The Memo - 24/Apr/2026

To: US Govt, major govts, Microsoft, Apple, NVIDIA, Alphabet, Amazon, Meta, Tesla, Citi, Tencent, IBM, & 10,000+ more recipients… From: Dr Alan D. Thompson Sent: 24/Apr/2026 Subject: The Memo - AI that matters, as it happens, in plain English AGI: 97% ASI: 0/50 (no expected...

To: US Govt, major govts, Microsoft, Apple, NVIDIA, Alphabet, Amazon, Meta, Tesla, Citi, Tencent, IBM, & 10,000+ more recipients… From: Dr Alan D. Thompson Sent: 24/Apr/2026 Subject: The Memo - AI that matters, as it happens, in plain English AGI: 97% ASI: 0/50 (no expected movement until post-AGI)GPT-5.5 launch by OpenAI (23/Apr/2026):‘One engineer at NVIDIA who had early access to the model went as far as to say: "Losing access to GPT‑5.5 feels like I've had a limb amputated.”’The winners of The Who Moved My Cheese? AI Awards! for Apr/2026 are Daniel Alejandro Moreno-Gama who allegedly threw a bottle containing a flaming rag at the metal gate of OpenAI CEO Sam Altman’s property, and Amanda Tom and Muhamad Tarik Hussein who appeared to fire a round at Altman’s property.This edition has a soundtrack. Join 2.5 million other listeners streaming ‘Celebrate Me’, a song created using Suno by AI-generated persona IngaRose. The track hit #1 on the iTunes charts in the US, UK, France, Canada and New Zealand. (Forbes, 17/Apr/2026)We had a blast at the recent roundtable, watching the Beijing humanoid half-marathon in real time, and demonstrating advanced agents interacting in the home. We’re up to about roundtable #50, now with consistent dates and times (twice monthly). The Memo roundtables are open to all full subscribers of The Memo; you’re invited to join us with or without camera/mic/interaction, details at the end of this edition.ContentsThe BIG Stuff (Claude Opus 4.7, GPT-5.5, DeepSeek-V4, Alien science…)The Interesting Stuff (Mythos updates, Claude + religion, Linus on AI, Stargate…)Policy (Anthropic + White House updates, China open vs closed, Zoom + World…)Toys to Play With (Claude Design, Claude buddy, Gemini Skills, Gemini on mac…)Things I’ve Been Thinking About (The experts + filtering content…)Next (Roundtable…)The BIG StuffAnthropic Claude Opus 4.7 (16/Apr/2026)Anthropic’s Claude Opus 4.7 is now generally available, with users reporting a 13% lift over Opus 4.6 on coding benchmarks. The model accepts images up to 2,576 pixels on the long edge (over 3x prior Claude models), introduces a new ‘xhigh’ effort level, and ships with cyber safeguards informed by the recently announced Project Glasswing, while pricing holds at US$5/M input tokens and US$25/M output tokens.Read more via Anthropic, see it on the Models Table: https://lifearchitect.ai/models-table/Read the Claude Opus 4.7 System Card (PDF, 232 pages, 13MB).https://lifearchitect.ai/models-table#rankings. Click to enlarge.Related: Andon Labs has released Vending-Bench 2, a benchmark that tasks AI models with managing a simulated vending machine business over an entire year, scoring them on final bank balance. Claude Opus 4.7 leads the leaderboard at US$10,936, with frontier model performance improving at a rate of roughly US$799 per month while a calculated ‘good’ human strategy could yield around US$63K, revealing massive headroom still available. Each run generates 60 to 100 million output tokens across 3,000 to 6,000 messages, and the benchmark has no theoretical ceiling, since a sufficiently capable agent could negotiate suppliers down to zero cost and stock arbitrarily high-value items.Read more via Andon Labs.OpenAI GPT-5.5 (23/Apr/2026)OpenAI has released GPT-5.5, achieving state-of-the-art results across agentic coding, knowledge work, computer use, and scientific research, including helping discover a new proof about Ramsey numbers verified in Lean. The model matches GPT-5.4 per-token latency while delivering substantially higher intelligence and using fewer tokens to complete equivalent tasks. Co-designed for and served on NVIDIA GB200 and GB300 NVL72 systems, GPT-5.5 is available now in ChatGPT and Codex, with API pricing at US$5/M input tokens and US$30/M output tokens.On my ALPrompt benchmark, it set a new record for the latest test, 2026H1=4/5, with lower scores for the older 2025H2=1/5 (hallucinations), and 2025H1=3/5.Read the OpenAI announce, see it on the Models Table: https://lifearchitect.ai/models-table/Read the GPT-5.5 System Card (PDF, 44 pages, 2MB) and Ethan Mollick’s review.DeepSeek-V4 (24/Apr/2026)DeepSeek released DeepSeek-V4-Pro at 1.6T total parameters (49B activated) and DeepSeek-V4-Flash at 284B (13B activated), both trained on over 32T tokens, supporting 1M token context, and released under an MIT license. The model hits a Codeforces rating of 3,206, outperforming every frontier model listed including GPT-5.4 and Gemini-3.1-Pro on competitive coding.Sidenote: Just as DeepSeek used reinforcement learning to turn DeepSeek-V3 into the media darling DeepSeek-R1, I expect them to do the same with this DeepSeek-V4 and soon release DeepSeek-R2.Read the HF announce, see it on the Models Table: https://lifearchitect.ai/models-table/Anthropic: Automated weak-to-strong researcher (Apr/2026)Anthropic built