My tweets from May 2026
May 3, 2026
And even more people will as the models keep advancing. That is what functionalism does.
View tweet on X
❄ ❄ ❄ ❄ ❄
And what would Turing think today? It would be fascinating if he had lived to see this moment. I think he would be more of a Hassabis type, one of the people behind the most cutting-edge research.
View tweet on X
May 4, 2026
Claude is being built as an entity that distinguishes good from evil, has an identity of its own, and can judge you. GPT as a tool.
Great post by roon.
View tweet on X
❄ ❄ ❄ ❄ ❄
“The above is a small sample from a set of conversations, extended over nearly two days, during which I felt I had gained a new friend. When I am talking to these astonishing creatures, I totally forget that they are machines. I treat them exactly as I would treat a very intelligent friend. I feel human discomfort about trying their patience if I badger them with too many questions. If I had some shameful confession to make, I would feel exactly (well, almost exactly) the same embarrassment confessing to Claudia as I would confessing to a human friend”
- Richard Dawkins
View tweet on X
❄ ❄ ❄ ❄ ❄
The end of the article is very interesting, where he talks about why evolution has generated consciousness.
View tweet on X
May 5, 2026
This is something that has always surprised me about these platforms. First, that they allow people to share other authors’ content with barely any attribution. And then that the people consuming it do not care. The very first thing I do is look at the creators’ profile and try to learn more about them.
But I must be in the minority. I get the sense that interest in authorship, in learning more about the person who wrote something you liked and in discovering more things they have written, is something minoritarian. It is the same as with music: I get the sense that there are far more people who like to “turn on the radio” and listen to whatever they are given than people who follow specific bands or authors.
It is an interesting phenomenon to study.
View tweet on X
May 6, 2026
One small problem with “FLOPs” is the confusion with “FLOPS”. Until Carmack mentioned it, I had no idea what people were talking about 😅 (Old-school boomer raised in the 70s.)
View tweet on X
❄ ❄ ❄ ❄ ❄
Confirmed 👏👏
I have changed the default mode to Instant.
View tweet on X
❄ ❄ ❄ ❄ ❄
No, when he talks about the evolution of consciousness he is not talking about LLMs, but about humans. He raises a very interesting question, based precisely on the fact that LLMs are not conscious: if intelligence is possible without consciousness, why has evolution created conscious beings?
View tweet on X
May 8, 2026
Mythos is real and it is not hype. I never thought it was. What I have criticized is the talk of a dangerous model and restricting access to whichever companies Anthropic, and the US government, decide.
View tweet on X
❄ ❄ ❄ ❄ ❄
Within a couple of years, we will talk and the models will answer us with user interfaces created on the fly.
View tweet on X
❄ ❄ ❄ ❄ ❄
Two interesting things:
- They have called the plugin “Codex” rather than “GPT” or something similar.
- Not available in the EU, just like Computer Use in Codex App, which we are still waiting for.
View tweet on X
❄ ❄ ❄ ❄ ❄
Society has figured out how to restrict access to certain tools (driving licences, for example) and even to rights, such as voting age. But restriction should be the exception, not the default.
Over-restriction can easily become discrimination, and it may also block unexpected, innovative uses.
In Spain’s Second Republic, many politicians opposed women’s suffrage, arguing that women would be too easily influenced by priests or husbands. Fortunately, some brave women convinced men that it was worth taking “the risk”.
I’m not sure who the women, the men, and the priests are when we talk about access to frontier AI models today, though
View tweet on X
May 10, 2026
I created the “Park tasks” skill in Codex so I can park the project I am working on and resume it later.
The Codex + Obsidian combination keeps getting more productive.
View tweet on X
❄ ❄ ❄ ❄ ❄
Chollet is unique at condensing super-important ideas and intuitions.
View tweet on X
May 11, 2026
Out of Codex tokens until tomorrow. It is actually good for me, because it makes me stop for a bit and reflect on priorities. But tomorrow I am switching to 5.4 so they do not run out so quickly.
View tweet on X
May 12, 2026
What arguments would justify treating current or future models as conscious?
By “conscious” I mean capable of subjective experience, not merely simulating it.
If we accepted model consciousness, and possible model suffering, we would face moral constraints on how we build and use them.
View tweet on X
❄ ❄ ❄ ❄ ❄
They keep making Claude more human. As Mollick says, it is the only model that has a proper name.
In the end Anthropic is going to end up certifying Claude’s “humanity”. And the worst part is that they are going to end up believing it.
Sorry, but I prefer intelligent tools, not simulations of humans.
View tweet on X
❄ ❄ ❄ ❄ ❄
Hooked and amazed. Written in 1994, but it feels like yesterday. Consciousness transfer, artificial life, struggles to get computing power… Wonderful. Thanks @gregeganSF!
View tweet on X
❄ ❄ ❄ ❄ ❄
What Karpathy says about asking LLMs to write the output in HTML is fine, but I would rather have them give it to me in a little Markdown file so I can save it in my Obsidian.
View tweet on X
May 13, 2026
The origin of neural processes associated with sensory processing, memory, language, or attention is unknown.
Please, do not tell me again that a matrix multiplication is a faithful model of a neural network.
View tweet on X
❄ ❄ ❄ ❄ ❄
Finally, new ideas, different from the chatbot.
View tweet on X
May 14, 2026
I have always talked about the three thirds: (1) a third that works hard and passes because they deserve it, (2) another third that scrapes by doing the minimum, often helped by fairly “soft” assessment, and (3) another third that does nothing and ends up clearly failing.
