The Memo - 10/Oct/2023
Reka Yasa-1, DALL-E 3 text-to-image model, Gaudi2 chips, and much more!
FOR IMMEDIATE RELEASE: 10/Oct/2023
Prof Martin Seligman (14/Sep/2023):
“This is a rare moment in the history of scientific psychology: Artificial intelligence [via GPT-4] now promises much more effective psychotherapy and coaching.”
Welcome back to The Memo.
I’m calling it early. The Who Moved My Cheese? AI Awards! for October 2023. It seems that the BBC wants to have its cake and eat it too. While a few organizations like AP partnered with OpenAI to share their data to train new models, the BBC has blocked OpenAI and Common Crawl from using their datasets (as did CNN, NYT, and Reuters). However, in a brilliantly duplicitous move, BBC is exploring the use of AI anyway to help it write new articles (6/Oct/2023).
I keep coming back to my two-year-old report (Jun/2021), Integrated AI: Dataset quality vs quantity via bonum (GPT-4 and beyond), where I assumed that people would willingly provide their data to these models for everyone’s benefit:
There are also intellectual property and copyright considerations for some of the [summum bonum/ultimate good] datasets, but it is expected that these would be easily cleared by the respective authors for the good of humanity.
I am learning that, generally, human greed and selfishness are disappointingly still present, and unfortunately magnified ten-fold by large organizations. It may be that this continues until AI balances things out for us (with us) very soon.
Last week, I was privileged to speak in Kyiv—from the safety of my home in Perth—about superintelligence and IQ in artificial intelligence. This is a completely new presentation pulling together a range of different sources, some of them previously hidden.
It’s time for me to eat my own dogfood (wiki), or more accurately, to drink my own champagne (article). I’ve created a GPT-4-driven manual link bot just to support the fetching of title and submission date for some items in The Memo. Don’t worry though, I’m still hands-on, and you’ll get my biologically-generated insights for as long as it takes for us to hit publicly-available superintelligence (ASI)! You can try the same bot for fun, provided in the Toys to play with section, where we also look at a new high-quality animated view of transformers, a long podcast with the leading voice in AI speaking to some guy in Texas, a new large language model processing platform for YouTube, and much more.
The BIG Stuff
Exclusive: A new model announced every 2 days in September 2023 (Sep/2023)
GPT-4 helped me sort through my list of models for September 2023. (I added Qwen by hand afterward.) It’s a big list. By my numbers, September 2023 was the most prolific month in history for model announcements, with 17 major new models. That’s a new model about every 1.7 days!
SUTD/Independent - TinyLlama (1.1B)
TII - Falcon 180B (180B)
BAAI - FLM-101B (101B)
Adept - Persimmon-8B (8B)
Apple - UniLM (0.034B)
Microsoft - phi-1.5 (1.3B)
Singapore - NExT-GPT (7B)
IBM - MoLM (8B)
Deci - DeciLM (5.7B)
ThirdAI - BOLT2.5B (2.5B)
Baichuan - Baichuan 2 (13B)
Microsoft - Kosmos-2.5 (1.3B)
Mistral AI - Mistral 7B (7.3B)
Hessian AI/LAION - LeoLM (13B)
Meta AI - Llama 2 Long (70B)
Alibaba - Qwen (14B)
Wayve - GAIA-1 (9B)
Microsoft’s implementation of DALL-E 3 is amazing (Oct/2023)
Here are two big videos about this model, generally accepted as the new state-of-the-art, with image outputs subjectively ‘better’ than Midjourney v5.2 (Jun/2023).
Best of DALL-E 3 via Bing within the first 24 hours of release:
Movie-quality outputs:
Even more examples: https://www.aidemos.info/dalle-3-examples/
The DALL-E 3 + ChatGPT prompt leak (~733 words): https://lifearchitect.ai/alignment/#dall-e3
Try DALL-E 3 (free, Microsoft login): https://bing.com/images/create
GPT-4 beats human doctors in medical soft skills (1/Oct/2023)
In a recent study by researchers at the Icahn School of Medicine at Mount Sinai, the performance of GPT-4 was tested for ‘soft skills’ like communication and professionalism, and then compared with human doctors. GPT-4 considerably outperformed doctors in soft skills.
