Ep. 59: WTF Is a Frontier Model?
In this episode I break down WTF a frontier model actually is, starting with what most people assume the term means versus the real technical definition. I cover how “frontier model” got coined in 2023, and why the term even came about at all, and how self-naming benefitted its creators. A bit of an etymology episode, this episode is more about the background of the term, and explores how new open weight models are challenging the definition.
Getting Curious
“Frontier model” is a term I’ve absolutely used before, and I don’t want to assume you actually know what it means, because in full transparency, prior to doing the background research for this episode, I only had a loose understanding of what it meant myself. It be like that sometimes.
So today we’re gonna keep the WTF series going and briefly cover WTF a frontier model is. There’s honestly not that much to say about them, so this will be a shorter episode, but curiosity has no size requirement.
Most people think the term frontier model refers to the most established AI models out right now (ChatGPT, Claude, Gemini), and technically it does, but that’s not actually the technical definition. So let’s start there.
A Brief History: The Frontier Model Forum
The term frontier model was first coined in July 2023 when Anthropic, Google, Microsoft, and OpenAI formed, wait for it, the Frontier Model Forum (FMF). They founded the FMF (which still exists today), as an industry body focused on ensuring safe and responsible development of AI models.
Per the FMF, a frontier model was defined as: a model that beats everything else that’s been widely deployed for the past 12 months. (If that’s not a brochacho definition I don’t know what is.)
About two years later the EU AI Act set for a more objective definition of a frontier model, and based the demarcation off of training-compute amount, basically how much math was done to train the model. This is notable, because now any new companies that want to build models that capable, and operate in the EU, have to build the safety evaluations, reporting, and compliance infrastructure required by the EU AI Act. Translation: you gonna have to pay if you want to play.
The FMF Today
So, to recap, the term frontier came about in 2023 when the biggest players in the AI game got together and said, “We’re going to form an AI safety committee that includes only us,” and we’re going to call it the Frontier Model Forum.
This forum was self-governed, self-named, and set the standard for what a frontier model was.
The FMF does still exist and has expanded from 4 to 6 members, with Amazon and Meta joining in 2024.
As for what the FMF actual does:
- The FMF publishes safety research and technical white papers on things like biosafety thresholds, cyber defense, and adversarial distillation.
- It runs the AI Safety Fund, which, and I could be wrong, feels like it’s for show more than anything given that the fund launched at $10 million in October of 2023 and has to grow despite the absolutely insane valuations of these AI companies.
- All members signed an agreement to share information about vulnerabilities, threats, and capabilities of concern specific to frontier AI.
But Why?
If you’re anything like me, aka hella suspicious and skeptical about these companies, you’re likely asking why these companies started the Frontier Model Forum in the first place. Aka, what did they have to gain?
You’d be right to ask that, because it did in fact benefit them:
- Regulatory capture — Governments were actively writing AI regulation at that time. Forming the FMF allowed these companies to define the category they’d be regulated under.
- Brand positioning — In 2023, AI was already being associated with deepfakes and bots. By calling themselves “frontier,” they positioned themselves as “the serious, safety-conscious ones.”
- Legal coordination — Antitrust law prevents competitors from sharing sensitive technical intel. By forming the FMF, OpenAI, Anthropic, and Google were legally able to share information about “threats” like Chinese “adversarial distillation” (systematic querying of a model to train a competing model).
So, yes, forming the FMF and categorizing themselves as “frontier” was absolutely a net positive for those founding companies.
What About Open Weight Models?
By the FMF’s own definition of frontier, a model that beats everything else that’s been widely deployed for the past 12 months, new open weight models like Kimi’s K3 (check out episode 55 to learn about open weight models) would qualify for this moniker.
What I personally am seeing in the AI space, however, are labels like “frontier-grade” or “near frontier,” which to me supports the idea that the term “frontier” is really more so tied to a specific club and the government relationships built around that club, not so much about benchmark scores. It’s giving ol’ boys club.
How I Used AI This Week
Each episode I share a quick example of how I used AI that week.
This week I used Claude to build a workout tracker for my mom! The whole thing took me like 2 hours, and I planned it out with Claude and then had Claude Code actually build it.
New to the build process this time, I used Google Sheets as the database.
If you want your personal web app to retain information across sessions (ex: remembering the workout you did last week), you need a database. I usually use Supabase for my database, but I’ve maxed out their generous free tier, and I’ve been wanting to play around with using Google Sheets as the database because #FREE99. I don’t need the functionality of Supabase, namely row level protection, which is great when you have multiple users, but since this is just a personal app for my mom, I decided to give Google Sheets a go and it worked out great.
Honestly, it was really cool to see how far I’ve come in my vibe coding abilities. With my first web app there were so many errors and issues and snags, and there were literally none this time. Yes, that also speaks to how much better Claude is than a year ago, but reps really do help SO much.
IMO, these personal web apps are definitely one of, if not the best use cases for vibe coding, and it was really cool to be able to create a solution for my mom. I customized it, personalized it, and had it waiting for her in her text messages in the morning, and the whole thing was literally no sweat off my back to make. What a time to be alive.
Da Wrap-up
So, to close the loop here: by definition, a frontier model is either one of the most capable AI models that already exists, or one big enough, measured by how much computing power went into training it, that regulators think it could cause serious harm if something goes wrong. So if you read the term “frontier”, your brain should automatically think: Claude, ChatGPT, Google, Meta.
However, given the politics behind all of this, you wouldn’t be “wrong” if you colloquially interpreted frontier model to mean “one of the most established models.”
AI being AI, we’ll see how and if this changes in the next five minutes.
As always, endlessly grateful for you and your curiosity.
Catch you next Thursday.
Maestro out.