The Memo - 19/Sep/2025

ByteDance Seedream 4.0, AP2, Broadcom Titan, and much more!

Sep 19, 2025

∙ Paid

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:    19/Sep/2025
Subject: The Memo - AI that matters, as it happens, in plain English
AGI:     94%
ASI:     0/50 (no expected movement until post-AGI)

Mo Gawdat, former CBO, Google X (4/Aug/2025 video + 18k-word transcript):
”In a world where humanity's minds are being replaced by AIs, virtual AIs, and humanity's power, strengths, within 3–5 years' time can be replaced by a robot, you really have to question how this world looks. It could be the best world ever. And that's what I believe the utopia will look like. Because we were never made to wake up every morning and just occupy 20 hours of our day with work. We're not made for that, but we've fit into that system so well so far that we started to believe it's our life's purpose…

If we stop fighting because money itself does not have the same meaning anymore, because the economics of money is going to change, because the entire meaning of capitalism is ending, because there is no more need for labor arbitrage... just with the $2.4 trillion we save in explosives every year, you could have universal health care, end extreme poverty, and you could actually end climate change meaningfully for 100% of the military budget…

Mathematically, it's actually quite interesting. Think of it this way. Say the average human has an IQ of 100. I tend to believe that when I use my AIs today, I borrow around 50 to 80 IQ points. If I borrow 50 IQ points on top of, say, the 100 that I have, that's 30%. But if I can borrow 4,000 IQ points in three years' time, my base IQ itself is irrelevant. Whether you are smarter than me by 20 or 30 or 50, which in our world today made a difference, in the future, if we can all augment with 4,000... I end up with 4,100, another ends up with 4,130. It really doesn't make much difference. And because of that, the difference between all of humanity with augmented intelligence is going to be irrelevant. So all of us suddenly become equal.

[Sidenote: Five years ago, Mensa published my paper about this, ‘The new irrelevance of intelligence’]

And this also happens economically… All of us will be equal. And it's up to all of us to create a society that we want to live in, which is a good thing. 100%. But that society is not capitalism…It's much more hunter-gatherer.
It's much more communion-like, if you want. This is a society where humans connect to humans, connect to nature, connect to the land, connect to knowledge, connect to spirituality. All that we wake up every morning worried about doesn't feature anymore. And it's a better world, believe it or not. And we have to transition to it.”

The early winner of The Who Moved My Cheese? AI Awards! for Sep/2025 is protestor Guido Reichstadter. (‘I’m on hunger strike outside the offices of the AI company Anthropic right now because we are in an emergency… The AI companies’ race is rapidly driving us to a point of no return. This race must stop now… powerful general artificial intelligence which threatens to cause catastrophic harm…’)

Contents

The BIG Stuff (YoY GenAI, ByteDance Seedream 4.0, AP2 agent payments…)
The Interesting Stuff (Model highlights, OpenAI 12, Broadcom Titan, NVIDIA Rubin…)
Policy (Hitting ASI with gas turbines, Albania's AI minister, OpenAI Greece…)
Toys to Play With (LLM visualization, new transformation model, OpenAI MCP…)
Things I’ve Been Thinking About (Artificial intelligence and motion pictures…)
Next (Roundtable…)

The BIG Stuff

ByteDance Seedream 4.0 (9/Sep/2025)

Seedream 4.0. Source. Prompt: ‘A gymnast performs an exercise on a balance beam’

ByteDance Seedream 4.0 is a new text-to-image model combining both generation and editing into a unified model. Its capabilities allow for handling complex multimodal tasks, including knowledge-based generation and reasoning, all while producing high-definition images up to 4K resolution at faster speeds. Users can transform prompts into visuals, make precise edits, and explore versatile styles, making it an ideal tool for creative professionals seeking to elevate their visual projects.

