จิตร์ทัศน์
@jittat
Followers
4K
Following
41K
Media
415
Statuses
32K
Mastodon: [email protected]
Joined October 2007
This paper is quietly one of the most damning findings about current LLM architecture. Google Research tested 7 models across 7 benchmarks. The intervention was embarrassingly simple: paste the prompt twice. The result: 47 wins out of 70 tests, zero losses. Gemini Flash-Lite
LLMs process text from left to right — each token can only look back at what came before it, never forward. This means that when you write a long prompt with context at the beginning and a question at the end, the model answers the question having "seen" the context, but the
75
214
2K
A young aspiring mathematicians asks about the future of math with all the AI advances. My advice: Embrace research but be ready to pivot as needed. We many be seeing a new age in mathematics, how could you not be part of it. https://t.co/FY3fXCLMX4
2
15
94
I started my career as a theorist, and am now an empirical LLM researcher. In today's blog post, I talk about the parallels between theory and empirical research:
7
40
408
a first: in rejecting an article I submitted to a journal, reviewer 2 noted I failed to engage the work of one Andre Pagliarini
390
2K
47K
Congratulations to Arthur Jacot (NYU) on winning the $50,000 AMR @AMathRes Prize in the Mathematics of AI! https://t.co/NGAxpDviEc From the citation: "His most influential contribution to the subject is the development of the Neural Tangent Kernel framework. This work,
6
28
212
A three-decade-old open problem of mine just got solved! Matt Kovacs-Deak, Daochen Wang and Rain Zimin Yang showed that if a Boolean function f is exactly computed by a low-degree rational function, the quotient of two polynomials, then f has low decision-tree complexity. 1/2
2
12
125
On the latest episode of the @ScienceMagazine podcast, @AlexKontorovich and I discuss the challenges of communicating mathematics among professional mathematicians: https://t.co/SfPUOPICqr
science.org
On this week’s show: Sending telescopes out beyond the Solar System to see other worlds, and solving the communication problem in math
2
16
58
What is good mathematics? - From Terence Tao His personal thoughts and opinions on what “good quality mathematics” is. "(i) Good mathematical problem-solving (e.g. a major breakthrough on an important mathematical problem); (ii) Good mathematical technique (e.g. a masterful
18
99
460
Reflecting on @geoffreyhinton's flawed view of math as a "closed system", here are 3 key aspects of math that the general public (including physics Nobel prize winners) tends to get wrong:⤵️
10
43
256
Real Analysis, The Game (v0.1) is DONE!! 44 Worlds 138 Levels All your old favorites like Bolzano-Weierstrass and Heine-Borel, Uniform Convergence and Riemann Sums, and the biggest Boss of all, the Intermediate Value Theorem! :) Play the game here: https://t.co/0FkIcFMk4i
46
228
1K
I’d say the 4 way clash beats asta vs Lucius in terms of choreography
Kagurabachi fans, I love that manga like crazy—I even glazed it at some point—but can y'all try reading other manga? Just the fight choreography from Asta and Yuno vs. Lucius clone v.2 clears any fight in Kagurabachi 😭
7
31
997
Had a conversation with Lee Sedol, the legendary Go master who played the historic match against AlphaGo. We discussed the past 10 years of AI evolution and the future. I am grateful to Lee Sedol for the insightful dialogues. He is currently a Professor at UNIST, focusing on
25
75
912
I have flipped flopped about these things over time. I was recently talking to my son, who was then learning derivatives of trigonometric functions (on Math Academy). He could quickly, efficiently, and accurately compute all kinds of things, but when I asked him *why*, say, the
@Tzaidecar @AlexKontorovich I might know what 6+3 is, but can I define addition (this is not the same as knowing how to add)? Can I define a natural number? Could I answer why infinity + 1 = infinity? Most people cannot; because most people have not been taught to even ask these kinds of questions
55
49
738
Open-weight models are becoming increasingly capable while creating significant risks beyond those that already exist for closed-weight models. To continue benefiting from the advantages of open-weight models, it is urgent to develop risk mitigation methodologies specifically for
🚨New paper🚨 From a technical perspective, safeguarding open-weight model safety is AI safety in hard mode. But there's still a lot of progress to be made. Our new paper covers 16 open problems. 🧵🧵🧵
14
41
220
My student @Yang_Liuu is on the job market looking for an industry lab position and she's fantastic! She's done rigorous experimental work on in-context learning and identified the value of looping early on for meta learning algorithms and demystified aspects of task vectors
"Looped Transformers are Better at Learning Learning Algorithms" in ICLR @Yang_Liuu offers a simple and clean message in this paper. When it comes to emulating learning algorithms, using a looped transformer (i.e., one where the iterative structure is hardcoded) helps a lot.
3
11
66
Sutskever's List In this book, you go through 30 key deep-learning papers and, in plain English, see what each did and the practical ideas you can apply. You’ll discover: - The technical breakthroughs that have shaped AI strategy, safety, and cultural shifts - How to decode
18
164
1K
Foundations of Responsible Computing (FORC) is a super exciting new conference focused on the intersection of mathematical research and society. It's also a fantastic and vibrant community. Check out the CfP, with two deadlines. Also follow the new Twitter account @FORCConf !
📢 In case you missed it: the first-cycle deadline for FORC 2026 is *tomorrow*, November 11. Submit your best work on mathematical research in computation and society, writ large. Too soon? We'll also have a second-cycle deadline on February 17, 2026. CfP link below!👇
1
5
12