Justin(as)

AI Research at @Stanford Graduate student in Mathematics @TU_Muenchen

[Justin(as) @JZaliaduonis]

I'm sorry, but "Open-weight models are inherently decelerationist" is just wrong

Economic/societal acceleration comes from new, better solutions to existing problems

Open-sourcing frontier LLMs gives new ideas for what's working and increases the combinatorial space of things to try to tackle the existing problems

This coming from someone at @OpenAI does not help the case :)

[Justin(as) @JZaliaduonis]

Insane content

Really love this new format by @dwarkesh_sp

[Dwarkesh Patel @dwarkesh_sp]

Adam Brown (@A_G_I_Joe) is back!

General relativity is said to be the most beautiful idea the human mind has ever produced.

Most of us will never get to fully appreciate its elegance by taking the 20-lecture graduate course Adam taught on it at Stanford.

But in the video below, Adam distills the key idea at its heart so clearly and compellingly that even I could keep up lol.

At the core of general relativity, Einstein is trying to figure out the principle behind a particular coincidence: that the mass that resists acceleration and the mass that gravity pulls on just happen to be exactly the same. Adam then leads us through the path of insight which Einstein called his “happiest thought.”

Then Adam lectures on black holes. First, by showing how even under special relativity you could create a perpetual motion machine if black holes weren't truly black. And then, by explaining why the observations of an infalling observer and a distant bystander to the black hole would be so radically different

Adam leads Blueshift, the team at Google DeepMind cracking science and reasoning.

Which gave us the opportunity to discuss at the very end how close we are to AIs that could rediscover general relativity from scratch. Stay till the close for some philosophy of science.

0:00:00 – The coincidence that led Einstein to general relativity
0:16:42 – Gravity is a consequence of curved spacetime, not a force
0:31:46 – Why black holes prevent unlimited energy extraction
0:47:12 – Black holes are the ultimate power plants
1:13:50 – What falling into a black hole would actually feel like
1:18:51 – The three ways we know black holes are real
1:24:21 – The first time we saw gravity bend light
1:29:33 – How far can AI get without experimental evidence?

Look up Dwarkesh Podcast on YouTube/Spotify to watch. Enjoy!

[Justin(as) @JZaliaduonis]

@RT_com @grok Where did this happen? Why would someone grow killer cobras? Why did people not think that this could happen?

[Justin(as) @JZaliaduonis]

Students that scored 75+ in a proctored exam is a really interesting case

Probably studied honestly, but relied too much on AI in topics they did not like or when cognitive fatigue kicked in

[Brendan Nyhan (@BrendanNyhan on 🟦☁️) @BrendanNyhan]

Friends don't let friends continue to give exams and papers that are not adapted to the current AI landscape (this is AFTER 27 bailed so the reality is even worse)

PS Props to students 1, 22, and 31

[Justin(as) @JZaliaduonis]

LMAO "A war broke out in Russia"

War did not break out in russia, war broke out in Ukraine. russia attacked Ukraine.

Very misleading wording

[Molly O’Shea @MollySOShea]

Matt Weigand (@Mweigand13) on why Accel invested in Nebius $NBIS Founder Arkady Volozh:

"He is the most quietly humble killer I've ever met. He is truly, truly special and an N-of-1 entrepreneur."

"Arkady had built Yandex, which was built to a $30B market cap on the Nasdaq."

"And then the war broke out in Russia, and many of the engineers followed Arkady out of Russia."

"Most people would've hung the cleats up and said, 'I've had a great run. I've made billions of dollars from Yandex, and I'm going to retire.'"

"He was thinking, 'How can I plow this all into infrastructure because I have this unfairly advantaged team to go take that market by storm?'"

[Justin(as) @JZaliaduonis]

@zmkzmkz Yeah, the intuition on second order optimization methods was insane. Always wondered why noone used them

[Justin(as) @JZaliaduonis]

Heading to ICML 2026 to present our work on mathematical foundations of Contrastive Learning

If you're interested in search, data-efficient learning, post-training, RL, and multimodal reasoning - would be happy to chat!

