The Memo - 17/Sep/2024
o1 analysis, GameGen-O, NotebookLM demos, and much more!
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: 17/Sep/2024
Subject: The Memo - AI that matters, as it happens, in plain English
AGI: 81%Robin Sloan, author (Mar/2024)
’I’m trying hard to imagine an average day [post ASI].
I can’t even come up with a half-decent scene.
Will people live in buildings? Will they wear clothes?
My imagination is almost physically straining.
It’s easier to write the defeat than the victory, isn’t it?
Easier to write the failure than the success.
For some reason, the success seems like it might be … boring.
…Turns out, no, it’s not boring at all. Plot gallops on, even at the outer limits of matter and energy. Even at the far reaches of freedom, the stories are only just beginning.’
Contents
The BIG Stuff (o1 analysis, humanoid robots value, Last Exam, OpenAI profit…)
The Interesting Stuff (Neuralink update from Noland, AlphaProteo, Tao, Oracle…)
Policy (AI in judgment, TSMC US, China H100s, datacenter=critical infrastructure…)
Toys to Play With (AI quiz, clone webpage via Claude, NotebookLM + demos…)
Flashback (Google DeepMind Gemini…)
Next (Roundtable…)
The Memo continues to be a primary source of AI analysis for all humanity, including major government and enterprise. The Memo features in foundational papers across organizations like Accenture, Brookings, NYU, Weights & Biases, and more (all paper links at the end of this edition).
Apple referenced The Memo for its new model paper and visualizations this month, leaning on analysis of model sizes and parameter counts provided in this publication.
Thanks to ‘LiliTheAdventurer’ for this AI-generated track via the new Suno v3.5, ‘Gone’. It’s challenging for me to remember that this has nothing to do with 1980s MIDI control signals, but actually a complete studio-performance-style song conceptualized from scratch. With vocals in Japanese…
The BIG Stuff
Exclusive: o1: Smarter than we think (17/Sep/2024)
Building on the special edition from last week (12/Sep/2024), I’ve released a new page highlighting the major advancements of OpenAI o1. The comprehensive analysis includes size estimates for the o1 model.
Ready for new discoveries and new inventions
Reasoning via reinforcement learning and increased test-time compute
Hidden CoT (chain of thought)
Increased self-awareness
The most dangerous model ever released (so far...)
Pricing
Observations
Conclusions
Solar system model (HTML animation)
Coding competitions (Codeforces)
Mathematics (Prof Terry Tao)
Read more: https://lifearchitect.ai/o1
How ARK is thinking about humanoid robotics (10/Sep/2024)
As Elon Musk said during Tesla’s first-quarter earnings call, “If you've got a sentient humanoid robot that is able to navigate reality and do tasks at request, there is no meaningful limit to the size of the economy.” Musk was sharing a vision that we believe is turbocharging the robotics industry, creating new robotics companies and increasing venture investments focused on its promise.
…if humanoid robots are able to operate at scale, they could generate ~US$24 trillion in revenues, split roughly equally between household and manufacturing robotics, according to ARK’s research…
At a cost of US$16,000, for example, a humanoid robot would have to deliver little more than a 5% gain in productivity relative to its human counterpart to become economically viable.
Read more via ARK Invest.
It’s worth remembering that robots can work 24×7, probably without air conditioning, and in the dark…
I still use ARK’s Jun/2022 AI figures in my keynotes:
1X NEO: I Lived With a Humanoid Robot for 48 Hours by S3 (8/Sep/2024)
In last week's S3 1X shared their plans to quickly deploy Neo, their new robot, in homes. So I asked if they'd put Neo in my home.
This is worth watching in full, and added an ‘info’ note to my AGI countdown due to how close it comes to fulfilling Woz’s AGI definition (strange house, make a cup of coffee).
Watch the video: https://youtu.be/Sb6LMPXRdVc
Reuters: AI experts ready ‘Humanity’s Last Exam’ to stump powerful tech (17/Sep/2024)
Scale AI has launched a global initiative called ‘Humanity’s Last Exam’ to challenge AI systems with difficult questions that extend beyond current benchmark tests.
"[o1] destroyed the most popular reasoning benchmarks… They're now crushed."
— Dr Dan Hendrycks, creator of MMLU and MATH (Reuters, 17/Sep/2024)
This project, organized by the Center for AI Safety and Scale AI, aims to evaluate AI's true capabilities, particularly in abstract reasoning, by crowdsourcing over 1,000 challenging questions.
Read more via Reuters, and source via Scale AI.
