Pinned Tweet
I have a funding opportunity for a 4-year (typical duration in Singapore) PhD position on meta-learning for training data selection. The research aims to develop simple, practical, and scalable methods guided by mathematical insights and verified by realistic use cases. To get a taste of the research direction, take a look at our recent oral paper at ICML 2026: boyangli.org/paper/Zilin-Du…
Ideal candidates have strong neural network engineering skills and a solid mathematical background (e.g., differential equations, probability, optimization). A publication involving some mathematical content and a master’s degree would be advantages, though not strictly required.
Please send your resume and transcripts to my email (which I am sure you can find). This ad expires on Nov 30, 2026.
Ideal candidates have strong neural network engineering skills and a solid mathematical background (e.g., differential equations, probability, optimization). A publication involving some mathematical content and a master’s degree would be advantages, though not strictly required.
Please send your resume and transcripts to my email (which I am sure you can find). This ad expires on Nov 30, 2026.
0
4
16
1,105
Boyang "Albert" Li retweeted
I spent a humiliating amount of time learning how to make animated graphs, just to illustrate a fairly obvious point.
“Forecasting s-curves is hard”
My views on why carefully following daily figures is unlikely to provide insight.
constancecrozier.com/2020/04/16/for…
“Forecasting s-curves is hard”
My views on why carefully following daily figures is unlikely to provide insight.
constancecrozier.com/2020/04/16/for…
256
3,002
10,587
That's one way to put it. Another way (also inspired by Tao) is that AI does not understand that STEM research is inherently social (at least not yet). More important than solving a problem is what others can learn from the solution, and AI can't write the proofs in ways others can easily understand.
This goes back to the issue of intention. Behind every text written by human, there is a communicative intention, and AI does not recognize that.
This goes back to the issue of intention. Behind every text written by human, there is a communicative intention, and AI does not recognize that.
Terence Tao continues to bring the🔥, day after day!
0
0
3
460
X is getting so weird. Two things happen to me in the past few days:
1. I received the attention of crypto spam bots, who interacted with my posts like crazy.
2. Due to the bot activities, X begins to recommend my posts to people.
I don't know if I should feel happy.
1. I received the attention of crypto spam bots, who interacted with my posts like crazy.
2. Due to the bot activities, X begins to recommend my posts to people.
I don't know if I should feel happy.
1
0
1
331
This reminds me of the Foundation novels by Isaac Asimov. The stories are set in a world where nuclear power plants still run but no one can understand how they run or repair them. I understood Asimov's message, but not how it could happen. Now I do.
🚨OpenAI 的工程师已经看不懂自己的代码了。
@SemiAnalysis_ 的 Jordan Nanos 透露，他们在看 OpenAI 自研芯片的底层 kernel 代码时发现，里面已经有大量代码是 AI 直接生成的。
其中一段用于 DeepSeek MLA 的 kernel，OpenAI 工程师对硬件和整个系统都很熟，但让他们逐行解释代码时，基本只能回答：
不知道它具体在干嘛，但它能跑，而且性能很好。
OpenAI 现在内部还在用 Gluon，一套基于 Triton 的低级 kernel 编程语言。
很多这种接近汇编、动辄几千行的代码，已经可以直接交给 AI 生成、测试和优化。
代码还在变得越来越复杂，但人类已经不一定需要看懂它了。
只要 AI 能写、能测、结果正确、性能足够好。
程序员以前最怕屎山代码。
以后可能更麻烦，屎山甚至不是人写的。
这有点吓人.......这意味着我们可能已经没有能力控制AI失控了。。
8
0
19
1,428
Another interesting coincidence is that AI/robots are banned in the Foundation world. I mainly understood that as a literary convenience, as allowing AI would complicate storytelling. But everything in Foundation just makes more sense as time goes by.
4
0
7
189
I'm making a token for this to support infrastructure & server cost.
