Bishal Santra (@mr_bytestorm) header
Bishal Santra (@mr_bytestorm) profile picture

Senior Researcher @Microsoft @MSFTResearch | IIT KGP | https://nitter.cf/t.co/oXDQhZx8mR | Fan of Machine Intelligence and Animal Intelligence

India
Joined June 2012
Photo by Bishal Santra (@mr_bytestorm)Photo by Bishal Santra (@mr_bytestorm)Photo by Bishal Santra (@mr_bytestorm)Photo by Bishal Santra (@mr_bytestorm)Photo by Bishal Santra (@mr_bytestorm)Photo by Bishal Santra (@mr_bytestorm)Photo by Bishal Santra (@mr_bytestorm)Photo by Bishal Santra (@mr_bytestorm)Photo by Bishal Santra (@mr_bytestorm)Photo by Bishal Santra (@mr_bytestorm)
pirateface.co/ @thepirateface Nice work! Really digging the model browser btw!
The solution exists... P2P for model weights
1
5
552
Bishal Santra retweeted
Had a Teammate run on Fable 5.2 for a few hours and build an endless zombie shooter that you can play on your browser. No use of any 3D assets. Just rawdogging it all in code. cordon9.pplx.app/
152
53
21
933
82,322
But how do you verify that the LLM generated lean interpretation of the code is correct? I.e. whether the lean verification is transferable to the codebase.
2
140
Bishal Santra retweeted
Bursting crackers while wearing a mask is a different level of irony ive seen today.
1
13
647
Bishal Santra retweeted
Sorry babe can’t talk right now. Claude was about to auto-compact
12
3
94
8,357
Bishal Santra retweeted
India is waging a war against its own researchers: In Careers360’s Research Summit, Prof. Ram Gopal Rao, Vice Chancellor of BITS Pilani described India’s lack of investment in research beautifully. He said “India’s Higher Education is described as African Budgets, American ambitions, Russian methods, Chinese opacity and Indian Bureaucracy. In the past decade, we have seen a slide in scientific temper and investments in research. 1. Non-NET fellowship are stuck at ₹8,000 a month since 2012, for 16,270 scholars in 45 central universities. 2. JRF stipend: In 2014, the JRF was Rs.25,000 per month. In 2019, it was revised to Rs.31m000. it is ₹37,000 PM. The hrowth hasn’t even kept pace with inflation and in real terms, it has FALLEN. IISc scholars even went on a hunger strike demanding an increase. 3. Maulana Azad fellowship: where more than 1,400 scholars have had stipends stalled since January 2025. 4. OBC fellowships: lists for three NET cycles in 2024-25 were held back, which the social justice ministry put down to insufficient funds. 5. INSPIRE: fellows have their stipends unpaid for 3 to 9 months. The schemes now under Vigyan Dhara fell 67.5% between 2016 and 2025. 6. KVPY: scrapped in 2022, after nearly two lakh students a year had taken its test. 7. JRF bottleneck: only 5,108 of 7,35,614 candidates (0.7%) got JRF in December 2025. Minister Sukanta Majumdar gave no year-wise data when asked. 8. The University of Hyderabad stopped Rohit Vemula’s ₹25,000 stipend in July 2015, and his last letter said he hadn’t been paid for seven months. He later committed suicide. 9. India’s R&D spending: fell from 0.82% of GDP in 2009-10 to 0.64% in 2020-21. PMRF spending was revised to ₹290 crore, less than half its ₹600 crore outlay. 10. SC fellowships: fell 61%, from 4,841 to 1,872, between 2020-21 and 2022-23. If India indeed wants to be a Vishwa Guru, India should be investing in research, in creating quality Higher-Ed institutions. We cannot be cutting down on stipends, not paying our researchers, holding back payments, scrapping scholarships, not increasing fellowship amounts even to match inflation and still expect students to choose research.