You can submit a question to win US$5,000. Scale AI suggests that items should only be designed by PhD experts(!) and:
Questions should not be simple trick questions. Questions should not be straightforward calculation/computation questions. The questions for this dataset can be so challenging that they are impractical for real-world human exams or problem sets. The answers should not be easily found through a Google search.
It should impress you if an AI correctly answers your question. We found questions written by undergraduates tend to be too easy for the models. As a rule of thumb, if a randomly selected undergraduate can understand what is being asked, it is likely too easy for the frontier LLMs of today and tomorrow.
Submit an item: https://agi.safe.ai/submit
Sam Altman told OpenAI staff the company’s non-profit corporate structure will change next year (13/Sep/2024)
Sam Altman, CEO of OpenAI, informed staff that the company will alter its non-profit corporate structure next year. The change is driven by the company's growth, which has outgrown its current structure where a nonprofit controls a for-profit entity. This shift towards a more traditional for-profit model aims to align OpenAI with other tech giants and provide a clearer return for investors.
Read more via Fortune.
The Interesting Stuff
OpenAI fundraising set to vault startup’s value to $150 billion (11/Sep/2024)
OpenAI is reportedly in discussions to secure US$6.5 billion from investors, propelling its valuation to US$150 billion. This new valuation marks a substantial increase from the US$86 billion valuation during the company's tender offer earlier this year, reinforcing its status as one of the world's most valuable startups.
Read more via Bloomberg.
People think it's like the Matrix: Neuralink's first patient on having a brain chip (14/Sep/2024)
Noland Arbaugh, the first recipient of Neuralink's brain implant, shares his experience of enhanced interaction with computers and learning languages post-surgery. The implant allows him to control a computer cursor using his mind, significantly improving his ability to play chess and learn languages like French and Japanese. Despite a setback with the implant's threads detaching, Arbaugh remains optimistic about the future potential of brain-computer interfaces to aid those with paralysis.
Read more via Euronews.
Prof Terry Tao: Experiments with OpenAI's o1 (14/Sep/2024)
Terence Tao conducted experiments where o1 successfully identified Cramer's theorem for a mathematical query and worked on complex analysis problems, albeit with some guidance. He also tested the model's ability to formalize results in Lean, finding it promising but limited by outdated training data.
Read Terry’s explanation: https://mathstodon.xyz/@tao/113132502735585408
See all three of Terry’s experiments in ChatGPT (no login; you may have to open these in an incognito browser window due to OpenAI's login handler):
ChatGPT Experiment 1, ChatGPT Experiment 2, ChatGPT Experiment 3
DeepMind AlphaProteo generates novel proteins for biology and health research (5/Sep/2024)
DeepMind has introduced AlphaProteo, an AI system designed to generate novel protein binders, accelerating advancements in drug design and biological research. AlphaProteo successfully creates high-affinity protein binders for various targets, including those associated with cancer and viral infections, demonstrating superior binding strengths compared to traditional methods. This breakthrough could significantly reduce experimental timelines in protein research, though challenges remain in developing binders for complex targets like TNFɑ.
Read more via Google DeepMind.
Read the paper: https://www.arxiv.org/abs/2409.08022
Oracle to deploy a supercluster of ~130,000 NVIDIA Blackwell GPUs, alludes to a ‘gigawatt’ capacity data center that will be powered by 3 nuclear reactors (11/Sep/2024)
Oracle has announced plans to deploy a supercluster featuring approximately 131,072 NVIDIA Blackwell GPUs, a move that underscores the company's commitment to advancing AI capabilities. This groundbreaking infrastructure aims to deliver 2.4 zettaflops of peak AI compute, aiding the training and deployment of next-gen AI models. Additionally, Oracle plans to construct a data center powered by three small-scale nuclear reactors, highlighting its ambitious expansion in response to the surging demand for AI and cloud services.
Read more via Wccftech.
Roblox is launching a generative AI that builds 3D environments in a snap (6/Sep/2024)
Roblox is set to introduce a generative AI tool that allows creators to build entire 3D scenes using simple text prompts. This innovation means developers can quickly generate environments like a desert race track by typing a description, significantly reducing the time and skill needed for 3D design. The tool works by ‘tokenizing’ 3D blocks in a manner similar to how large language models process words, making it easier for developers with minimal 3D art skills to create engaging game spaces.
Read more via MIT Technology Review.
Apple Intelligence notes (10/Sep/2024)
The latest iPhone 16 Pro introduces several advanced features that leverage AI for enhanced functionality:
Photos/Album Search: Now includes video understanding, allowing users to scroll to the exact time of described moments, which has shown promising results in initial examples.
