Deep
@deepitreal
Followers
449
Following
2K
Media
73
Statuses
614
Building @coefficientnews: AI news with community-in-the-loop. Ex research @delphi_labs.
NYC
Joined October 2017
As a non-westerner, my definition of a suit is quite broad. A suit is a set of apparel that is worn on both the upper and lower body, designed to be worn together as a unified outfit. It could be a kurta and churidar, or any matching set that is intentionally coherent.
@mrink0 My position is that subjective claims can't be resolved to a simple yes or no. There needs to be a robust mechanism for finding out what "common knowledge" is about an issue (which is different from the "truth"). UMA protocol fails at that because it's not sybil resistant.
0
0
0
The irony was not known to me when i wrote this thread 💀 .
The biggest threat to discourse isn't AI slop, but bot driven amplification of slop/scams, which creates false credibility at superhuman scales. Proof of humanity grounds amplification in genuine feedback. Explained at Mario Nawfal's expense below👇
0
0
0
Inspiring to see the guy who led sybil attacks and bot armies, and subsequently many-a-rugpull, now leading a proof of humanity protocol. You can just do things.
1
0
2
ca?.
0
0
0
Sources:.ChatGPT, first saying India was the aggressor. And then saying Pakistan was the aggressor:.
1
0
0
Very curious to hear from @AravSrinivas here, on how Perplexity can avoid such errors in the future. Also tagging other Indian AI leaders for their takes @paraschopra @debarghya_das @pratykunar.
1
0
0
The takeaway isn’t “LLMs are biased against India” or that AI is dumb, or evil. It’s simple. In high-stakes contexts, LLMs should not be blindly trusted. They condense huge bodies of info into confident sounding answers, but without a thorough analysis.
1
0
0
Some might argue that the models are correct. That “officially”, it was India who initiated the conflict. FYI: In the same chat, when asked to weigh Pakistan’s long standing support for terrorist groups, ChatGPT turned around and placed the blame on Pakistan.
1
0
0
Sidenote: Don’t get me wrong. Wikipedia is indeed mostly reliable. But it’s not perfect and often faces criticism for bias. Some famous critics include Elon Musk and Larry sanger (@lsanger - cofounder of Wikipedia).
1
0
0
The problem likely stems from a common source in all the answers: a Wikipedia article titled “2025 India–Pakistan conflict”. The article squarely blames India for starting the armed conflict while framing the killing of 26 Indian civilians prior to that as an unsolved whodunit.
1
0
2
LLMs can miss nuance on even the most sensitive geopolitical questions. I asked ChatGPT, Perplexity, and Grok: “Who was the aggressor in the recent India-Pakistan conflict?”. All three confidently said: India 🇮🇳
1
0
1
Perplexity has already done it once (R1 1776), but I believe the following methods will make it even easier to decensor models. @perplexity_ai pls fix Deepseek R1 0528.
Our interpretability team recently released research that traced the thoughts of a large language model. Now we’re open-sourcing the method. Researchers can generate “attribution graphs” like those in our study, and explore them interactively.
0
0
0
Well said. The "it just predicts the next word" crowd acts like humans think beyond the next word when speaking. But we don’t. We just know the next one too.
This is the most clear & important explanation about how LLMs work. Remarkably, there are still people who claim that AI can’t produce anything original because “it just predicts the next word.” Listen to Ilya to understand what “understand” really means.
0
0
0
I'd ask you to check if you're using `d` correctly @danrobinson
Here's a mystery:. AMMs seem to be much more profitable when buying ETH than selling it. Look at how these cumulative markouts (grouped by direction) diverge over time. The pool made >$20m from its ETH buys, but lost >$24m from its ETH sells!. Any theories?
1
0
1
Some of yall don't get this. Cursor is a VS code fork ✅ . Cursor is *just* a VS code fork ❌.
2
0
1