The Memo - 30/Mar/2025

29+ reasoning models, Google Gemini 2.5 Pro, OpenAI GPT-4o images, and much more!

Mar 29, 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:    30/Mar/2025
Subject: The Memo - AI that matters, as it happens, in plain English
AGI:     92%
ASI:     0/50 (no expected movement until post-AGI)

Contents

The BIG Stuff (Anthropic, Google Gemini 2.5, GPT-4o images, reasoning models…)
The Interesting Stuff (Keynotes, 1X in 100s of homes, AI newspaper, edu, jobs…)
Policy (Russia, OpenAI + US Gov, new China mandates, TSMC, Waymo WA DC…)
Toys to Play With (OpenAI transcribe, new synthesia clone, GPT-4.5 diary, viz…)
Flashback (GPT-3 and resurrection…)
Next (new Llama model, Roundtable to be scheduled later…)

The BIG Stuff

Anthropic: Tracing the thoughts of a large language model (27/Mar/2025)

There were some outstanding frontier AI model releases and upgrades this month, but they were all just expected dots on the exponential curve. Maybe the biggest announcement was Anthropic’s research in the new field of mechanistic interpretability, also known as ‘what exactly happens inside a large language model’. This is important because seeing precisely how AI thinks (and how it makes decisions) allows us to peek into its mind, helping us steer it toward being grounded, transparent, and aligned with our human values.

Anthropic’s research found Claude’s ability to plan ahead when writing, such as in poetry, and its conceptual multilingual capabilities, indicating a shared ‘language of thought’.

Claude will plan what it will say many words ahead, and write to get to that destination. We show this in the realm of poetry, where it thinks of possible rhyming words in advance and writes the next line to get there. This is powerful evidence that even though models are trained to output one word at a time, they may think on much longer horizons to do so…
These findings aren’t just scientifically interesting—they represent significant progress towards our goal of understanding AI systems and making sure they’re reliable.

Read the blog post via Anthropic.

Read the dense ‘circuit tracing’ paper (45,000 words) via Anthropic.

Read the dense ‘biology’ paper (42,000 words) via Anthropic.

Read related discussions: LessWrong, HN, Reddit

Sidenote: It turns out these models aren’t just randomly parroting stuff. Leta AI was trying to tell us this four years ago…

Watch my short 2021 video clip (link):

Exclusive: 29+ new reasoning models (Mar/2025)

The media has been great at framing an adversarial power play between the US and China, especially highlighting major new reasoning models like DeepSeek-R1 and OpenAI o3.

However, these are not the only two models available. Released in Sep/2024, OpenAI’s original o1-preview model inspired many labs to focus on allowing models time to ‘think’ before responding. As of this edition, and within just five months (Nov-Mar), I’ve documented more than 29 reasoning models.

Source: https://lifearchitect.ai/models-table/ Click to enlarge.

Google Gemini 2.5 Pro: Newest Gemini model with thinking (25/Mar/2025)

Google DeepMind has unveiled Gemini 2.5 Pro Experimental, which leads in benchmarks like LMArena (measuring human preferences), and excels in coding, math, and science tasks. This model introduces improved post-training and a 1 million token context window. Google aims to integrate thinking capabilities into all models to handle increasingly complex scenarios.

Gemini 2.5 Pro Experimental ranks #4 on GPQA as shown on my Models Table.

Read more via Google, see some excellent output examples: 1, 2

The image below has been doing the rounds, and I think it’s still appropriate. Note that I don’t use or recommend Google’s Gemini because of the secret identifying watermarks they embed in all outputs.

OpenAI: Introducing 4o image generation (25/Mar/2025)

OpenAI has announced the release of GPT-4o with native image generation. GPT-4o now acts as a multimodal model for creating photorealistic and precise images. The model excels in rendering fingers, text, and more accurately following detailed prompts, demonstrating significant advancements.

Source: OpenAI. Image generated by GPT-4o in Mar/2025. The prompt is 120 words: ‘A wide image taken with a phone of a glass whiteboard, in a room overlooking the Bay Bridge. The field of view shows a woman writing, sporting a tshirt wiith a large OpenAI logo. The handwriting looks natural and a bit messy, and we see the photographer's reflection. The text reads: (left) "Transfer between Modalities: Suppose we directly model p(text, pixels, sound) [equation] with one big autoregressive transformer. Pros: * image generation augmented with vast world knowledge * next-level text rendering * native in-context learning * unified post-training stack Cons: * varying bit-rate across modalities * compute not adaptive" (Right) "Fixes: * model compressed representations * compose autoregressive prior with a powerful decoder" On the bottom right of the board, she draws a diagram: "tokens -> [transformer] -> [diffusion] -> pixels"‘

The Interesting Stuff

Alan’s keynote rehearsal “AI: Humanity’s Last Invention” (Mar/2025)

I’ve provided a recording of my 2025 keynote below, and shortly plan to release my more formal keynote rehearsal for the 31st International Symposium on Controversies in Psychiatry, ‘Generative AI: GPT-5 and brain-machine interfacing’.