The Memo - 30/Dec/2022
IQ testing GPT-3.5 with Raven's, OpenAI CEO/Bill Gates/Ray Kurzweil on 2023 AI, Med-PaLM, and much more!
FOR IMMEDIATE RELEASE: 30/Dec/2022
Welcome back to The Memo.
I’m based in California for the next few months. If your organization is interested in a hands-on workshop or keynote on AI and language models, you can reply to this email and we’ll see what we can cook up...
In the Toys to play with section, we look at a new ChatGPT clone with web search, a text-to-Powerpoint platform (text-davinci-003 + DALL-E 2), 24/7 AI-generated radio, entire magazines generated with GPT-3 and DALL-E 2/SD for 1 cent per article(!), and a new playground for complex maths using questions from Google PaLM Minerva.
If you were hoping for a quiet end to the year, this is not it!
The BIG Stuff
Latest GPT-3.5 model outperforming humans on real IQ tests (19/Dec/2022)
For brief context, I entered this AI field nearly three years ago while serving as Chairman for Mensa International’s Gifted Families committee, serving prodigies and gifted children through coaching via Harvard Medical School, and speaking at gifted education conferences around the world. That all ended with my May/2020 article and presentation, The New Irrelevance of Intelligence. I immediately began working with these large language models, and co-presented a session to the 2021 World Gifted Conference with Leta AI (now at Episode 65!).
Given Leta’s smarts, I had approached several of my psychologist colleagues in 2021 to run formal IQ assessments with GPT-3. It was viewed as a daunting challenge, as the standard assessments from Stanford-Binet or Wechsler rely on several subtests including kinaesthetic testing, which a large language model simply cannot undertake.
Researchers at UCLA weren’t as phased as we were, and flipped the Raven’s Progressive Matrices into numbers and letters to feed to GPT-3.5:
They found that the latest GPT-3.5 model (text-davinci-003) from 28/Nov/2022 outperformed humans in most settings.
GPT-3 [actually GPT-3.5, InstructGPT, text-davinci-003] displayed a surprisingly strong capacity for abstract pattern induction, matching or even surpassing human capabilities in most settings… achieves the analogical capacity that is often considered the core of human intelligence… it has been forced to develop mechanisms similar to those thought to underlie human analogical reasoning – despite not being explicitly trained to do so… through a radically different route than that taken by biological intelligence.
My longer article on this (26/Dec/2022): https://lifearchitect.ai/ravens/
Summary and viz of IQ + AI: https://lifearchitect.ai/iq-testing-ai/
Thread by UCLA: https://twitter.com/TaylorWWebb/status/1605395851081224193
OpenAI invests in four new companies (Dec/2022)
When we launched the OpenAI Startup Fund, we committed to invest early in a small number of startups that are pushing the boundaries of how powerful AI can positively impact the world and profoundly change people’s lives.
Descript (video editor).
Harvey AI (legal workflows).
Mem (workspace manager).
Speak (language learning).
GPT-3 can detect dementia in voice samples (22/Dec/2022)
Any successful application of large language models to health and medicine belongs in ‘big stuff’!
Researchers at Drexel used GPT-3 embeddings to analyze voice samples, and found that GPT-3 can provide early diagnosis of dementia, with 80% accuracy on a challenge dataset of speech recordings.
Google announces Med-PaLM 540B (26/Dec/2022)
Also in the medical field, Google further fine-tuned Flan-PaLM (Finetuned Language model, Pathways Language Model) on clinical data, to produce Med-PaLM.
This became an interesting collab between Google and DeepMind: While the project was mainly Google's, Dr Nenad Tomasev from DeepMind was also involved and cited in the paper. This model builds on the enormous Pathways family driven by Dr Jeff Dean. See my long paper about all the different models in this family: https://lifearchitect.ai/pathways
Here’s my take…
In 2016, ‘Johns Hopkins patient safety experts calculated that more than 250,000 deaths per year are due to medical error in the U.S… third leading cause of death.’
In the new Med-PaLM paper, the model was outperforming its nearest leading model significantly in minimising answers potentially leading to harmful outcomes (pp3):
Flan-PaLM: 29.7% harmful outcomes
Human clinician: 6.5% harmful outcomes
Med-PaLM: 5.8% harmful outcomes (a 10% improvement on human doctors)
I am not a medical doctor, but as a 2017 Fellow of the IOC at Harvard Medical School, I am reminded of the dictum, ‘First, do no harm’…
To my mind, just as self-driving cars can save thousands of lives, if post-2020 large language models are improving clinical outcomes by even a few percentage points (compared to human doctors), and that had a flow-on effect directly impacting whether a patient lives or dies, I wonder if we are on our way to having current AI with the capacity to save even more human lives…
Med-PaLM paper: https://arxiv.org/abs/2212.13138
The Interesting Stuff
ChatGPT public acceptance (28/Dec/2022)
When ChatGPT was announced on 30/Nov/2022, I started a livestream demo as soon as the notification hit my inbox. I remember googling for the term in quotes, “ChatGPT”. There were zero results, and Google actually tried to correct me to the term “chatbot”.
