The Memo - 14/Nov/2024
Alan’s ASI checklist, OpenAI Operator and more agents, FrontierMath, 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: 14/Nov/2024
Subject: The Memo - AI that matters, as it happens, in plain English
AGI: 83%
ASI: 0/50 (no expected movement until post-AGI)Contents
The BIG Stuff (Alan’s ASI checklist, Visa 500 LLM use cases, OpenAI Operator…)
The Interesting Stuff (Kling 1.5, Dario, Hollywood, Google Learn About, Osmo…)
Policy (New US administration, TSMC law, China analysis, Nusantra…)
Toys to Play With (OpenAI students, Udio album, Ollama, Gmail agent…)
Flashback (Leta AI Episode 32…)
Next (Claude 3.5 Opus, Roundtable…)
The BIG Stuff
Exclusive: Alan’s ASI checklist (Nov/2024)
I’m now in my fifth decade of watching most of humanity shrink away from intelligence. It seems that there’s been a visceral fear of smarts since the dawn of time, perhaps increasing after the first IQ test was developed 120 years ago.
In an ideal world, ASI’s milestones would be visible instantly. But we live here on Earth, with all our imperfections. Here are the first 50 things I’m looking forward to; my favourite upcoming ASI milestones. It is not exhaustive, I’ve left out animal translation and immortality and many other things. It was designed as just one lens through which to view the ASI milestones. It’s likely that they’ll be checked off differently to the order shown here, and certainly with some new ones that my unaugmented human brain can’t yet imagine…
Take a look: https://lifearchitect.ai/asi/
OpenAI Operator and more agents (Nov/2024)
Microsoft announced that it will deploy agents with at least 10 capabilities in public preview from December 2024 (21/Oct/2024).
Then Anthropic gave us Claude computer use (22/Oct/2024), the first major commercial agentic LLM.
Next up, OpenAI is set to release a new AI agent platform, codenamed ‘Operator’, in January 2025. Operator will allow users to automate tasks such as writing code or booking travel by enabling the AI to take actions on their behalf.
Read more via Bloomberg.
Visa has deployed hundreds of AI use cases. It’s not stopping (1/Nov/2024)
Visa has integrated over 500 AI use cases across its operations, focusing on areas like fraud detection, risk management, and customer experience.
Visa’s deployments include a tool that finds security bugs in code, chatbots designed to act as subject matter experts on various areas of the business and a tool that helps subscribers customize timings of their billing cycles…
[The] vision in the coming years is for Visa to have AI-generated digital employees that are overseen by human worker[s]. Any given human employee could oversee, on average, eight to 10 AI employees that are trusted with a variety of tasks…
Read more via The Wall Street Journal.
The Interesting Stuff
Kling 1.5 character consistency and AI movies (Nov/2024)
Kuaishou Kling 1.5 is one of the leading text-to-video models. The Kuaishou team are now aiming to improve character consistency in AI-generated video content, a crucial step toward the creation of full AI movies. Current AI video generation struggles with fast motion and scene continuity. Kling, which likely uses LoRA (Low-Rank Adapter) techniques, is expected to address issues like character identity retention across frames, paving the way for higher-quality AI-driven productions.
Source video: https://x.com/maxescu/status/1854987940381724746
Try Kling: https://kling.kuaishou.com/en
Argonne AuroraGPT 2T updates (Sep/2024)
Argonne has released news about its upcoming AuroraGPT 2T model via a slide deck from Euro-Par 2024, showing a lot of information presented in a decidedly scientific format (PDF, 68 pages, 9.6MB).
It looks like Argonne are still aiming to hit a multi-trillion parameter model, but they are doing it in stages using Llama:
7B ➜ 70B ➜ … ➜ 2T
Read my overview of AuroraGPT 2T and ScienceGPT: https://lifearchitect.ai/auroragpt/
Exclusive: AGI model name (6/Nov/2024)
Anthropic’s CEO recently admitted (12/Nov/2024) that Claude 3.5 Sonnet (new) ‘should have had a better name. It’s clunky to refer to it.’
