What did you do in the AI Revolution, Daddy?
Pre-GPT-5, Post AGI? For too many, hope is the largest part of their AI strategy. We are still under-estimating the speed of AI progress, and the profundity of its implications.
This article is really ‘just’ an update on AI progress. All the developments here are from the last three weeks or so. I would go back meticulously to the last time I posted, but the article would be too long and the point is made more neatly here by not trying to cover everything. Progress is still astonishing.
Why now? Well, rumours are rife that GPT-5, the model that OpenAI CEO Sam Altman says ‘scares him’ that seems to have convinced him we are past AGI, and now the target is ASI- super-intelligence (variously defined1), is said to be ready for roll-out and release this month (August 2025). Indeed GPT-5 and various derivatives may have been trial-released in LLM Arena last weekend, as Zenith, Summit, possibly Lobster, Nectarine, Starfish, and o3-alpha, drawing breathless commentary for their performance2.
After GPT-5 release happens, there’s a very real risk and high likelihood the remarkable recent progress is eclipsed and under-discussed. As Rob Bassett-Cross (CEO Adarga AI) and I wrote in our eponymous article ‘Don’t Blink’ - AI watchers daren’t look away for a moment such is the pace of progress. My concern remains that as a country, a society, and in many of our companies, we in the UK are still under-estimating the speed of AI progress, and the profundity of its implications.
This under-estimation is despite the discussion on X and Substack beginning to break into the mainstream. Tom Whipple wrote in the Times last week (25 July):
‘I know the riposte. Humans always fear machines. …AI won’t take your job, goes the mantra, but if you don’t get with the revolution, someone using AI will. My response? You’re not listening.’
But even with views like this in the thumpingly centrist Times, nothing really changes in UK policy circles. But it changes everywhere else.
In the past few weeks for example, an essay collating the thoughts of leading Chinese thinkers on AI - shows their emphasis on Open Source, diffusion, ‘talent’ as espionage vector, and regulatory standards as the route to values dominance. A hyper-pragmatic, and believable approach to the otherwise unequal competition with US Big Tech. It is free of the delusions and not-really-even half-measures dressed up as solutions that characterise the UK’s response (and that get almost no scrutiny or criticism from our media for which, Whipple not withstanding, this remains a fringe issue - as opposed to perhaps the most disruptive technological revolution in human history).
We’ve seen the Chinese Premier Li Qiang announce a plan (consistent with the above) for China to launch an ‘Organisation for AI Cooperation' based out of China but with global aspirations.
We see in China a growing focus on ‘embodied AI’ - accelerating diffusion of AI into its industrial base as a way to advantage, perhaps also making the vast amounts of data it collects available, to influence future models.
That’s just China. RUSI reports on the scale of US deals with the Middle-East on AI investments as ‘staggering’. But these investments are not staggering - they are a proportionate reaction to the continued acceleration of AI progress, an anticipation of the rate, direction and implications of this progress - they are other nations’ recognition that they can’t afford to be left behind.
Yesterday (31 July 25) we read that Norway has struck a deal with the US to build one of its Stargate AI projects (the US ‘Manhattan Project for AI’ announced in January) in country, a gigafactory, an AI data centre entirely powered by renewable energy, to supply Europe. Norway and US business are investing $1bn in the first phase of this project. ‘Europe needs more compute’, Sam Altman notes. More than just ‘more compute’, we - UK and Europe alike - need a credible plan proportionate to the threat and opportunity posed by AI.
Here’s why. In the past few weeks:
Nature Chemical Engineering reports that self-driving (AI/automated) labs are producing new materials, in the field of materials science, ten times faster, processing ten times the data, accelerating discovery exponentially.
Nature Communications reports that self-driving labs have discovered in the venom of snakes and spiders, 2000 new antibacterial proteins that could serve as alternatives to current antibiotic treatments, 58 of these have been tested so far, 53 had medical applications such as killing the currently drug-resistant bacteria e-coli and staph - vital in a world of growing anti-microbial resistance.
