subj
Points on the board
make that "63 AND COUNTING" winners of The Nobel Prize - say the stickers on our football helmets for our game against UNC tomorrow...
Cal just won FOUR more Nobel Prizes - Go Bears!
Big brains in these helmets too 🧠🏆
Cal just won FOUR more Nobel Prizes - Go Bears!
Big brains in these helmets too 🧠🏆
subj
Go Bears! For realz
Comparison of North American bear claws.
Fun fact: kodiak brown bears and grizzly bears are the same species (Ursus arctos) with kodiaks being considered a sub-species of the north american grizzly. The main difference is kodiak bears are isolated on the islands off alaska and bc of the abundance of food (think salmon run) and lack of competition theyve become huge (island gigantism).
More interestingly is that because of this kodiak bears are generally a lot more docile towards humans than grizzlies especially during the salmon run. Theres such an abundance of food during this time they don’t bother with anything they have to chase and they’re even picky with the salmon, only eating the heads and skin of the fish.
Bears learn their behavior from their parents instead of it being instinctive so grizzlies learn to be aggressive since theres more competition in the mainland US, where as kodiak bears learn to be fairly tolerant of people.
Fun fact: kodiak brown bears and grizzly bears are the same species (Ursus arctos) with kodiaks being considered a sub-species of the north american grizzly. The main difference is kodiak bears are isolated on the islands off alaska and bc of the abundance of food (think salmon run) and lack of competition theyve become huge (island gigantism).
More interestingly is that because of this kodiak bears are generally a lot more docile towards humans than grizzlies especially during the salmon run. Theres such an abundance of food during this time they don’t bother with anything they have to chase and they’re even picky with the salmon, only eating the heads and skin of the fish.
Bears learn their behavior from their parents instead of it being instinctive so grizzlies learn to be aggressive since theres more competition in the mainland US, where as kodiak bears learn to be fairly tolerant of people.
subj
deep learning may be deterministic at heart
At the dawn of the 21st century, mathematicians faced a paradox: the networks they wanted to study had grown too large to grasp node by node. Social graphs, protein interactions, energy grids - each held millions of links. What was needed was a new language, a way of seeing the forest instead of the trees. Out of this came the theory of 𝐠𝐫𝐚𝐩𝐡 𝐥𝐢𝐦𝐢𝐭𝐬.
The key insight was to replace a massive network with a 𝐜𝐨𝐧𝐭𝐢𝐧𝐮𝐨𝐮𝐬 𝐥𝐚𝐧𝐝𝐬𝐜𝐚𝐩𝐞, called a 𝐠𝐫𝐚𝐩𝐡𝐨𝐧. Imagine a heatmap: red where connections are likely, blue where they are rare. Suddenly the incomprehensible becomes visible. Graphons let researchers describe infinite networks the way physicists describe gases - without tracking every particle. This was a breakthrough: complexity translated into geometry.
But not all networks are dense. Many real systems are sparse. For these, graphons were insufficient. The next step was 𝐚𝐜𝐭𝐢𝐨𝐧 𝐜𝐨𝐧𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞. Instead of asking only “how often do links appear?”, it asked: "𝐡𝐨𝐰 𝐝𝐨𝐞𝐬 𝐭𝐡𝐞 𝐬𝐲𝐬𝐭𝐞𝐦 𝐫𝐞𝐚𝐜𝐭 𝐰𝐡𝐞𝐧 𝐩𝐮𝐬𝐡𝐞𝐝?" Two networks could be equivalent if their responses to stimuli matched, even if their details differed. This gave mathematicians a universal yardstick - a way to compare networks by their 𝐝𝐲𝐧𝐚𝐦𝐢𝐜𝐬 𝐨𝐟 𝐢𝐧𝐟𝐨𝐫𝐦𝐚𝐭𝐢𝐨𝐧 𝐟𝐥𝐨𝐰.
This move from static maps to 𝐨𝐩𝐞𝐫𝐚𝐭𝐨𝐫𝐬 bridged to artificial intelligence. Neural networks are vast weighted graphs where information cascades layer by layer. For years, researchers feared instability: does a network’s worldview depend heavily on random initialization? The consensus leaned toward fragility.