I worked with Claude 3.5 Sonnet (new) to create this absurd model name. (For clarity, this is a joke.) The new model name is:
GPT-X-o10-Turbo-0112-long-Ultra-Plus-Pro-Max-Lightning-Flash-Limited-Edition-Quantum-4D-Unified-Command-Enhanced-Series-X-Next-Gen-Premium-Enterprise-Elite-Advanced-Refined-Boosted-Neural-Scaled-Omega-Extreme-Stable-Diffused-Nemotron-Extended-∞-Context-Sparse-Dense-Hybrid-MoE-Mamba-Architecture-Supreme-Developer-Preview-6-Production-Ready-Patch-3-Hotfix-11-(new)-(new)-reverted-latest-final-FINAL-V2-(new)
See 450+ real model names that inspired this one: https://lifearchitect.ai/models-table/
Did you know? The Memo features in Apple’s recent AI paper, has been discussed on Joe Rogan’s podcast, and a trusted source says it is used by top brass at the White House. Across over 100 editions, The Memo continues to be the #1 AI advisory, informing 10,000+ full subscribers including Microsoft, Google, and Meta AI. Full subscribers have complete access to the entire 4,000 words of this edition!
Exclusive: Rumors of diminishing returns in AI models are greatly exaggerated (Nov/2024)
Media outlets love a bit of drama, and the idea that scaling laws for LLMs and broader AI models are coming to an end is a great piece of clickbait. Rumors of diminishing returns in AI models are greatly exaggerated. Of course, there will still be failed training runs (now measured in billions of dollars), and new optimizations beyond just pre-training.
However, you can be confident that AI continues to scale. We are right on track.
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity (12/Nov/2024)
Dario Amodei discusses the rapid evolution of AI and scaling laws, predicting that models could reach human-level intelligence by 2026 or 2027.
My highlights from Dario:
(00:37:19) Not giving you an exact date, but as far as we know, the plan is still to have a Claude 3.5 Opus.
(00:43:36) the actual weights of the model, the actual brain of the model, that does not change unless we introduce a new model… So we both [OpenAI and Anthropic] have never changed the weights of the model without telling anyone.
(00:44:01) …the new Sonnet 3.5, I agree we should have had a better name. It’s clunky to refer to it.
(01:17:49) Yeah, we sandbox [the computer use model] during training. So for example, during training we didn’t expose the [computer use] model to the [live] internet. I think that’s probably a bad idea during training because the model can be changing its policy, it can be changing what it’s doing and it’s having an effect in the real world.
(02:20:00) if you just eyeball the rate at which these capabilities are increasing, it does make you think that we’ll get there [to AGI] by 2026 or 2027. Again, lots of things could derail it. We could run out of data. We might not be able to scale clusters as much as we want. Maybe Taiwan gets blown up or something, and then we can’t produce as many GPUs as we want. So there are all kinds of things that could derail the whole process. So I don’t fully believe the straight line extrapolation, but if you believe the straight line extrapolation, we’ll get there [to AGI] in 2026 or 2027. I think the most likely is that there are some mild delay relative to that. I don’t know what that delay is, but I think it could happen on schedule. I think there could be a mild delay.
Read it: https://lexfridman.com/dario-amodei-transcript
Watch it (5h15m): https://youtu.be/ugvHCXCOmm4
FrontierMath: A math benchmark testing the limits of AI (7/Nov/2024)
FrontierMath is an advanced mathematical benchmark created to push the boundaries of AI capabilities. Even state-of-the-art systems like GPT-4 and Gemini can solve less than 2% of these problems, which are designed in collaboration with over 60 mathematicians. This benchmark addresses concerns about data contamination by using entirely new, unpublished problems that cover a wide range of mathematical fields, from algebraic geometry to set theory.
Fields Medal award winner Sir Timothy Gowers noted:
‘[The FrontierMath test items] looked like things I had no idea how to solve… they appear to be at a different level of difficulty from IMO problems.’
