The Memo - 24/Jan/2024

DeepMind AlphaGeometry, AI political robocalls begin, Zhipu GLM-4, and much more!

Jan 23, 2024

∙ Paid

To:      US Govt, major govts, Microsoft, Apple, NVIDIA, Alphabet, Amazon, Meta, Tesla, Citi, Tencent, IBM, & 10,000+ more recipients…
From:    Dr Alan D. Thompson <LifeArchitect.ai>
Sent:    24/Jan/2024
Subject: The Memo - AI that matters, as it happens, in plain English
AGI:     64%

Ilya Sutskever, OpenAI (15/Mar/2023):
Say you read a detective novel… complicated plot, a storyline, different characters, lots of events, mysteries, clues, it’s unclear. Then, let’s say that at the last page of the book, the detective has gathered all the clues, gathered all the people and saying: ‘Okay, I’m going to reveal the identity of whoever committed the crime and that person’s name is…’ Predict that word. Now, there are many different words. But predicting those words better and better, the understanding of the text keeps on increasing. GPT-4 predicts the next word better.

Welcome back to The Memo.

You’re joining full subscribers from Alaska, Berkeley, Brown, Buffalo, Columbia, Cornell, Duke, Harvard, Hawaii, MIT, Notre Dame, Princeton, Rice, Stanford, and many more institutions.

The BIG Stuff

DeepMind AlphaGeometry (17/Jan/2024)

AlphaGeometry solving an Olympiad problem: Problem 3 of the 2015 International Mathematics Olympiad (left) and a condensed version of AlphaGeometry’s solution (right). The blue elements are added constructs. AlphaGeometry’s solution has 109 logical steps. Source

AlphaGeometry is an AI system that solves complex geometry problems at a level approaching a human Olympiad gold-medalist. It was trained using 100% synthetic data, and is completely open source.

Notably, in the International Mathematical Olympiad (IMO), AlphaGeometry ‘produces human-readable proofs, solves all geometry problems… under human expert evaluation and discovers a generalized version of a translated IMO theorem…’

DeepMind CEO Demis Hassabis said: ‘[AlphaGeometry is] Another step on the road to AGI.’ (Twitter)

The Metaculus prediction of an open-source AI winning IMO Gold Medal in Jan/2028 is much closer to being achieved, and may have already been achieved with this system. Human crowd-sourced estimates about exponential growth may be becoming irrelevant.

This system moved my conservative countdown to AGI from 64% → 65%.

Read the DeepMind blog.

Read the paper: https://www.nature.com/articles/s41586-023-06747-5

View the repo: https://github.com/google-deepmind/alphageometry

Watch the author explanation video.

Meta CEO: Llama 3 now training, Meta buying 350,000 H100s (18/Jan/2024)

Meta CEO: We're currently training our next-gen model Llama 3, and we're building massive compute infrastructure to support our future roadmap, including 350k H100s by the end of this year…

See not very much more on a noisy Instagram video post by Meta’s CEO: https://www.instagram.com/reel/C2QARHJR1sZ/

Read a summary by PCMag.

Open-sourced Llama 3 was first mentioned in The Memo edition 30/Aug/2023. We also introduced NVIDIA’s H100 Hopper chips—up to 6x faster than the A100 chips—in The Memo edition 17/Jul/2022, explored the NVIDIA/Microsoft 10,000× H100 supercomputer in The Memo edition 24/Nov/2022, and covered recent H100 purchases by major AI labs in The Memo edition 2/Dec/2023.

Zhipu AI GLM-4 (16/Jan/2024)

GLM-4 benchmarks from Zhipu DevDay slide. Click to enlarge.

Chinese AI labs continue to play catch-up with model benchmarks from OpenAI and Google. Zhipu AI, a commercialization of Tsinghua University’s output, has announced a new version of their GLM-3 model, unsurprisingly called GLM-4. They have admitted to closely following OpenAI’s roadmap.

2024 GLM-4 Analysis

10.6MB ∙ PDF file

Download

Source (Chinese): https://zhuanlan.zhihu.com/p/678179798

Download

Chinese commentators have called this the best Chinese model to date, with an MMLU score of 81.5 (GPT-4=86.4), and a context window of 128k (same as GPT-4 Turbo; GPT-4 was 8k/32k). I continue to assert that models out of China are coming at a faster pace and (in many cases) with better performance than models out of the US.

According to text on the official playground site ChatGLM.cn, the GLM-4 model is ‘hundreds of billions of parameters’. I’ve estimated this model at 200B parameters trained on 4T tokens, though it is completely multimodal and tied into a new version of CogView for image generation and GLM-4V for ‘seeing’ like OpenAI’s GPT-4V.

Read not very much more via PanDaily (English).

See it mentioned in the older official GLM-3 repo (English).

Read a related Twitter discussion (English).

See it on the Models Table.

The Interesting Stuff

Robocall with artificial Joe Biden voice tells Democrats not to vote (22/Jan/2024)

In The Memo edition 16/Dec/2023, we looked at Ashley, the world’s first AI-powered political campaign caller, using ‘over 20 different AI models, some proprietary and some open source’. That misuse of AI has now been applied in the US for the 2024 elections.

An anti-voting robocall using a suspected AI-generated and fake version of US President Joe Biden's voice directed New Hampshire Democrats to skip the Presidential Primary, triggering an investigation by the state Attorney General's office.

Listen:

AI-Biden: What a bunch of malarkey. You know the value of voting Democratic when our votes count. It's important that you save your vote for the November election. We'll need your help in electing Democrats up and down the ticket. Voting this Tuesday only enables the Republicans in their quest to elect Donald Trump again. Your vote makes a difference in November, not this Tuesday. (NBC).

The Memo by LifeArchitect.ai

The Memo - 24/Jan/2024

DeepMind AlphaGeometry, AI political robocalls begin, Zhipu GLM-4, and much more!

The BIG Stuff

The Interesting Stuff

This post is for paid subscribers