I wrote an updated analysis on tariffs for 2025 - important clarifications on who actually pays for it. The methodology is synthetic control
medium.com/@vishalmisra...
@vishalmisra.bsky.social
Vice Dean Computing and AI @CUSEAS. Dean of Cricket Analytics @SFOUnicorns. Tweets on academia, cricket, dad/bad/wry jokes. Opinions of guy am pointing at.
I wrote an updated analysis on tariffs for 2025 - important clarifications on who actually pays for it. The methodology is synthetic control
medium.com/@vishalmisra...
A look back at the fractal modeling craze of the β90s, why it fell short, and what Internet traffic actually looks like: layered, stochastic, human.
https://medium.com/@vishalmisra/fractals-fables-and-the-nature-of-internet-traffic-81ed2d736eae
Sometimes itβs just layers, not fractals.
Iβve spent the last 20+ years accidentally building things that turned out to be right.
Live sports on the web. Control theory for TCP. Net neutrality. RAG.
Wrote up 5 of those stories here:
https://medium.com/@vishalmisra/five-ideas-that-were-early-but-turned-out-to-be-right-d6f51ab09574
Next time I go to a restaurant instead of paying the bill in cash I will teach them a class on stochastic differential equations.
To get rid of the trade imbalance.
New research: Cluster Synthetic Control (ClusterSC) addresses the curse of dimensionality in causal inference with large datasets (AISTATS 2025).
Blog: https://medium.com/@vishalmisra/a-smarter-way-to-compare-apples-to-apples-how-cluster-synthetic-control-improves-causal-inference-d8684dce6abc
Some people buy Teslas, I buy tomatoes
14.03.2025 00:13 β π 4 π 0 π¬ 1 π 1The next distinguished lecture in AI organized by @CUSEAS is on 3/28 by Ion Stoica - a legendary researcher and entrepreneur. Here is the registration link for the talk
https://www.eventbrite.com/e/lecture-series-in-ai-ion-stoica-uc-berkeley-tickets-1277915703269
Looking forward to you visit Ion!
This is pitch perfect by Dr. Katrina Armstrong, our President and leader. She speaks to all our values.
12.03.2025 14:33 β π 1 π 0 π¬ 0 π 0SHSAT results came out and our son joins our daughter for high school at Bronx Science. I told them no pressure but itβd be nice if they can add to this list of 9 Nobel Prize winners from Bronx Science.
No pressure.
A must read on whatβs happening in that other place
https://www.richardhanania.com/p/liberals-only-censor-musk-seeks-to
Well isnβt this fantastic
https://www.scmp.com/news/china/science/article/3299395/americas-loss-chinas-gain-top-chinese-universities-welcome-phd-refugees-us
Yesterdayβs announcement of funding cuts makes us more committed than ever in supporting our students, our
faculty, and our research mission. @Columbia has a reputation for being one of the most important and influential research institutes in the world and it will continue.
Mumbai Tech Week was a fantastic event - had the pleasure of speaking on both days. Once with Amit Sharma - the CTO of Dream Sports on the Columbia Dream Sports AI Innovation Center - #MumbAI
08.03.2025 16:39 β π 0 π 0 π¬ 0 π 0On Day 1 of Mumbai Tech Week. Appearing on a panel on all things AI moderated by Varun Mayya with Niki Parmar (co-inventor of the transformer) and Siddharth Srinivasan, head of Eleven Labs India. #MumbAI
08.03.2025 16:37 β π 0 π 0 π¬ 0 π 0Aaaand here we go
https://www.thefp.com/p/exclusive-trump-administration-cancels
Found this on my college WhatsApp group. At least I have been consistent with my views!
06.03.2025 19:47 β π 1 π 0 π¬ 0 π 0My first time at AAAI - giving an invited talk on βLarge Language Models as Approximate Bayesian Learners: A Unified Viewβ
Joint work with Naman Agarwal and Siddhartha Dalal
Thanks B Ravi from IITM gfor the invitation!
Key insight: LLMs encode algorithmic reasoning during training. Chain-of-thought prompting isn't 'teaching' new skillsβit's triggering existing computational pathways. DeepSeek's (and similar) approaches made these pathways the default, while others require explicit activation
19.02.2025 02:42 β π 0 π 0 π¬ 0 π 0TokenProbe visualization shows how prompting unlocks these 'latent algorithms': When asked to solve 33Γ117 directly, tokens show uncertainty (lighter green). But prompt for step-by-step, and high-confidence tokens (green) reveal systematic computation emerging.
19.02.2025 02:41 β π 0 π 0 π¬ 1 π 0A simple experiment in arithmetic on our tool TokenProbe reveals the power of inference time scaling in LLMs: Given '33Γ117', token probabilities show uncertain guessing for direct computation, but near-certainty for each step when solving methodically.
19.02.2025 02:08 β π 1 π 0 π¬ 1 π 0Nobody voted for DOGE. People certainly didnβt vote for the Middle East Rivera. There is considerable evidence that it was, in fact, the price of eggs that drove the outcome and the Harris campaign nearly overcame that.
https://medium.com/@vishalmisra/thoughts-on-the-2024-us-elections-a228086d9fba
Two words instantly doubled shampoo sales: Rinse, Repeat
Four words have quadrupled Kumbh business: Once in 144 years
This is what genius marketing looks like.
Because this is getting downplayed in some of the coverage, the racist staffer DOGE rehired isnβt a case of βstupid things he wrote years ago as a teen.β He posted about how we should βnormalize Indian hateβ IN SEPTEMBER.
08.02.2025 19:03 β π 9181 π 2670 π¬ 174 π 86What are the best books on how democracies fail?
02.02.2025 18:45 β π 929 π 85 π¬ 319 π 20We all like to yell at the Democrats, but itβs really Speaker Mike Johnson and Senate Majority Leader John Thune who are letting Elon Musk and his faceless goon squad make a mockery of their power as leaders of a coequal branch without uttering a peep or lifting a finger.
02.02.2025 17:34 β π 15713 π 3426 π¬ 600 π 298Sunny G was making the point over and over again how Abhishek Sharmaβs innings was full of classical strokes - and the evidence bears out. Almost all in the V - no βreverseβ or βscoopβ anythings
02.02.2025 17:42 β π 1 π 0 π¬ 0 π 0Dean Shih-Fu Chang and Vice Dean @vishalmisra.bsky.social discuss the future of #AI in education as a part of the Winter 2025 issue of Columbia Engineering Magazine.
Read their conversation: bit.ly/40FIhCa
@columbiauniversity.bsky.social
#ColumbiaEngineering #E4H
My thoughts on DeepSeek sent to a reporter asking for comments
27.01.2025 16:20 β π 1 π 1 π¬ 0 π 0AI at Columbia Engineering https://www.engineering.columbia.edu/about/magazine/winter25/artificial-intelligence/drafting-blueprint-ai-columbia-engineering
16.01.2025 21:25 β π 3 π 0 π¬ 0 π 0π Ever wanted to peek inside an LLM's "brain" while it thinks? Now you can share those insights with anyone!
TokenProbe's new sharing feature lets you capture and share LLM behavior with just one click. Watch probability distributions evolve in real-time π§
Try it: https://shorturl.at/yR8EQ