Mail and Notifications: These will now display summaries, improving readability and efficiency over previous truncated formats.
Siri Enhancements: Siri now acts as an extensive manual for iOS 18, able to read texts and suggest actions through Personal Context Understanding. It may also recommend Apple TV shows based on user preferences.
AI-Powered Camera Control: The new camera control button utilizes AI to prioritize Apple's visual search, enhancing user interaction by positioning it ahead of OpenAI/ChatGPT and Google search.
Automatic Camera Event Integration: The camera can automatically add detected events to the calendar, streamlining user scheduling processes.
On-Device and Cloud Processing: Features are executed both on-device and via cloud services, ensuring seamless and efficient operation, though this dual capability is not explicitly highlighted.
Source: https://x.com/swyx/status/1833234781221622022
Video source (5mins): https://x.com/swyx/status/1833231875537850659
The real power of Apple Intelligence will show up in third-party apps (10/Sep/2024)
Apple's new AI capabilities in iOS 18, known as Apple Intelligence, are set to enhance third-party apps by allowing deeper integration with Siri and other smart features. This will enable users to execute tasks with simple voice commands, such as asking Siri to perform specific actions within apps without navigating through menus. Additionally, developers can leverage these features to improve app functionality and user engagement, potentially transforming how apps are used and accessed in the Apple ecosystem.
Read more via TechCrunch.
Introducing Chai-1: Decoding the molecular interactions of life (9/Sep/2024)
Chai Discovery has unveiled Chai-1, a cutting-edge multi-modal foundation model for predicting molecular structures, pivotal in drug discovery. Chai-1 surpasses existing benchmarks, achieving a 77% success rate on PoseBusters and outperforming models like AlphaFold3 and ESM3-98B. It uniquely predicts multimer structures using single sequences without MSAs, marking a significant advancement in biomolecular interaction modeling. Chai-1 is available for free via a web interface for commercial use, with its model weights and inference code released for non-commercial purposes.
Read more via Chai Discovery.
A day in the life of the world’s fastest supercomputer (4/Sep/2024)
Here’s an in-depth look at Frontier, the world's fastest supercomputer, located in Oak Ridge, Tennessee, USA. Frontier is equipped with nearly 50,000 processors and uses an impressive amount of power, equivalent to that needed for 10,000 homes. It excels in executing complex simulations for scientific research, including climate modeling and the development of AI models. Frontier's role is likened to that of a scientific instrument, providing researchers with unparalleled computational resources to advance a wide range of scientific fields.
Read more via Nature.
GameGen-O: Open-world video game generation (13/Sep/2024)
It was only a couple of weeks ago that we discussed Google’s GameNGen (the first game engine fully powered by a neural model) in The Memo edition 5/Sep/2024.
Now China has introduced a diffusion transformer model designed specifically for generating open-world video games. Called ‘GameGen-O,’ it simulates various game engine features, offering high-quality, open-domain generation with interactive controllability. The model is supported by the comprehensive OGameData dataset, which was meticulously curated to enhance the training process and enable innovative character and environment generation, as well as dynamic actions and events.
See the repo: https://gamegen-o.github.io/
This GameGen-O output demo video reminded me of the original Tenchu games (updated link):
AI version of Michael Jackson’s History Tour in Munich 1997 (Jun/2024)
In a previous life, I was a live sound engineer for some of the biggest acts around (at various times, I was system engineer for Simply Red, Korn, k.d. lang, Buzzcocks, Sinéad O’Connor, and several symphony orchestras). I also worked with Michael Jackson’s live engineer, James ‘Trip’ Khalaf (Queen, Fleetwood Mac, Steely Dan, Elton John, Madonna, Roger Waters, Kiss, Mariah Carey).
I wonder what Trip would have to say about the use of AI across some of his old live concerts. This one is more than a quarter century old (1997), and brings AI across Jackson’s vocals, and upscales the entire band’s audio quality to 2024 levels.
It brings me back to the questions at the end of my 2023 AI retrospective, The sky is comforting:
Rather than try to spell out this unlimited evolution, the conclusion of this report is a set of big questions for your consideration. Even if you don’t come up with answers right now, they’ll be answered for you (and society) very soon anyway. So I encourage you to set aside some time, take a deep breath, and allow your responses to bubble up…
Post-2020 AI currently has the ability to amplify and augment your output by about 2×, and this will increase to 1,000× soon. What does this look like for you?
Given that meaning and accomplishment are vital parts of human flourishing, what are you choosing to do with your life in a world without work?
Similarly, in a world without money or human labor, and with all of your survival and comfort needs met for ‘free,’ what will you choose as your primary motivation?
