My tweets from June 2026
June 1, 2026
Yes, I am sure the advances will end up in everyone’s hands. As has always happened. The advantage here will be Mythos-style: being first and having the “genie” exclusively for a few months. But there will not be too much difference between that AGI and what we are using now, or will be using by the end of the year. I am one of those who think that what we already have is almost AGI, with the right harness and prompts.
View tweet on X
June 2, 2026
Richard Sutton on the limits of generative AI and the need for discovery.
Sutton draws a distinction here between generative AI and an AI truly capable of discovery. Like Chollet, his thesis is that current generative models are, in essence, imitation systems. They learn from large numbers of examples and produce text, images, or video resembling the training data.
That is not necessarily a flaw. When we ask an AI to summarize a document, answer from sources, or help us with existing information, we do not want novelty. We want fidelity. If the model adds something of its own, we call it a hallucination.
The exception appears when we are looking for fiction, entertainment, or creative variation. There we do ask for novelty.
For Sutton, real creativity requires something more: discovery. He defines it as the combination of three steps: variation, evaluation, and selective retention.
It is the pattern of evolution, the scientific method, and ordinary learning: try things, evaluate which ones work, and keep the best.
His central criticism is that generative AI and supervised learning lack evaluation at runtime. Without evaluation there is no selective retention. Without selective retention there is no discovery.
His proposal is to look for an AI that does not merely imitate, but shares goals, tries things, evaluates, and discovers. A scientific AI, not just a generative one.
View tweet on X
❄ ❄ ❄ ❄ ❄
Great project by Jason Snell @jsnell & Mike Hurley. Very happy to contribute and support them, after many years listening to them and enjoying Upgrade.
“Designed in California: An Apple history podcast by Jason Snell” on @Kickstarter
View tweet on X
June 3, 2026
“Do not fall into this insular, fearful, protectionist way of thinking. Programming is evolving. We do not know exactly what its final form will be, but giving more people access to the fruits of the freedoms computing offers is worth resisting the temptation to close the doors of participation.”
View tweet on X
❄ ❄ ❄ ❄ ❄
It looks like they are not going to rename the ChatGPT + Codex super app.
View tweet on X
❄ ❄ ❄ ❄ ❄
Thank you very much! Time to read them.
View tweet on X
June 4, 2026
Ted Chiang, another one joining what we have been saying around here for years: LLMs are not conscious, and playing with that idea can bring us problems.
View tweet on X
June 5, 2026
I remember to open it and clean it!
View tweet on X
June 7, 2026
I did not know that! Same here, stuck with OpenAI out of convenience. In the end the moat is the same as always: convenience.
View tweet on X
June 8, 2026
Gwern has just published another one of those essential articles of his.
View tweet on X
June 9, 2026
What an intelligent human would do is not guess your intentions, but explore them: ask questions, contrast, request examples, detect ambiguities.
Perhaps the next leap is not only that LLMs understand better, but that they learn to ask better questions.
View tweet on X
❄ ❄ ❄ ❄ ❄
A super-interesting thread showing that this is not a matter of one or two people, but that an entire network of corrupt influence and patronage was woven, or that the one that already existed was used. A real and deep cleanup is urgent: everyone who has had contact with this network should resign now.
View tweet on X
❄ ❄ ❄ ❄ ❄
And bravo to @ajsalvador70 for using the few cracks we citizens have to exercise oversight. How nice it would be to live in Sweden, where tax returns are public. Let us see when something similar is done here.
View tweet on X
June 10, 2026
And with new command-line utilities for using models from the terminal and from Python!
View tweet on X
❄ ❄ ❄ ❄ ❄
Tomorrow, Thursday: demonstration to get decent access to Parque del Mar. Remove the tracks! ADIF and the City Council should stop passing the buck.
View tweet on X
June 11, 2026
Apple has presented the third generation of its Apple Foundation Models, AFM 3, at WWDC 2026.
The family includes five models: AFM 3 Core, AFM 3 Core Advanced, AFM 3 Cloud, ADM 3 Cloud for images, and AFM 3 Cloud Pro.
AFM 3 Core is the local base model, with around 3 billion parameters. AFM 3 Core Advanced is a 20-billion-parameter multimodal model designed to run on Apple Silicon devices through a sparse architecture.
In AFM 3 Core Advanced, the full model can reside in flash memory, while only part of the experts are loaded into DRAM for each request. Apple calls this technique Instruction-Following Pruning, IFP. Expert selection is based on the user’s instruction and can be updated during generation.
