Over the last few months I've been in a number of conversations with business executives, media and policy professionals and researchers, who in the sincerest of intentions have said something like "but what even is open source AI" or the scarier version "oh but we don’t have a definition of open source". I think people intend for these statements to be provocations to start a conversation but, in my humble opinion, these become more like confounders.

Let me explain my frustration using the metaphor of a chilli. Yes, that ordinary ingredient used in most kitchens across the world.

By Takeaway, licensed under CC BY-SA 4.0 via Wikimedia Commons

Anyone who spends time in a kitchen- most mothers, some fathers, the born-in-pandemic amateur chefs and botanists know how to identify a chilli. If we want to get technical, it belongs to the genus Capsicum and is eaten raw, dried, fried, crushed. It is one versatile little thing that packs a punch in a dish. There are a bunch of dishes- Mexican Chilli, Chilli Chicken, Honey Chilli potato- where chilli, strictly by volume or weight, isn’t the main ingredient. But chilli is added as an adjective because it adds something unique and important to the dish. Chilli is so important that its influence has spilled far outside the culinary stage. Iconic rock bands are named after it. Restaurants ranging from American fast food chains to small dhabas in India fight over calling themselves Chilli. Whether their food uses chilli or black pepper as the primary spice is a different matter. It has spun off crazy subcultures of chilli eating competitions that serve no functional purpose. None of this, however, confuses anybody about the organic edible thing called chilli. No one is going into malls buying a plastic toy in the shape of a chilli, mistaking it for the real thing, and adding it to their food.

There are a lot of other kinds of plants such as white pepper, black pepper and ginger that can be added to spice food. Sometimes the black pepper growers association might feel envious of the popularity of chilli and crash the chilli growers association conference and scream, “But we do exactly what you do. We are also a great spice!” This doesn’t make black pepper a chilli. A plant must contain capsaicinoids to be called a chilli. Sorry black pepper, you just ain’t chilli.

Now in the crazed decade that is the 2020s, an evil/mad/genius food manufacturer- let’s call it X for simplicity- produces a new range of hot sauces. X pushes these sauces to Michelin star restaurants and grocery stores alike, promising a culinary revolution. The sauce never spoils, it is cheaper than your spices and you can forget your worries about salting your food right. The sauce with its special adaptive properties will figure out exactly how much saltiness to express. So says X.

X does an aggressive push for their chilli sauce to all the kitchens doling out Mexican Chilli, Honey Chilli Potato and Chilli Chicken. But the chefs in this kitchen are circumspect. They don’t know the exact ingredients of the sauce but the professionals have probed enough to know that the sauce also contains some sichuan pepper, ginger, black pepper and other non-spicy things like salt and sugar. Given all these other ingredients, can they use this magical sauce and advertise their dishes with the word chilli in it? They are hesitant to buy it. So X pushes the loose association of botanists to come up with rules against which sauce can be labeled chilli sauce. They, after all, classified plants just okay. Is it a chilli sauce if 20% of the ingredients are chilli? 30%?

In the meantime, the mothers, some fathers and the born-in-pandemic amateur chefs merrily go about cooking in their respective kitchen knowing exactly how to identify a chilli. Some of them give the sauce a try but trust their own expertise over X’s marketing spiel and mix it with other spices- ultimately they just care about a tasty and healthy meal. Some purists eschew the sauce altogether.

Now switch out:

• Chilli for open source Software
• Spice for software
• Sauce for AI

And I think we get a pretty good picture of what is happening with open source and AI.

After that spicy background, I offer four hot takes:

1. The definition of open source software is fairly settled. Any classification exercise, be it of a plant, software or AI, is a heavily social process. There are power dynamics in whose voices count in naming something¹. What happens to objects on the boundaries of classes can be arbitrary (remember when Pluto was still a planet?). But time becomes a good test of any classification system and with open source software, we’ve had two decades. The definition is pretty clear. Just as chilli is identifiable by some chemical properties (presence of capsaicin), open source software is identifiable by legally testable properties.
   None of us working with FOSS have any doubt about how to identify free and open source software. Even when the black pepper folks (DPGs) say that they are brethren of chilli (open source software), it is easy to factually check this. Even if the success of open source software has inspired domains outside of software, the ‘vibe’ of openness is tied to material practices in software. There is a checklist of the four freedoms- to use, study, modify, distribute software- that software licenses must meet.

