Misses a few interesting early models: GPT-J (by Eleuther, using gpt2 arch) was the first-ish model runnable on consumer hardware. I actually had a thing running for a while in prod with real users on this. And GPT-NeoX was their attempt to scale to gpt3 levels. It was 20b and was maybe the first glimpse that local models might someday be usable (although local at the time was questionable, quantisation wasn't as widely used, etc).
GPT-J was the one that made me really interested in LLMs, as I could run it on a 3090.
Some details on the timeline are not quite precise, and would benefit from linking to a source so that everyone can verify it. For example, HyperClOVA is listed as 204B parameters, but it seems it used 560B parameters (https://aclanthology.org/2021.emnlp-main.274/).
Great idea! Thanks
Great catches — just added GPT-Neo (2.7B, Mar 2021), GPT-J (6B, Jun 2021), and GPT-NeoX (20B, Apr 2022). Thanks!
This would be interesting if each of them had a high-level picture of the NN, "to scale", perhaps color coding the components somehow. OnMouseScroll it would scroll through the models, and you could see the networks become deeper, wider, colors change, almost animated. That'd be cool.
Thanks! Great idea
Why is it hard in the times where AI itself can do it to add a light mode to those blacks websites!? There are people that just can't read dark mode!
Visual presentation has been a weak point of AI generation for me. There isn't a lot of support for them seeing how a potential presentation might appear to a human.
Models that take visual input seem more focused on identifying what is in the image compared to what a human might perceive is in an image, and most interfaces lack any form of automated feedback mechanism for them to look at what it has made.
In short, I have made some fun things with AI but I still end up doing CSS by hand.
Thank you! Sorry for the inconvenience. I'll add it a bit later
750+ here:
Great resource — Dr. Thompson's table is exhaustive. llm-timeline.com takes a different angle: visual timeline format, focused on base/foundation models only, filterable by open/closed source. Different tools for different needs.
Calling this "The complete history of AI" seems wrong. LLM's are not all AI there is, and it has existed for way longer than people realize.
Fair point — updated the tagline to 'The complete history of LLMs'. AI as a field goes back decades; this is specifically tracking the transformer/LLM era from 2017 onward
Most of "AI" before ChatGPT was just researchers wasting public grant money, eg BLOOM.
Easy to forget but there was a ton of industry+investor excitement around computer vision from ~2015-2021, to the extent that the "MLops" niche sprung up around it. This was called AI at the time, and mostly went out the window when general-pupose pretrained models arrived.
And now it's private companies wasting investor money. Not sure there's much difference between the two.
Would be nice to see some charts and perhaps an average of the cycles with a prediction of the next one based on it
Thanks! I'll add some charts
It misses almost every milestones, and lists Llama 3.1 as milestone. T5 was much bigger milestone than almost everything in the list.
Fair point on T5 — just marked it as a milestone. On Llama 3.1: it's there as a milestone because it was the first open model to match GPT-4 at 405B, which felt like a genuine inflection point. Happy to debate the milestone criteria though — what would you add?
That was llama 3, which is marked as milestone already.
Also I would say add apple/DCLM-7B(not as milestone imo) as it was kind of the first fully open model which was at least somewhat competitive with closed data model.
> T5 was much bigger milestone than almost everything in the list.
It's in the timeline though? Or are you saying that one should somehow be highlighted, even though none of the other ones are? Seems it's just chronological order, with no one being more or less visible than others, as far as I can see.
Some are highlighted and listed as milestones.
The models used for apps like Codex, are they designed to mimic human behaviour - as in they deliberately create errors in code that then you have to spend time debugging and fixing or it is natural flaw and that humans also do it is a coincidence?
This keeps bothering me, why they need several iterations to arrive at correct solution instead of doing it first time. The prompts like "repeat solving it until it is correct" don't help.
> as in they deliberately create errors in code that then you have to spend time debugging and fixing
No, all the models are designed to be "helpful", but different companies see that as different things.
If you're seeing the model deliberately creating errors so you have something to fix, then that sounds like something is fundamentally wrong in your prompt.
Besides that, I'm guessing "repeat solving it until it is correct" is a concise version of your actual prompt, or is that verbatim what you prompt the model? If so, you need to give it more details to actually be able to execute something like that.
Great site! I noticed a minor visual glitch where the tooltips seem to be rendering below their container on the z-axis, possibly getting clipped or hidden.
Thanks for the feedback! I'll fix it asap.