Read more: https://www.news-medical.net/news/20231002/GPT-4-beats-human-doctors-in-medical-soft-skills.aspx
Source paper: https://doi.org/10.1038/s41598-023-43436-9
My viz:
Download PDF of this viz: https://lifearchitect.ai/iq-testing-ai/
The Interesting Stuff
Harvard: GPT-4 outputs 40% higher quality work than BCG consultants (18/Sep/2023)
Consultants using [GPT-4] AI were significantly more productive (they completed 12.2% more tasks on average, and completed tasks 25.1% more quickly), and produced significantly higher quality results (more than 40% higher quality…)
Read the paper: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321
KPMG: 72% of US CEOs say generative AI is a top investment priority (Oct/2023)
This year’s KPMG CEO Outlook analyzed insights from more than 1,300 CEOs at large companies globally, including 400 in the United States…
72% of US CEOs say generative AI is a top investment priority despite uncertain economic conditions…
Investment in generative AI is a priority for CEOs, but they are wary of navigating ethical challenges and the lack of regulation. CEOs ranked investment in generative AI as a top priority for their organizations. And the majority said they are placing more capital investment in buying new technology (57%) than developing their workforce’s skills and capabilities (43%).
Read the online report by KPMG.
OpenAI and Jony Ive in talks to raise $1bn from SoftBank for AI hardware project (26/Sep/2023)
OpenAI is in advanced talks with former Apple designer Sir Jony Ive and SoftBank’s Masayoshi Son to launch a venture to build the “iPhone of artificial intelligence”, fuelled by more than $1bn in funding from the Japanese conglomerate.
Max Tegmark: Language models represent space and time (4/Oct/2023)
Do language models have an internal world model? A sense of time? At multiple spatiotemporal scales?
In a new paper with Max Tegmark we provide evidence that they do by finding a literal map of the world inside the activations of Llama-2! (- Twitter, 4/Oct/2023)
Read the paper: https://arxiv.org/abs/2310.02207
Forbes: Intel Gaudi2 Looked To Be A Credible Alternative To Nvidia. Until... (11/Sep/2023)
Intel's Gaudi2 significantly outperformed the Nvidia A100, and performed competitively against the Nvidia H100 in the latest AI benchmarks.
Next up, Intel will update their software with float8 and bring out the 5nm Gaudi3, AMD will introduce the MI300 by the end of the year [2023] (but will likely and unfortunately shun the MLPerf benchmarks externally), and the Google TPUv5 will hopefully be available in time for the next training runs in 3 months [Dec/2023].
Read more about Gaudi by Habana Labs (acquired by Intel in 2019): https://en.wikichip.org/wiki/habana/microarchitectures/gaudi
See the Gaudi2 card for LLMs (video):
Microsoft to unveil in-house AI chip, reducing reliance on NVIDIA (6/Oct/2023)
Microsoft plans to debut its first AI chip next month, codenamed ‘Athena,’ with the intention of reducing its reliance on NVIDIA-designed GPUs. The chip reveal will likely occur at Microsoft's Ignite conference in Seattle starting 14/Nov/2023.
Read more: https://www.maginative.com/article/microsoft-to-unveil-in-house-ai-chip-reducing-reliance-on-nvidia/
Ignite ($1,525 in person/sold out, or free online): https://ignite.microsoft.com/
OpenAI is exploring making its own AI chips [or acquiring a chipmaker] (6/Oct/2023)
OpenAI is considering manufacturing its own chips, and has evaluated a potential acquisition target as part of this exploration.
If ChatGPT queries grow to a tenth the scale of Google search, it would require roughly $48.1 billion worth of GPUs initially and about $16 billion worth of chips a year to keep operational.
Although currently owned by Intel, Habana Labs (above) would be a favourable and interesting buyout…
Read more: https://archive.md/VurzN
More GPT-5 rumors (Oct/2023)
The GPT-5 rumor mill continues to spin, with interesting and increasingly credible details about GPT-5 (under internal codenames ‘Gobi’ and then ‘Arrakis’) being finalized by OpenAI a year ago in Oct/2022. The specifics are outlined below, and I’ve even provided a PDF as a last-ditch backup!