It’s possible that you don’t prefer this particular AI version of Jackson. But you also have unlimited control over what you’d tweak to bring the King of Pop’s magic into visceral reality. Do you want it louder? In a different style? Do you want to experience being at that concert in all five senses? The early live recordings don’t do him justice, and the application of AI to bring the experience into the present is just one tiny—but tangible—way that AI amplifies our lives.
Watch and listen to Jackson live at Munich via AI: https://youtu.be/ho3NYUKSYFU
Compare with the non-AI version here: https://youtu.be/jgyTyJiCYlM
Policy
US judge runs mini-experiment with AI to help decide case (6/Sep/2024)
US Circuit Judge Kevin Newsom used AI models to assist in interpreting the legal term ‘physically restrained’ during the appeal of a man's 11-year prison sentence.
But he said because there was no dictionary definition for "physically restrained" as a combined phrase, he "couldn't help himself" and asked OpenAI's ChatGPT and two other generative AI programs what the phrase's ordinary meaning is.
Newsom said his "humble little mini-experiment" revealed the programs produced answers similar to an initial one ChatGPT generated, which said the phrase "refers to the act of limiting or preventing someone’s movement by using physical force or some kind of device."…
He concluded ultimately that the slight variations "accurately reflects real people's everyday speech patterns," which he said showed the AI programs' models could accurately predict the ordinary meaning of words.
As a result, he said he increasingly believed such programs "may well serve a valuable auxiliary role as we aim to triangulate ordinary meaning."
Despite initial concerns over slight variances in AI-generated answers, Newsom found that AI could be a valuable tool in understanding ordinary meanings of words. This case, United States v. Deleon, highlights the potential of AI to serve as a supplementary aid in legal interpretations, although human oversight remains crucial.
Read more via Reuters.
This is an interesting development in the field of judicial practice, so I’m going to spell out some of the instances of judgments assisted by post-2020 AI from previous editions of The Memo:
The Memo 5/Feb/2023: ChatGPT helps a judge with a verdict (31/Jan/2023)
On January 31, the first labor court of Cartagena, Colombia, resolved a guardianship action with the help of the famous artificial intelligence known as ChatGPT, arguing that it applied Law 2213 of 2022, which says that in certain cases these virtual tools can be used.
English: https://interestingengineering.com/innovation/chatgpt-makes-humane-decision-columbia
Spanish source: https://www.bluradio.com/judicial/sentencia-la-tome-yo-chatgpt-respaldo-argumentacion-juez-de-cartagena-uso-inteligencia-artificial-pr30
The Memo 3/Apr/2023: Judge asks GPT to decide bail in murder trial (29/Mar/2023)
You may be experiencing déjà vu, as a judge also used ChatGPT for a ruling in Colombia back in Jan/2023. This one is from India.
Prompt (human): What is the jurisprudence on bail when the assailant assaulted with cruelty?
Response (GPT): If the assailants have been charged with a violent crime that involves cruelty, they may be considered a danger to the community and a flight risk. In such cases, the judge may be less inclined to grant bail or may set the bail amount very high to ensure that the defendant appears in court and does not pose a risk to public safety. However, in general, if the assailants have been charged with a violent crime that involves cruelty, such as murder, aggravated assault, or torture, they may be considered a danger to the community and a flight risk.
The revolutionary decision occurred at the Punjab and Haryana High Court during the trial of Jaswinder Singh, who was arrested in 2020 for allegedly assaulting and killing an individual, Bar and Bench reported. He was reportedly seeking a stay of imprisonment during the trial. Left at a loss for how to handle the situation, judge Anoop Chitkara decided to seek counsel from GPT-4… Armed with [GPT’s] legal expertise, Chitkara ultimately rejected the defendant’s bail bid on the grounds that they did act cruelly before the victim died.
See my full list of ChatGPT capabilities (Google Sheet).
The Memo 26/Sep/2023: UK Court of appeal judge praises ‘jolly useful’ ChatGPT after asking it for legal summary (15/Sep/2023)
“I’m taking full personal responsibility for what I put in my judgment, I am not trying to give the responsibility to somebody else. All it did was a task which I was about to do and which I knew the answer and could recognise as being acceptable.”
This is the first known use of ChatGPT by a British judge to write part of a judgment.
