. We only expect the list to grow.

• Cross-platform AI frameworks that prevent vendor lock-in (like TinyGrad, and Modular AI).

• Compute Marketplaces that make it easier for buyers and sellers to transact compute, like Compute.Exchange

• New compute options that break free from a single chip or vendor (like groq).

• Transparent compute market data that helps developers optimize pricing and access (like SiliconData).

• Open, Neutral Standards: An Ethernet-like standard for AI infrastructure to avoid closed ecosystems.

• New funding models that enable open research with high equity upside.

• Decentralized Training - infrastructure that lets you get more out of your compute (like PrimeIntellect)

Get involved

Join our mailing list below to keep a pulse on upcoming events, and how the ecosystem grows over time. Please feel free to directly email kyle@demi.so with any feedback and ideas.

It’s time to build actually open AI.