How many AI models is too many? It depends on how you look at it, but 10 a week is probably a bit much. That’s about as many as we’ve seen roll out over the past few days, and it’s getting harder and harder to tell if and how these models compare to each other, if it was ever possible to begin with. So what’s the point?
We’re in a weird time in the evolution of artificial intelligence, though of course it’s been pretty weird all along. We are seeing a proliferation of models large and small, from niche developers to large, well-funded ones.
Let’s run down the list from this week, shall we? I tried to summarize what makes each model stand out.
- LLaMa-3: Meta’s latest “open” model of large languages. (The term “open” is currently disputed, but this project is widely used by the community regardless.)
- Mistral 8×22: A “specialty mix” model, on the large side, from a French outfit that has eschewed the openness they once embraced.
- Stable Diffusion 3 Turbo: An upgraded SD3 to match the new open stability API. Borrowing “turbo” from OpenAI’s model nomenclature is a bit odd, but ok.
- Adobe Acrobat AI Assistant: “Talk to your document” from the 800 lb document gorilla. Surely this is mostly a wrapper for ChatGPT, though.
- Reka Core: From a small team formerly employed by Big AI, a multimodal model built from scratch that is at least nominally competitive with the big dogs.
- Idefics2: A more open multimodal model, built on top of recent, smaller Mistral and Google models.
- OLMo-1.7-7B: A larger version of AI2’s LLM, one of the most open out there and a stepping stone to a future 70B scale model.
- Pile-T5: A version of the ol’ reliable T5 model optimized in the Pile code database. The same T5 you know and love but better encoding.
- Cohere Compass: An “integration model” (if you don’t already know it, don’t worry about it) that focuses on integrating multiple data types to cover more use cases.
- Imagine the Flash: Meta’s newest image production model, based on a new distillation method to speed up diffusion without too much compromise in quality.
- Unlimited: “A personalized AI powered by what you’ve seen, said or heard. Iit’s a web app, a Mac app, a Windows app, and a mobile app.” 😬
That’s 11, because one was announced while I was writing this. And not all models were released or previewed this week! It’s just what we saw and discussed. If we relaxed the inclusion conditions a bit, there would be dozens: some perfected existing models, some combinations like Idefics 2, some experimental or specialized, and so on. Not to mention this week’s new tools for crafting (torch) and fight against (Ganoma 2.0) genetic AI!
What to do about this endless avalanche? We can’t “review” everything. So how can we help you, our readers, understand and track all these things?
The truth is, you don’t have to go on. Some models such as ChatGPT and Gemini have evolved into full web platforms, covering multiple use cases and access points. Other large language models like LLaMa or OLMo — although they technically share a basic architecture — don’t actually fulfill the same role. They are meant to live in the background as a service or component, not in the foreground as a brand.
There is some deliberate confusion about these two things because the model developers want to borrow some of the fanfare associated with big AI platform releases like GPT-4V or Gemini Ultra. Everyone wants you to believe that their release is important. And while it’s probably important to someone, it’s almost certainly not you.
Think of it in terms of another broad, diverse category like cars. When they were first invented, you just bought “a car”. Then a bit later you could choose between a big car, a small car and a tractor. Today, there are hundreds of cars on the market every year, but you probably don’t need to know about one in ten of them, because nine out of ten aren’t a car you need, or even a car as you understand the term. Similarly, we are moving from the big/small/tractor era of AI to the proliferation era, and even AI experts can’t keep up and test all the models out there.
The other side of this story is that we were already at this stage long before ChatGPT and the other big models came out. Far fewer people were reading about it 7 or 8 years ago, but we covered it anyway because it was clearly a technology waiting for its moment to explode. There were papers, models, and research coming out all the time, and conferences like SIGGRAPH and NeurIPS were full of machine learning engineers comparing notes and building on each other’s work. Here is a visual comprehension story I wrote in 2011!
This activity continues daily. But since AI has become big business—arguably the biggest in tech right now—these developments have added a little extra weight, as people wonder if one of them might be such a big leap over ChatGPT that the ChatGPT was compared to its predecessors.
The simple truth is that none of these models are going to be that big of a step, since OpenAI’s progress was based on a fundamental change in machine learning architecture that every other company has now adopted and that hasn’t been superseded. Incremental improvements, like a point or two better on a synthetic benchmark, or marginally more persuasive language or imagery, are all we have to wait for now.
Does this mean that none of these models matter? They certainly do. You don’t get from 2.0 to 3.0 without 2.1, 2.2, 2.2.1 and so on. And sometimes these developments make sense, address serious shortcomings, or expose unexpected vulnerabilities. We try to cover the interests, but this is only a fraction of the full number. We’re actually working on a piece that collects all the models we think ML geeks should know about, and it’s in the order of a dozen.
Don’t worry: when a big one comes along, you’ll know it, and not just because TechCrunch covers it. It will be as obvious to you as it is to us.
