# Three years of estrategIA: the time we are letting slip away

Author: Fernando Nieto Lobato
Original publication: 2026-10-07
Spanish original: https://estrategiabyaleph.substack.com/p/estrategia-158-de-una-ia-que-fallaba
English URL: https://elcontemplador.github.io/estrategia-english/essays/158/
Status: Published translation

This is a translation of the original Spanish essay published on 7 October 2026. Its claims, examples and forecasts retain that historical context.

English publication: 2026-10-07

Three years ago, in October 2023, we started estrategIA with a conviction: what was happening with artificial intelligence deserved far more political attention. Even then, when very few of us were using it intensively, AI was showing signs of becoming a disruptive technology that would spread incredibly quickly and change the world. From 2023 onwards, week after week, we began asking how AI would transform our lives, our politics and our governments, and how to adapt well to it. That remains this newsletter's main purpose. What has changed is the urgency.

In 2023, we marvelled at systems that could write fluently and [get a simple addition wrong](https://aclanthology.org/2023.rocling-1.22/). The gap between their fluency and their mistakes invited scepticism. That was reasonable. What is becoming less reasonable is still having that AI in mind, as many people do, when we talk about the AI in front of us today.

On 8 September, [OpenAI presented a proposed solution to one formulation of the Navier–Stokes problem](https://openai.com/index/navier-stokes-solution/), obtained by around 10,000 agents in 88 hours and formalised in Lean. It still requires independent mathematical scrutiny, but the leap is striking: in three years, we have gone from pointing out arithmetic errors to discussing whether an AI has solved, as it appears to have done, a Millennium Prize Problem.

And this is no exception. As the next image shows, the latest AI models are completing benchmark tasks that some of the world's best mathematicians predicted, barely a year and a half ago, would withstand AI for years.

![Screenshot of a post by Haider comparing mathematicians' earlier expectations for FrontierMath with a later Tier 4 leaderboard.](https://elcontemplador.github.io/estrategia-english/assets/images/158-01.png)

*English transcription of the Spanish post in the original screenshot; translated from the displayed Spanish, not verified against the original English post:*

> Your daily reminder that exponential progress can fool even Fields medallists.
>
> December 2024: Terence Tao and Timothy Gowers called FrontierMath Tier 3 exceptionally challenging.
>
> Tao said it could “resist AIs for several years”. Gowers said that solving even a few problems would be “a genuine breakthrough”.
>
> Less than two years later, even Tier 4 is saturated.

*The embedded excerpts and leaderboard are already in English. The screenshot identifies the account as Haider, @haider1, and dates the post to 5 October 2026. These statements are reproduced as part of the original article's illustration.*

Only a few hours before we send this newsletter, [OpenAI has published 372 results in a GitHub repository covering advances across a great many branches of mathematics](https://openai.com/index/sharing-ai-progress-in-mathematics/). On social media, many experts are already calling it the most important day in the history of the discipline (we say more in the news section of the original newsletter).

And the leap is beginning to extend beyond the screen. [This year, an agent system developed by Chinese researchers and Alibaba's DAMO Academy analysed 2.4 million crystal structures and helped identify four new superconducting materials that were subsequently synthesised and experimentally verified](https://arxiv.org/abs/2604.23758). AI is not only beginning to solve problems: it is also starting to guide which experiments are worth carrying out in the physical world.

For our [second anniversary](https://estrategiabyaleph.substack.com/p/estrategia-105-dos-anos-de-estrategia), we tried to illustrate that leap with five charts. Looking back at them, I am struck by how much the scale of the conversation has changed in barely a year. In [issue 100](https://estrategiabyaleph.substack.com/p/estrategia-100-tras-un-centenar-de), we wrote that when AI accelerates, politics cannot wait. I wish that warning had aged less well.

Today, even the release calendar is bewildering: [OpenAI launched GPT-6 Astra on 3 September, GPT-6 Sol on the 22nd and GPT-6.1 Sol on the 29th](https://developers.openai.com/api/docs/changelog). One week between the two versions of Sol. Each release has a different scope, but their succession helps explain how difficult it is to keep up, and how the pace keeps accelerating.