#ICML2026

[Justin(as) @JZaliaduonis]

One of my favorite facts about financial markets is that credit predates money

Credit came BEFORE money

[Justin(as) @JZaliaduonis]

@harjtaggar @t_blom @grok How many?

[Justin(as) @JZaliaduonis]

I guess their Matrices are too dangerous to be multiplied by me

[Anthropic @AnthropicAI]

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.

Access to all other Claude models is not affected.

We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.

Read our full statement: https://t.co/bwn0sximKZ

[Justin(as) @JZaliaduonis]

@steipete Built this: https://t.co/d97OqdBmr3

Might be useful, lmk if you need credits

[Justin(as) @JZaliaduonis]

Literally was in the middle of a coding session

[Anthropic @AnthropicAI]

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.

Access to all other Claude models is not affected.

We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.

Read our full statement: https://t.co/bwn0sximKZ

[Justin(as) @JZaliaduonis]

@GoogleDeepMind Super excited about co-scientist!

Could have named it John though :) (in the name of John von Neumann)

[Justin(as) @JZaliaduonis]

This is incredible.

Science used to have genius polymaths like von Neumann, da Vinci, and Simon.

Then the knowledge base grew very large and complex, and it practically became in possible to have a great career in science without hyper-specializing, since getting to the frontier of knowledge required years of studying and mastery.

Super excited about co-scientist by DeepMind to bring back the age of polymaths and advance the frontier of scientific knowledge!

The impact of AI on the real world we've seen so far is a drop in the bucket compared to what we will see in the upcoming years!

[Google DeepMind @GoogleDeepMind]

We believe AI can be a dedicated research partner to help discover the next breakthrough.

Enter Co-Scientist: our latest Gemini-based multi-agent system that can generate, debate and evolve novel hypotheses for complex scientific problems 🧵

[Justin(as) @JZaliaduonis]

Would play this!

I think there’s insane potential for AI-generated cross-overs/makeovers of old video games

[GORM THE OLD @GORMTHEOLD25]

GTA: The Dark Ages (A.I. Video)

[Justin(as) @JZaliaduonis]

San Francisco housing market is a nightmare.

Between Craigslist, Facebook groups, random spreadsheets, and listings disappearing in hours, finding a place can easily turn into a second full-time job.

This weekend I teamed up with @kaarelkaarelson and built a platform to agentically search housing listings in San Francisco based on your current needs.

You can simply type queries like:

- "pet-friendly homes with less than 3 roommates"
- "homes for stays from June 12th to August 19th"
- "Appartaments in safe neighborhoods with parks nearby"

And cut straight to the relevant listings, saving hours of manually scrolling through craigslist

https://t.co/iPYD0NWx53

[Justin(as) @JZaliaduonis]

Took this weekend to finish expriments with self-supervised representation learning methods for images.

Implemented the 5 most widely used frameworks.

Autoencoders: reconstruct the input through a bottleneck https://t.co/ILIIsISeLZ

SimCLR: pull augmented views of the same image together, push others apart https://t.co/i4NLtrwqtZ

BYOL: predict one view's representation from another, no negatives needed https://t.co/WdvDEri40N

Masked Autoencoders: mask large patches, reconstruct the missing pixels https://t.co/vZFpbuoItp

I-JEPA: predict masked patches in latent space (pioneered by @ylecun) https://t.co/8j6pYto71F

The arc from reconstructing pixels to predicting latent representations is an interesting mental walk through how AI researchers' thinking has evolved.

https://t.co/eJEOeNjsIe

[Justin(as) @JZaliaduonis]

Trained a ColBert encoder for our product.

https://t.co/6XaALMHCGy

it's one of the most underrated technical ideas of the last 5 years.

Representing variable length documents with variable length encodings seems like a cruicial idea to scaling precise search methods

https://t.co/xNc8atGW82