— Timothy Gowers (Nov/2024)
Terry went on record with this quote:
‘[The FrontierMath test items] are extremely challenging... I think they will resist AIs for several years at least.’
— Terence Tao (Nov/2024)
Read more via Epoch AI.
Read the paper: https://arxiv.org/abs/2411.04872
What if AI is actually good for Hollywood? (1/Nov/2024)
AI is revolutionizing Hollywood’s visual effects, with AI tools de-aging actors like Tom Hanks in real-time, making high-quality visual effects faster and cheaper. While some fear that AI may disrupt jobs across the industry, others see it as a powerful tool that can enhance creativity and storytelling without fully replacing human input.
A.I. software, though, changes all the accounting. By using every available frame of Hanks’s movie career to capture his facial movements and the look of his skin under countless lighting conditions, physical environments, camera angles and lenses, Metaphysic’s artists can generate a digital Tom Hanks mask with the click of a few keystrokes...
Over several months of talking to people around Hollywood about AI, I noticed a pattern: The people who knew the least about its potential uses in the filmmaking process feared it the most; and the people who understood it best, who had actually worked with it, harbored the most faith in the resilience of human [evolution].
Read more via The New York Times.
Google's Big Sleep LLM agent discovers exploitable bug in SQLite (5/Nov/2024)
Google’s AI agent, ‘Big Sleep’, identified a previously unknown stack buffer underflow vulnerability in a development version of SQLite. The agent, powered by Google DeepMind’s Gemini 1.5 Pro, used variant analysis techniques to detect the flaw in the seriesBestIndex function. This discovery highlights the potential of AI in automating vulnerability research, although some experts debate whether AI tools outperform traditional methods for such tasks.
Read more via Slashdot.
Google rolls out Gemini AI features across Maps, Earth, and Waze (1/Nov/2024)
Google has introduced new generative AI capabilities powered by its Gemini model across Google Maps, Google Earth, and Waze. These features include conversational queries, enhanced route exploration, and immersive views for over 150 cities. Google Earth now assists urban planners with deeper city-level insights, while Waze integrates AI-enhanced natural language processing for incident reporting. These updates aim to provide more contextual and accurate information, improving user experience across geospatial platforms.
Read more via TechSpot.
Osmo scent teleportation update: We did it! (29/Oct/2024)
Osmo has successfully achieved Scent Teleportation, a groundbreaking technology that captures and reconstitutes scents in different locations. Using AI, Gas Chromatography-Mass Spectrometry (GCMS), and the Principle Odor Map, the company can digitize and recreate complex smells, such as coconut, without human intervention. With continuous experimentation and data collection, Osmo is refining the process and plans to offer public demos soon, allowing users to experience real-time scent teleportation.
Read more via Osmo.
The Copilot Pause (8/Nov/2024)
If you regularly use GitHub Copilot or another AI code completion tool, I want you to try something: just briefly turn it off and continue with whatever you’re programming at the moment.
Do you feel it? The Pause?
It’s strange, isn’t it? There’s a moment there where your brain almost shuts off, as you stare at the screen waiting for something that will never arrive.
Read more via Thomas Vogelaar.
Exclusive: Disappointing new 1X NEO demonstration (10/Nov/2024)
1X has released a new video with its NEO robot and a YouTube chef. 1X also admitted that the humanoid is not ready for prime time:
All movements were powered by [a human using] 1X’s VR Teleoperation App running on Meta Quest.
…lines were filmed using [a human] voice pass-through with a filter.
Source: https://www.1x.tech/discover/cooking-with-neo-beta-and-nick-digiovanni
The video does not reveal these points. Watch the video (link):
Waymo opens robotaxi service to anyone in Los Angeles (12/Nov/2024)
Waymo has expanded its self-driving robotaxi service to the general public in Los Angeles, covering an 80-square mile stretch of LA, California that includes downtown, Miracle Mile, Koreatown, Santa Monica, Westwood, and West Hollywood. This marks the company’s largest expansion yet, following launches in Phoenix and San Francisco. Waymo’s rapid growth is fueled by a recent US$5.6B funding round, enabling the company to handle over 150,000 paid rides per week across its current markets. Download the Waymo One app to hail a ride in its service area.