75
724
82
2,248
109,720
Diff: 80's expert systems were trying to figure out both knowledge engineering and process engineering. With LLMs, the knowledge part is taken care of (either through LLM weights or an external search toolset), so today's harnesses can focus more on process engineering.
Ngl, LLM harness engineering gives me 1980s expert-system vibes.
1
131
Iykyk 🤣
> good evening, how are you? ● Bash(cd ~/repo && git status --short | head -20 ...
1
113
I didn't get time to read this yet, but this exactly was the question I had. Did it write the weights directly (this would be insane!), or used some tools to pinpoint issues in the weight space and fix them? Cool work nevertheless.
This is so insanely misleading. No, it did not edit the raw weights. It used existing tooling to identify a problem area for the model, then it found the direction associated with it, and ablated that from the model. The oversell here is downright intellectually dishonest.
6
581
Bishal Santra retweeted
Introducing Honest Thumbnails A Chrome Extension to unslop YouTube thumbnails with stills from the video. Brings back the early 2010s vibe of Youtube where not everything was about clickbait. While looking into how @karpathy's HackerTube works, I learnt that YouTube has a CDN that provides still frames from the following chunks ¼, ½, ¾ of each video. So as you can imagine, it was only a matter of time until Claude turned this into a Chrome Extension. Honestly, it took me longer to make the icon than it took Claude to write the code. Open Source and MIT. Fork it, steal it, do whatever your heart wants with it. github.com/851-labs/honest-t…
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5: hackertube.org/ - Excellent recs without any logins (state is kept in local browser session) - Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads - Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version - HackerTube Longs ✊ A bit more on the process: I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing. Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
76
46
21
1,399
120,037
While people are fighting over how good or bad the project is, they are missing out on the key point in the last para: > "Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of."
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5: hackertube.org/ - Excellent recs without any logins (state is kept in local browser session) - Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads - Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version - HackerTube Longs ✊ A bit more on the process: I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing. Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
4
313
xPRT: Cross-Passage Retrieval Pretraining Danush Khanna, Himanshu Singhal, Pranjal A. Chitale, Bishal Santra arxiv.org/abs/2610.40122 [𝚌𝚜.𝙸𝚁]
3
4
76
In our team, we have been really curious about how to break the barriers of the standard pretraining methods for retrieval (or even LLM training for that matter). There are couple of challenges that piqued our interest. Contrastive learning-based supervision is surface level for most training data points, esp. with random in-batch negatives. This question and subsequent explorations led us to xPRT. On top of E5, existing retrieval pretraining methods barely move MTEB. xPRT adds +1.80. arxiv.org/abs/2610.40122 More on how we got here, and what we’re still stuck on 🧵
1
3
17
684
Another question that we are still interested in and have not fully solved is how to fix the slowness of the standard pretraining process. This is most severe in long training runs. After the initial training phase, the data efficiency goes down badly. This is something that we’re still exploring.
1
1
1
121
Work with the fantastic team at Microsoft Research India w/ @pranjalchitale, Himanshu Singhal, @danushkhanna7 @MSFTResearch Thanks for the shoutout @_reachsumit ♥️ nitter.cf/_reachsumit/status/211… arxiv.org/abs/2610.40122
xPRT: Cross-Passage Retrieval Pretraining Microsoft presents an unsupervised pretraining method for dense retrievers that learns from how passages within a document relate, with a passage's embedding predicting related passages the encoder never read. 📝 arxiv.org/abs/2610.40122
1
2
106
Bishal Santra retweeted
xPRT: Cross-Passage Retrieval Pretraining Microsoft presents an unsupervised pretraining method for dense retrievers that learns from how passages within a document relate, with a passage's embedding predicting related passages the encoder never read. 📝 arxiv.org/abs/2610.40122
8
1
27
1,116
This is a whole different level of "attention is all you need" 😵‍💫
1/4 Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM. It runs C and MicroPython in its forward passes, no external runtime. On DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%.
1
185
LLM contexts have a really bad design right now. Serial, full of noise, no structure...
89