Burn the forest until it works
Am I the only one finding out about this so late? There was a source code leakage from Claude Code, 2 months ago. It is quite interesting to see the inner details of what makes Anthropic's Claude persistently smarter than OpenAI's Codex. Most important takeaways:
- Claude Code was written, yes, by Claude Code itself. A random walkthrough shows how the system is cumbersome and lacks logical precision - signature of LLM generated code. Anthropic's way to build: Generate the code. If it doesnt work, generate again. So, burn the forest until the code works. Have a look at this article for more details.*
- The whole system is not pure deep learning. Some parts are old-school symbolic AI. And this is important. Apparently, it is a proof that making the models ever bigger is not gonna solve problems of LLMs.
*it is ironic how the company that talks most about safety is not taking safety seriously
Am I the only one finding out about this so late? There was a source code leakage from Claude Code, 2 months ago. It is quite interesting to see the inner details of what makes Anthropic's Claude persistently smarter than OpenAI's Codex. Most important takeaways:
- Claude Code was written, yes, by Claude Code itself. A random walkthrough shows how the system is cumbersome and lacks logical precision - signature of LLM generated code. Anthropic's way to build: Generate the code. If it doesnt work, generate again. So, burn the forest until the code works. Have a look at this article for more details.*
- The whole system is not pure deep learning. Some parts are old-school symbolic AI. And this is important. Apparently, it is a proof that making the models ever bigger is not gonna solve problems of LLMs.
*it is ironic how the company that talks most about safety is not taking safety seriously
π₯5β€1
Had the first meeting with my Stanford supervisors.
Impression - overwhelmed with their thought lead. You feel the difference.
Impression - overwhelmed with their thought lead. You feel the difference.
π₯13β€βπ₯5π1
"Study hard what interests you the most in the most undisciplined, irreverent and original manner possible."
Richard Feynman
Richard Feynman
β€17β€βπ₯4π1
Intuition is a compressed way of looking at some phenomenon, an easy way to understand something. You observe again and again that people who do real innovations often have very robust intuition on the topic. So you wonder how could one get more of it?
Many physicists think of Feynman when the word intuition comes up. He was very good at giving intuitions for complex phenomena. One wonders how Feynman did it. But you barely find anything from his own words, except that he was good at imagination.
So, what to do? There is a scientist called Stephen Wolfram (I was listening to his appearance on Lex Fridman's podcast). It turns out Wolfram was a younger colleague and friend of Feynman's at Caltech. He gave some insights about Feynman's intuition that you cannot find elsewhere:
- First, Feynman knew that you cannot compress everything into simple intuitions β not everything can be reduced to a child-level analogy. (This is similar to the line often attributed to Einstein: everything should be made as simple as possible, but not ever simpler.)
- Then, Feynman was very good at calculating. He would do complex calculations/simulations on paper or in his head, fast, then see the overall picture, then come up with a post-hoc intuition for the phenomenon. BUT because he thought calculation was easy, he would simply not mention the many days he spent calculating.
So the important point: to get intuition you need to work in the weeds / fight in the trenches, then you understand what is happening, then you get the intuition. Usually there is no shortcut, and often the intuition you end up with is not child-level easy.
Many physicists think of Feynman when the word intuition comes up. He was very good at giving intuitions for complex phenomena. One wonders how Feynman did it. But you barely find anything from his own words, except that he was good at imagination.
So, what to do? There is a scientist called Stephen Wolfram (I was listening to his appearance on Lex Fridman's podcast). It turns out Wolfram was a younger colleague and friend of Feynman's at Caltech. He gave some insights about Feynman's intuition that you cannot find elsewhere:
- First, Feynman knew that you cannot compress everything into simple intuitions β not everything can be reduced to a child-level analogy. (This is similar to the line often attributed to Einstein: everything should be made as simple as possible, but not ever simpler.)
- Then, Feynman was very good at calculating. He would do complex calculations/simulations on paper or in his head, fast, then see the overall picture, then come up with a post-hoc intuition for the phenomenon. BUT because he thought calculation was easy, he would simply not mention the many days he spent calculating.