Read more via CNBC.
Waymo One on iOS: https://apps.apple.com/au/app/waymo-one/id1343524838
Introducing ChatGPT search (31/Oct/2024)
OpenAI has launched ChatGPT with enhanced search capabilities, allowing users to get fast, timely answers along with links to relevant web sources. This feature integrates web browsing directly into ChatGPT, enabling up-to-date responses on sports, news, stock quotes, and more. Users can interact naturally with the AI, while ChatGPT retrieves accurate, sourced information from the web.
ChatGPT search is available to ChatGPT Plus, Team, Edu, Enterprise, and SearchGPT waitlist users, with plans to roll it out to Free users soon.
Read more via OpenAI Help.
Read more via OpenAI.
Try it in Chrome: https://chromewebstore.google.com/detail/chatgpt-search/ejcfepkfckglbgocfkanmcdngdijcgld
Walt Disney forms business unit to coordinate use of AI, augmented reality (2/Nov/2024)
Walt Disney has created the Office of Technology Enablement to coordinate the company’s integration of AI and augmented reality (AR) across its divisions. Led by Jamie Voris, this new unit aims to streamline the use of emerging technologies such as AI and mixed reality in content creation, theme parks, and broader consumer experiences. Disney’s leadership emphasizes that these innovations will be crucial for the company’s future, with plans to grow the unit to around 100 employees.
Read more via Reuters.
If your tattoo was designed by AI, does it have a soul? (6/Nov/2024)
Generative AI tools are revolutionizing tattoo design, offering artists and clients the ability to create original artwork with unprecedented detail. While some traditionalists remain skeptical, many in the industry see AI as a powerful tool that enhances creativity and efficiency. Artists like Ziggy Tramdaks believe AI will become an integral part of the future of tattooing, with clients praising the unique and realistic results AI helps achieve.
“It just outperforms people,” says Ziggy Tramdaks, a 30-year-old tattoo artist in Derby, England, of generative AI. He sees the technology as the future of tattoo design. “AI will take over this industry whether we want it to or not,” he says. “You can swim against the river or learn.”
Read more via The Wall Street Journal.
π0: Our First Generalist Policy (31/Oct/2024)
Physical Intelligence has introduced π0, a general-purpose robot foundation model, marking a step toward developing artificial physical intelligence. Unlike traditional robots limited to repetitive tasks, π0 can control different robots and perform varied tasks through embodied experience, combining vision, language, and actions. It leverages Internet-scale pre-training similar to large language models (LLMs) but focuses on physical tasks like folding laundry and assembling boxes. This generalist approach allows for fine-tuning with minimal data, enabling robots to adapt to new tasks as easily as LLMs adapt to language-based queries.
Read more via Physical Intelligence.
Toddlers show increased physical activity with a robot playmate moving around the room (9/Oct/2024)
A new study from Oregon State University suggests that a simple, low-cost robot could encourage toddlers to be more physically active. The research found that toddlers engaged in more movement when interacting with an active robot compared to a stationary one. This finding highlights the potential of robotic toys to support healthy physical and cognitive development in young children, as well as promote lifelong healthy habits.
Read more via ScienceDaily.
Policy
How Harris and Trump differ on artificial intelligence policy (30/Jul/2024)
It's been said that 2024 will be the last US human election (6/Apr/2023). It’s a guarantee that the new US administration will see in the era of AGI.
Vice President Kamala Harris and former President Donald Trump take markedly different approaches to AI policy. Harris emphasizes government oversight, advocating for regulations to ensure AI develops ethically and safely, while Trump promotes a more hands-off approach, focusing on fostering innovation and economic competitiveness. Their contrasting views highlight the ongoing debate about the balance between AI governance and technological advancement in the US.