On the cloud side, Apple introduces AFM 3 Cloud and AFM 3 Cloud Pro within Private Cloud Compute. AFM 3 Cloud Pro is aimed at more complex reasoning tasks and tool-use scenarios. Apple also says that this model runs on NVIDIA GPUs in Google Cloud, inside its private computing infrastructure.
The Foundation Models framework will allow developers to use Apple’s local models, Private Cloud Compute models, and third-party models that implement the LanguageModel protocol.
Apple will soon publish a full technical report, with updated evaluations and benchmarks.
View tweet on X
❄ ❄ ❄ ❄ ❄
Completely agree with Pablo Simón @kanciller
View tweet on X
June 14, 2026
A recent paper on dendritic computation raises a very suggestive idea: a single pyramidal neuron, modeled with considerable biophysical detail, could solve complex nonlinear tasks if a suitable configuration of weights and synaptic locations is found.
The work is very interesting because it proposes some novel ideas and, for me, because it illustrates the large number of open questions in the field of neuronal modeling.
In a biological neuron there is dendritic morphology, electrical compartments, ion channels, excitatory and inhibitory synapses, NMDA receptors, temporal dynamics, structural plasticity, and distributed local nonlinearities.
The work shows that a functional configuration exists in the model, capable of performing computational tasks such as distinguishing dogs from cats in images.
But it leaves many questions open:
It does not show that a real neuron could find that configuration through biological learning mechanisms.
Nor does it answer questions about the stability of that configuration: how much does it tolerate changes in synaptic weights, dendritic locations, molecular noise, spontaneous plasticity, homeostasis, or spine turnover?
Another key question is whether those solutions are very narrow points in parameter space or broad, robust regions. The difference is enormous: a mathematically possible solution can be biologically fragile.
And, in addition, a real neuron is not isolated or dedicated to a single task. It lives embedded in recurrent circuits, with neuromodulation, spontaneous activity, changing brain states, and multiple simultaneous functions.
That is why I find it interesting, because of the questions raised by its proposals: dendrites may greatly expand the computational repertoire of a neuron, but we still know little about how that potential is learned, stabilized, and actually used in the living brain.
It also serves as a reminder of something that is sometimes forgotten: so-called “artificial neural networks” are not realistic models of biological neural networks. They are highly simplified mathematical abstractions, distantly inspired by the nervous system, but far from the biophysical richness of a real neuron.
This is what I have always criticized about the “fathers” of AI: their insistence on using biological metaphors, which only leads to more confusion. “Artificial neural networks” are just matrices implementing a differentiable function. They have nothing to do with real neural networks. It is becoming increasingly important to stress this when many people are playing with the possibility that AIs might be conscious at some point.
View tweet on X
❄ ❄ ❄ ❄ ❄
I have just listened to the interview with Satya Nadella and it is excellent. You could take notes on every question and answer and spend a long time reflecting and researching each one.
View tweet on X
❄ ❄ ❄ ❄ ❄
For those of you who were my students in Programming Languages and Paradigms: this interview with Simon Peyton Jones, co-creator of Haskell, is a good way to return to one of the central ideas of the course: functional programming.
Peyton Jones does not present it as an academic oddity, but as a different way of thinking about programs: less mutable state, fewer hidden effects, more composition, more immutability, and more explicit information in types.
One of the key ideas is that purity forces side effects to be made visible. In Haskell, a pure function and a function that performs input/output do not have the same type: Int -> Int and Int -> IO Int say different things. Effects do not disappear, but the language forces you to declare and sequence them.
He also talks about lazy evaluation, the difference with OCaml, the relationship between functional programming and safety, and why type systems are important for maintaining large programs over years.
And there is a particularly curious ending: Peyton Jones notes that Excel can be seen as a functional language. Each cell is defined by an expression that depends on other cells, without directly modifying its internal state; and with the addition of LAMBDA, Excel can even express user-defined functions.
In the end, the interview recalls something we tried to see in the course: functional programming is not just “using functions,” but a way of reasoning about state, effects, composition, and the guarantees a language can offer us.
View tweet on X
❄ ❄ ❄ ❄ ❄
Very cool, I did not know SPJ until I listened to the podcast. Unlike you, I am not very deep into the Haskell world. I will follow him more from now on. The episode is excellent, and you can tell he is a great communicator, someone who knows a lot and knows how to explain it very well.