2. This is a relatively minor point- the creation of a derivative product shouldn’t be pulling the legitimacy of an original ingredient into question. Just because DPG/black pepper is gaining popularity (no less because chilli made spices popular), and we now have AI/sauces doesn’t mean we need to be confused about what open source software/chilli is.

3. Just as the creation of sauce involves a bunch of chemical processes (glycation, acidulation) where ingredients bond in a way that makes it challenging to talk about properties coming from constituent ingredients, so it is with AI. AI is a new kind of technical artifact. This whole affair of defining open source AI started because corporations were using ‘open source’ as a marketing push for AI. There was a fair bit of discussion and deliberation that happened in the process of OSI coming up with OSAID. That definition might not last but we are far beyond the point of “what even is open source AI.” By and large, everyone recognizes that it won’t be as easy to certify AI products against open source licenses as it is to certify software, because AI is software and data amalgamated in a way that makes it hard to separate the contribution of either. More importantly, arguing about the chilli quotient of a sauce might not be the most productive use of our time, when all other aspects of the other production process are also concerning. The content and quality of other ingredients matters. The process and quality of the facility in which the sauce was made also matters. With AI, all aspects of the production process need investigation. This is why I think that the question of “what even is open source AI?” is a distraction or a confounder. If Meta or Anthropic trains their models on books and papers scraped from Libgen but then opens all the data and weights and code for LLama, they would have met all aspects of their model open. But should we celebrate this as a success?
   The open source software definition was supposed to be a mental shortcut for four tangible things that a piece of code enables. It is clear that an open source AI definition won’t map onto a similarly simple check list of four freedoms because data and compute are significantly expensive for any person to exercise the four freedoms with AI. I think people are right to watch the licenses that models are released with, with hawk eyes. But, this won't be enough to ensure autonomy to tinker and innovate, which is what the four freedoms enable. FOSS communities need to get more involved in conversations of data governance (beyond open data) and politics of hardware supply chains. All of that, at present, seems more important than splitting hair over what should constitute an Open Source AI definition.

4. And last, after all this confusion, maybe we should stop invoking ‘open source’ as a synonym of ‘good’. People use open source as an adjective for digital IDs, governments, infrastructures, charity... Just as the band Red Hot Chilli Pepper shares little in common with the plant outside of the name, many of these concepts that might be inspired by FOSS share little in common with FOSS software. The material difference between government and software, money and software, IDs and software matters. Just applying the suffix ‘open’ to big concepts doesn’t make them all unanimously good. This also holds for AI.

It is worth stating that the most egregious misuses of the open source language don’t (usually) come from FOSS practitioners. But a narrative battle around open source has been underway for a while, and we need to take some responsibility for not tackling it with more force. Perhaps we were so used to fighting like the underdogs that we took every adaptation of the language, however loose, to be a win. Or perhaps if we had taken more effort to build public understanding of OSS, a narrative spin would have been less successful. To go back to the chilli metaphor, imagine an alternate world where only some people could distinguish a chilli from sichuan pepper, but the average population understanding stayed at the level of spice. And so when black pepper or sichuan pepper sellers said they are the same as chilli, people had no reason to doubt it. After all, the chilli sellers hadn’t said anything earlier.

FOSS practitioners (and the general public) are caught in a maelstrom of jargon. The one thing we need is to defend and steady the turf of open source software- the definition of open source software is not contested and we need to make this point as loudly and as frequently as possible, and in forums outside our FOSS cliques. We need to up our game as public communicators to bring back the precision of language used to describe technical objects. In chasing that precision we will automatically find ourselves beyond the boundary of open source AI, examining the properties of all the constituents that come together to become AI.

There is also a provocation for the ‘tech for good’ and policy folks (in the odd chance that anyone gets this far)- can we challenge ourselves to write and talk about national infrastructures and digital sovereignty in the age of AI without using the terms Digital Public Goods and Open Source AI? This one little change alone could do a lot to improve the quality of the conversation.

A special thanks to Denny George and Srravya Chandramowli for revieweing an earlier, less hinged version of this blog.

Footnotes

1. In this context, Srravya Chandramowli recommended the book: Sorting Things Out ↩