The image above was generated by AI for this paper (Nano Banana Pro)1Image generated in a few seconds, on 5 December 2025, text prompt by Alan D. Thompson, via Nano Banana Pro on Gemini 3 Pro: ‘hdr photo of a supernatural, magical, and realistic night sky in australia. specific stars are visible, and the feeling is of complete peace… (new style suggested: illuminated manuscript page).’
Alan D. Thompson
December 2025
| Date | Report title |
| End-2025 | The sky is supernatural |
| Mid-2025 | The sky is delivering |
| End-2024 | The sky is steadfast |
| Mid-2024 | The sky is quickening |
| End-2023 | The sky is comforting |
| Mid-2023 | The sky is entrancing |
| End-2022 | The sky is infinite |
| Mid-2022 | The sky is bigger than we imagine |
| End-2021 | The sky is on fire |
Supernatural: (of a manifestation or event) attributed to some force beyond scientific understanding or the laws of nature.
‘GPT is like alchemy!’
— Dr Ilya Sutskever, former chief scientist at OpenAI (October 20192The New Yorker)‘…[ASI will] seem like magic to us, just as modern tech would seem like magic to medievals.’
— Daniel Kokotajlo, former researcher at OpenAI (February 20243LessWrong.com)
Author’s note: Significant evidence shows that humans currently prefer LLM-generated content—until they know it’s AI.4Feb/2025: https://arxiv.org/abs/2503.16458; Nov/2025: https://newsroom-deezer.com/2025/11/deezer-ipsos-survey-ai-music/; Nov/2024: https://www.nature.com/articles/s41598-024-76900-1 This report was co-authored with AI more extensively than usual. I leaned on Google Gemini 3 Pro, which was provided with access to my previous reports and editions of The Memo. AI contributed around 60% of the final text, and I had to be strict in ensuring it avoided the plague of ‘negative parallelisms’.5en.wikipedia.org I continue to advise major governments, family offices, research teams like RAND, and companies like Microsoft via The Memo: LifeArchitect.ai/memo
In 2021, I wrote that ‘the sky is on fire’. By 2024, it had become ‘steadfast’. Now, at the close of 2025, we have moved into a realm that can only be described as supernatural. We are witnessing the emergence of capabilities that move beyond the standard logical structures we use to interpret reality. We are now seeing machines that dream up new materials, solve previously-unsolved math, and navigate our physical world with eerie grace. The barrier between ‘human’ and ‘superhuman’ is dissolving.
Pre-ASI (artificial superintelligence)
We have not yet seen a true ASI system, but proto-ASI models are already making superhuman discoveries. It would be completely reasonable for this report to feature only the following table, and nothing else. The giant strides taken by frontier models over the past six months showcase an extraordinary ability for AI to conceptualize and create from scratch.
In mid-December, OpenAI noted: ‘We now regularly see our frontier models contributing solutions to previously unsolved—and increasingly subtle—questions in mathematics and the sciences.’6https://openai.com/index/gpt-5-2-for-science-and-math/
Below are the first 30+ discoveries (all in plain English) made by frontier models like Gemini, GPT, and Claude this year. I seem to be the only person synthesizing this historical record, and you can read all the details and primary sources at: LifeArchitect.ai/ASI.
