{"id":19192,"date":"2023-10-03T15:40:15","date_gmt":"2023-10-03T15:40:15","guid":{"rendered":"https:\/\/www.michaeldempsey.me\/blog\/?p=19192"},"modified":"2023-10-03T15:41:41","modified_gmt":"2023-10-03T15:41:41","slug":"open-source-ai-business-models-brand-moats","status":"publish","type":"post","link":"https:\/\/www.michaeldempsey.me\/blog\/2023\/10\/03\/open-source-ai-business-models-brand-moats\/","title":{"rendered":"Open Source AI Business Models & Brand Moats"},"content":{"rendered":"\n

Typically in open-source software the business model for venture-backed companies has been to release a public good and build managed services on top of it. This has worked with low capex and a set of contributors working for long periods of time. These companies and individuals maintain the core repo and build a community, both of which builds structural resilience amongst a group of core developers and becomes some form of a standard.1<\/a><\/sup>We can debate the semantics of this but let\u2019s not.<\/span><\/p>\n\n\n\n

With open-source AI (OSAI) labs you are running super high capex in order to build a public good (model) for some period of time, that you hope the community perhaps then goes to iterate upon. While certain labs create small communities, very few have broken out in language, instead coalescing around subreddits like r\/LocalLLaMa<\/a> and Hugging Face. However as models become more modular and hot-swappable at the inference layer, it\u2019s likely that a given model (of which we can assume $100k-$1M of capex each at least) will become stale and be swapped out. <\/p>\n\n\n\n

While this dynamic can and will likely push forward the overall rate of progress within AI, it\u2019s unclear if it\u2019s a dominant economic strategy. And depending on how horizontal your model is, these dollar figures can creep higher and result in 7-8 figures to \u201cmaintain\u201d your place in the OSAI ecosystem if playing at the model layer.<\/p>\n\n\n\n

This brings obvious complexity issues as to how a variety of labs predicated on open-source last long-term unless the goal is to do radical things like start open-source and then move closed-source, or build new forms of appstores where a series of models are commercialized and the OSAI lab becomes a take-rate driven model at the API layer. Both of these feel like a losing battle vs. OpenAI, Anthropic, Meta, Apple, Google, MSFT etc. especially as Meta allows free commercial usage of LLaMa-2<\/a> up until ~700 million monthly active users.2<\/a><\/sup>You could also go the path of building novel hardware around your open-source models. I like this approach and there’s currently a battle between Rewind and Tab on this right now, both of which I worry will end up hitting some limitations.<\/span><\/p>\n\n\n\n

The obvious other answer is of course to build tools around your models and own the infra\/ middleware layer and\/or building custom models for enterprises or governments. This is compelling as we see more \u201cnovel\u201d approaches to performance pop up that can increase inference quality and speed, including architectures like Mixture of Experts or deployment choices like Agents, each of which have basically no discernible and approachable infrastructure at scale today.3<\/a><\/sup>There are efforts to try this recently with OpenMoE<\/a> amongst others which we are bullish on.<\/span><\/p>\n\n\n\n

That said, the enterprise approach is one that a variety of proprietary, closed-source model providers will go after, especially as the Frontier Model companies start to see divergence in success. One can easily imagine a possible scenario where if OpenAI and\/or Anthropic steamroll everyone on horizontal models, the next pitch for second-tier “Frontier” language model companies will be the custom-trained enterprise path in order to grow into multi-billion dollar valuations.<\/p>\n\n\n\n

I\u2019m net incredibly bullish on the open-source side of AI and my answer to many of these questions comes down to me perpetually screaming that this becomes a view on who can build the best process to advance proprietary model performance, the infrastructure to be malleable on the model side, and the best (and most opinionated version defining \u201cbest\u201d) product to utilize both proprietary and finetuned open-source models over time.<\/p>\n\n\n\n

In this world, I\u2019m not sure the labs doing open-source model development win at scale, but maybe I\u2019ll be wrong, so please challenge me on this.<\/p>\n\n\n\n

Brand Moats<\/h2>\n\n\n\n

This leads me to my second view on OSAI models and broadly all AI labs and companies today, which is the increasing importance of brand moats.<\/p>\n\n\n\n

While not a novel paradigm in open-source, developers have historically been far less fleeting in how they utilize fairly solidified OSS layers of the stack because of some form of network effect that spins up. This network effect is often a slow build and creates a superior product, that generates minimal revenue on its own, and thus becomes a public good that faces moderate but not intense competition for some short to mid-term period.4<\/a><\/sup>Again, there\u2019s nuance here but just let me have this, i\u2019m writing this in a hotel lobby at a conference while I recharge my social battery.<\/span><\/p>\n\n\n\n

If we try to parse how this could play out in OSAI, crypto protocol development of all places may have some important lessons for companies to learn from.<\/p>\n\n\n\n

What AI Labs can learn from Crypto Protocols<\/h3>\n\n\n\n

Crypto has faced quite different dynamics to traditional OSS, with protocols facing consistent forks as the crypto industry tries to financialize public goods and open-source value creation. This leads to a rolling ball of money that is incentivized to rotate between projects and usage that is mercenary at best.<\/p>\n\n\n\n

However, as we\u2019ve seen over the past few years, this hardcore financial incentivization rarely pulls developers or users for a long period of time. This is because protocols that are longer-standing or truly novel have done a great job at generating brand moats in a fairly recursive and tight-knit space. Sometimes this is due to superior technology, but other times it is merely elegance of design, Lindy effect, community, and a variety of other intangible<\/em> factors.<\/p>\n\n\n\n

\"\"
Friend.Tech Protocol Fee chart which shows daily USD protocol fee revenue (cumulative y-axis is off)<\/figcaption><\/figure>\n\n\n\n

Most recently a great example of this is Friend.tech<\/a>, a project that has done over $15M in revenue<\/a> in the past ~2 months. As Friend.Tech has faced a variety of mercenaries forking<\/a> their protocol, it largely has been unaffected due to a brand moat (and some product development pace) that is unassailable despite the simple mechanism of speculating on access to group chats via Keys being the core underpinning of the product.<\/p>\n\n\n\n

Put more simply, crypto participants just like and trust Racer<\/a>, the co-creator the application, and respect their native understanding of the community and their brand cache that comes with having now shipped multiple<\/em> consumer crypto products that capture the zeitgeist of the industry. The dollars and users thus won\u2019t flow as freely just because of incremental progress.<\/p>\n\n\n\n

AI Brand Moats<\/h3>\n\n\n\n

At this stage of the OSAI cycle, the dynamics for open-source models are far more closely related to those that the crypto industry deals with than more traditional OSS software.<\/p>\n\n\n\n

In a prior post<\/a> I wrote:<\/p>\n\n\n\n

\n

What happens when you add personality or an opinionated aesthetic to a given AI tool is you create a community that feels allegiance to your product in a deeper way than just ROI, perhaps allowing for products to compete with larger, more general purpose models, or just better-funded organizations.<\/em><\/p>\n

This community can be an immensely beneficial flywheel and perhaps a more important moat than we appreciated, as we\u2019ve seen with the thousands of people contributing to the Midjourney and Runway communities over time, pulling out new experiences with prompts and showing everyone what is possible. These are one of the core things that compound within these startups outside of technical knowledge, data, and compute.<\/em><\/p>\n