So the important point: to get intuition you need to work in the weeds / fight in the trenches, then you understand what is happening, then you get the intuition. Usually there is no shortcut, and often the intuition you end up with is not child-level easy.
β€9π₯3π2β€βπ₯1
Research opportunity
I am starting a research project and am looking for assistants. The topic is on AI for education. If you are an undergraduate or even high school student looking to publish as a first author, this is your chance.
Traits we are looking for: critical thinking; grit (in Uzbek "tishlaganini uzib oladigan"π); some math, tutoring and coding background are nice to have.
Interested - apply here:
https://forms.gle/BadWxohKuaUY39G26
Deadline - July 15th
I am starting a research project and am looking for assistants. The topic is on AI for education. If you are an undergraduate or even high school student looking to publish as a first author, this is your chance.
Traits we are looking for: critical thinking; grit (in Uzbek "tishlaganini uzib oladigan"π); some math, tutoring and coding background are nice to have.
Interested - apply here:
https://forms.gle/BadWxohKuaUY39G26
Deadline - July 15th
π7β€βπ₯2β€2
It is very hard to beat the person who is doing it for fun.
Peter Steinberger
Developer of OpenClaw
p.s. Peter built solo what a battery of startups were trying to build and couldn't - an agent present everywhere in your devices. His explanation for why the startups couldn't beat him: "They take themselves too seriously".
Peter Steinberger
Developer of OpenClaw
p.s. Peter built solo what a battery of startups were trying to build and couldn't - an agent present everywhere in your devices. His explanation for why the startups couldn't beat him: "They take themselves too seriously".
β€7β€βπ₯6π3π₯1
Oh mani yorim dardi-darmonim,
Soβra jonim berayin borim.
Oh mani yorim san begβuborim,
Mani doim senda hayolim.
--β----------β-β-----------------------
Oh my beloved, the cure of my pain,
Ask, and I'll give you my life, all I have.
Oh my beloved - my pure one, you,
Always, my thoughts are with you.
Soβra jonim berayin borim.
Oh mani yorim san begβuborim,
Mani doim senda hayolim.
--β----------β-β-----------------------
Oh my beloved, the cure of my pain,
Ask, and I'll give you my life, all I have.
Oh my beloved - my pure one, you,
Always, my thoughts are with you.
π9π₯4β€2π€1π1
A person is a city
We grew up hearing tales of evil characters that do everything only with bad intentions. But as we grow - we realize that people are not white or black; a person is never singular. Rather he is a city of multiple characters. City where many entities take over in different times. City where both good and bad coexist. City where the bad vies for power and pays cheap tribute to the revolting good, and yet - always claims the title of nobility. City where the good is always present and never disappears. Like a lantern that might get covered in dust. But continues to shine in the hopes that one day - it will light all the city.
We grew up hearing tales of evil characters that do everything only with bad intentions. But as we grow - we realize that people are not white or black; a person is never singular. Rather he is a city of multiple characters. City where many entities take over in different times. City where both good and bad coexist. City where the bad vies for power and pays cheap tribute to the revolting good, and yet - always claims the title of nobility. City where the good is always present and never disappears. Like a lantern that might get covered in dust. But continues to shine in the hopes that one day - it will light all the city.
π₯8β€7π―3
This media is not supported in your browser
VIEW IN TELEGRAM
There it starts!
Y Combinator + San Francisco vibes
Y Combinator + San Francisco vibes
π₯19β€4π4
Case for building a startup
Building a startup, if you can, beats almost anything else on the table. The Cursor guys make the case clean: co-founder Sualeh Asif's stake is now worth $2.7 billion, four years after starting Anysphere. Do the math and you see he was doing $56 million per month. Compare that to a $20k capped salary at Google. This is a 3000x difference.