TSMC's $65 billion Arizona facility can now match Taiwan production yields according to early trials (7/Sep/2024)
TSMC's new Arizona facility has reached a significant milestone by achieving production yields comparable to its plants in Taiwan during early trials. This development is crucial for TSMC's ambitious US project, which involves a massive US$65 billion investment in three plants. The first fab [fabrication plant] is set to commence mass production next year, with subsequent fabs planned to utilize advanced 2nm and 3nm process technologies by the end of the decade. The project has faced challenges, including workforce shortages and cultural differences, but recent successes in trial production have alleviated some concerns.
Read more via TechSpot.
UK datacenters to enjoy special protections with CNI status (12/Sep/2024)
The UK government has designated datacenters as critical national infrastructure (CNI), offering them special protection to prevent disruptions from cyberattacks, IT outages, and extreme weather. This support includes a dedicated CNI data infrastructure team to monitor threats and coordinate emergency responses. The decision is expected to bolster confidence in the sector and attract investments, such as the proposed £3.75B investment in Europe's largest datacenter in Hertfordshire. This move aims to ensure datacenters' resilience, recognizing their vital role in the digital economy.
Read more via The Register.
Read the press release: https://www.gov.uk/government/news/data-centres-to-be-given-massive-boost-and-protections-from-cyber-criminals-and-it-blackouts
Readout of White House roundtable on US leadership in AI infrastructure (12/Sep/2024)
The Biden-Harris Administration convened a roundtable with leaders from AI companies, datacenter operators, and utility companies to discuss maintaining US leadership in AI infrastructure. The meeting focused on strategies for clean energy and workforce requirements needed for AI datacenter development. The White House announced the creation of a Task Force on AI Datacenter Infrastructure to streamline policy coordination and ensure AI systems are safe, secure, and trustworthy, while fostering job creation and economic growth.
Read more via The White House.
Read more via CNBC.
NVIDIA’s AI chips are cheaper to rent in China than US (6/Sep/2024)
Despite US export restrictions, NVIDIA's AI chips, including the A100 and H100, are more affordable to rent in China than in the US. Chinese cloud providers offer these chips at lower prices due to a plentiful supply, facilitated by a black market that circumvents US restrictions. This situation underscores the challenges of enforcing export controls and the competitive dynamics within the global AI technology market.
He estimated there were more than 100,000 NVIDIA H100 processors in the country based on their widespread availability in the market. The NVIDIA chips are each roughly the size of a book, making them relatively easy for smugglers to ferry across borders, undermining Washington’s efforts to limit China’s AI progress.
“We bought our H100s from a company that smuggled them in from Japan,” said a start-up founder in the automation field who paid about RMB500,000 (US$70,000) for two cards this year. “They etched off the serial numbers.”
NVIDIA said it sold its processors “primarily to well-known partners… who work with us to ensure that all sales comply with US export control rules”.
Read more via Financial Times.
Oprah just had an AI special with Sam Altman and Bill Gates — here are the highlights (12/Sep/2024)
Following on from The Memo edition 5/Sep/2024, Oprah Winfrey hosted a special titled ‘AI and the Future of Us’, featuring guests like OpenAI CEO Sam Altman and Microsoft founder Bill Gates. The discussion focused on AI's potential and challenges, with Oprah emphasizing the need to adapt to the technology's rapid advancement. Altman discussed AI safety testing and the importance of regulation, while Gates highlighted AI's potential in education and healthcare. However, concerns about AI biases and misinformation, especially with deepfakes, were also addressed, underscoring the need for cautious optimism.
Read more via TechCrunch.
Watch a brief clip with OpenAI CEO and Oprah.
NIST draft guidance inherently hostile to open-source AI models (10/Sep/2024)
TechFreedom has criticized the National Institute of Standards and Technology's (NIST) draft guidelines on AI safety and security, asserting they are unfavorably biased against open-source AI models. The guidelines are perceived to impose stricter restrictions on open-source models compared to closed-source systems, potentially hindering their development. TechFreedom suggests NIST adopt a marginal risk and benefit analysis, similar to the approach used by the National Telecommunications and Information Administration (NTIA), to ensure balanced and fair guidance for open-source AI development.
Read more via TechFreedom.
Australian government acts to make AI safer (5/Sep/2024)
The Australian government, under Prime Minister Anthony Albanese, announced new measures to ensure the safe and ethical use of artificial intelligence. The framework is aligned with ISO 42001:2023.
Read the announce.
Read document 1: Safe and responsible AI in Australia (69 pages, PDF).
Read document 2: Voluntary AI Safety Standard (also 69 pages, PDF).
I appeared on Australian media for this one, original source or watch:
This is another comprehensive edition of The Memo, coming in at nearly 5,000 words. And that’s after filtering out ~95% of the AI ‘news’ that comes across my desk. Let’s jump in…