28 days later, on 28/Dec/2022, Google shows 77,300,000 results for the term “ChatGPT”.
77.3 million new references to a new technology within less than 30 days…
AI is now essential infrastructure (26/Dec/2022)
Very generally, the mainstream media are not good at recognizing and documenting artificial intelligence. They settle for cliche, usually with a header image of a killer robot, and can’t seem to grasp the societal implications of this natural evolution for humanity. Wired did a decent job with this short article, though:
For governments to fully deliver on the promise of AI, however, they will need to invest. Soon, a comprehensive digital infrastructure… will be as important to a country as roads, rail, and public water supply. In 2023, more and more countries will accelerate the building of such nationwide digital architectures, allowing them to deliver more AI-powered responsive services that cater to the individual and help the population at large. In 2023, bold governments will be making this move—and they will be examples to follow for the rest of the world.
OpenAI CEO on AI in 2023 (26/Dec/2022)
there will be scary moments as we move towards AGI-level systems, and significant disruptions, but the upsides can be so amazing that it's well worth overcoming the great challenges to get there.
and we’ll see great benefits all along the way—they will make ChatGPT look like a boring toy.
in particular, there are going to be significant problems with the use of OpenAI tech over time; we will do our best but will not successfully anticipate every issue.
we will run a very tight feedback loop for improvement, and try to make our mistakes while the stakes are low.
nut its impossible to get this right with no contact with reality. we are learning so much from ChatGPT; its going to get a lot better, less annoying, and more useful fast. — via Twitter, 26/Dec/2022
Bill Gates on AI in 2023 (21/Dec/2022)
…the positive impact that artificial intelligence can have. Over the next few years, people are going to see the incredible progress in AI. I’m committed to making sure that these advances benefit everyone, and not just people in rich countries—for example, by using AI to help with developing new drugs, diagnosing diseases, and supporting students and teachers…
First AI-generated speech in British parliament via ChatGPT (22/Dec/2022)
Luke Evans has become the first MP to deliver a Commons speech written entirely by Artificial Intelligence (AI) via ChatGPT. Speaking just before the House adjourned for Christmas, Evans read the 324-word speech.
The prompt was: Write a Churchillian speech on the state of the United Kingdom over the past 12 months.
The response began:
Ladies and gentlemen,
Over the past 12 months, the United Kingdom has faced many challenges and has made many important decisions. In the face of these challenges, the country has remained steadfast in its commitment to its values and principles.
Read the whole speech: https://order-order.com/2022/12/22/luke-evans-delivers-first-ai-generated-speech-in-commons/
Writing books with ChatGPT (25/Dec/2022)
I’ve been covering books written by AI (GPT-3) for more than two years. ChatGPT streamlines this process even further; you can literally give it instructions, and it will write the title, plot, and can then do the whole book by chapter.
In an interview on 25/Dec/2022, Leanne Leeds revealed one of her prompts, and the fact that ChatGPT has written the plots of six new murder mystery books for her!
The prompt is in The ChatGPT Prompt Book from a recent edition of The Memo.
Dr Ray Kurzweil gives an update on AI (Dec/2022)
– ChatGPT is a ‘sizeable advance’, but ‘not quite right’.
– Large language models (LLMs) are moving in the direction of sentience.
– LLMs carry risks, just as all tech does, including railroads.
Transcription (by OpenAI Whisper). (Note: I used OpenAI Whisper this time instead of Otter.ai, and the output was great! Whisper does punctuation, grammar, capitalization, and line breaks. Here’s the Google colab notebook if you’d like to generate your own transcripts of YouTube videos for free. It took about 5mins of processing for a 20min video.)
Midjourney Toddler Series (21/Dec/2022)
Yep, that’s Harry Potter as a toddler.
The prompt is: cute baby [character name], hyper realistic, unreal engine, studio lighting, HDR, 8k, cute toddler face expression
Take a look at the full series by Ben: https://www.instagram.com/benmornin/
AI for UI/UX design (23/Dec/2022)
50+ examples of UI design by Midjourney v4. Just beautiful!
Poly AI art asset generation library (28/Dec/2022)
You’re going to hear more about this AI-generated asset library. I can see them driving assets for Pixar and Weta and more very soon. Poly AI is Y-Combinator backed, and even though they are mainly 3D textures right now, Poly’s CEO has alluded to ‘perfect icon, illustration, font or sound effect’ generation.
Check them out: https://withpoly.com/
South Park creators secure $20M funding for AI media (22/Dec/2022)
Trey Parker and Matt Stone have been using a new-ish synthetic media generation platform for two years. There first clip was of Donald Trump, but the platform also famously generated Kendrick Lamar’s “The Heart Part 5” music video, released in May 2022, in which Lamar’s face transforms into the visages of OJ Simpson, Jussie Smollett, Nipsey Hussle, Kobe Bryant and Kanye West.