The Memo - 11/Apr/2023

Anthropic's Claude-Next, OpenAI to capture $100T of world's wealth, the Stanford report, EleutherAI's Pythia 12B, and much more!

Apr 10, 2023

∙ Paid

FOR IMMEDIATE RELEASE: 11/Apr/2023

Welcome back to The Memo.

AI is advancing much more quickly than any previous technology or invention. It is now nearly a certainty that artificial general intelligence (AGI)—and the runaway of the singularity (wiki)—will occur long before governmental or intergovernmental organizations have put effective policy in place.

As with nearly all major new human inventions post-GPT-4, AI should be directly engaged to solve this puzzle. Applying limited human minds to ‘super wicked problems’ (wiki) like global climate change, economic overhauls, AI policy, and even AI alignment, is both inefficient and ineffective. Instead, these conundrums are best given to AI itself.

In this edition, the Policy section focuses on the OpenAI/CIA link, AI science versus policy time lag, and new AI guidance out of Washington as the US Government heats up. We also feature an exclusive look at my confidential draft video on AI alignment.

We’ve had a couple of queries about searching The Memo archives for keywords, and you can do that in the platform here:
Search The Memo archives.

In the Toys to play with section, we look at a new (free) interface for ChatGPT with voice-in and voice-out, a clever Midjourney v5 prompt set, a new chatbot for the M1 Macbook, and much more…

We’re also piloting a private discussion channel for paid subscribers, details at the end of this edition.

The BIG Stuff

New viz: AI Race 2023! (7/Apr/2023)

Sources:

BritGPT: ‘“Because AI needs computing horsepower, I today commit around £900m of funding… for an exascale supercomputer,” said the chancellor, Jeremy Hunt.’
Anthropic Claude-Next (see below): ‘we have been convinced of the necessity of commercialization, which we fully committed to in September [2022]…We’ve developed a strategy for go-to-market/’
Meta AI: ‘the company will commercialize its generative artificial intelligence by December [2023].’
Google/DeepMind Gemini: ‘Gemini’s goal is to develop a large language model (a computer program that can understand and generate human-like language) and will use up to 1 trillion parameters.’
OpenAI GPT-5: ‘[rumor only] GPT-5 is scheduled to complete training this December [2023] and that OpenAI expects it to achieve AGI.‘

It is possible that NVIDIA, Cerebras, and Tesla AI could also be in this list, though I haven’t seen any hard evidence that any of them are aiming for 2T+ parameters.

Watch the video:

Anthropic training Claude-Next (7/Apr/2023)

Anthropic says that it plans to build a “frontier model” — tentatively called “Claude-Next” — 10 times more capable than today’s most powerful AI [Alan: 5-10T parameters?], but that this will require a billion dollars in spending over the next 18 months.
Anthropic describes the frontier model as a “next-gen algorithm for AI self-teaching,” making reference to an AI training technique it developed called “constitutional AI.” At a high level, constitutional AI seeks to provide a way to align AI with human intentions — letting systems respond to questions and perform tasks using a simple set of guiding principles.
Anthropic estimates its frontier model will require on the order of 10^25 FLOPs, or floating point operations — several orders of magnitude larger than even the biggest models today. Of course, how this translates to computation time depends on the speed and scale of the system doing the computation; Anthropic implies (in the deck) it relies on clusters with “tens of thousands of GPUs.” [Alan: Anthropic uses Google Cloud exclusively (press release 3/Feb/2023), which uses proprietary TPU v4 chips.]…
“These models could begin to automate large portions of the economy,” the pitch deck reads. “We believe that companies that train the best 2025/26 models will be too far ahead for anyone to catch up in subsequent cycles.”

The Interesting Stuff

Microsoft researcher lectures on GPT-4 and AGI (7/Apr/2023)

Sébastien Bubeck leads the Machine Learning Foundations group at Microsoft Research. He authored Microsoft’s recent paper about GPT-4, originally titled ‘First Contact With an AGI System’ (The Memo edition 24/Mar/2023). This video is worth watching.

Watch (48m32s):

Stanford’s 2023 AI report for 2022 (4/Apr/2023)

Top takeaways:

Industry races ahead of academia.

