Read the show’s notes in full
Cloning HGI for AGI
Human intelligence got to where it is today through evolution. Some argue that to get to AGI, we will approximate all the “FLOPs” that went into that process, an approach most famously mapped out by Ajeya Cotra’s Biological Anchors report:
The early days of OpenAI were very reinforcement learning-driven with the Dota project, but that's a very inefficient way for these models to re-learn everything. (Kanjun from Imbue shared similar ideas in her episode).
David argues that there’s a shortcut. We can bootstrap from existing intelligence.
“Years ago, I had a debate with a Berkeley professor as to what will it actually take to build AGI. And his view is basically that you have to reproduce all the flops that went into evolution in order to be able to get there… I think we are ignoring the fact that you have a giant shortcut, which is you can behaviorally clone everything humans already know. And that's what we solved with LLMs!”
LLMs today basically model intelligence using all (good!) written knowledge (see our Datasets 101 episode), and have now expanded to non-verbal knowledge (see our HuggingFace episode on multimodality). The SOTA self-supervised pre-training process is surprisingly data-efficient in taking large amounts of unstructured data, and approximating reasoning without overfitting.
But how do you cross the gap from the LLMs of today to building the AGI we all want?
This is why David & friends left to start Adept.
“We believe the clearest framing of general intelligence is a system that can do anything a human can do in front of a computer. A foundation model for actions, trained to use every software tool, API, and webapp that exists, is a practical path to this ambitious goal” — ACT-1 Blogpost
Critical Path: Abstraction with Reliability
The AGI dream is fully autonomous agents, but there are levels to autonomy that we are comfortable giving our agents, based on how reliable they are. In David’s word choice, we always want higher levels of “abstractions” (aka autonomy), but our need for “reliability” is the practical limit on how high of an abstraction we can use.
“The critical path for Adept is we want to build agents that can do a higher and higher level abstraction things over time, all while keeping an insanely high reliability standard. Because that's what turns us from research into something that customers want. And if you build agents with really high reliability standard, but are continuing pushing a level of abstraction, you then learn from your users how to get that next level of abstraction faster. So that's how you actually build the data flow.
That's the critical path for the company. Everything we do is in service of that.”
We saw how Adept thinks about different levels of abstraction at the 2023 Summit:
The highest abstraction is the “AI Employee”, but we’ll get there with “AI enabled employees”. Alessio recently gave a talk about the future of work with “services as software” at this week’s Nvidia GTC (slides).
No APIs
Unlike a lot of large research labs, Adept's framing of AGI as "being able to use your computer like a human" carries with it a useful environmental constraint:
“Having a human robot lets you do things that humans do without changing everything along the way. It's the same thing for software, right? If you go itemize out the number of things you want to do on your computer for which every step has an API, those numbers of workflows add up pretty close to zero. And so then many points along the way, you need the ability to actually control your computer like a human. It also lets you learn from human usage of computers as a source of training data that you don't get if you have to somehow figure out how every particular step needs to be some particular custom private API thing. And so I think this is actually the most practical path (to economic value).”
This realization and conviction means that multimodal modals are the way to go. Instead of using function calling to call APIs to build agents, which is what OpenAI and most of the open LLM industry have done to date, Adept wants to “drive by vision”, (aka see the screen as a human sees it) and pinpoint where to click and type as a human does. No APIs needed, because most software don’t expose APIs.
Extra context for readers: You can see the DeepMind SIMA model in the same light:
One system that learned to play a diverse set of games (instead of one dedicated model per game) using only pixel inputs and keyboard-and-mouse action outputs!
The OpenInterpreter team is working on a “Computer API” that also does the same.
To do this, Adept had to double down on a special kind of multimodality for knowledge work:
“A giant thing that was really necessary is really fast multimodal models that are really good at understanding knowledge work and really good at understanding screens. And that is needs to kind of be the base for some of these agents…
…I think one big hangover primarily academic focus for multimodal models is most multimodal models are primarily trained on like natural images, cat and dog photos, stuff that's come out of the camera… (but) where are they going to be the most useful? They're going to be most useful in knowledge work tasks. That's where the majority of economic value is going to be. It's not in cat and dogs.
And so if that's what it is, what do you need to train? I need to train on like charts, graphs, tables, invoices, PDFs, receipts, unstructured data, UIs. That's just a totally different pre-training corpus. And so Adept spent a lot of time building that.”
With this context, you can now understand the full path of Adept’s public releases:
* ACT-1 (Sept 2022): a large Transformers model optimized for browser interactions. It has a custom rendering of the browser viewport that allows it to better understand it and take actions.
* Persimmon-8B (Sept 2023): a permissive open LLM (weights and code here)
* Fuyu-8B (Oct 2023): a small version of the multimodal model that powers Adept. Vanilla decoder-only transformer with no specialized image encoder, which allows it to handle input images of varying resolutions without downsampling.
* Adept Experiments (Nov 2023): A public tool to build automations in the browser. This is powered by Adept's core technology but it's just a piece of their enterprise platform. They use it as a way to try various design ideas.
* Fuyu Heavy (Jan 2024) - a new multimodal model designed specifically for digital agents and the world’s third-most-capable multimodal model (beating Gemini Pro on MMMU, AI2D, and ChartQA), “behind only GPT4-V and Gemini Ultra, which are 10-20 times bigger”
The Fuyu-8B po
