The Memo - 11/Jan/2025
NVIDIA Cosmos dataset 9 quadrillion tokens, ALPrompt 2025H1, OpenAI Reflections, 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: 11/Jan/2025
Subject: The Memo - AI that matters, as it happens, in plain English
AGI: 88%
ASI: 0/50 (no expected movement until post-AGI)Dr Jim Fan (1/Jan/2025):
It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. Our children will grow up as ‘robot natives’. They will have humanoids cook Michelin dinner, robot teddy bears tell bedtime stories, and FSD drive them to school.
We are the generation of ‘robot immigrants’, en route to a new world of ubiquitous Physical AI, much like our parents are ‘digital immigrants’, learning to realign their lives on 6 inches of touch screen. It’s a journey of both inventing sci-fi tech and reinventing ourselves. Everything that moves will be autonomous. Every year from now on will be the Year of Robotics. Here’s to a wild 2025 ahead.
Welcome to 2025! This is the big one, for sure, and it’s already started with a bang. I’ve completed a comprehensive analysis of the xAI Grok series of models, and will send that 30-page report out to full subscribers shortly. We recently hit 500 total large language model highlights on the Models Table, with 27 new major models released last month, all detailed towards the end of the edition.
Contents
The BIG Stuff (NVIDIA Cosmos, ALPrompt 2025H1, brain atrophy due to LLMs…)
The Interesting Stuff (Roborock Saros Z70, Foxconn robots, more Veo 2 demos…)
Policy (2025 US AI legislation, Altman interview, ByteDance $7B training budget…)
Toys to Play With (Sonus-1, prompt challenge, infinite chat, SF test, MERPHI…)
Flashback (Jan/1967: All watched over by machines of loving grace…)
Next (Roundtable…)
The BIG Stuff
NVIDIA Cosmos + video dataset calculations (8/Jan/2025)
Announced at CES in January 2025, NVIDIA Cosmos is a platform designed to accelerate the development of physical AI systems using state-of-the-art generative world foundation models (WFMs). These models are trained on extensive robotics and world video data and are available under an open model license, making physical AI development more accessible.
AI labs are leveraging the Cosmos platform ‘to accelerate the development of [their] humanoid robot[s]’ via ‘Foresight and “multiverse” simulation, using Cosmos and Omniverse to generate every possible future outcome an AI model could take to help it select the best and most accurate path.’ 14 humanoids were shown, country of origin added (missing Tesla Optimus and others):
Apptronik Apollo (USA)
ROBOTERA Star1 (China)
Agibot A2 (China)
NEURA Robotics 4NE-1 (Germany)
Fourier GR-2 (China)
Boston Dynamics E-Atlas (USA)
Agility Robotics Digit (USA)
Figure 02 (USA)
1X NEO (Norway)
Galbot G1 (China)
Mentee MenteeBot (Israel)
Unitree H1 (China)
XPENG Iron (China)
Sanctuary AI Phoenix (Canada)
Watch the launch (timecode): https://youtu.be/k82RwXqZHY8?t=4580
The NVIDIA Cosmos video dataset is now the largest publicly-known training dataset in the world. With 20M hours of video (est 2.5B videos, avg 30sec), it has a raw size of ~45PB or ~9 quadrillion tokens (9Qa, 9,000T tokens). It surpassed the previous record holder from Jun/2024—the Common Crawl text dataset DCLM-Pool with 240T tokens—by 37.5×. All working by LifeArchitect.ai, assisted by OpenAI o1 on poe.com.
Read my Cosmos dataset calculations: https://lifearchitect.ai/cosmos/
Read the Cosmos paper (75 pages, 80MB).
Try the related Cosmos Nemotron 34B VLM (free, no login): https://build.nvidia.com/nvidia/cosmos-nemotron-34b
Exclusive: Empirically, LLM use is already leading to brain atrophy (9/Jan/2025)
Drawing from my background in human intelligence research and as a former chairman of Mensa International’s gifted families, the evidence suggests that widespread use of large language models (LLMs) such as ChatGPT and Claude for everyday tasks—emails, trivia, research, ‘life admin’—seems to already be fostering distinct mental shortcuts that may lower our cognitive fitness. This ‘lazy brain’ phenomenon might persist until we seamlessly merge AI into our neural pathways via brain-machine interfaces (BMIs).
We’ve seen big hits to human intelligence before, from lead in petrol to fluoride in drinking water. The average lead-linked loss of 2.6 IQ points per person totals a staggering 824 million points worldwide. Certain birth cohorts in the ’60s and ’70s were hit especially hard, with some individuals losing 7.4 IQ points—enough to shift them into the range of intellectual disability (Princeton, Mar/2022). Fluoride has also been implicated: this month, a systematic review found an IQ drop of 1.63 points for every 1 mg/L increase in urinary fluoride. That’s a decrease of 6.5 IQ points if kids in one region are exposed to 4 mg/L more fluoride than their peers (NIH, Jan/2025).