Until 2014, most significant machine learning models were released by academia. Since then, industry has taken over. In 2022, there were 32 significant industry-produced machine learning models compared to just three produced by academia. Building state-of-the-art AI systems increasingly requires large amounts of data, compute, and money, resources that industry actors inherently possess in greater amounts compared to nonprofits and academia.

Performance saturation on traditional benchmarks.

AI continued to post state-of-the-art results, but year-over-year improvement on many benchmarks continues to be marginal. Moreover, the speed at which benchmark saturation is being reached is increasing. However, new, more comprehensive benchmarking suites such as BIG-bench and HELM are being released.

The world’s best new scientist … AI?

AI models are starting to rapidly accelerate scientific progress and in 2022 were used to aid hydrogen fusion, improve the efficiency of matrix manipulation, and generate new antibodies.

The demand for AI-related professional skills is increasing across virtually every American industrial sector.

Across every sector in the United States for which there is data (with the exception of agriculture, forestry, fishery and hunting), the number of AI-related job postings has increased on average from 1.7% in 2021 to 1.9% in 2022. Employers in the United States are increasingly looking for workers with AI-related skills.

While the proportion of companies adopting AI has plateaued, the companies that have adopted AI continue to pull ahead.

The proportion of companies adopting AI in 2022 has more than doubled since 2017, though it has plateaued in recent years between 50% and 60%, according to the results of McKinsey’s annual research survey. Organizations that have adopted AI report realizing meaningful cost decreases and revenue increases.

Download full report (386 pages, 25.3MB).

AI Achievements Unlocked: Emergent abilities of LLMs (6/Apr/2023)

Unpredictable abilities that have been observed in large language models but that were not present in simpler models (and that were not explicitly designed into the model) are usually called “emergent abilities”. Researchers note that such abilities “cannot be predicted simply by extrapolating the performance of smaller models”.
(- wiki)

Watch my video:

EleutherAI releases Pythia 12B (6/Apr/2023)

How do large language models (LLMs) develop and evolve over the course of training? How do these patterns change as models scale? To answer these questions, we introduce Pythia, a suite of 16 LLMs all trained on public data seen in the exact same order and ranging in size from 70M to 12B parameters [Alan: Chinchilla-aligned with 300B tokens trained]. We provide public access to 154 checkpoints for each one of the 16 models, alongside tools to download and reconstruct their exact training dataloaders for further study.

Read the announce by EleutherAI.

Repo: https://github.com/EleutherAI/pythia

Demo: https://huggingface.co/EleutherAI/pythia-12b

Midjourney /describe (5/Apr/2023)

Text-to-image models take a prompt written in plain English and then generate a new image. Midjourney’s new /describe feature works in reverse: upload an image, and it gives you a prompt! Great for reverse-engineering images, and generating new iterations of existing images. Here’s an example; you can see the the picture uploaded of a train below, and the mode figuring out four examples of what the prompt may have been.

Policy

OpenAI: Our approach to AI safety (5/Apr/2023)

Building increasingly safe AI systems.
Learning from real-world use to improve safeguards.
Protecting children.
Respecting privacy.
Improving factual accuracy.

US Govt on AI policy (5/Apr/2023)

President Biden has mentioned AI (one of only a handful of times in history) in a tweet:

‘Artificial Intelligence has enormous potential to tackle some of our toughest challenges. But we must address its risks. That's why last year, we proposed an AI Bill of Rights to ensure that important protections for the American people are built into AI systems from the start.’ - via Twitter

President Biden @POTUS

Artificial Intelligence has enormous potential to tackle some of our toughest challenges. But we must address its risks. That's why last year, we proposed an AI Bill of Rights to ensure that important protections for the American people are built into AI systems from the start.

Read the US AI Bill of Rights: https://www.whitehouse.gov/ostp/ai-bill-of-rights/

Download the report PDF (73 pages).

Read a summary of the Bill by the World Economic Forum.

