subj
schoolboy studying afar, from the past
Across nearly two millennia, a boy’s voice still carries a very familiar frustration. On a scrap of papyrus from 3rd-century CE Egypt, a schoolboy named Thonis wrote to his father:
“Look, this is my 5th letter to you and you’ve only written to me once—not even a word about your welfare, nor come to see me.”
The letter, Papyrus 1575, survives in the British Library, and it offers a deeply human glimpse into everyday life of the Roman Empire. Thonis was studying away from home, and like any child separated from family, he longed for care, attention, and a simple reply. His words echo the universal cry of neglected children everywhere: “Write back!”
What makes this papyrus so extraordinary isn’t just its age, but its tone. Amid grand histories of emperors and wars, here is a boy’s impatience and hurt preserved for 1,700 years. It collapses the distance between then and now: the Roman world, often imagined as marble and legions, is suddenly filled with the ordinary rhythms of family, education, and neglect.
We may never know if Thonis’s father answered him, or if his son’s irritation softened into forgiveness. But the papyrus survives as proof that ancient lives were not so different from ours. Even across centuries, the bonds—and silences—between parent and child remain strikingly familiar.
Would you hear this letter as a complaint, a plea for love—or just the timeless sigh of a boy missing home?
https://papyri.info/ddbdp/sb;3;6262
“Look, this is my 5th letter to you and you’ve only written to me once—not even a word about your welfare, nor come to see me.”
The letter, Papyrus 1575, survives in the British Library, and it offers a deeply human glimpse into everyday life of the Roman Empire. Thonis was studying away from home, and like any child separated from family, he longed for care, attention, and a simple reply. His words echo the universal cry of neglected children everywhere: “Write back!”
What makes this papyrus so extraordinary isn’t just its age, but its tone. Amid grand histories of emperors and wars, here is a boy’s impatience and hurt preserved for 1,700 years. It collapses the distance between then and now: the Roman world, often imagined as marble and legions, is suddenly filled with the ordinary rhythms of family, education, and neglect.
We may never know if Thonis’s father answered him, or if his son’s irritation softened into forgiveness. But the papyrus survives as proof that ancient lives were not so different from ours. Even across centuries, the bonds—and silences—between parent and child remain strikingly familiar.
Would you hear this letter as a complaint, a plea for love—or just the timeless sigh of a boy missing home?
https://papyri.info/ddbdp/sb;3;6262
More Emails to the World
subj
Art <-> Life
This is a steel and bronze sculpture of the Androgen Receptor (ligand binding domain) bound to enzalutamide, a drug that targets the protein and is used to treat prostate cancer. The sculpture was for Charles Sawyers at Memorial Sloan Kettering and was commissioned by Rodrigo Romero, a postdoctoral fellow in the lab.
The protein is made from polished steel. Enzalutamide was made from bronze and the base of the sculpture was made from steel and bronze. The stand is granite.
The protein is made from polished steel. Enzalutamide was made from bronze and the base of the sculpture was made from steel and bronze. The stand is granite.
subj
If All We Build Are LLMs, AI Progress Will Stall, Hit a Wall, and Fall Away
Update on LLMs from Gary Marcus: https://garymarcus.substack.com/p/game-over-for-pure-llms-even-turing
Yann LeCun’s recent comments reflect what many top researchers are quietly acknowledging: large language models (LLMs) like ChatGPT or DeepSeek are powerful tools — but they capture only a thin slice of what true intelligence is.
LLMs are masters of language, not meaning. They weave words with astonishing fluency, yet remain disconnected from the real world — from perception, intention, and consequence. They don’t understand what they’re saying, only how language statistically tends to unfold. In that sense, they’re mirrors of human expression rather than minds that grasp reality.
As LeCun, Yoshua Bengio, Gary Marcus, and others have argued, building real, reliable artificial general intelligence (AGI) will require something very different — architectures designed not just to imitate thought, but to generate understanding. That means integrating:
1. World models that can predict and reason about real situations, not just text (LeCun’s Joint Embedding Predictive Architecture, 2022).
2. Autonomous learning that discovers causal structure instead of memorizing patterns.
3. Energy-based or modular systems that reason, plan, and act coherently within physical and ethical boundaries.
4. Embodied sentience and salience — systems grounded in sensory experience, capable of focusing on what truly matters and aligning ethically with human values.
