OSS vs OpenAI
The race goes not always to the swift ... but that remains the way to bet.
On Twitter/X today one can witness an ongoing war of words1 between proprietary LLMs like OpenAI’s, and open-source models like Llama-22 and Mistral. Some say real generative AI applications will be built on proprietary LLMs, because they’re, well, better. Others claim that real, defensible, LLM apps will require custom-trained models, often based on OSS; anything less is referred to contemptuously as mere “OpenAI wrappers,” doomed to be superseded.
It would probably be good if open source won. As it often does, in software! …But alas, history suggests that in this arena, it most likely will not. The reason being that but LLMs are far more like hardware than software.
Intense, research-heavy, bleeding-edge tech, with its own approach nothing like that of “traditional” software? Check. (Modern AI is so different from procedural software that “traditional” practitioners often can’t even understand conversations between AI practitioners.) Enormous upfront cost to (fabricate|train) the technology? Check. Further, very significant expenses to actually (buy|run) the (chip|model)? Check. Analogy is of course always suspect … but if you analogize today to, say, the birth of modern computing in the 80s, OpenAI currently has far more in common with Intel than with Microsoft.
That in turn makes “OpenAI wrapper companies have no moat” sound like “Software companies are just hardware wrappers with no moat.” Which today sounds not just wrong but hilariously wrong … but people actually said that! More precisely, they said “People who are really serious about software should make their own hardware.” This was a very popular belief. It was successful-ish for a little while at places like Sun and SGI, and it of course remains extremely successful at Apple. Custom ASICs remain a multibillion-dollar market. But it is clear in hindsight that pure software companies were, he understated, not exactly doomed.
The killer flip side to “don’t be an Intel wrapper, build your own hardware” was not just the enormous expense but the subsequent Red Queen’s race with Intel/Motorola. What do you say to customers who demand to use their superior chip instead of your homebrew crap; and what do you do, other than quietly die, when next year’s Pentium blows your feeble efforts out of the water?
Similarly, “nobody will want to send OpenAI their data” and “But you don’t have control over your own platform” is just cloud computing all over again. It turns out people and enterprises alike are happy for Amazon and Google and Microsoft to host their data. And just as the ability to switch among those Big Three (and smaller upstarts like Digital Ocean) keeps the cloud relatively open and competitive, OpenAI vs Anthropic vs DeepMind vs other upstarts should do the same for LLMs.
This in turn is what made last week’s OpenAI Dev Day announcements so interesting. They released GPTs, sort of no-code customizable / shareable ChatGPTs, and those are indeed very cool and interesting. But they also released their “Assistants API,” which goes far beyond what any of their competitors (currently) offer … and I think in the long run that’s much more interesting.
LLMs all fundamentally have exactly the same interface; shovel input tokens into their context window, siphon output tokens out. But the Assistants API adds another layer of abstraction. It runs code, does custom function calling, and runs RAG on your files for you. In other words, it’s a step towards not just an LLM interface — but an LLM operating system.
Imagine if Intel had pre-empted Microsoft by building and offering an OS themselves, way back in the “In The Beginning Was The Command Line” days. Would Microsoft have even existed? Would Intel have become the all-conqueror of both hardware and software? That seems to me to be the play OpenAI is making today.
Even in that somewhat disconcerting future, there will be many places for open-source LLMs, to do the relatively unsophisticated model gruntwork at scale, or on-device, or far more cheaply. But for the flagship applications, the hard analysis, the serious pattern recognition, the actual … well … artificial intelligence?3 That’s what people most want … and that’s where, while I hate to say it, it’s currently hard to see OSS becoming seriously competitive, and easy to see OpenAI aiming at becoming not just the Intel but also the Microsoft of modern AI.
because hilariously this arguably needs clarification, I mean a war of words between the champions of the two types of LLMs, not between LLMs themselves.
from Meta/Facebook, the surprise new champion of open source!
Stipulated that it’s not “real” intelligence, but if it’s technology used in lieu of real intelligence, “artificial intelligence” doesn’t seem like the worst term, y’know?