Human-like large language models have been available since GPT-3 in 2020, and their cognitive impact coincides with an alarming technological gap. Brain-machine interfaces (BMIs) are currently limited to being unidirectional (from human to machine, not vice versa), and are not yet ready for the public. In the time between our present use of LLMs via chat and the future use of fully integrated AI via brain-machine interfaces, I anticipate a measurable decline in human cognitive ability.
This erosion of mental capabilities through AI use will stem from reduced mental practice and diminished active engagement. While the exact magnitude is yet to be quantified, I welcome research in this area. There are many similar studies on smartphone and internet overuse (Feb/2024), and even on measuring how GPS navigation decreases spatial memory and navigation skills (McGill, Apr/2020).
The good news is that any LLM-induced cognitive atrophy doesn’t have to be permanent. When researchers discovered lead’s harmful effects, public health campaigns systematically reduced it. Looking ahead, BMIs will also offer a solution by enabling bidirectional engagement with AI systems. BMIs should then augment our cognitive processes while potentially expanding the neural pathways that make us uniquely human. Early BMI research suggests that precisely targeted neural stimulation via BMIs could even promote neurogenesis and synaptic plasticity, effectively growing new neural pathways (McKnight, May/2015).
I’ll explore some of these points within the broader context of LLMs and BMIs in my invited keynote for the 31st International Symposium on Controversies in Psychiatry in Barcelona in Apr/2025.
Edit just after publication of this edition: First major study released: ‘AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking’ SBS, Jan/2025. https://www.mdpi.com/2075-4698/15/1/6
The Interesting Stuff
ALPrompt 2025H1: latest quick benchmark prompt for LLMs (Jan/2025)
ALPrompt 2025H1 is now available, providing a quick resource for evaluating large language models (LLMs). Updated twice a year, ALPrompt supports researchers and developers in the AI community. Despite previous challenges with the 2024H2 prompt being exposed via Twitter, ALPrompt will continue to be available for free to all humanity.
Try it: https://lifearchitect.ai/alprompt/
Roborock Saros Z70 with robotic arm (Jan/2025)
The Saros Z70 by Roborock is an advanced robotic vacuum equipped with a foldable five-axis mechanical arm, designed to assist cleaning by moving obstacles. Roborock says: ‘It performs simple household tasks and even serves as a companion’. Its ultra-slim 7.98cm design allows it to clean previously inaccessible areas. The Saros Z70 features AI-powered customization, a dual camera system for enhanced object recognition, and a multifunctional dock with auto water refill and drainage for hands-free convenience, signaling a new era in robotic vacuum technology.
Read more via Roborock Global Official.
I prompted NVIDIA Cosmos Nemotron 34B (see the beginning of this edition) to watch the official Saros Z70 video above, and it responded: ‘The robotic vacuum cleaner is shown picking up a shoe and then dropping it back down. This action suggests that the vacuum cleaner is capable of handling objects, possibly as part of its cleaning or maintenance routine. The fact that it's able to pick up and drop the shoe without any visible issues implies that the vacuum cleaner is designed to be versatile and adaptable to different tasks.’
Watch (link):
OpenAI CEO: Reflections (6/Jan/2025)
Leading in to 2025, OpenAI’s CEO reflected on GPT, AGI, ASI, leadership, governance, and much more.
‘We are now confident we know how to build AGI as we have traditionally understood it. We believe that, in 2025, we may see the first AI agents… We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word… Superintelligent tools could massively accelerate scientific discovery and innovation well beyond what we are capable of doing on our own, and in turn massively increase abundance and prosperity.‘
Read it: https://blog.samaltman.com/reflections
Taking the Unitree G1 Humanoid Robot for a Spin (7/Jan/2025)
At CES 2025, CNET was given a hands-on experience with the Unitree G1 Humanoid and Go2 Canine robots. Unitree is headquartered in Hangzhou, China. The event highlights the G1's details, customization options, and controller features, as well as its unique ability to dance. The Go2 Canine robots, including the Go2-W variant, are demonstrated with features like Lidar mapping and a handstand mode, illustrating Unitree's advancements in robotics.
Watch (link):
Foxconn teams up with NVIDIA to develop humanoid robots (4/Jan/2025)
Taiwan-based Hon Hai Precision Industry Co., also known as Foxconn, has partnered with NVIDIA to advance humanoid robot development amidst the AI surge. Foxconn is leveraging NVIDIA’s software and hardware technologies for this initiative, particularly in Kaohsiung, Taiwan (one of my favourite cities). The collaboration aims to integrate humanoid robots into services including healthcare, and reflects Foxconn's strategy to diversify its product offerings beyond its traditional manufacturing role.
Read more via Focus Taiwan.
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!
Google Veo 2: Even more video samples (Jan/2025)