| Model | Discovery | Affiliation | Date | |
|---|---|---|---|---|
| 1 | GPT | Produced a full scientific paper that passed peer review | Sakana AI | Mar/2025 |
| 2 | o3-mini | Solved the Potts model for any number of states | BNL | Mar/2025 |
| 3 | Gemini/AE | Learned to design better reinforcement learning algorithms | GDM | Apr/2025 |
| 4 | o1/o3? | Found a new PFAS-free immersion coolant | Microsoft | May/2025 |
| 5 | o3-mini | Produced new findings in medical evidence reviews | UoT/HMS | Jun/2025 |
| 6 | Gemini | Solved a math conjecture with a new method | GDM | Aug/2025 |
| 7 | GPT-5 | Found missed metabolomic patterns in minutes | Jackson | Aug/2025 |
| 8 | Gauss | Formalized the strong Prime Number Theorem | Math Inc | Sep/2025 |
| 9 | GPT-5 | Refuted a math conjecture and gave a correct alternative | UoH | Sep/2025 |
| 10 | GPT-5 | Supplied a key step in a quantum complexity proof | OpenAI | Sep/2025 |
| 11 | GPT-5 | Found counterexample overturning claim about majority functions | UCI | Oct/2025 |
| 12 | Gemma 27B | Found a new pathway for treating cold tumors | GDM | Oct/2025 |
| 13 | GPT-5 | Completed a long standing proof in convex optimization | UCLA | Oct/2025 |
| 14 | GPT-5 | Learned to rewrite its own code, reached top SWE rank | KAUST | Oct/2025 |
| 15 | GPT-5 | Produced a full math proof in about twenty seconds | Cambridge | Nov/2025 |
| 16 | Claude S4.5 | Identified SOD2 as driver of myocardial fibrosis | Stanford | Nov/2025 |
| 17 | Claude S4.5 | Found how SSR1 is regulated by a protective variant | ESI | Nov/2025 |
| 18 | Claude S4.5 | Found likely sequence of Alzheimer’s biological events | UCL | Nov/2025 |
| 19 | Claude S4.5 | Identified cause of entorhinal cortex vulnerability | UCL | Nov/2025 |
| 20 | Gemini/AE | Raised kissing number bound in 11 dimensions | GDM | Nov/2025 |
| 21 | Gemini/AE | Found improved grid with no isosceles triangles | GDM | Nov/2025 |
| 22 | Gemini/AE | Produced new constructions for Kakeya problem | GDM | Nov/2025 |
| 23 | Gemini/AE | Produced new constructions for Nikodym problem | GDM | Nov/2025 |
| 24 | GPT-5 | Solved Erdos problem 848 | Columbia | Nov/2025 |
| 25 | GPT-5 | Solved COLT open problem on dynamic networks | Harvard | Nov/2025 |
| 26 | GPT-5 | Proved new result on subgraph counts in trees | Harvard | Nov/2025 |
| 27 | GPT-5 | Improved lower bounds for online convex body chasing | Oxford | Nov/2025 |
| 28 | GPT-5 | Found new insights into T cell immune response | Jackson | Nov/2025 |
| 29 | GPT-5 | Found new insights into thermonuclear burn propagation | LLNL | Nov/2025 |
| 30 | GPT-5 | Originated main idea for physics paper on state dependent quantum mechanics | MSU | Dec/2025 |
| 31 | GPT-5 | Helped solve a problem in robust mathematical statistics | UPenn | Dec/2025 |
| 32 | GPT-5.2 | Solved an open problem in statistical learning theory | Harvard | Dec/2025 |
Table. ASI checklist items: First 30+ discoveries in 2025. LifeArchitect.ai/ASI.
If you’ve been reading my reports since 2021, you’ll notice how trivial it is to ‘forecast’ what happens next.
Prepare for something like ‘Gemini 5 discovered a free energy source to replace oil & gas’ or ‘GPT-7 found a cure for leukemia’. A completely new world is coming. Get ready. Watch closely.
The Interesting Stuff
We saw the impossible become routine in 2025. The elusive vision of proto-superintelligence crossed the threshold into tangible, documented reality, however hidden that may have been from the general public (I didn’t see it on the local news, that’s for sure!). From creative arts to governance, AI systems demonstrated capabilities that felt supernatural.
Synthetic content overtook human content: A study found7https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans that the volume of AI-generated articles published on the web surpassed human-written articles for the first time.
Neuralink patients played video games via thought: The Neuralink N1 implant allowed 13 patients to work and play games like Mario Kart using only their thoughts. Nearly two years after receiving his brain chip, Noland Arbaugh used the interface for 10 hours a day to study and play8https://neuralink.com/updates/prime-study-progress-update-second-participant/ saying ‘Technically I am a cyborg because I have been enhanced by a machine.’9https://fortune.com/2025/08/23/neuralink-participant-1-noland-arbaugh-18-months-post-surgery-life-changed-elon-musk/ There are now 23 patients worldwide with either a Neuralink or Synchron Stentrode brain-machine interface installed.