![English-language post by josh, @jfonsecarivera, showing a chart in which the median gap between model releases from Anthropic and OpenAI falls from 70 days to 11 days.](https://elcontemplador.github.io/estrategia-english/assets/images/158-02.png)

And it is not just OpenAI and Anthropic. Across the companies covered by [AI Release Tracker](https://aireleasetracker.com/analytics), recorded releases rose from 22 in 2023 to 92 in 2025. By 1 October 2026, the total for that year had already reached 83. The count includes updates and variants, but it gives a sense of how much activity in the sector has intensified.

![AI Release Tracker screenshot with an English-language stacked chart of releases by lab over time, reporting 22 releases in 2023, 92 in 2025 and 83 in 2026 through 5 October.](https://elcontemplador.github.io/estrategia-english/assets/images/158-03.png)

But September also brought a decision with clear political implications: [OpenAI cancelled the GPT-6.1 Astra launch planned for early October because it had not met its internal safety standards](https://www.reuters.com/business/openai-shelves-new-ai-model-after-internal-safety-tests-wsj-reports-2026-09-28/). The decision brings back a question we have raised here before: who should assess these risks, and against what public criteria?

Beneath the releases, something deeper is happening: AI is beginning to [improve algorithms and components of the infrastructure used to develop it](https://deepmind.google/blog/alphaevolve-impact/). A concrete example: Google DeepMind explained that [AlphaEvolve made a key operation in Gemini's training 23% faster, reducing total training time by 1%](https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/). That may seem a small percentage, but these are processes that consume enormous resources.

These are the first steps towards RSI (recursive self-improvement): one AI helps build a more capable AI, which in turn can improve the next. This possibility of accelerating its own development is one of artificial intelligence's most transformative characteristics. My conviction, in line with ideas put forward by [Demis Hassabis](https://www.gsb.stanford.edu/insights/demis-hassabis-thinks-were-foothills-singularity), [Sam Altman](https://blog.samaltman.com/the-gentle-singularity) and [Jack Clark, co-founder of Anthropic](https://jack-clark.net/2026/05/26/import-ai-458-reckoning-with-the-future-and-a-singularity-story/), is that we are entering the early stages of what is usually called the singularity: a dynamic in which AI progress feeds its own acceleration and makes what comes next increasingly difficult to anticipate. That is my interpretation of the signals. Available energy, chips and system reliability continue to impose limits. No one has a reliable timetable for its consequences, but it seems very clear that we are accelerating towards a world that, in a few years, will probably look unrecognisable to us in many respects.

## Preparing to govern in the fog

That is where the political problem lies. We can be convinced of the direction of change while remaining deeply uncertain about what will happen in six months. There will be uneven advances, failures and obstacles. Meanwhile, training professionals, transforming a public administration or agreeing how to distribute wealth will still take years of work. A new model version does not, on its own, shorten these processes.

In [issue 138, “Governing in the fog”](https://estrategiabyaleph.substack.com/p/estrategia-138-inteligencia-artificial), we brought into politics an argument that [Toby E. Stuart made for business in Harvard Business Review](https://hbr.org/2026/04/the-future-is-shrouded-in-an-ai-fog). Governing in these conditions requires institutions able to observe, learn and correct course much faster, while maintaining a recognisable democratic direction: protecting rights, expanding opportunities and distributing the benefits. We can prepare several responses, test them and agree which changes in hiring or productivity would justify putting them into effect. Waiting for all uncertainty to disappear means surrendering the initiative to those who are already acting.

We could start this very year with three things: tracking, through public data, how employment is changing in the most exposed sectors; launching pilot projects in public administrations with results that can be evaluated; and strengthening public capacity to examine models, their risks and their opportunities. This work requires trained people, budgets and continuity. The longer we take to begin, the more we will depend on other people's decisions.

