Siva Narayanan (@K2_181) profile picture

Co-founder CTO @FyleHQ (acquired by Sage). ex-Qubole, ex-Greenplum. Ph.D. in Big Data. @sushkini's lesser-half. Thought follower, synergy explorer.

Bangalore, India
Joined July 2009
Photo by Siva Narayanan (@K2_181)Photo by Siva Narayanan (@K2_181)Photo by Siva Narayanan (@K2_181)Photo by Siva Narayanan (@K2_181)Photo by Siva Narayanan (@K2_181)Photo by Siva Narayanan (@K2_181)Photo by Siva Narayanan (@K2_181)Photo by Siva Narayanan (@K2_181)Photo by Siva Narayanan (@K2_181)Photo by Siva Narayanan (@K2_181)
Wow!
The first 100,000 CVEs took 18 and a half years to publish (Sep 1999 to Apr 2018). 2026 published 100,000 in 322 days.
212
It hits harder as a catchy song.
1
2
11
485
This is the only way. @krishnabgowda ji - pls invite CM and others to a different part of the city every month.
Power of Common Man ❌ Power of Chief Minister ✅ Varthur-Gunjur Road in Bengaluru was in a pathetic condition for years. Nobody bothered to fix it. But as soon as news came that the Karnataka CM would be visiting, the entire stretch got fresh asphalt and proper lane markings overnight. People have been struggling on this road for years but one CM visit was enough to get the work done. Good that the road is finally fixed, but why does it take a VIP visit for basic road infrastructure?
2
5
57
3,632
This is sooo cool
in case you were curious: fable 5.2 and I confirmed that if you record yourself singing and whistling every note in your range, you can play basically any song using only your voice no song is playing in this video. every sound is one of 64 clips of me holding a single note
9
509
This cannot be true, right?
The bug is like a joke. On Alpine Linux, type 257 to 1024 characters of anything and you get root. 💀
2
335
Siva Narayanan retweeted
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools. It chats in Tamil, Tanglish and English, translates, and can call tools. We're releasing the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop. The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records. Thank you to everyone who spent time with that early version. The video walks through the model. You can try it here: hf.co/spaces/mechramc/agaram… நன்றி 🙏
113
421
36
2,269
93,669
Hey @Swiggy took a nice pic with your logo in case you want to use it
1
27
2,846
massive update! claiming victory on the most dirtiest spot i’ve seen in bangalore. we transformed it in less than 10 days. wind tunnel road is no longer a garbage blackspot. here's exactly how we did it: step 0: when i saw the video post by @madhuriadnal, i called mr. bhimesh (DGM) and mr. ravi (AGM) at BSWML. there was immediate action. @arunpaiblrwalks joined the next day, and we had an approximate plan for execution ready. step 1: a one time clean up of all the legacy waste. this isn't the fix but it gives you a start so you can track how much waste gets dumped overnight. BSWML also made sure garbage collection happened every day from the areas around. step 2: study the content and quantity of garbage. figure if it is household, restaurant or street vendor waste. step 3: basics of broken windows theory. we illuminated the spot with a floodlight and parked a garbage auto there at night. people started dropping waste bags in the auto instead of on the road. by step 3 (~day 4) we knew most of the waste came from street vendors on wind tunnel road and some of it from residents who miss the morning pickup. step 4: early beautification. a green net along the bridge, the walls painted white. people don't dump when a place is maintained, illuminated, and beautified. step 5: enforcement. a marshal was on site for a couple of nights, fining and educating people who dumped. we started seeing results. overnight dumping dropped sharply, and more than once we saw people pick up what they'd thrown and put it in the auto. step 6: take over the space by final beautification, a kerb, reflective bollards, and a footpath. on the final day everyone showed up. @GBA_office, BSWML, arun and his team, regular volunteers, and people who'd followed along here on X. we painted the wall and planted saplings. (the great volunteers @amitpgupta @frunkad @__jstar @anurag25phadnis @shbhmmshr)
hi @madhuriadnal, thanks for recording this. i just spoke to mr. bhimesh (BWSML deputy general manager) to get this cleared up asap. BWSML said this will be done early morning tomorrow because the garbage collecting folks work until 2:30 PM. they’ve promised to send pictures/videos after getting this done. i got in touch with him using iknowaguy.help/bangalore/fin……. let’s see what happens.
47
79
18
656
33,840
Siva Narayanan retweeted
we have received a lot of dms asking: how 2 people, without a research lab, trained hobbylm using roughly 100 - 400× fewer gpu-hours than many models in this space so here is the short answer: we spent 76 hours training on 4 h200s, bringing the full base run to roughly 304 gpu-hours. for context: sarvam-1: ~122,880 gpu-hours zoho zia 7b: ~153,600 gpu-hours llama 3.2 1b: 370,000 gpu-hours llama 3.2 3b: 460,000 gpu-hours i will share the engineering decisions, failed experiments, what worked, and the full training story over the next few days. built with @hariharanms95.
We are releasing HobbyLM from @Fuellabsai : a sparse MoE trained from scratch, with 1.037B total parameters and ~305M active per token, on ~100B tokens. Here's how we built 🤞
27
42
8
375
98,665
Bengaluru last elected its ward councillors in 2015. Since then the city has spent about ₹38,455 crore a year with nobody local to blame for the pothole outside your gate. The ward elections are coming soon, hopefully 🤞. A few of us are building something for them. 🧵
11
66
5
257
8,052
Special shoutout to Claude (@claudeai @AnthropicAI) and Codex (@OpenAIDevs @OpenAI) which enabled us to build this.
1
1
4
289