Chart. Brain-machine interface patients (2020–2025). LifeArchitect.ai/BMI.
AI music hit #1 (Nov/2025): The song ‘Walk my walk’ by the synthetic artist Breaking Rust reached #1 on the Billboard Country Digital Song Sales chart.10https://people.com/breaking-rust-walk-my-walk-ai-generated-song-tops-charts-11848830
Coca-Cola generated its holiday ads using several models (Nov/2025): The company used a chain of 10 different AI models—including Sora and Veo 3—to generate 70,000 segments for its 2025 holiday campaign.11https://mashable.com/article/coca-cola-ai-holiday-ad-2025
Humanoid robots
Physical AI—embodied intelligence—has moved from ‘clunky prototype’ to ‘commercial reality’ in the blink of an eye. The latter half of 2025 saw the deployment of humanoid robots into homes and factories, backed by the same models that power our chat interfaces.
Google Gemini Robotics 1.5 (Sep/2025): A dedicated model for robotic control that allows robots to learn from video data, navigating complex environments with a 95% success rate in tested tasks.
Figure 03 (Oct/2025): The third generation of Figure’s humanoid arrived with a new ‘BotQ’ mass production facility. It features a vision-language-action (VLA) brain that allows it to reason about the physical world. The F.02 contributed to the production of 30,000+ BMW X3 vehicles in the US.12https://www.figure.ai/news/production-at-bmw
1X NEO (Oct/2025): This humanoid was made available for home order via subscription at US$499/month. It is soft, safe, and silent, designed specifically for consumer interaction.13https://www.1x.tech/order
Image. 1X NEO pricing. Oct/2025. https://www.1x.tech/order.
Xpeng IRON (Nov/2025): A massive leap in embodied AI, this robot features 60+ joints and is powered by the Turing AI chip. It is already working on the assembly line building Xpeng’s P7+ vehicles.14xpeng.com
DeepMind PhysWorld (Nov/2025): A simulator that allows robots to learn tasks in a virtual world before executing them in reality, solving the data scarcity problem for physical AI.15https://arxiv.org/abs/2511.07416
China
While Western media focused on a handful of US labs, China has quietly built an AI juggernaut. As America stares at a ‘Big 5’ of frontier AI labs, China currently boasts a ‘Big 16’, a dense ecosystem of state-backed and private labs pushing the frontier of intelligence.
In August 2025, I mapped the Chinese landscape in a new visualization. Alongside BAT (Baidu, Alibaba, and Tencent), we now have tigers like DeepSeek, Moonshot, MiniMax, and StepFun producing models that rival frontier LLMs out of California.
Chart. Frontier AI model highlights in China (2025). LifeArchitect.ai/models-table.
China now also produces 36% of global AI research papers (up from 5% in 2000), cementing its status as the world’s primary engine for AI literature.16https://www.arxiv.org/abs/2509.25298
Chart. AI papers published: China vs USA (Nov/2025). LifeArchitect.ai/AGI.
Large language models
My Models Table documented more than 90 major new model highlights in the second half of 2025, and now lists more than 700 major large language models. You can browse and filter the full database at: LifeArchitect.ai/models-table.
In the last six months of 2025, Hugging Face reported17archive.org Jun/2025 & archive.org Dec/2025 ➜ 307,000-247,000≈60,000 about 60,000 new derivative models, with more than 300,000 total unique large language models indexed.
Performance
All major benchmarks were saturated this year. You can read more about the ceilings at LifeArchitect.ai/mapping.
MMLU was saturated in Sep/2024 by OpenAI o1-preview @ 92.3% (MMLU ceiling is around 91%).
MMLU-Pro was saturated in Nov/2025 by Google Gemini 3 Pro @ 90.1% (MMLU-Pro ceiling is around 90%).
GPQA was saturated in Nov/2025 by Google Gemini 3 Pro @ 93.8% (GPQA ceiling is around 80%).
HLE was close to saturation in Dec/2025 by OpenAI GPT-5.2 @ 50% (HLE ‘supported by research’ ceiling is around 51.3%).