A survey published back in March showed Agent-based AI models’ increasing abilities in completing software development tasks. In 2022 they could complete at the touch of a button tasks that took humans 30 seconds. In 2025 they are completing tasks that take 1-hour. Their performance was doubling every 12-months, but is now doubling every 7 months. On this current trend, even without further acceleration in the rate, by 2027 they will be able to complete tasks that take humans a month. By 2029 tasks that take humans a working year.
I re-report that because now, further research shows these improvements in Agentic AI ‘task scaling’ are also being seen with Agentic AI in scientific reasoning, maths, robotics, the use of computers for tasks, and in self-driving systems.
Most significantly, in my view, since last I wrote, mathematicians have had their Gary Kasparov/Lee Sedol moment, when AI surpasses them and they wonder what they are for. This was triggered by ‘general’ models like OpenAI’s latest reasoning model and an advanced form of Google’s Gemini winning ‘gold medals’ at the International Maths Olympiad (IMO). The achievement prompted both amusement and amazement. One Data Science Professor posted about the poor performance and limitations of general AI models on Olympiad tasks just hours before their success was announced and was much mocked. In his defence, this was faster progress than some of the most bullish experts had predicted, and is another example where progress has been consistently under-estimated. The question ‘When will an AI win a Gold Medal in the International Math Olympiad?’ was forecast as follows:
July 2021: 2043 (22 years away)
July 2022: 2029 (7 years away)
July 2023: 2028 (5 years away)
July 2024: 2026 (2 years away)
July 2025: AI wins a gold medal in the International Math Olympiad.
This is not the first time I have written out these kinds of estimates to illustrate how difficult humans find it to understand exponentials, to forecast AI progress.
On Friday, seven days before I clicked ‘publish’ on this post, AI performance on the ARC AGI 2 benchmark [‘Abstraction & Reasoning Corpus for Artificial General Intelligence 2’ – an AGI test redesigned after ARC AGI 1 was surpassed in December 2024] jumped from 16.5% to 19%. ‘Only’ a 2.5-point gain, it represents a >15% relative improvement in capability. If we assume exponential growth here too, given this benchmark was launched end March 2025, and models were scoring ~12% by end April, it could be fully solved by ~August 2026.
We’ve seen Deepmind’s Aeneas AI historian decode and complete inscriptions and text, adding nuance and contextualised understanding. We saw Deepmind launch AlphaEarth - where the impact on espionage and military operations should be obvious - take this summary from one commentator:
'AlphaEarth Foundations does something clever -- instead of drowning in petabytes of Earth observation data, it creates compact summaries of every 10x10m square on Earth by fusing optical, radar, LiDAR, and climate data. ...it can see through clouds in Ecuador and reveal hidden agricultural patterns in Canada.’
As I’ve written before, only in the world of AI could developments like these two new Deepmind models and milestones be essentially addendum to the main points. More likely than not, all the developments reported herein will soon be eclipsed by and quickly forgotten after the release of GPT-5.
A reminder - as I wrote in In Athena’s Arms - ‘…this is not a millenarian declaration of deterministic certainty.’ Maybe progress will stop. Maybe someone will declare ‘AI will never’ - like chess, Go, the IMO - and this time be right. Maybe someone will point to a fundamental barrier to AI progress that we can’t innovate our away around, through or over. But right now, the ‘straight lines on graphs’ that have driven and reflect AI progress for almost two decades, show no signs of stopping. Very few people seem to be taking this as seriously as you might expect given the evidence - trying to plan, build, invest, govern, write, report, or converse as if AGI could be with us imminently, or in the next few years. Hope is for too many, the entirety of their AI strategy.
We are still under-estimating the speed of AI progress, and the profundity of its implications.
There is a great tyranny of definitions here, explained at length in previous posts’ footnotes. You’re on your own this time, unless you want to go back and check the previous caveats and explanations.
e.g. https://x.com/kimmonismus/status/1949037129163547083, https://www.reddit.com/r/singularity/comments/1m9v5s5/what_model_is_summit_on_lmarena/ & https://www.linkedin.com/posts/emollick_kinda-amazing-the-mystery-model-summit-activity-7355073850738413568-H_pc