Washington vows to tackle AI (8/Apr/2023)

The Washington Post reported on 8/Apr/2023:

President Biden on Tuesday held a meeting on the risks and opportunities of artificial intelligence, where he heard from a variety of experts on the Council of Advisors on Science and Technology, including Microsoft and Google executives…
The Justice Department’s top antitrust enforcer, Jonathan Kanter, said at South by Southwest last month that his office had launched an initiative called “Project Gretzky” to stay ahead of the curve on competition issues in artificial intelligence markets. The project’s name is a reference to hockey star Wayne Gretzky’s famous quote about skating to “where the puck is going.”…
Lieu is working on legislation to build a government commission to assess artificial intelligence risks and create a federal agency that would oversee the technology, similar to how the Food and Drug Administration reviews drugs coming to market…
He warned that Congress alone is not equipped to move quickly enough to develop laws regulating artificial intelligence…
Harris and Raskin compared the current moment to the advent of nuclear weapons in 1944, and Harris called on policymakers to consider extreme steps to slow the rollout of AI, including an executive order…

OpenAI’s strong ties to the US Government (Apr/2023)

As previously reported in The Memo, OpenAI’s CEO visited Washington several times in Jan/2023:

In the meetings, Altman told policymakers that OpenAI is on the path to creating “artificial general intelligence,” a term used to describe an artificial intelligence that can think and understand on the level of the human brain.
New technologies are usually around for years before they enter the radar of D.C. lawmakers. The fact that ChatGPT is already a topic of conversation in the halls of Congress just a few months after it was made available to the public, speaks to its potential for massive impact and disruption…
But there’s another reason Altman is likely in D.C. this week: Unlike most technology companies of the past, OpenAI began focusing on its potential impacts on society long before it ever launched a consumer product…
It makes sense that Altman would be on a tour of Washington right now. He is likely less concerned about competitors taking pot shots at his company, and more worried that the technology’s rapidly-advancing progress may require swift action on the part of regulators.
If that happens, the more information regulators have now, the better. - via Semafor.

Regular contact: On 17/Mar/2023, OpenAI’s CEO told ABC that he is in ‘regular contact’ with government officials.

CIA on OpenAI board: Around two years ago on 3/May/2021, former CIA operative and Congressman Will Hurd joined the board of directors for OpenAI. Will served three terms in the U.S. House of Representatives, has been a leading voice on technology policy, and coauthored bipartisan legislation outlining a national strategy for artificial intelligence. On 17/Dec/2020, Will told NPR:

I left the job at the CIA because I thought I could help the intelligence community a different way by running for Congress… the opportunity to continue working on these - what I consider generation-defining challenges in other ways is exciting… I've gotten more legislation signed into law in six years than most people do in a couple of decades - I'm pretty proud of my legislative accomplishments.

Alan’s take on AGI + Policy lag (Apr/2023)

Well, the date above says April 2023, but I started talking about time lags a little while ago, and documented it in my 2022 report Roadmap: AI’s next big steps in the world.

In that report, I referenced a 2015 chart by the United Nations, showing a massive historical gap between scientific advice and policy action.

Chart: Time lags between science and policy (in years). Expert group meeting for the Global Sustainable Development Report: A systematic approach to science and technology issues for the attention of policy makers, Geneva, 28-29 May 2015 https://sdgs.un.org/sites/default/files/documents/9587Presentation%2520RAR%2520Geneva%252028-29%2520May.pdf

In essence, the chart says that major scientific discoveries take decades to make their way through policy, and to finally affect humanity. One example is the health risks associated with tobacco, giving a fairly rapid scientific consensus that it was not a good thing, followed by 40 years of delayed national policy action and then a further 10 years of global policy action before the harm reduction was put in place. More recently in the public spotlight, climate change features a 100-year gap between early warning and the global policy action happening right now in 2022.

Here’s my take…

As with nearly all major new human inventions post-GPT-4, AI should be directly engaged to solve this puzzle. Applying limited human minds to ‘super wicked problems’ (wiki) like global climate change, economic overhauls, AI policy, and even AI alignment, is both inefficient and ineffective. Instead, these conundrums are best given to AI itself.

Toys to Play With

The AI Alignment Problem (9/Apr/2023)

The Memo by LifeArchitect.ai