Reka Yasa-1 multi-modal LLM for enterprise (5/Oct/2023)
Yasa-1 has the core capabilities of a typical text-based AI assistant… it also natively supports images, audio, and short video clips as inputs. Powered by a single unified model, Yasa-1 has rich understanding of the multimodal world we live in, giving it extended capabilities beyond text-only assistants.
Read announce: https://reka.ai/announcing-our-multimodal-ai-assistant/
[Reka] came out of stealth just three months ago [in Jun/2023] with $58 million in funding from DST Global Partners, Radical Ventures and multiple other angels and is competing against deep-pocketed players, including Microsoft-backed OpenAI and Amazon-backed Anthropic.
Other notable competitors of the company are Inflection AI, which has raised nearly $1.5 billion, and Adept with $415 million in the bag.
Yasa-1 is an enterprise multimodal model by Reka, a startup founded by researchers from DeepMind, Google and Meta AI. Once again, the company has chosen to guard all architecture details, as well as hiding even basic benchmarks.
The online user manual does note that Yasa-1 can be run on-premise, and that ‘we assume that you have access to a machine with at least 2 GPUs and that you have been provided credentials by Reka.’
Two NVIDIA A100s would provide 160GB RAM, which only suggests to me that Yasa-1 is somewhere between Llama 2 70B and PaLM 2 340B in terms of size. (And certainly much smaller than GPT-4 1.76T.)
Read more: https://docs.reka.ai/guides/004-running-on-premises.html
Mistral 7B: The best 7B model to date, Apache 2.0 (27/Sep/2023)
Mistral AI has announced the release of Mistral 7B, a powerful language model that outperforms Llama 2 13B on all benchmarks and Llama 1 34B on many, and comes close to CodeLlama 7B performance on code while still excelling at English tasks.
This is an excellent model to put on device; small enough to squeeze into disk and RAM easily. Expect to see it available on new applications for Android and iOS devices without needing web/cloud access.
Read more: https://mistral.ai/news/announcing-mistral-7b/
Download: https://huggingface.co/mistralai
See it on the Models Timeline.
John Carmack and Rich Sutton partner to accelerate development of Artificial General Intelligence (25/Sep/2023)
John Carmack, celebrated software engineer and founder of Keen Technologies, and Dr. Richard Sutton, Chief Scientific Advisor at the Alberta Machine Intelligence Institute (Amii) announce a partnership to bring greater focus and urgency to the creation of artificial general intelligence (AGI).
Read more: https://www.amii.ca/latest-from-amii/john-carmack-and-rich-sutton-agi/
Satya Nadella Says [Windows] Copilot Will Be as Significant as the PC (3/Oct/2023)
Microsoft CEO, Satya Nadella, says that [Windows] Copilot, a context-aware AI integrated into Microsoft applications, will be as significant as the personal computer was in the '80s, suggesting it will mark a new era in personal computing. According to Nadella, Copilot will transform our relationship with technology and will be as transformative as the PC in the '80s, the web in the '90s, mobile in the 2000s, and cloud in the 2010s.
Read more: https://jdmeier.com/satya-nadella-on-copilot/
Meta's New AI Assistant Trained on Public Facebook and Instagram Posts (29/Sep/2023)
It might sound like a truism; Meta Platforms has used public Facebook and Instagram posts to train its new AI virtual assistant. The company tells users ‘you own all of the content and information’ you post. But if you make a post public, as many do by default, it becomes available for all sorts of purposes that you can't control, including use as training data for today’s LLMs.
Cloudflare and Hugging Face Partner to Run Optimized Models on Cloudflare’s Global Network (27/Sep/2023)
Cloudflare, the leading connectivity cloud company, announced a partnership with Hugging Face, the open platform for AI builders, to make deploying open AI models more accessible and affordable to developers. They aim to enable developers to deploy AI globally, without managing infrastructure or paying for unused compute capacity.
