The Memo - 2/Mar/2023
Microsoft Kosmos-1, OpenAI's 736 staff, Tesla Optimus progress, and much more!
FOR IMMEDIATE RELEASE: 2/Mar/2023
Welcome back to The Memo.
In the Toys to play with section, we look at a new and better version of Replika.ai (which reminds me of Leta AI), a Gmail client with new GPT-3 functionality, an advanced prompt cookbook by OpenAI, and much more.
The BIG Stuff
OpenAI’s 736 employees: 59 from Google AI, 34 from Meta AI (23/Feb/2023)
The heading might read ‘59 stolen from Google AI, 34 stolen from Meta AI’!
A report by a US-based thinktank reveals even more detail. ‘Punks and pinstripes is a private network of the most impactful, rebel executives in the Fortune 500.’ Their latest report looks at the most recent hires at OpenAI, and it’s a useful read to understand where the PhD-level staff training the world’s most popular models (OpenAI’s GPT-4 and GPT-5, DALL-E2, and many more) have come from…
Some companies are already replacing workers with ChatGPT (25/Feb/2023)
Earlier this month, job advice platform Resumebuilder.com surveyed 1,000 business leaders who either use or plan to use ChatGPT. It found that nearly half of their companies have implemented the chatbot. And roughly half of this cohort say ChatGPT has already replaced workers at their companies…
Business leaders already using ChatGPT told ResumeBuilders.com their companies already use ChatGPT for a variety of reasons, including 66% for writing code, 58% for copywriting and content creation, 57% for customer support, and 52% for meeting summaries and other documents. - via Fortune (25/Feb/2023)
Read more via the primary source; it is very much worth reading if you are a government policy writer, an AI lab, or a human who cares about what is happening!: https://www.resumebuilder.com/1-in-4-companies-have-already-replaced-workers-with-chatgpt/
Stability.ai’s 70B Chinchilla-aligned language model is training now (26/Feb/2023)
CEO and Founder of Stability AI admitted that they ‘have the new language and code ones training [now]… Should outperform [Meta AI’s latest 65B model] llama in Lm side at least. Doubt anyone needs more than 70bn parameters’
Analysis of OpenAI’s Foundry and GPT-4 (Feb/2023)
Nathan Labenz provides some insights on Foundry and GPT-4 in that horrible Twitter-thread-format. (I don’t know what Twitter is for, but it’s not for this. Just write a report or article, please!)
…economically transformative AI is not only here, but OpenAI is already selling it…
Read it unrolled: https://threadreaderapp.com/thread/1630284912853917697.html
Tesla’s Optimus robot progress: now walking (2/Mar/2023)
Yes, this is ‘big stuff’! Announced a few hours ago at Tesla Investor Day 2023.
Elon: The latest video shows Optimus walking around. It's weird seeing the arms and legs just separate. We have a whole lab full of arms and legs.
During AI day, this version of Optimus didn't work for walking at all, so the rate of improvement here is significant. It's obviously not doing parkour, but it is walking around, and we have multiple copies of Optimus.
The thing that Tesla brings to the table that others don't have is that we have real-world AI. We're the most advanced in real-world AI. The same AI that drives the car, which you can think of as a robot on wheels, goes into Optimus.
It's not that helpful to have a humanoid robot if you have to program every individual action. It needs to be able to walk around autonomously and solve tasks. You should be able to instruct it in simple things by showing visually what the robot needs to do or just telling it what to do. So, that's a key advantage that we have.
We are also good at designing things for manufacturing, and then manufacturing itself. The actuators in Optimus are all custom designed Tesla actuators. We designed the electric motor, the gearbox, the power electronics, obviously the battery pack, and everything else that goes into Optimus. We were quite surprised to find how little was available off the shelf because there's a vast number of electric motors, gearboxes, and whatnot available in the world, and we found none of them were useful in a humanoid robot. You have to custom design the actuators for a humanoid robot.
For practical purposes, this means that we should be able to bring an actual product to market at scale that is useful far faster than anyone else. Assuming the things I'm saying are true, or at least I think they are true, it's just a question of timing.
You start getting into interesting questions of like: ‘what's the ratio of humans to humanoid robots?’ I think it might be greater than one to one because you could see a use for robots in homes and certain industrial uses for humanoid robots. I think we might exceed a one-to-one ratio of humanoid robots to humans. It's not even clear what an economy means at that point because an economy is output per person times persons, but if output is much higher and there's no limit on persons, then what's the actual limit on the economy?
The Interesting Stuff
Microsoft Kosmos-1: a Multimodal Large Language Model (MLLM) (27/Feb/2023)
The model can see the images in the real Raven’s Progressive Matrix tests! https://lifearchitect.ai/ravens/#kosmos-1
AND, the model outperforms random chance with these Raven’s Progressive Matrices. That is a big advance. Compare what they did for text-davinci-003 with the much simpler flattened digits: https://lifearchitect.ai/ravens/
Alibaba + Ant release Composer text-to-image model (26/Feb/2023)
Composer is a new large (5B) controllable diffusion model.
‘Composer exponentially expands the control space through composition, leading to an enormous number of ways to generate and manipulate images, i.e., making "the infinite use of finite means".’ Code and models will be made available.
View the project page: https://damo-vilab.github.io/composer-page/
View the repo: https://github.com/damo-vilab/composer
ChatGPT API released (1/Mar/2023)
As expected, OpenAI have finally released an API to ChatGPT, with the model called gpt-3.5-turbo. A couple of surprises:
The ChatGPT temperature is not revealed; they recommend a temp of 0.2-0.8.
The cost of using gpt-3.5-turbo is 10% of text-davinci-003: $0.002 vs $0.02 per 1,000 tokens or ~750 words. This means that $1 would get you roughly 350,000 words in and out; about 10x lower than using the next best model.
Maybe not a surprise, but the API setup is based on a conversation and adds new message parameters: prompt/priming (‘system’), and then ‘user’, and ‘assistant’ turns.
Read more: https://platform.openai.com/docs/guides/chat
Waymo adds LA to Phoenix, San Francisco, and more (27/Feb/2023)
We talk a lot about Google’s self-driving cars, Waymo, in The Memo, because I love using driverless cars (driven by AI) wherever I can. They recently added Los Angeles to their service cities. ‘Following a rigorous cycle of validation and safety readiness evaluation, Waymo is starting fully-autonomous (no human driver) testing in LA. Thrilled by the data confirming, once again, how well our ML-based 5th-gen Driver generalizes across cities!’
Google Imagen referenced in new Android build (25/Feb/2023)