Take universal basic income as an example. I think we should begin to consider it seriously as one possible response and, at the very least, have it studied and ready, as I argued in [issue 156](https://estrategiabyaleph.substack.com/p/estrategia-156-como-hacer-segura). An AI capable of producing more wealth should enable us to live with greater security and freedom, including the freedom to care for others, study or leave a job we hate. Whether that happens will depend on political decisions about who receives that wealth.

Preparing a universal basic income, or policies of comparable significance, means discussing how to fund it, its effects on different households and the additional support some people will still need. It also means defining how to implement it and under what conditions to launch it. We can do that work in advance while observing how employment changes. If we wait until a crisis forces us to act, we will have lost the room to test, correct and build agreement.

That is the mismatch that worries me. Public debate can spend months caught up in a controversy while the conditions on which the next parliamentary term's decisions will depend are being transformed. In my notes for this anniversary issue, I put it less diplomatically: “AI moving at an incredible pace, and politicians asleep at the wheel.”

Democracy needs deliberation, listening and safeguards. Precisely because making good decisions takes time, we should start sooner. Preparing now may allow us to discuss these options freely, before urgency or fear ([as we discussed two weeks ago](https://estrategiabyaleph.substack.com/p/estrategia-156-como-hacer-segura)) narrows our possibilities.

But before closing this issue, let me offer some thanks on our third anniversary. Thank you to our editor, Pablo Martín, and our reviewer and proofreader, Sofía García, and also to those who preceded her in that role: Natalia Benito and Andrea Molina. Thank you to all the excellent guest authors who have shared their ideas here, and to the whole team at Institución Educativa ALEPH who have made this journey possible. And above all, thank you to you, who read us, challenge us and recommend us every week. A newsletter is also a sustained conversation. Celebrating three years of keeping it going together brings me a joy that is difficult to sum up in a text.

We are three years old, and I still believe it is worth explaining this transformation with curiosity, rigour and hope. I am excited by what we could do with it. I am also impatient with how little we are preparing together. Helping people understand better and decide in time is the best reason I can find for continuing estrategIA. Thank you for being with us. I would be delighted if that note about politicians being asleep at the wheel soon became outdated.

**[Fernando Nieto Lobato](https://www.fernandonieto.es/)**  
Director of Digital Innovation at [Institución Educativa ALEPH](https://institucioneducativaaleph.com/) and director of the estrategIA newsletter.

![Original Spanish infographic summarising three years of AI progress, the gap between technological change and political preparation, and three areas of public action. Full English transcription follows.](https://elcontemplador.github.io/estrategia-english/assets/images/158-04.png)

### English transcription of the original infographic

**Three years of estrategIA: the time we are letting slip away**

**The AI of 2023 is no longer the reference point**

- **2023:** It wrote fluently, but got addition wrong.
- **2026:** It achieves superhuman results on advanced mathematical tasks and opens new avenues for science.

**AI is beginning to accelerate its own development**

AI → better algorithms and infrastructure → more capable AI → back to AI.

We believe we are entering the early stages of the singularity: AI progress is beginning to feed its own acceleration.

**The political mismatch**

- **Technology: weeks.** Releases follow one another; their scope is not the same.
- **Public preparation: years.** Training professionals, transforming public administrations and agreeing how to distribute wealth require continuity.
- **Governing in the fog.** Observe, learn and correct course quickly, without waiting for uncertainty to disappear. Maintain the direction: protect rights, expand opportunities and distribute benefits.

**Three areas of work to begin now**

- **Observe employment.** Use public data to track changes in the most exposed sectors.
- **Test and evaluate.** Promote pilots in public administrations with results that can be evaluated.
- **Strengthen public capacity.** Examine models, risks and opportunities: who assesses them, and against what public criteria?

This work needs trained people, funding and continuity.

**Prepare options before the crisis**

We propose studying universal basic income as one possible response: funding, effects on households, additional support and conditions for implementation.

Doing this work in advance allows us to test, correct and build agreement.

*“Precisely because making good decisions takes time, we should start sooner.”* — translated from the Spanish infographic.

*estrategIA*