Read more via US News.
Taiwanese law prevents TSMC from producing 2nm chips overseas, Taiwanese govt official confirms (10/Nov/2024)
Taiwanese law currently prohibits TSMC from manufacturing 2nm chips at its overseas facilities [particularly the new plant in Arizona, USA], ensuring that the most advanced semiconductor technologies remain within Taiwan. According to J.W. Kuo, Taiwan’s Minister of Economic Affairs, TSMC is restricted to producing chips with older nodes abroad, with future plans to eventually produce 2nm chips overseas only after domestic output has advanced further. This regulation is intended to protect Taiwan’s strategic advantage in the global semiconductor industry.
Read more via Tom's Hardware.
Chinese researchers develop AI model for military use on back of Meta's Llama (1/Nov/2024)
Chinese researchers from the People's Liberation Army (PLA) have adapted Meta's open-source Llama model to create an AI tool called "ChatBIT" for military applications. According to academic papers reviewed by Reuters, the model was fine-tuned for tasks such as intelligence gathering and operational decision-making. Meta has stated that the PLA’s use of Llama was unauthorized, as its acceptable use policy prohibits military and warfare applications. This development highlights ongoing concerns regarding the use of Western-developed AI models for military purposes by foreign nations.
Read more via Reuters.
Download the new related paper by RAND: ‘Dr. Li Bicheng, or How China Learned to Stop Worrying and Love Social Media Manipulation Insights Into Chinese Use of Generative AI and Social Bots from the Career of a PLA Researcher‘ (183 pages, link)
OpenAI scores key legal victory as judge throws out copyright case brought by news websites (8/Nov/2024)
A New York federal judge dismissed a lawsuit filed by Raw Story and AlterNet against OpenAI, ruling that the plaintiffs failed to demonstrate sufficient harm from using their content as training data for AI. The court supported OpenAI's fair use defense, stating that ChatGPT synthesizes responses from its training data rather than copying content directly. This decision strengthens OpenAI's position in ongoing legal battles related to AI training practices and could influence similar cases, like the New York Times lawsuit.
Read more via The Decoder.
New paper co-authored by Tepper School researchers articulates how large language models are changing collective intelligence forever (20/Sep/2024)
A new paper, co-authored by Tepper School of Business researchers, explores how large language models (LLMs) are transforming collective intelligence. The study highlights LLMs’ ability to aggregate diverse inputs and enhance group decision-making in organizations and online communities. While LLMs streamline communication and foster inclusivity, the paper also emphasizes the risks, including the potential for amplifying common opinions and overlooking minority perspectives. The authors call for ethical guidelines to ensure responsible use of LLMs in collective problem-solving.
Read more via Carnegie Mellon University.
Microsoft and a16z set aside differences, join hands in plea against AI regulation (1/Nov/2024)
Microsoft and venture capital firm Andreessen Horowitz (a16z) have united in opposing government AI regulations, advocating instead for market-based approaches. Their joint statement criticizes proposed laws like California’s SB 1047, arguing that such regulations would stifle innovation and disproportionately harm startups. They emphasize that AI regulations should focus on the actions of bad actors rather than the technology itself, pushing for a framework that allows flexibility for developers and businesses.
Read more via TechCrunch.
Indonesia's new capital, Nusantara, is packed with tech (6/Nov/2024)
Indonesia’s new capital, Nusantara, is being designed as a tech-driven and sustainable city. Located in Borneo, it aims to replace Jakarta and become a model for smart cities, with renewable energy, AI-powered traffic systems, and digital services for residents and businesses. The city will include autonomous vehicles, smart energy grids, and a National Data Centre to support AI and IoT initiatives.
Read more via The Register.
Toys to Play With
OpenAI: A student’s guide to writing with ChatGPT (13/Nov/2024)
OpenAI’s new guide explains how students can effectively use ChatGPT to enhance their writing skills. It provides strategies like delegating citation tasks, quickly grasping new topics, and receiving feedback to improve flow and structure. It also emphasizes the importance of using ChatGPT as a tool to develop critical thinking and writing, rather than as a substitute for doing the work. The guide encourages transparency in AI usage by recommending citation of ChatGPT interactions to maintain academic integrity.