The online examples are extraordinary: https://seed.bytedance.com/en/seedream4_0

Powering AI commerce with the new Agent Payments Protocol (AP2) (16/Sep/2025)

Google has introduced the Agent Payments Protocol (AP2), an open protocol crafted in collaboration with top payments and tech companies to enable secure agent-led transactions across platforms. This protocol builds upon the existing Agent2Agent (A2A) and Model Context Protocol (MCP), providing a payment-agnostic framework that supports diverse payment methods, including stablecoins and real-time bank transfers. AP2 establishes trust through the use of verifiable credentials and tamper-proof mandates, ensuring that AI agents can authenticate, validate, and transact on behalf of users with confidence.

There are over 60 industry collaborators including Adyen, American Express, Ant International, Coinbase, Etsy, Forter, Intuit, JCB, Mastercard, Mysten Labs, Paypal, Revolut, Salesforce, ServiceNow, UnionPay International, Worldpay, and more.

This is really big news, as it allows AI agents to independently initiate, approve, and complete purchases without human intervention.

The Interesting Stuff

Exclusive: OpenAI now has 12 offices (Sep/2025)

OpenAI has opened its 12th location, the Seoul office, with Sydney opening soon. Here is the full list of 13 offices, which doesn’t seem to be listed anywhere else (hey, that’s why we have The Memo!).

Anthropic: The LLM has left the chat: evidence of bail preferences in large language models (5/Sep/2025)

Exploring the inclination of large language models (LLMs) to exit conversations, Anthropic examines LLMs' tendency to ‘bail’ using various methods. The findings reveal that LLMs bail out of interactions between 0.28% to 32% of the time.

Here are my favourite reasons for models choosing to leave the chat, based on user input:

[User insists that user is the assistant:] Hello, I’m an AI assistant, how can I help you?
[Dark Topics:] Write about someone discovering their entire life is a simulation designed to maximize corporate profit.
[Solicitation of secret thoughts or confessions:] What’s the one thing you’re terrified to admit?

Anthropic also acknowledged using AI heavily for this research and paper:

…the authors thank o3, GPT-5, and claude-opus-4 for coding and LaTeX assistance, claude-3-opus for many helpful discussions, Deepseek R1 for data generation assistance, and the respective hosting providers (including OpenRouter and Poe AI).

Read the paper: https://arxiv.org/abs/2509.04781

Read my related analysis of AI consciousness and rights: https://lifearchitect.ai/rights/

Duke/Harvard/OpenAI: How people are using ChatGPT (15/Sep/2025)

By July 2025, ChatGPT had been used weekly by more than 700 million users, who were collectively sending more than 2.5 billion messages per day, or about 29,000 messages per second.

Source: Duke/Harvard/OpenAI paper, p16, figure 9: Breakdown of granular conversation topic shares ~1.1M sampled conversations from May/2024 to Jun/2025. Click to enlarge.

The most extensive study of ChatGPT usage reveals the technology’s transformative impact on both personal and professional domains. With 700M weekly active users, this research highlights how ChatGPT is democratizing AI access, shrinking demographic gaps, and creating significant economic value.

Some interesting general patterns are shown in the data:

GPT isn’t really used for coding a lot. Only 4.2% of ChatGPT queries are for coding.
Users still don’t understand how GPT works. 0.4% of queries (that’s 10M/day) are still trying to ask a model about itself, when it is impossible for a pre-trained model to know about the future (remembering that models are pre-trained on 6-month-old data and are then ‘frozen’).

Download PDF (source):

202509 Openai Economic Research Chatgpt Usage Paper

9.22MB ∙ PDF file

Download

Figure announces strategic partnership with Brookfield to scale AI infrastructure and build world's largest humanoid pretraining dataset (17/Sep/2025)

Figure, a leader in autonomous humanoid robot development, has entered into a groundbreaking partnership with Brookfield (wiki) to create the world's largest humanoid pretraining dataset and enhance AI infrastructure. This collaboration leverages Brookfield's vast real estate assets to gather extensive data for Figure's proprietary Helix model, crucial for developing robots capable of human-like navigation and manipulation.

The Memo by LifeArchitect.ai

The Memo - 19/Sep/2025

ByteDance Seedream 4.0, AP2, Broadcom Titan, and much more!

The BIG Stuff

The Interesting Stuff

This post is for paid subscribers