Sualeh is Pakistani. He did math olympiads before MIT. So, he has a lesson to share with us, and the lesson is not - become the hustling, talkative, self-proclaimed Steve Jobs who talks a vision without substance. It is - study, build real depth in something hard. Bring an actual skill to the table, math, engineering, research, whatever it is, and pair with people who have equally strong skills. A startup is real skills compounding into something big, not a personality performance. That's the trade worth making.
Building a startup, if you can, beats almost anything else on the table. The Cursor guys make the case clean: co-founder Sualeh Asif's stake is now worth $2.7 billion, four years after starting Anysphere. Do the math and you see he was doing $56 million per month. Compare that to a $20k capped salary at Google. This is a 3000x difference.
Sualeh is Pakistani. He did math olympiads before MIT. So, he has a lesson to share with us, and the lesson is not - become the hustling, talkative, self-proclaimed Steve Jobs who talks a vision without substance. It is - study, build real depth in something hard. Bring an actual skill to the table, math, engineering, research, whatever it is, and pair with people who have equally strong skills. A startup is real skills compounding into something big, not a personality performance. That's the trade worth making.
π₯21π2β€βπ₯1π1
Universities need reform
What I am realizing is - I am almost finishing my masterβs and I didnβt attend almost any course for more than its beginning 1-2 weeks - for both bachelorβs and masterβs. This was made possible thanks to recorded lectures and AI chatbots.
So my university experience says - something should change in education system. You donβt necessarily have to sit in the lectures. At the end of the day, learning is about students making sense of the content, not teachers pounding in their ears. That making sense is much better served if the student can pause the lecture, ask his questions, do more exercises, and think together with mentors. This would also give more time for lecturers to actually do more research rather than repeating the same thing again and again. This would be win-win case for both sides. In any case, presence of AI should rewire the education system and whoever does it first reaps the biggest harvest.
I found this interesting article that delves deeper into this topic.
What I am realizing is - I am almost finishing my masterβs and I didnβt attend almost any course for more than its beginning 1-2 weeks - for both bachelorβs and masterβs. This was made possible thanks to recorded lectures and AI chatbots.
So my university experience says - something should change in education system. You donβt necessarily have to sit in the lectures. At the end of the day, learning is about students making sense of the content, not teachers pounding in their ears. That making sense is much better served if the student can pause the lecture, ask his questions, do more exercises, and think together with mentors. This would also give more time for lecturers to actually do more research rather than repeating the same thing again and again. This would be win-win case for both sides. In any case, presence of AI should rewire the education system and whoever does it first reaps the biggest harvest.
I found this interesting article that delves deeper into this topic.
β€15π₯6π€1
Technology & Islam
Many people tend to see technology and Islam as separate things. Muslims are often seen as more spiritually inclined, and even techno-averse. The history too backs this up on the surface, like the Ottomans effectively blocking the printing press for around 300 years. I had this same unconscious view, until something clicked:
You see, in the Qur'an, reading, writing, and books are praised heavily. God emphasizes giving humans the ability to write with a pen. Now think about it - pen, book, writing, and reading - these are technology and technical skills. Meaning Islam actually encourages Muslims to engage with technology. Early Muslims lived this out too. Historically, they were the ones who made the mass production of paper a reality, enabling mass education, which directly led to the Islamic Golden Age. Now, what remains is to keep the same spirit.
p.s. this view is somehow rewiring how I approach my work - making AI much closer to my beliefs.
Many people tend to see technology and Islam as separate things. Muslims are often seen as more spiritually inclined, and even techno-averse. The history too backs this up on the surface, like the Ottomans effectively blocking the printing press for around 300 years. I had this same unconscious view, until something clicked:
You see, in the Qur'an, reading, writing, and books are praised heavily. God emphasizes giving humans the ability to write with a pen. Now think about it - pen, book, writing, and reading - these are technology and technical skills. Meaning Islam actually encourages Muslims to engage with technology. Early Muslims lived this out too. Historically, they were the ones who made the mass production of paper a reality, enabling mass education, which directly led to the Islamic Golden Age. Now, what remains is to keep the same spirit.
p.s. this view is somehow rewiring how I approach my work - making AI much closer to my beliefs.