A new result challenges this. Szegedy Balázs and his team trained networks with identical architectures but different seeds. To their surprise, the final internal representations were far more alike than expected. Despite noisy starts, the models converged to nearly the same abstract features. It was as if they wandered on different trails yet always reached the same valley.
This suggests deep learning is not chaotic but follows hidden mathematical structures in the data. The representation of the world is not invented but discovered again and again, robust to randomness. In plain terms: deep learning may be deterministic at heart.
If proven, the consequences are seismic. Deterministic deep learning would mean that AI systems build essentially the same “world models” regardless of start, because reality’s patterns exert a gravitational pull. The implications: stronger explainability, greater stability, and perhaps even laws of learning as universal as Newton’s laws of motion. What looks like a black box could become a new kind of physics.
The proof is not yet there. But the trajectory is clear: from graphons to action convergence to Szegedy’s discovery, a decades-long chain of thought is converging on one possibility. If determinism in deep learning holds, AI will not just be engineered - it will be understood.
Act
The key insight was to replace a massive network with a 𝐜𝐨𝐧𝐭𝐢𝐧𝐮𝐨𝐮𝐬 𝐥𝐚𝐧𝐝𝐬𝐜𝐚𝐩𝐞, called a 𝐠𝐫𝐚𝐩𝐡𝐨𝐧. Imagine a heatmap: red where connections are likely, blue where they are rare. Suddenly the incomprehensible becomes visible. Graphons let researchers describe infinite networks the way physicists describe gases - without tracking every particle. This was a breakthrough: complexity translated into geometry.
But not all networks are dense. Many real systems are sparse. For these, graphons were insufficient. The next step was 𝐚𝐜𝐭𝐢𝐨𝐧 𝐜𝐨𝐧𝐯𝐞𝐫𝐠𝐞𝐧𝐜𝐞. Instead of asking only “how often do links appear?”, it asked: "𝐡𝐨𝐰 𝐝𝐨𝐞𝐬 𝐭𝐡𝐞 𝐬𝐲𝐬𝐭𝐞𝐦 𝐫𝐞𝐚𝐜𝐭 𝐰𝐡𝐞𝐧 𝐩𝐮𝐬𝐡𝐞𝐝?" Two networks could be equivalent if their responses to stimuli matched, even if their details differed. This gave mathematicians a universal yardstick - a way to compare networks by their 𝐝𝐲𝐧𝐚𝐦𝐢𝐜𝐬 𝐨𝐟 𝐢𝐧𝐟𝐨𝐫𝐦𝐚𝐭𝐢𝐨𝐧 𝐟𝐥𝐨𝐰.
This move from static maps to 𝐨𝐩𝐞𝐫𝐚𝐭𝐨𝐫𝐬 bridged to artificial intelligence. Neural networks are vast weighted graphs where information cascades layer by layer. For years, researchers feared instability: does a network’s worldview depend heavily on random initialization? The consensus leaned toward fragility.
A new result challenges this. Szegedy Balázs and his team trained networks with identical architectures but different seeds. To their surprise, the final internal representations were far more alike than expected. Despite noisy starts, the models converged to nearly the same abstract features. It was as if they wandered on different trails yet always reached the same valley.
This suggests deep learning is not chaotic but follows hidden mathematical structures in the data. The representation of the world is not invented but discovered again and again, robust to randomness. In plain terms: deep learning may be deterministic at heart.
If proven, the consequences are seismic. Deterministic deep learning would mean that AI systems build essentially the same “world models” regardless of start, because reality’s patterns exert a gravitational pull. The implications: stronger explainability, greater stability, and perhaps even laws of learning as universal as Newton’s laws of motion. What looks like a black box could become a new kind of physics.
The proof is not yet there. But the trajectory is clear: from graphons to action convergence to Szegedy’s discovery, a decades-long chain of thought is converging on one possibility. If determinism in deep learning holds, AI will not just be engineered - it will be understood.
Act
subj
Vintage Ann Arbor
Ann Arbor Directory (September 10, 1986)
express, not stress