5. Cognitive world models and evolutionary learning modules — hybrid systems that combine:
• Common-sense reasoning about space, time, and agency,
• Evolutionary and meta-learning algorithms that improve over generations of experience, and
• Analog–digital integration layers that bridge symbolic reasoning with continuous perception.
Together, these advances would give AI something LLMs entirely lack: a living, adaptive model of the world — capable of understanding, learning, and moral reasoning.
Without these deeper ingredients, LLMs will remain syntactic but not semantic — brilliant compressors of language, not genuine thinkers.
And when it comes to Artificial Superintelligence (ASI), the gap grows even wider. Safety and reliability demand systems that can explain themselves, stay grounded in context, and correct their own mistakes. LLMs hallucinate, lack self-awareness, and can’t verify truth — traits that make them fundamentally unsuitable as the core of safe, autonomous intelligence.
In short, LLMs are the spreadsheets of cognition: quite useful, but not the essence of mind. The next breakthrough in AI won’t come from training on more tokens — it will come from discovering the deeper principles of how analog and digital processes integrate, how reasoning connects to the world, and how coherence emerges from interaction.
References
• LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. arXiv:2209.06212.
Yann LeCun’s recent comments reflect what many top researchers are quietly acknowledging: large language models (LLMs) like ChatGPT or DeepSeek are powerful tools — but they capture only a thin slice of what true intelligence is.
LLMs are masters of language, not meaning. They weave words with astonishing fluency, yet remain disconnected from the real world — from perception, intention, and consequence. They don’t understand what they’re saying, only how language statistically tends to unfold. In that sense, they’re mirrors of human expression rather than minds that grasp reality.
As LeCun, Yoshua Bengio, Gary Marcus, and others have argued, building real, reliable artificial general intelligence (AGI) will require something very different — architectures designed not just to imitate thought, but to generate understanding. That means integrating:
1. World models that can predict and reason about real situations, not just text (LeCun’s Joint Embedding Predictive Architecture, 2022).
2. Autonomous learning that discovers causal structure instead of memorizing patterns.
3. Energy-based or modular systems that reason, plan, and act coherently within physical and ethical boundaries.
4. Embodied sentience and salience — systems grounded in sensory experience, capable of focusing on what truly matters and aligning ethically with human values.
5. Cognitive world models and evolutionary learning modules — hybrid systems that combine:
• Common-sense reasoning about space, time, and agency,
• Evolutionary and meta-learning algorithms that improve over generations of experience, and
• Analog–digital integration layers that bridge symbolic reasoning with continuous perception.
Together, these advances would give AI something LLMs entirely lack: a living, adaptive model of the world — capable of understanding, learning, and moral reasoning.
Without these deeper ingredients, LLMs will remain syntactic but not semantic — brilliant compressors of language, not genuine thinkers.
And when it comes to Artificial Superintelligence (ASI), the gap grows even wider. Safety and reliability demand systems that can explain themselves, stay grounded in context, and correct their own mistakes. LLMs hallucinate, lack self-awareness, and can’t verify truth — traits that make them fundamentally unsuitable as the core of safe, autonomous intelligence.
In short, LLMs are the spreadsheets of cognition: quite useful, but not the essence of mind. The next breakthrough in AI won’t come from training on more tokens — it will come from discovering the deeper principles of how analog and digital processes integrate, how reasoning connects to the world, and how coherence emerges from interaction.
References
• LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. arXiv:2209.06212.
subj
LLM and diminishing returns
We are doing the same thing with LLMs. Increasing the model size by 10x will only give diminishing returns.
You have a 3D bust and you want to approximate it as close as possible with a fixed amount of triangles.
Having 600 instead of 60 triangles makes a huge difference, but having 6m instead of 60k triangles doesn’t make a real difference other than being 100x more expensive to compute.
This is exactly the same of what we are seeing with LLMs nowadays. The bust is the data we are scraping from the web, and the triangles are the model weights. And we have the same problem: Increasing the model size by 100x will be 100x times more expensive but we will get improvements that are really hard to notice.
Why? The problem is inherently to the LLM structure.
Models spend the same amount of compute on filler words like “however”, “therefore” or variables when coding or doing math. No one cares which filler word to use, but if you even just flip a sign in a math formula it’s just wrong.
It’s funny how in the model has high entropy on filler words not because it doesn’t know, but because it’s not important which filler word to use (this point extends also to synonyms).