β€10π4π2β‘1π1
Forwarded from Abdumalik Abdukayumov
Terence Tao: Mathematics in the Age of AI
(August 17, 2026)
Tao's essay based on his ICM 2026 lecture. Instead of arguing about whether AI can do research math, he conditions on the assumption that it will, and asks a different question: what do mathematicians actually value?
Key ideas:
1. A second foundational crisis. 1900-1930 forced math to make its foundations explicit (set theory, formal proof). AI will force us to make our values explicit: what counts as a contribution, what we reward, who gets credit.
2. Goodhart's law hits math. Historically, all goals of mathematics (solving problems, building theory, training students, community) were correlated, so "solve open problems" worked as a proxy for everything else. AI breaks this: it optimizes for the appearance of success, and the AI industry is financially rewarded for exactly the benchmarkable metrics we used as proxies. Over-optimize one goal and the goals diverge.
3. The problem-solving pipeline is 5 stages, not 1. Generation β verification β exposition β community acceptance β canonicalization (getting into the textbooks). We only ever made the first stage explicit. AI accelerates stage 1 massively, barely touches stage 5, which Tao calls the most valuable part. Fun point: AI training data is the output of canonicalization.
4. Proof scarcity β proof abundance. Journals, priority norms, hiring, prizes were all designed for scarcity. Expect "proof indigestion": verified proofs nobody understands, correct results nobody has refereed. This is already happening on the ErdΕs problems database.
5. Exposition can be over-optimized too. Human proofs have "natural friction" where the author struggled, and that friction teaches you where to pay attention. AI-polished proofs sand it away, making texts easy to read but hard to learn from. He shows a Bourgain page he annotated in frustration as a grad student.
His rule of thumb: if the authors can't give a clear expert-level talk on their result, it shouldn't be published. A verified proof no human can explain is incomplete.
Also cites the First Proof project: 7 of 10 novel research problems got an essentially correct AI solution at $10-100s of compute per problem.
@abdumalik_abdukayumov
(August 17, 2026)
Tao's essay based on his ICM 2026 lecture. Instead of arguing about whether AI can do research math, he conditions on the assumption that it will, and asks a different question: what do mathematicians actually value?
Key ideas:
1. A second foundational crisis. 1900-1930 forced math to make its foundations explicit (set theory, formal proof). AI will force us to make our values explicit: what counts as a contribution, what we reward, who gets credit.
2. Goodhart's law hits math. Historically, all goals of mathematics (solving problems, building theory, training students, community) were correlated, so "solve open problems" worked as a proxy for everything else. AI breaks this: it optimizes for the appearance of success, and the AI industry is financially rewarded for exactly the benchmarkable metrics we used as proxies. Over-optimize one goal and the goals diverge.
3. The problem-solving pipeline is 5 stages, not 1. Generation β verification β exposition β community acceptance β canonicalization (getting into the textbooks). We only ever made the first stage explicit. AI accelerates stage 1 massively, barely touches stage 5, which Tao calls the most valuable part. Fun point: AI training data is the output of canonicalization.
4. Proof scarcity β proof abundance. Journals, priority norms, hiring, prizes were all designed for scarcity. Expect "proof indigestion": verified proofs nobody understands, correct results nobody has refereed. This is already happening on the ErdΕs problems database.
5. Exposition can be over-optimized too. Human proofs have "natural friction" where the author struggled, and that friction teaches you where to pay attention. AI-polished proofs sand it away, making texts easy to read but hard to learn from. He shows a Bourgain page he annotated in frustration as a grad student.
His rule of thumb: if the authors can't give a clear expert-level talk on their result, it shouldn't be published. A verified proof no human can explain is incomplete.
Also cites the First Proof project: 7 of 10 novel research problems got an essentially correct AI solution at $10-100s of compute per problem.
@abdumalik_abdukayumov
arXiv.org
Mathematics in the age of AI
An essay, based on a public lecture delivered at the 2026 International Congress of Mathematicians, on how the mathematical community might respond to the arrival of artificial intelligence tools...