The Memo - 27/Dec/2024
NVIDIA GB300, OpenAI o3, Microsoft’s 485,000 Hopper chips, 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: 27/Dec/2024
Subject: The Memo - AI that matters, as it happens, in plain English
AGI: 84 ➜ 88%
ASI: 0/50 (no expected movement until post-AGI)It was a nice surprise to see my GPT-5 dataset paper findings featured in the G7 AI document. G7 (wiki) is an intergovernmental organization of Canada, France, Germany, Italy, Japan, the UK, and the US. The document is available for download in the Policy section of this edition.
The Memo subscription fee for new subscribers will increase from 1/Jan/2025. If you’re a current subscriber, you’ll always be on your old/original rate while you’re subbed. Free readers can become a full subscriber at the old/current rate before New Year’s Day.
Contents
The BIG Stuff (AI report, o3 reasoning model, DeepSeek-V3, 485k Hopper chips…)
The Interesting Stuff (Veo 2 examples, Falcon 3, The Well, Harvard Books…)
Policy (G7 AI document, NIST Claude 3.5S new, NIST o1…)
Toys to Play With (new free OpenAI course…)
Flashback (My Aurora experiments from 2021…)
Next (Roundtable…)
The BIG Stuff
2024 AI retrospective: The sky is steadfast (21/Dec/2024)
Full subscribers to The Memo received my end-of-year AI report last week. Headings include: The BIG Stuff, Large language models, Datasets, China, and Performance.
The final section, ‘How I'm preparing for AGI and ASI,’ lists six practical actions I’ve been personally taking for the last few years.
Read the report: https://lifearchitect.ai/the-sky-is-steadfast/
o3 reasoning model (19/Dec/2024)
Many people are freaking out about the superhuman performance of OpenAI’s latest reasoning model, o3. In my view, this is just a natural point on an exponential curve. This thing is smart, outperforming most people on the planet, and much closer to ASI (beyond the most gifted human) than just plain ol’ AGI (average human). This model increased my AGI countdown from 84% ➜ 88%.
My latest viz shows most of what you need to know:
Read more: https://lifearchitect.ai/o3/
DeepSeek-V3: Strong MoE language model (2024)
Hangzhou-based AI lab DeepSeek-AI (深度求索, Shēndù qiúsuǒ) released a Christmas gift with the open-source DeepSeek-V3 model. This thing is huge: 685B parameters MoE, of which 37B are activated per token, trained on 14.8T tokens (22:1).
GPQA=59.1, MMLU=87.1.
The model is available immediately on the free chat.deepseek.com playground, scoring 5/5 for the 2024H1 ALPrompt, but 0/5 for the 2024H2 ALPrompt.
See it on the Models Table: https://lifearchitect.ai/models-table/
NVIDIA GB300 'Blackwell Ultra' 288GB memory, 1,400W power (23/Dec/2024)
NVIDIA's upcoming GB300 AI server is set to redefine performance standards with its B300 GPU chip, boasting 288GB of HBM3E memory and a 1,400W TDP [Thermal Design Power].
SemiAnalysis commented (25/Dec/2024) on the move from H100 ➜ H200 ➜ B300:
Reasoning models [like o1 and o3] can be a poor user experience due to significant waiting time between requests and responses. If you can offer significantly faster reasoning time, this will increase the user’s propensity to use and pay for them.
A 3x difference in cost is massive. Hardware delivering 3x with a mid-generation memory upgrade is frankly insane, way faster than Moore’s law, Huang’s Law, or any other pace of hardware improvement we’ve seen.
We have observed that the most capable and differentiated models are able to charge a significant premium over even slightly less capable models. Gross margins on frontier models are north of 70%, but on trailing models with open source competition, margins are below 20%. Reasoning models don’t have to be 1 chain of thought. Search exists and can be scaled up to improve performance as it has in o1 Pro and o3. This enables smarter models that can solve more problems and generate significantly more revenue per GPU.
