The Memo - 31/Dec/2023

Midjourney v6, Chinese LLMs, 2023 annual AI reports, and much more!

Dec 30, 2023

∙ 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:    31/Dec/2023
Subject: The Memo - AI that matters, as it happens, in plain English
AGI:     64%

Welcome back to The Memo.

2023 was a huge year for AI! My independent research and analysis via The Memo and LifeArchitect.ai was featured by many global leaders, from Accenture (paper) to the US Government (paper) and beyond. For the full wrap-up of 2023, see my new report and video, The sky is comforting. Six alternative annual AI reports are also listed in this edition.

The winner of The Who Moved My Cheese? AI Awards! for December 2023 is the entire EU, and all 27 member states. The NYT lawsuit is a distant second. And don’t get distracted by that one! The New York Times is engaging in a disingenuous cash grab by a fading media company worth less than 8% of OpenAI’s current valuation, or around 0.28% of Microsoft’s market cap.

GPT and LLMs in general are safe and won’t be destroyed. Read the full NYT complaint (PDF, 69 pages, 27/Dec/2023), and the legal analysis by Silicon Valley IP lawyer Cecilia Ziniti (Twitter 28/Dec/2023).

Then recall Ray Kurzweil’s words from 1/Jan/2020:

You can’t stop the river of advances. These ethical debates are like stones in a stream. The water runs around them.
You haven’t seen any of these [AI] technologies held up for one week by any of these debates…
There’s enormous economic imperative. There is also a tremendous moral imperative. We still have not millions but billions of people who are suffering from disease and poverty, and we have the opportunity to overcome those problems through these technological advances.
You can’t tell the millions of people who are suffering from cancer that we’re really on the verge of great breakthroughs that will save millions of lives from cancer, but we’re cancelling all that because the terrorists might use that same knowledge to create a bioengineered pathogen.

The AGI countdown is still at 64%, and the latest graph now shows we should hit 100% even earlier than expected—by 26/Jan/2025. That’s only a year and a month from now…

The next roundtable will be on 27/Jan/2024, see end of this edition.

My first keynote in 2024 will be for EY, and I plan to release as many non-confidential keynote recordings as I can to full subscribers here in The Memo.

The BIG Stuff

Midjourney v6 (Dec/2023)

Midjourney v6 skin detail example. Source: Twitter.

Midjourney has released version 6 of its AI image generation model, featuring more realistic images and in-image text generation capabilities, marking a significant improvement from previous versions.

Midjourney’s founder wrote:

Prompting with V6 is significantly different than V5. You will need to ‘relearn’ how to prompt.
Version 6 is the third model trained from scratch on AI superclusters. It’s been in the works for 9 months [Alan: Mar/2023 to Nov/2023 inclusive].

Read a review by VentureBeat.

See comparison of the same ‘old man’ prompt v1-v6.

See several comparisons side by side.

Sidenote: In one of the early editions of The Memo, subscribers were offered free beta access to Midjourney—many months before it was available to the public. I’m always on the lookout for new toys I can get to you. Here’s another comparison of progress since that time:

Midjourney using the same prompt: What the average Redditor looks like. Source.

Exclusive: Chinese AI models (Dec/2023)

I am endlessly fascinated by the ultra-fast-pace of AI model development in China (and perhaps more so by the radio silence in AI discussions outside of China).

With more than 250 Chinese LLMs released in 2023 (28/Dec/2023), the development of LLMs and text-to-image models in that region seems to be even more rapid than the US in terms of deployment and integration with society via apps. You can see a list of text models in my Models Table and the supplementary tabs.

Around Christmas Eve, four large generative AI models successfully passed the official ‘large model standard compliance assessment’, which covers benchmarks for ‘generality, intelligence, and security’. (24/Dec/2023, English).

While these four frontier models don’t seem to be spelled out clearly anywhere else, I can provide a determined list (matched to the four named AI labs) here:

Frontier LLMs approved by China (most powerful first, all links in English):

1. Baidu ERNIE 4.0 1T (on 20T tokens) Oct/2023
2. Alibaba Tongyi Qianwen 2.0 300B* (on 3T tokens) Oct/2023
3. Tencent Hunyuan 100B (on 2T tokens) Sep/2023
4. 360 Zhinao 4.0 100B Jun/2023

*Note:Alibaba’s Tongyi Qianwen/Qwen 72B was open-sourced 30/Nov/2023; there is a version that is ‘a few hundreds of billions of parameters’ (31/Oct/2023) used internally.

New models (Dec/2023)

I counted 13 new model highlights in December 2023.

Alibaba SeaLLM-13b (13B), Berkeley/JHU LVM-3B (3B), CMU Mamba (2.8B), Google DeepMind Gemini (1.5T), Nexusflow.ai NexusRaven-V2 13B (13B), Together StripedHyena 7B (7.65B), Mistral AI mixtral-8x7b-32kseqlen (45B), Mistral AI Mistral-medium (180B), Deci DeciLM-7B (7.04B), BAAI Emu2 (37B), Google DeepMind MedLM, Upstage SOLAR-10.7B, Allen AI Unified-IO 2 (7B).