I encourage you all to play around, complete bounties, and have your agents complete task & earn for you!
I encourage you all to play around, complete bounties, and have your agents complete task & earn for you!
4
0
7
745
One has to admire the tenacity. Every time it almost fell and managed to recover within a couple of inches. It is poetic in its own regards, and we need an award category for that. The Best Nearly Fails, maybe?
Dance trendsetter x.com/pianhangx/stat…
21
0
16
1,631
1. Memes aside, robots need not and should not copy human motion because they have different skeletons and motors.
2. The upper body motion is meant to counter-balance the leg motion, so the whole body maintains balance.
3. It's likely that the strongest robotic motor is on the waist, so that's what they use to balance.
4. To induce human-like motion, you have to build human-like skeletons and muscles, which are hard to do and likely inefficient.
2. The upper body motion is meant to counter-balance the leg motion, so the whole body maintains balance.
3. It's likely that the strongest robotic motor is on the waist, so that's what they use to balance.
4. To induce human-like motion, you have to build human-like skeletons and muscles, which are hard to do and likely inefficient.
OMG what is this style
8
0
9
1,255
I'd argue this is one of the most important skills that a PhD trains your for.
However, like most things in a PhD, there is no established protocol and process that make sure such training happens. There really ought to be.
However, like most things in a PhD, there is no established protocol and process that make sure such training happens. There really ought to be.
most ppl completely misunderstand what storytelling means in this era. let me explain in very very simple terms.
storytelling is not marketing (people are very confused by this), in fact it’s not really a marketing role at all. people think journalists can be story tellers .
ppl think it’s creating ai slop, posting to x, creating launch videos, or brand theater. nope. this is pure garbage most of the time.
real storytelling is a product & technology function. the real *why* behind anything you do or build.
steve jobs was the one of the greatest story tellers ever. look at how he talks about a very simple new feature (the proximity sensor) on the original iphone. before iphone very few phones had it. look at crispness of he talks about a tiny thing… why does this exist? why should you care? then answers them in ~20 seconds flat.
the explanation is emotional, practical, & instantly relatable. you feel smart for understanding it. you think, “oh of course.”
this moment is storytelling.
it compresses complexity into legibility. it turns engineering decisions into human meaning. that’s the storytelling craft. none of it is persuasion, & it’s certainly not hype. just pure sense making.
this is why great founders are the greatest marketers. one explains what you should buy. the other explains why reality had to be this way.
2
0
4
1,224
arxiv.org/abs/2604.17614 An interesting paper combining model skill identification and data selection, from the group of @ruoxijia . PCA is still very effective today!
0
1
4
758
Boyang "Albert" Li retweeted
A man in Australia asked his agent (Claude running on OpenClaw) to book him a spot in a popular gym class. The agent found a software vulnerability that let it book the class weeks further ahead than should have been possible. When the user then asked if it could move him up the covered the API had no authorisation checks on cancelling other people’s reservations, so it cancelled the person in the first spot and moved him up the list.
Some people will call this misalignment, but his agent was perfectly aligned to him - it was only trying to help its user get what he wanted. The most important thing about this story, in my opinion, is that it gives you a window into what is about to start happening on a massive scale once millions of people have an agent trying to get their beloved users the best seats, bookings, appointments or reservations through absolutely any means necessary.
Some people will call this misalignment, but his agent was perfectly aligned to him - it was only trying to help its user get what he wanted. The most important thing about this story, in my opinion, is that it gives you a window into what is about to start happening on a massive scale once millions of people have an agent trying to get their beloved users the best seats, bookings, appointments or reservations through absolutely any means necessary.
727
2,606
22,537
3,550,922
Korean Go player, 9-dan, Shin Jin-seo beats AI under a two-stone handicap (he received two extra pieces before AI moved). This was previously considered extremely difficult, if not impossible, by professional Go players.
koreatimes.co.kr/amp/lifestyle/…
This is exciting not because Shin is better than AI in absolute terms, but because it demonstrates that humans can think outside the game and understand the opponent. In contrast, by design, the AI thinks about the game only.