Read more via TechPowerUp.
Microsoft acquires twice as many NVIDIA AI chips as tech rivals (18/Dec/2024)
Microsoft has emerged as a leader in AI infrastructure by purchasing 485,000 NVIDIA ‘Hopper’ chips this year, significantly outpacing rivals like Meta, Amazon, and Google. This strategic move, driven by its substantial investment in OpenAI, positions Microsoft at the forefront of developing advanced AI systems, leveraging its Azure cloud infrastructure to support innovations such as OpenAI’s latest models.
Sidenote: In 2025, the AI lab with the most chips will be more likely to be the AGI winner. If Microsoft has 500k+ H100-equivalent chips, and xAI has 1M+, it’s going to be a very interesting race.
Read more via Financial Times.
The Interesting Stuff
Happily ever after: A children’s story on post-ASI scenario (23/Dec/2024)
My latest piece is a children’s story about the upcoming post-scarcity world.
Some grown-ups were scared about their piggy banks and money. “If the humanoids do everything for free,” they said, “what will happen to all our savings? What about our jobs?”
Read it: https://lifearchitect.ai/happily-ever-after/
Watch it (link):
Nobel Minds 2024: Prof Geoffrey Hinton on post-AGI scenario (19/Dec/2024)
This was a beautiful setting for the 2024 Nobel prize winners to speak. AI godfather Geoffrey Hinton had an interesting and pessimistic point about a possible post-AGI scenario (to offset my ‘hyper-optimistic’ children’s story above!)…
Interviewer: Geoffrey Hinton, do you think that this increase in productivity, essentially, that will come with automation and so on and so forth, is a good thing for society?
Hinton: Well, it ought to be, right? It's crazy; we're talking about having a huge increase in productivity, so there's going to be more goods and services for everybody, so everybody ought to be better off…
But actually, it's going to be the other way around. And it's because we live in a capitalist society. And so what's going to happen is this huge increase in productivity is going to make much more money for the big companies and the rich, and it's going to increase the gap between the rich and the people who lose their jobs. And as soon as you increase that gap, you get fertile ground for fascism.
And so it's very scary that we may be at a point where we're just making things worse and worse. And it's crazy because we're doing something that should help everybody, and obviously will help in healthcare, and help in education. But if the profits just go to the rich, that's going to make society worse.
Watch (link):
South Korea considers creating 'KSMC' chipmaker to compete with TSMC, $13.9B investment (24/Dec/2024)
South Korea is considering establishing a government-funded contract chipmaker, the Korea Semiconductor Manufacturing Company (KSMC), to compete with TSMC. The initiative aims to address the country's semiconductor industry's structural weaknesses and reliance on Samsung, proposing a $13.9 billion investment that could yield $208.7 billion in economic benefits by 2045. The plan highlights the need for increased R&D, financial incentives, and reduced regulatory burdens to enhance Korea's competitive edge and support smaller semiconductor firms.
Read more via Tom's Hardware.
Google Veo 2: Chopping tomatoes (17/Dec/2024)
Human or AI? This one is the very latest text-to-video AI model from Google.
Source: https://x.com/agrimgupta92/status/1868745017571131582
Google Veo 2: The Heist (23/Dec/2024)
This 2-min film is unbelievable, and with notable consistency across assets.
Watch (link):
OpenAI ‘considered’ building a humanoid robot: report (24/Dec/2024)
OpenAI has reportedly explored the idea of building its own humanoid robot, as revealed by sources cited by The Information. Although OpenAI had previously closed its robotics division in 2021, recent advances in hardware and AI systems have rekindled interest in this area. The company has been active in the robotics space through financial investments in companies like Figure and 1X, and the general-purpose AI firm Physical Intelligence.
Read more via TechCrunch.
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 this edition!
TII Falcon 3: UAE’s Technology Innovation Institute launches small models (17/Dec/2024)