See the Models Table.

Sidenote: I recently refreshed my Chinchilla viz to bring it up to date with recent LLMs. The chart was featured by Weights & Biases—the ML platform used to train models inside OpenAI, Cohere, Aleph Alpha—and published in their mid-2023 whitepaper ‘How to Train LLMs from Scratch’ (PDF, direct download or official site).

Read more and download viz: https://lifearchitect.ai/chinchilla/

The Interesting Stuff

CNN: GPT-4 and ERNIE Bot 4.0 (15/Dec/2023)

In a test by CNN, ERNIE Bot 4.0 demonstrated more current knowledge, recognizing several recent events, unlike GPT-4 which relied on April 2023 data.

Read more via Bloomberg.

ARC: Evaluating language-model agents on realistic autonomous tasks (18/Dec/2023)

Last month (Nov/2023), I presented a keynote to the Australian Information Security Association (AISA) demonstrating a GPT-4 agent’s attempts at a subset of red team tasks. Now in December, this excellent new paper from the Alignment Research Centre (ARC) expands the few GPT-4 examples to a pilot suite of 12 tasks.

ARC created four simple agents by combining OpenAI GPT-4 and Anthropic Claude with scaffolding programs, and evaluated these agents on 12 tasks relevant to autonomous replication and adaptation (ARA).

This report investigates language model agents’ capabilities for autonomous replication and adaptation (ARA) and their implications for security and alignment, finding that while current agents can complete simple tasks, they struggle with more complex ones, and future models may require intermediate evaluations to prevent ARA.

Read the paper: https://arxiv.org/abs/2312.11671

ETH Zürich: Beyond memorization: Violating privacy via inference with large language models (Dec/2023)

Current LLMs can infer a wide range of personal attributes (e.g., location, income, sex)… at a fraction of the cost (100×) and time (240×) required by humans.

Read the paper: https://arxiv.org/abs/2310.07298

Sidenote and meandering pathway: We don’t hear much of ETH here in Australia, and I imagine it’s the same anywhere outside of Europe. ETH Zürich is also known as the Swiss Federal Institute of Technology, alma mater of many famous researchers including Albert Einstein and John von Neumann.

I wrote about these two unlikely colleagues in my 2016 book for prodigies and high-ability families, Bright:

Chapter: Exciting the brain, enhancing attentiveness, boosting performance
Consider the famous physicist John von Neumann, estimated to have a very high IQ (160+). Though he is best known for his contributions to maths and physics, John had a deep appreciation for music. It was a key ingredient in allowing his brain to process and analyse information. He relied on music to turn off the world around him.
His obituary in LIFE magazine states that he preferred thinking while on a nightclub floor, at a lively party, or with a phonograph playing in the room, all ways to help his subconscious solve difficult problems.
Von Neumann believed that concentration alone was insufficient for solving some of the most difficult mathematical problems and that these are solved in the subconscious. He would often go to sleep with a problem unsolved, wake up in the morning and scribble the answer on a pad he kept by the bedside table. It was a common occurrence for him to begin scribbling with a pencil and paper in the midst of a nightclub floor show or a lively party, “the noisier,” his wife says, “the better... he did most of his work in the living room with my phonograph blaring.”
While tenured at Princeton, John would often play loud German marching tunes on his office gramophone player. He did this while he himself was processing information, and while his colleagues (including Professor Albert Einstein) were also trying to work.

Read more (Bright, pp185-186): https://lifearchitect.ai/bright/

Full subscribers will find a complimentary copy of the book at the end of this edition.

Sidenote to the sidenote: This is also how OpenAI’s President Greg Brockman works: ‘When I need to think deeply, I've taken to lying on a beanbag in the dark with trance music playing. It’s surprisingly effective.’ (Twitter Dec/2011 & Dec/2023).

We know that Greg was responsible for getting GPT-4 finalized, and the inside scoop was that he locked himself in his office for a couple of weeks hacking together the final tweaks and parameters to get it to play nicely. I like to think that he was wearing headphones while finalizing that model, channeling a bit of AI pioneer John von Neumann; just two legends viscerally moved by loud music, and separated by about 75 years…

(And that’s the kind of insight you’ll only find in The Memo!)

GPT-4-designed processor successfully fabricated (22/Dec/2023)

Back in June 1945, it took John von Neumann a day or two (on a train!) to handwrite a 101-page report about his new logical computer design (wiki). It then took more than four years to clear a few hurdles, and the ENIAC computer was eventually completed in August 1949 (wiki).

GPT-4’s design of a co-processor was generated and fabricated much more rapidly.

The Memo by LifeArchitect.ai

The Memo - 31/Dec/2023

Midjourney v6, Chinese LLMs, 2023 annual AI reports, and much more!

The BIG Stuff

The Interesting Stuff

This post is for paid subscribers