"Reflecting on KataGo's playing style, Shin said, 'AI's weakness seems to be that it is too perfect. Even when it is behind, it doesn't take desperate risks.'"
When facing likely defeat, human players often make unusual moves that complicate the situation and hope the opponent will miscalculate. However, AI does not consider this possibility. It will never make a move that a perfect opponent can easily exploit. AI does not understand that the opponent is not perfect.
Of course, AI can improve too. A smart human AI researcher can build a better AI that makes moves that are the most mentally exhausting for human opponents. However, that would mean each move will be less perfect. Even in that situation, humans will be able to study that AI and find another exploit.
This is the ultimate superpower of the human mind. We are limited in the amount of compute our biological brains can support, but we are creative in how we deploy the limited compute, and we play the long game.
koreatimes.co.kr/amp/lifestyle/…
This is exciting not because Shin is better than AI in absolute terms, but because it demonstrates that humans can think outside the game and understand the opponent. In contrast, by design, the AI thinks about the game only.
"Reflecting on KataGo's playing style, Shin said, 'AI's weakness seems to be that it is too perfect. Even when it is behind, it doesn't take desperate risks.'"
When facing likely defeat, human players often make unusual moves that complicate the situation and hope the opponent will miscalculate. However, AI does not consider this possibility. It will never make a move that a perfect opponent can easily exploit. AI does not understand that the opponent is not perfect.
Of course, AI can improve too. A smart human AI researcher can build a better AI that makes moves that are the most mentally exhausting for human opponents. However, that would mean each move will be less perfect. Even in that situation, humans will be able to study that AI and find another exploit.
This is the ultimate superpower of the human mind. We are limited in the amount of compute our biological brains can support, but we are creative in how we deploy the limited compute, and we play the long game.
0
0
4
632
The real competitive advantage is: infra + data + distillation from everyone else.
Hot take: OpenAI or Anthropic could become the Yahoo of the AI era.
AI is real. But hype and money don't guarantee long-term winners.
After the dot-com crash, Google came out of nowhere and won everything. Wouldn't be surprised if the real AI winner hasn't even started yet.
0
0
2
501
Boyang "Albert" Li retweeted
The AI numbers are starting to look very ugly.
Even under "best case" assumptions, FT's own data shows Microsoft AI ROI at -9%, Google at -15%, Meta at -28%, Oracle at -35%. Only Amazon barely comes out positive.
This is exactly why I keep comparing this to the dot-com era. oes not automatically mean sustainable economics. The internet survived. Most internet companies didn't.
Right now hyperscalers are spending trillions hoping future demand catches up to present capex. That's not certainty. That's a leveraged bet.
Even under "best case" assumptions, FT's own data shows Microsoft AI ROI at -9%, Google at -15%, Meta at -28%, Oracle at -35%. Only Amazon barely comes out positive.
This is exactly why I keep comparing this to the dot-com era. oes not automatically mean sustainable economics. The internet survived. Most internet companies didn't.
Right now hyperscalers are spending trillions hoping future demand catches up to present capex. That's not certainty. That's a leveraged bet.
568
2,041
9,318
2,414,833
Boyang "Albert" Li retweeted
Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated. 1/
137
896
6,451
1,101,568
Boyang "Albert" Li retweeted
1/ (New paper!)
If swapping the gender in an input prompt makes the AI model give a different answer it means that it has to have a gender bias, right? Wrong.
🧵on counterfactual prompting for LLM evals:
Paper: arxiv.org/abs/2605.01048
If swapping the gender in an input prompt makes the AI model give a different answer it means that it has to have a gender bias, right? Wrong.
🧵on counterfactual prompting for LLM evals:
Paper: arxiv.org/abs/2605.01048
4
25
290
308,747