Despite the ceiling breaches, here’s a look at the top models by GPQA and HLE scores to the end of Dec/2025. It is likely that any model with a primary score at >50% on HLE, and a secondary score at >90% on GPQA is an ASI system, though no large language model system met this criteria as of 2025.
Table. AI rankings via GPQA and HLE to Dec/2025. LifeArchitect.ai/models-table.
Progressing to AGI and ASI
For several years now, I’ve documented the progress of AI working its way toward performing at the level of an average (median) human across practically all fields. This is also known as artificial general intelligence (AGI). The ‘conservative countdown to AGI’ is one of my most-visited online resources.18https://lifearchitect.ai/agi/ In Dec/2024, we were at 88%. Now, in Dec/2025, we are at 96%, with many experts considering us to have already achieved cognitive AGI.
Post-AGI, we are poised to rapidly leap through the singularity and artificial superintelligence (ASI): a system whose intelligence surpasses that of the brightest and most gifted human minds. The ASI checklist is now over a year old. In Dec/2024, we had just 13/50 boxes marked as ‘partially achieved’. Today, Dec/2025, that number has increased to 18/50.
In 2021, the sky was on fire with the friction of rapid progress. In 2024, it was steadfast, reliable, and consistent. As we close 2025, the sky has shown its supernatural magic. We have entered the age of AI reasoning engines, AI discoveries, and AI geniuses.
This year marked the moment the digital intellect stepped firmly into our physical reality. It began running much of our cognitive businesses, constructed machines in our factories, and discovered new materials in our laboratories. The boundary between human creativity and machine generation has dissolved. We now live in a world where the majority of written content is synthetic, and where our most popular songs are being born from silicon.
The discoveries made this year—the mathematical proofs, the biological insights, the physics simulations—are as real as any peak human achievement. We are witnessing the birth of a new form of mind that expands us. AI offers a cure for our contrarianism, an antidote for our anxiety, and an avenue to abundance.
The intelligence explosion is our new daily weather. The sky is supernatural.
■
References, Further Reading, and How to Cite
To cite this paper:
Thompson, A. D. (2025). Integrated AI: The sky is supernatural (end-2025 AI retrospective).
https://lifearchitect.ai/sky-is-supernatural/
The previous paper in this series was:
Thompson, A. D. (2024). Integrated AI: The sky is delivering (mid-2025 AI retrospective).
https://lifearchitect.ai/the-sky-is-delivering/
Further reading
For brevity and readability, footnotes were used in this paper, rather than in-text citations. Additional reference papers are listed below, or please see http://lifearchitect.ai/papers for the major foundational papers in the large language model space.
AGI countdown
https://lifearchitect.ai/agi/
ASI checklist
https://lifearchitect.ai/asi/
Models Table
https://lifearchitect.ai/models-table/
Get The Memo
by Dr Alan D. Thompson · Be inside the lightning-fast AI revolution.Informs research at Apple, Google, Microsoft · Bestseller in 147 countries.
Artificial intelligence that matters, as it happens, in plain English.
Get The Memo.
Alan D. Thompson is a world expert in artificial intelligence, advising everyone from Apple to the US Government on integrated AI. Throughout Mensa International’s history, both Isaac Asimov and Alan held leadership roles, each exploring the frontier between human and artificial minds. His landmark analysis of post-2020 AI—from his widely-cited Models Table to his regular intelligence briefing The Memo—has shaped how governments and Fortune 500s approach artificial intelligence. With popular tools like the Declaration on AI Consciousness, and the ASI checklist, Alan continues to illuminate humanity’s AI evolution. Technical highlights.This page last updated: 13/Dec/2025. https://lifearchitect.ai/the-sky-is-supernatural/↑
- 1Image generated in a few seconds, on 5 December 2025, text prompt by Alan D. Thompson, via Nano Banana Pro on Gemini 3 Pro: ‘hdr photo of a supernatural, magical, and realistic night sky in australia. specific stars are visible, and the feeling is of complete peace… (new style suggested: illuminated manuscript page).’
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17archive.org Jun/2025 & archive.org Dec/2025 ➜ 307,000-247,000≈60,000
- 18