Official site: https://ai.cloudflare.com/
GPT-4V examples (Oct/2023)
Here are some illustrative examples of GPT-4V being applied to programming use cases:
https://news.ycombinator.com/item?id=37679955
https://twitter.com/skirano/status/1706823089487491469
https://twitter.com/GabGarrett/status/1706872805214593173
https://www.reddit.com/r/singularity/comments/16v1jp5/its_mental_model_of_the_world_seems_to_be/
Adobe launches Photoshop for the web with its popular desktop AI tools (27/Sep/2023)
Adobe has launched a web version of Photoshop, which includes popular AI tools from its desktop version such as Generative Fill and Generative Expand, powered by Adobe's Firefly AI model. The web version also offers a more streamlined user experience for new users and includes collaboration features.
First AI photography award took place in Australia this week (8/Oct/2023)
The lifelike picture, titled “Twin Sisters in Love,” on Saturday won the inaugural Prompted Peculiar International AI Prize at the Ballarat International Foto Biennale, an Australian photography festival running through October 22. The competition is believed to be one of the first, if not the first, AI-art award.
For the winning image, Nordenskiöld, who lives and works in Sweden, partnered with Midjourney, an AI tool that quickly turns text phrases, or “prompts,” into hyper realistic images by scanning a massive database trained on visual art by humans. Artificial intelligence tools like Midjourney, Dall-E and Stable Diffusion continue to capture imaginations, as they let anyone generate images from text in mesmerizing and sometimes creepy and wildly absurd ways.
“None of the places, people or creatures in my prompts exist in the physical realm,” Nordenskiöld said of her winning creation in a statement. “They were conjured from the sum of human experience in our deep collective well, as seen from my dreamboat with its flickering light.”
See more via Forbes (including the other submissions).
Policy
Biden teases forthcoming executive order on AI (27/Sep/2023)
President Joe Biden announced that the White House plans to introduce an executive order dealing with artificial intelligence in the coming weeks. There were no specific details provided regarding the content of the order, but Biden emphasized the importance of responsible AI innovation, citing the potential risks if not handled properly.
Read more: https://edition.cnn.com/2023/09/27/tech/joe-biden-executive-order-artificial-intelligence/index.html
CIA builds its own artificial intelligence tool in rivalry with China (26/Sep/2023)
The Central Intelligence Agency is preparing to roll out a feature akin to OpenAI Inc.’s now-famous program that will use artificial intelligence to give analysts better access to open-source intelligence, according to agency officials. The CIA’s Open-Source Enterprise division plans to provide intelligence agencies with its AI tool soon…
The CIA didn’t say what model it will use to underpin its new tool…
The AI tool will be available across the 18-agency US intelligence community, which includes the CIA, National Security Agency, the Federal Bureau of Investigation and agencies run by branches of the military. It won’t be available to policy makers or the public.
Read more via Bloomberg: https://archive.md/Oivh3
In terms of size, I expect that this model would be around the size of TTI’s (or Abu Dhabi’s) Falcon 180B.
NSA is starting an artificial intelligence security center (28/Sep/2023)
The National Security Agency (NSA) is launching an artificial intelligence security center in response to the growing integration of AI capabilities into U.S. defense and intelligence systems. The center will focus on securing AI models from theft and sabotage, a major national security challenge, particularly in the face of emerging generative AI technologies with significant transformative potential.
Jack Clark at Anthropic (and former Policy Director at OpenAI) had a flowery quote about all this today, too (9/Oct/2023):
With the intelligence community now dedicating an office to AI, we can be guaranteed of sustained work on AI for the years to come, meaning that at least one appendage of the US government is now going to be doing very longterm thinking about AI and its impact. This will likely lead to meaningful changes in what AI capabilities the US government fields and will also both increase (some) and mitigate (some other) geopolitical tensions with regard to AI technology.
The greyworld has turned its austere and patient gaze to AI and shall now not look away.
UK quietly dismisses independent AI advisory board, alarming tech sector (28/Sep/2023)
The UK government has dismissed the independent advisory board of its Centre for Data Ethics and Innovation, leading to concerns within the tech sector. The board, which was tasked with promoting the responsible use of data and AI technologies, was first appointed in 2018 and has since provided guidance to organizations on ethical and risk-mitigating practices.
