The Memo - 26/April/2025
Pangu Ultra 135B trained on Chinese Ascend chips, Gemini 350M users + ChatGPT 600M users, America's Superintelligence Project, 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: 26/Apr/2025
Subject: The Memo - AI that matters, as it happens, in plain English
AGI: 94%
ASI: 0/50 (no expected movement until post-AGI)Dr Eric Schmidt to Congress (9/Apr/2025):
What we need from [the US Govt] is to meet energy needs in all forms (renewable, non-renewable, whatever). It needs to be there, and it needs to be quickly… People are planning 10-gigawatt data centers… an average nuclear power plant in the US produces one gigawatt… This gives you a sense of how big this crisis is. Many people think that the demand for energy in our industry [AI] will increase from 3% to 99% of total generation… data centers will require an additional 29GW of power by 2027 and 67 more gigawatts by 2030. This gives you a sense of the scale we're talking about.
A quick reminder that AI scaling laws are still in full effect.
The runner-up for this month’s Who Moved My Cheese? AI Awards! is British author Philip Womack (‘AI will never write good fiction’). However, the real winner for Apr/2025 is colleague Ege Erdil (pronounced ‘egg-ay’). Despite being extremely well-informed, he’s gone on record to say that AGI will take another 20-30 years, with a prediction of around 2045 for ‘remote worker replacement’. As I continue to analyze and document, we’ve already hit significant milestones for this expanded economic definition, with Harvard finding GPT-4 performed at 138% of the level of human workers at Boston Consulting Group (Sep/2023), Google admitting that 25% of its new code is written by Gemini (Oct/2024) increasing to 30% today (25/Apr/2025), GPT-4.5 being judged more human than humans in conversation (Mar/2025), AI writing human-peer-reviewed scientific papers with new discoveries (Mar/2025), and much, much, much more.
In the context of current AI progress in 2025, an AGI date prediction of two more decades stands out as especially negligent. This reinforces my strongest possible recommendation for caution when evaluating expert commentary.
This month, the IEEE featured my quote and GPT-4.5 analysis in its article ‘Google Succeeds With LLMs While Meta and OpenAI Stumble‘ (21/Apr/2025).
Contents
The BIG Stuff (US AI + edu, OpenAI internal AI, XPeng Iron locomotion…)
The Interesting Stuff (Pangu Ultra 135B, Shopify, conscious models, Colossus 2…)
Policy (UAE AI laws, China AI + edu, GPT-4o personalized spam + deepfakes…)
Toys to Play With (Codex CLI, GPT-4.1 prompting…)
Flashback (Information and attention…)
Next (DeepSeek-R2, Roundtable…)
The BIG Stuff
Advancing artificial intelligence education for American youth (23/Apr/2025)
Matching China (see the Policy section below), the US President has issued an executive order to enhance AI education for American children, recognizing AI's transformative impact on the global landscape. The order establishes a White House Task Force on AI education, aiming to integrate AI literacy into K-12 education and beyond. This initiative also seeks to empower educators through AI-focused professional development and promote collaboration across government, academia, and industry to ensure the US remains a leader in AI innovation.
Sec. 7. Enhancing Training for Educators on Artificial Intelligence.
(a)…prioritize the use of AI in discretionary grant programs for teacher training… for:
(i) reducing time-intensive administrative tasks;
(ii) improving teacher training and evaluation;
(iii) providing professional development for all educators, so they can integrate the fundamentals of AI into all subject areas; and
(iv) providing professional development in foundational computer science and AI, preparing educators to effectively teach AI in stand-alone computer science and other relevant courses.
(b) Within 120 days of the date of this order, the Director of the NSF shall take steps to prioritize research on the use of AI in education.
Read more via The White House.
Perhaps to celebrate the fifth anniversary of GPT-3 (announced May/2020), this executive order is a full five (5) years late.
You can read my early work in this area: ‘The new irrelevance of intelligence’ for Mensa (May/2020), or my seminal work on Elon Musk’s gifted school published back in 2017.
Exclusive: How OpenAI applies AI internally (Apr/2025)
How we automate our own work at OpenAI
At OpenAI, we live with AI every day, so we’re often spotting new ways to automate our own work.
An example: Our support teams were getting bogged down, spending time accessing systems, trying to understand context, craft responses, and take the right actions for customers…
Our first use case: working on top of Gmail to craft customer responses and trigger actions. Using our automation platform, our teams can instantly access customer data and relevant knowledge articles, then incorporate the results into response emails or specific actions—such as updating accounts or opening support tickets…
This platform handles hundreds of thousands of tasks every month, freeing people to do more high-impact work. Not surprisingly, the system is now spreading across other departments.
Read ‘AI in the Enterprise’ (PDF, 25 pages, source page):
XPeng Iron walking/locomotion (Apr/2025)
The XPeng Iron humanoid robot seems to have solved walking/locomotion. Originally announced 6/Oct/2024, it is 1.73m (5′8″) tall, weighs 70kg, and has 60 articulating joints and 200 degrees of freedom (DoF). Estimated retail pricing is US$150k.
XPeng Iron is powered by XPeng’s proprietary 40-core Turing AI chip with a 30B parameter local/embedded model.
Watch the video (link):
The Interesting Stuff
Huawei Pangu Ultra 135B: Pushing the limits of dense large language models on Ascend NPUs (11/Apr/2025)
Huawei's Pangu Ultra is a new dense large language model with 135 billion parameters. This is one of the first major large language models trained exclusively on Chinese hardware (non-NVIDIA and non-Google chips). Pangu Ultra was trained on 8,192 Ascend NPUs (neural processing units, Kunpeng 920 processors in Huawei Atlas 800T A2 servers).
This is a pretty serious advance, meaning that for the first time*, AI model training hardware is not US-led (both NVIDIA and Google are headquartered in California, USA), but instead 100% Chinese-led (Huawei is headquartered in Shenzhen, China).
*Baidu continues to use Korean-led Samsung Kunlun chips, though they along with Alibaba, Bytedance, and Tencent recently bought tens of thousands of US-led NVIDIA H100s each.
From this point forward, considering the current political and technological climate, it is becoming increasingly likely that China will speed past the US in AI development.
Read the paper: https://arxiv.org/abs/2504.07866
See it on the Models Table: https://lifearchitect.ai/models-table/
Google Gemini has 350 million monthly users, reveals court hearing (23/Apr/2025)
Google's AI chatbot Gemini has achieved significant growth, reaching 350M monthly active users globally as of Mar/2025, according to information disclosed during an antitrust court case. This marks a rapid increase from 9M daily active users in Oct/2024 to 35M daily users recently. Despite this growth, Gemini still trails behind ChatGPT, which reportedly had around 600M monthly users.
Read more via TechCrunch.
The Memo features in recent AI papers by Microsoft and Apple, 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 RAND, Google, and Meta AI. Full subscribers have complete access to all 25+ AI analysis items in this edition!
Exclusive: AI responses are already killing traditional search (17/Apr/2025)
…the presence of AI Overviews [via Gemini] reduces the click-through rate for position 1 by ~34.5%.