View tweet on X
❄ ❄ ❄ ❄ ❄
Antonio, I think your point about an opaque market is exactly right — and I also agree with Bayesian here: what ultimately matters is the readers, not an unreliable detector score.
Ethan Mollick has repeatedly warned that AI detectors are easy to bypass, produce false positives, and can penalize non-native English writers or formal technical writing.
View tweet on X
June 16, 2026
Maybe the “risk” is that OTHERS can fix the security holes that they are using
View tweet on X
❄ ❄ ❄ ❄ ❄
Satya Nadella has rewritten the Pope’s encyclical in LinkedIn form. Where the Pope talks about “humanity”, Nadella talks about “enterprise”. But in the end they are talking about something common: preserving agency versus handing it over to AI.
View tweet on X
❄ ❄ ❄ ❄ ❄
Or, more cynically, preserving market share.
View tweet on X
❄ ❄ ❄ ❄ ❄
New post on my website: my tweets from the first half of June, grouped on a single page.
So anyone, and LLMs, can read my comments without having to go through X.
View tweet on X
❄ ❄ ❄ ❄ ❄
The page is generated automatically from the tweets I collect with my Docflow tool:
View tweet on X
❄ ❄ ❄ ❄ ❄
The “jailbreak” that got Fable shut down is asking it to “fix this code”. What if what they want is for the NSA to keep being able to get into non-American systems? It is conspiratorial nonsense, but…
View tweet on X
June 18, 2026
I read someone, I do not remember whether it was Gwern or Roon, saying that we would know the singularity was approaching when exceptional things started happening.
We already have two so far this month: the US government banning a model, and AI leaders meeting with G7 leaders.
View tweet on X
❄ ❄ ❄ ❄ ❄
18/06 - Kevin Roose leaves the NYTimes and, together with Casey Newton, is taking Hard Fork out of the Times and making it independent.
View tweet on X
June 19, 2026
You show Codex how you do a task on the computer, it records it, understands it as it goes, and at the end generalizes the steps you took so you can save them as a skill. Then you can use that skill so Codex performs the task when you ask it to.
Incredible, astonishing. Another of those things that two years ago we thought were science fiction. I am sure there will be times when it does not quite work, but I am also sure it will keep getting better. Thinking about how this feature can evolve, and what it might be like in another two years, makes your head explode.
For now we have to wait to try it; it is not available in the EU.
View tweet on X
❄ ❄ ❄ ❄ ❄
Hahaha, exactly. I have always seen Hard Fork as the bad English-language copy of Monos. They have even tried to imitate the “puerta grande” sometimes, but they cannot pull it off.
View tweet on X
❄ ❄ ❄ ❄ ❄
Excellent. And if you have not read it, I recommend Greg Egan’s Permutation City. Incredible, amazing. Written in 1994, and it fits the consciousness topic perfectly. I have just read it in English, but there is a Nova translation by Pedro Jorge.
View tweet on X
June 20, 2026
Oooooh
View tweet on X
June 22, 2026
After the Jot Down case and the Leire Diez, Abalos, Zapatero, etc. affair, I am amazed by the lack of critical and ethical capacity among those who are supposed to be society’s “reference points”. We urgently need a generational handover.
View tweet on X
June 23, 2026
Codex tasks run by default on the lighter 5.4 model, and that can sometimes cause problems. A tip: review whether there has been any friction and ask the better model, 5.5 High, to fix the issues so the next run is more correct.
View tweet on X
June 25, 2026
Encouragement to the neighbors, and let us hope that @oscar_puente_ gives more explanations than he did on June 23 in the Senate, where he did not mention anything about the tracks around Casa Mediterraneo.
Remove the tracks and give us access to Parque del Mar now!
View tweet on X
June 30, 2026
We will hold you to your word, minister.
@oscar_puente_ said in the interview:
“These are tracks that have been abandoned for a long time, with a protected area that is Casa Mediterraneo. And what can be developed there is a large park. I think there are few homes planned; I do not know if there are 100 homes there. I do not think this is something we should drag out too much, and let us see if by the end of this year we reach an agreement and can get started.”
At the end of the year we will see whether ADIF and the City Council reach an agreement on the 100 homes, and whether the neighbors and Casa Mediterraneo finally get what we have been fighting for for years.
View tweet on X
❄ ❄ ❄ ❄ ❄
30/06 Apple raises prices on many products, and a widespread memory shortage is looming in 2027.
View tweet on X