And if you don’t do good data curation, which is currently being neglected by academics and industry (the industry is slowly going more and more) than you will spend a ton of compute during training on things like filler words.
Where do we go from here?
First big step is data curation. Our current datasets are so massive that we rarely do more than a few epochs when training. So we get rid of the easy examples to shrink the dataset to give more compute to the hard samples. This might require something like curriculum learning during pretraining.
Many people thing we have LLMs because of the transformer architecture or some other model idea, but actually it has been data all along. First ImageNet, then web scraping, large scale human preference data, and finally data from verifier systems. You can have a LLM without the transformer architecture, but not without the data. (read https://lnkd.in/eS2YbZgC)
Later down the line we will need to change the model architecture. Token-by-token prediction is slow and error-prone. Language Diffusion models are already showing impressive speed gains because they can unmask multiple tokens per step and have the ability to fix token errors. Now do that in latent space, get rid of the tokenization bias, add variable compute per step and some form of hierarchical representations and voilà, you might get yourself the next era of models.
Act
You have a 3D bust and you want to approximate it as close as possible with a fixed amount of triangles.
Having 600 instead of 60 triangles makes a huge difference, but having 6m instead of 60k triangles doesn’t make a real difference other than being 100x more expensive to compute.
This is exactly the same of what we are seeing with LLMs nowadays. The bust is the data we are scraping from the web, and the triangles are the model weights. And we have the same problem: Increasing the model size by 100x will be 100x times more expensive but we will get improvements that are really hard to notice.
Why? The problem is inherently to the LLM structure.
Models spend the same amount of compute on filler words like “however”, “therefore” or variables when coding or doing math. No one cares which filler word to use, but if you even just flip a sign in a math formula it’s just wrong.
It’s funny how in the model has high entropy on filler words not because it doesn’t know, but because it’s not important which filler word to use (this point extends also to synonyms).
And if you don’t do good data curation, which is currently being neglected by academics and industry (the industry is slowly going more and more) than you will spend a ton of compute during training on things like filler words.
Where do we go from here?
First big step is data curation. Our current datasets are so massive that we rarely do more than a few epochs when training. So we get rid of the easy examples to shrink the dataset to give more compute to the hard samples. This might require something like curriculum learning during pretraining.
Many people thing we have LLMs because of the transformer architecture or some other model idea, but actually it has been data all along. First ImageNet, then web scraping, large scale human preference data, and finally data from verifier systems. You can have a LLM without the transformer architecture, but not without the data. (read https://lnkd.in/eS2YbZgC)
Later down the line we will need to change the model architecture. Token-by-token prediction is slow and error-prone. Language Diffusion models are already showing impressive speed gains because they can unmask multiple tokens per step and have the ability to fix token errors. Now do that in latent space, get rid of the tokenization bias, add variable compute per step and some form of hierarchical representations and voilà, you might get yourself the next era of models.
Act
subj
Jensen spittin' wisdom again
Jensen Huang flipped the script on career advice: he says it's easier to fall in love with what you do than to find what you love 😳
“I loved it when I was a dishwasher. I loved it when I was a busboy. I loved it when I was waiting tables. And I’ve loved every single day at NVIDIA” - Jensen Huang
His point is simple:
It’s nearly impossible to stumble on a perfect “dream job.”
But it’s entirely possible to fall in love with whatever’s in front of you - by doing it well.
That mindset turned a dishwasher into the CEO of a $4 trillion AI giant.
Sure, he’s biased, and he’s a billionaire now.
But the lesson still hits:
Passion isn’t a prerequisite for great work. Great work creates passion.
So forget chasing the “right job.”
Keep showing up, taking pride, and letting mastery fuel meaning, and great things will happen.
“I loved it when I was a dishwasher. I loved it when I was a busboy. I loved it when I was waiting tables. And I’ve loved every single day at NVIDIA” - Jensen Huang
His point is simple:
It’s nearly impossible to stumble on a perfect “dream job.”
But it’s entirely possible to fall in love with whatever’s in front of you - by doing it well.
That mindset turned a dishwasher into the CEO of a $4 trillion AI giant.
Sure, he’s biased, and he’s a billionaire now.
But the lesson still hits:
Passion isn’t a prerequisite for great work. Great work creates passion.
So forget chasing the “right job.”
Keep showing up, taking pride, and letting mastery fuel meaning, and great things will happen.
express, not stress