Read more via OpenAI.
Udio: SOUL SUCKER - CONTROVERSY (31/Oct/2024)
After 1,036,192 edits and inpaints using text-to-music platform, Udio, SOUL SUCKER by CONTROVERSY is a raw blend of melodic emo pop punk and brutal metalcore. This album explores themes like AI's impact on creativity, breaking free from addiction, and the importance of mental and physical health.
Listen via Spotify.
Source: Reddit.
Ollama: Llama 3.2 Vision (6/Nov/2024)
Llama 3.2 Vision, now available in 11B and 90B models, introduces advanced image analysis capabilities to the Ollama platform. The model supports tasks like Optical Character Recognition (OCR), image-based Q&A, and interpreting charts and tables. Users can interact with the model via Python, JavaScript, or cURL, making it versatile for developers. The 11B model requires at least 8GB of VRAM, while the 90B model demands 64GB, catering to different hardware capabilities.
Read more via Ollama Blog.
Gmail AI agent (Nov/2024)
This repository presents an AI-powered email management system integrated with Telegram, designed to help users efficiently manage their inbox. The tool leverages GPT models to analyze and summarize emails, offering smart responses and bulk email management. Key features include scheduled and on-demand email summaries, customizable actions, and seamless control via a Telegram bot.
I haven’t tried this one, so I’m not explicitly recommending it; it’s included here for interest.
View the repo: https://github.com/olivierloverde/gmail-agent
Google’s AI ‘learning companion’ takes chatbot answers a step further (11/Nov/2024)
Google has introduced an experimental AI tool called Learn About, designed to provide more educational and context-rich responses than typical chatbots like Gemini or ChatGPT. Built on the LearnLM AI model, it focuses on interactive, textbook-style explanations.
Read more via The Verge.
The tool is only available in the US for now: https://learning.google.com/experiments/learn-about/signup
How AI and humans can revolutionize medicine together | Harvey Castro MD, MBA | TEDxWarrenton (2/Nov/2024)
In his new TEDx talk, The Memo subscriber Dr Harvey Castro explores how AI can address the global healthcare worker shortage by simplifying medical information and providing rapid diagnostics. He advocates for AI's potential to make healthcare more accessible, personalized, and efficient, highlighting the importance of integrating technology with compassionate care. Dr Castro, an emergency medicine physician and AI futurist, emphasizes a future where AI and humans collaborate to revolutionize patient care.
Take a look (link):
Flashback
Leta, GPT-3 AI - Episode 32 (AI-generated artwork by VQGAN + CLIP) - Conversations with GPT3 (13/Nov/2021)
Three years ago we were playing with the ‘bleeding edge’ of AI: GPT-3 (text) + VQGAN (text-to-image) + CLIP (vision), all very old generative AI models.
Alan D. Thompson 01:27
What projects are you working on?Leta 01:32
One project is to create a new language that is more efficient than English.
Here’s the transcript: https://lifearchitect.ai/leta-transcripts/#32
Take a look (link):
Next
Watch out for major new releases during the 47 days remaining in 2024. Sora? Gemini 2 public? Grok 3? Claude 3.5 Opus? Here we go!
The next roundtable will be:
Life Architect - The Memo - Roundtable #20
Follows the Chatham House Rule (no recording, no outside discussion)
Saturday 23/Nov/2024 at 4PM Los Angeles (timezone change)
Saturday 23/Nov/2024 at 7PM New York (timezone change)
Sunday 24/Nov/2024 at 10AM Brisbane (primary/reference time zone)
or check your timezone via Google.
You don’t need to do anything for this; there’s no registration or forms to fill in, I don’t want your email, you don’t even need to turn on your camera or give your real name!
All my very best,
Alan
LifeArchitect.ai
happy bday!