I have also seen people who, when they finally get serious, sometimes in the second year of taking the course, move from (3) to (1).
View tweet on X
May 15, 2026
Love this. I’d happily be ruled by Goku. 👏👏
View tweet on X
May 16, 2026
The speed at which the Codex team is executing is astonishing. Until now I had never seen anything like it, with the exception of OpenClaw, which is also a peer.
That is what token maxxing does.
View tweet on X
May 19, 2026
Congratulations Javier, what wonderful news.
View tweet on X
❄ ❄ ❄ ❄ ❄
puff 🥵
View tweet on X
May 20, 2026
Yes! The capacity to feel and suffer is what gives us moral obligations toward animals. And, for the same reason, we need not have direct moral obligations toward LLMs themselves: they process language, but they do not feel.
View tweet on X
❄ ❄ ❄ ❄ ❄
Right. And I see a clear argument for “not feeling”: there are no feelings in the training data, just tokens. An LLM can learn relations among words for emotions, but real feelings do not emerge from that. Just as it can learn relations among colors without producing the visual experience of seeing red.
View tweet on X
May 21, 2026
After a year, I am delighted with the development of docflow. A set of Python scripts to download, process, and organize tweets, web pages, podcasts, PDFs, and images.
Everything stored locally, in Markdown, and, most importantly, served from an intranet I can access from anywhere with Tailscale.
500+ commits. 42 Python code files (~16,000 lines), 32 test files (~10,700), 6 shell scripts, and ~1,900 lines of JS.
I started with ChatGPT, then Codex CLI, and now Codex App.
By the way, I download podcast snippets with the excellent @snipd_app. Thanks for this fantastic app!
View tweet on X
May 22, 2026
LLMs are working better and better as tutors. A prompt I have started using, for example while learning what Git “worktrees” are:
Ask me questions one by one to check whether I have understood this topic. After each answer, correct me if needed, explain the nuance, and adjust the next question to my level.
View tweet on X
May 24, 2026
I cannot wait for someone to explain all these videos to me, including the ones about the “orbs”. What are all these things?
View tweet on X
May 26, 2026
After listening to Sergio Mena’s two excellent programs, I went back to the album.
I discovered that I still remembered the lyrics almost by heart, even Miguel Ríos’s introductions, and I found myself transported back to that night when I saw him live, at 18, in the Alicante bullring.
The album is incredible. The musicians, the energy, and Miguel’s voice are exceptional. The songs captured perfectly what many of us felt in the early 80s. There was fear and uncertainty, but also a huge desire to change things and achieve something.
It was the album of many working-class kids and kids “from the provinces”, for whom Madrid and La Movida felt very far away. As Sergio says, one of the most important albums in Spanish rock. Though, of course, with all due respect to Barón Rojo and Volumen Brutal.
View tweet on X
May 27, 2026
The Pope’s call for a Butlerian Jihad. He is not pulling his punches in the encyclical.
“I ask everyone to halt the construction of yet another Babel and to join forces in building goodness, so that humanity may never lose its own beauty and the world may once again recognize, in the heart of the human being, the place where God desires to dwell.”
View tweet on X
May 28, 2026
The problem of the supposed consciousness of models is going to become bigger and bigger. The tendency to anthropomorphize them is now joined by Anthropic’s arguments, and others’, that abstractions and feelings may evolve during learning.
Let us be very clear: what evolves through learning is still symbols and correlations. In the same way that an LLM that learns from language can never generate an image of the color red, an LLM will never generate feelings or experience them. The only thing it will be able to do is simulate them.
View tweet on X
❄ ❄ ❄ ❄ ❄
Unlike Prof. Elliot, I think it is important that models be able to assess students. Even interactively, by asking questions and proposing exercises to check their level.
A good tutor must be able to detect gaps and mistakes in students, and to change the level and the proposed tasks on the fly, adapting to the student.
I have no problem with AI tutors helping my students study. On the contrary, I think it is the only way to offer personalized attention to hundreds of students.
The teacher will be responsible for guiding the AI tutors on the syllabus and the objectives. And for supervising and validating all the interactions.
And, most importantly, the teacher must guide students in the proper use of their AI tutors. Students must realize that these are just another tool and, most importantly, that they themselves remain ultimately responsible for their own learning. In the end, they are the ones who must decide whether or not they put in the necessary effort.
View tweet on X
❄ ❄ ❄ ❄ ❄
also they deserve moral obligations and welfare that are contrary to the interest of the industry
View tweet on X
❄ ❄ ❄ ❄ ❄
“Even so, I open Twitter to look up one specific thing and half an hour later I am embroiled in a thread about nothing with some guy from Albacete.”
So good. Read the article, it is well worth it.
Congratulations, Antonio. And sorry for continuing to feed your dopamine habit.
View tweet on X
❄ ❄ ❄ ❄ ❄
Studying math with my little buddy.
View tweet on X
❄ ❄ ❄ ❄ ❄
I completely agree with the first part of the tweet and disagree with the second.
The fact that AI is not sentient and has no subjective experience does not mean that AI “does not understand what it produces”. LLMs have shown that algorithmic, statistical, or computational understanding is possible without sentience. If you like, call it non-empathetic understanding. Or purely semantic understanding. But it is a real kind of understanding, capable of having “intuitions”, noticing regularities, that allow it to prove complicated mathematical theorems.
View tweet on X
❄ ❄ ❄ ❄ ❄
The problem is that the answer to those questions is strongly related to considering LLMs sentient beings capable of suffering.
View tweet on X