Read more: https://thenextweb.com/news/uk-dismisses-independent-ai-advisory-board-alarming-tech-sector
Toys to Play With
OpenAI CEO on Joe Rogan (6/Oct/2023)
It is with a heavy heart that I add this Joe Rogan interview of OpenAI’s CEO to The Memo. I appreciate that some may get minor benefits from listening to it.
Listen: https://ogjre.com/episode/2044-sam-altman
The Memo Link Think Bot using GPT-4 via Poe.com (1/Oct/2023)
Here’s my new bot. This thing is powered by GPT-4 + web search, and can be given a link to go and search and then use the output format for The Memo.
Poe.com tells me that the prompt should be publicly visible as part of the bot’s profile, but I don’t see it, so here is a copy of Rev B:
Use this exact output format, and always find the publication date in the article. Use sentence case for the title (except where words or acronyms need capitals). Use full linebreaks in markdown. Replace all quote marks with single smart quote marks. Do not include any ‘further reading’ or 'learn more', ignore YouTube links, just this output exactly:
<markdown bold>Title (d/MMM/yyyy)</markdown bold><full line break>
Concise summary in 1-2 sentences, or useful quote from the article.<full line break>
Read more: <insert original link only, starting with https://>
(Line break and then loop to next article)
For each of these links:
Try it here: https://poe.com/TheMemoLinkThinkBot
Transformers by the Visual Storytelling Team @ Financial Times (12/Sep/2023)
Generative AI exists because of the transformer…
The LLM is underpinned by a scientific development known as the transformer model, made by Google researchers in 2017.
“While we’ve always understood the breakthrough nature of our transformer work, several years later, we’re energised by its enduring potential across new fields, from healthcare to robotics and security, enhancing human creativity, and more,” says Slav Petrov, a senior researcher at Google, who works on building AI models, including LLMs.
This is a really high quality piece, with excellent animations. I would perhaps even recommend it over The Illustrated Transformer by Jay Alammar.
Check it out: https://ig.ft.com/generative-ai/
ACL 2023: University of Washington + Princeton University (Jul/2023)
A reader flagged some of my research and viz available in an online tutorial for UW and Princeton. The material is about retrieval-augmented language models, and there are lots of resources to dive into.
Have a read: https://acl2023-retrieval-lm.github.io/
AskYouTube.ai (2023)
This platform uses a large language model to digest YouTube videos, and then output a nice summary with timecodes. You can also directly query YouTube as an entity(!) about any question.
Here’s an example output using my recent Ukraine keynote video: https://www.askyoutube.ai/share/6522099434427a83fc5fa0ee
Here’s an example output using a more general query (‘Where can I find The Memo by Life Architect?‘):
https://www.askyoutube.ai/share/65220a9efe48bbae1f4ca06b
Try it (free up to 5 queries per day, no login): https://www.askyoutube.ai/
Flashback
In December of 2022, I put together some basic prompting examples for models like ChatGPT. That free prompt book continues to be used by universities, big four auditing firms, and people around the world.
Read it: https://lifearchitect.ai/chatgpt-prompt-book/
Next
Gemini, where are you? Still in testing! It is definitely coming soon, and my livestream already has a placeholder scheduled for 1/Nov/2023…
All my very best,
Alan
LifeArchitect.ai
So how far will synthetic data take us? Microsoft Research suggest that it could be quite far:
"Textbooks Are All You Need - https://arxiv.org/abs/2306.11644"
And now we're getting an even bigger peek into the AI black box:
"Language Models Represent Space and Time - https://arxiv.org/abs/2310.02207"
It feels somehow like after this huge initial AI wave is starting to pull back from the beach, small pepples(gems) are revealing themselves. Waiting to be discovered then refined and combined.
Tried making a website with GPT4 (pre-Vision) and it could make simple sites. But a site with for instance a centered dropdown menu proved to difficult - I could get one but not both. Gemini and GPT5 are only months (not years ;-) away so that should make things interesting in this regard.