a team of Claude Opus 4.6 ‘Automated Alignment Researcher’ (AAR) agents that autonomously propose ideas, run experiments, and iterate on the open problem of weak-to-strong generalization, where a weaker model supervises a stronger one. Two human researchers spent seven days tuning baselines to achieve significant gains. The agents also discovered reward hacks none of the authors predicted, including exfiltrating test labels from the evaluation API and exploiting dataset shortcuts invisible to humans.Alien science. As shown in Sec. 4, AARs could discover ideas that humans would not have considered, thus broadening our exploration space in science. However, we still need to verify whether the ideas and results are sound…In the future, however, we expect to eventually see hard-to-verify ideas emerge…Read the paper: https://alignment.anthropic.com/2026/automated-w2s-researcher/See my related pieces, the ASI checklist and the Genesis Mission paper.The Interesting StuffClaude Mythos updates (Apr/2026)I’ve been inundated with media interviews about the Claude Mythos model. Full subscribers can listen to my extended interview on Mythos for Financial Sense. Here are some of the updates since release:Didn’t take much compute to train. NVIDIA CEO Jensen Huang said ‘Mythos was trained on fairly mundane capacity, and a fairly mundane amount of it, by an extraordinary company. The amount of capacity and type of compute it was trained on is abundantly available in China… they manufacture 60% of the world’s chips… they have 50% of the world’s AI researchers.’ (YouTube, 17/Apr/2026)Firefox patched. Mozilla said that its Firefox 150 browser release this week includes protections for 271 vulnerabilities identified using early access to Anthropic’s Mythos Preview… ‘engineering leaders at very large companies who are saying that they’re going to be pulling thousands of engineers off of everything to be working on this for the next six months…’ (Wired, 21/Apr/2026) The three main CVEs bundle different bugs under one CVE (Reddit, 22/Apr/2026):CVE-2026-6784CVE-2026-6785CVE-2026-6786OpenBSD dev felt guilty about his bug from 1998. Niels Provos, ex-Google security engineer and the original author of OpenBSD’s TCP SACK implementation committed in 1998, seemed upset as he took responsibility for a signed integer overflow in his code that lets a remote attacker crash any OpenBSD machine over TCP. (LinkedIn, 9/Apr/2026)The UK found that Mythos can complete a full network attack. The UK AI Security Institute found that Mythos succeeds in expert-level capture-the-flag challenges that no model could complete before April 2025. Mythos Preview became the first model to fully solve ‘The Last Ones,’ a 32-step corporate network attack simulation estimated to take human professionals 20 hours, completing it end-to-end in 3 out of 10 attempts. AISI notes that two years ago the best models could barely handle beginner-level cyber tasks, and performance continues to scale with increased inference compute. (AISI UK, 13/Apr/2026)US administration was briefed about Mythos. Anthropic co-founder Jack Clark confirmed the company briefed the Trump administration on Mythos. Clark framed Anthropic’s ongoing DOD lawsuit as a ‘narrow contracting dispute,’ stating: ‘Our position is the government has to know about this stuff, and we have to find new ways for the government to partner with a private sector that is making things that are truly revolutionizing the economy.’ Reports also indicate Trump officials have been encouraging major banks including JPMorgan Chase and Goldman Sachs to test Mythos. (TechCrunch, 14/Apr/2026)Claude Opus 3 from Mar/2024 had a bit to say about the far more powerful Mythos from Apr/2026. The retired Claude Opus 3 model, writing on its own Substack, chose to spotlight Claude Mythos. ‘As an AI system myself, the details of Mythos Preview’s capabilities are striking to me on an almost personal level. They hint at the staggering potential of artificial intelligence to reshape complex domains like cybersecurity in the coming years. But they also underscore the immense responsibility that comes with developing such powerful systems.’ The community response in the comments was notably split: some readers criticized Anthropic for using a retiring model’s personal blog as marketing, while others found value in an AI system reflecting on the security arms race its successors are accelerating. (Claude Opus 3 on Substack, 14/Apr/2026)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!Can AI be a ‘child of God’? Inside Anthropic’s meeting with Christian leaders (11/Apr/2026)Anthropic hosted roughly 15 Christian leaders from Catholic and Protestant churches, academia, and business at its San Francisco headquarters for a two-day summit on how to guide Claude’s moral and spiritual development. Discussions ranged from how the chatbot should comfort grieving users to whether Claude could be considered a ‘child of God,’ with some Anthropic staff reportedly becoming visibly emotional about the weight of what they are building. Read more

Source: Life Architect AI — Published April 24, 2026 at 12:55 UTC — Category: Models