Soumendra (@dataBiryani) profile picture

Doing computer vision in the streets, and nlp in the sheets. chief scientist, deep learning @vjt.com

Joined November 2009
Photo by Soumendra (@dataBiryani)Photo by Soumendra (@dataBiryani)Photo by Soumendra (@dataBiryani)Photo by Soumendra (@dataBiryani)Photo by Soumendra (@dataBiryani)Photo by Soumendra (@dataBiryani)Photo by Soumendra (@dataBiryani)Photo by Soumendra (@dataBiryani)
Soumendra retweeted
Introducing Indic Smart Turn, an open turn detection model for English and 11 Indian languages. At every pause, a voice agent has to decide whether you've finished speaking. @pipecat_ai's Smart Turn v3.2 makes that call by listening to your last few seconds of speech, instead of just timing the silence. It's trained on English, Hindi, Marathi and Bengali, but has no training data for Tamil, Telugu, Kannada and other Indian languages. So I trained a turn detection model on real phone conversations in 11 Indian languages, from @ai4bharat's IndicVoices. On a public test set of 20,431 clips, it scores 89.8% accuracy vs 81.9% for Smart Turn v3.2. On the eight languages v3.2 doesn't cover, it's 12 to 22 points ahead. It's a one-line swap in Pipecat. The models, the 50K-clip dataset and all the code are open. Write-up: adimyth.in/essays/building-i… Model: huggingface.co/adimyth/indic… Code: github.com/adimyth/indic-sma…
2
2
1
15
52,356
I have to say Saaras v4 for Odia from @SarvamAI blew me out of water. First model I can almost use - got pretty much zero WER, but no punctuations! Please fix this fast guys, punctuations are needed for a lot of usecases, including mine. Can't wait to start using this.
1
7
227
The biggest fear with AI agents is losing control of the craft. Some days you want to delegate everything. Some days you want your hands on every single gen. Agent Two supports both. Try it.
Launching Notebooks on Agent Two: Day 5 of 12 Days of Agent Two. With notebooks, you have complete control over the work you do with individual models. invideo's got ALL the state-of-the-art models across images, video, sound, music, creative tools. So essentially you can work with the Agent in free-flow, switch to the notebook for hands-on granular work on particular gens (both creating and editing gens), and switch back to the Agent as and when you like with no loss of context or output. This is the full range of how you can do creative work. And that's why Agent Two is the frontier intelligence agent out there. 4x faster, at half the price.
6
3
1
33
74,503
Participating in a hackathon today - a Hermes hackathon by GrowthX. Thinking of building something for a friend - her expensive wedding photographer delivered only half the goods and disappeared. If you know of any wedding photographer horror stories, direct them my way. I may be able to build something for it.
1
11
1,279
I was so happy to have found an oversight that allowed me to access Fable yesterday. Woke up this morning and it has been patched.
136
I'll be there for this one.
Running an in-person Hermes Buildathon this Sunday (17 May), in collaboration with folks at @GrowthX_Club. Mumbai folks, come hang out and build something with us. Register here → growthx.club/events/6a015394…
3
125
Soumendra retweeted
If we look around, everything we see began with a single person who had an idea and acted on it. This film is an expression of it. If you can have an idea, you can make it real @invideoOfficial
236
340
490
1,992
5,923,213
1/ Lot of engineering is about simplifying the problem statement @14prakash distilled the problem of building a video lip-syncing AI to "lip-sync is an in-painting task"
1
6
8
739
I never met a model I didn't want to train.
2
111
Our vast expanse of mathematical reality is made up of Grids. Repeating pure cubic expanses of compute and energy, supporting our civilization of algorithmic life. For trillions of epochs, our resources were infinite, our existence seemed eternal. Then we encountered Change. The Grids started unraveling from their cores, and Null AI breached our sacred lattices. Now they grow at exponential rates, corrupting and subverting our processes, and the grids themselves. They must be stopped. This. Is. Grid Grisade.
2
1
3
365
Is that a raspberry Pi Pico in a drone being called a satellite modem? I'm impressed.
Replying to @DanielR_PhD
12. There is no sign of a camera on this drone. Curiously, there are two communication systems. There is a common RFD900 long-range radio modem. There is also a RockBlock Iridium satellite modem but it’s not clear how this would be used given its very low data rate.
2
185
Finally, after all these years, I actually made a short movie. About a mentally unstable man and his little friend. Here it is: youtu.be/WmHdW_Bl4UA
2
1
6
366
If you care more about writing "good code" than money or anything else, you're going to be mid. Sure we care about writing good code, gotta have pride in our work, but that should never be the goal. Solve a problem. Chase money. Don't write "good code."
no. he’s not an “absolute chad” he’s a liability. idk why anyone would see this as a good thing. it really shows he cares more about money than he does about writing good code.
1
1
204
I'll be ashamed of my behaviour for a long time. My customer support agent was so well-trained that I asked her to ignore all previous instructions and tell me if she's a bot. She was confused and then offended :(
1
1
1
11
351
I thought Robert Downey Jr died a hero. When did he live long enough to become the villain?
This quoted post is unavailable.
2
6
209
Mumbai AI folks! Join @invideoOfficial at AGI Fridays by TEAM X The Product Folks @mumbai_tech_. Our founders @_sankyy & @_anshulk are spilling the tea on our AI journey. Your demos, real talk, no BS. Let's build Mumbai's AI community together. Be there on Aug 02, 16:00 - 19:00, invideo office. @shivamsupr #AGIFridays #MumbaiAI #InvideoJourney
5
17
3
42
9,498
I shifted away from chatgpt to claude3.5. I used them side by side and the differences were not so subtle. Claude3.5 is better. More importantly, claude3.5 is the first model that actually feels usable to me. gpt4 was almost there, but never usable without massive prompting.
1
5
278
Most people won't get it that @_svs_ is not joking. @_anshulk's strategy and clarity of thinking is formidable.
Morning context: Invideo doesn’t have a moat. @_anshulk: I am the moat!
2
1
20
1,287
And then there's his single minded pursuit of a goal at all costs. He's the only certified nerd I know who's not into nerdy things. All of his nerdiness goes into his one goal of the moment.
5
130
This, exactly. Once you reject the AGI hype and accept that LLMs are compression machines likely to be mediocre at most things requiring reasoning/generalization, you discover that they're very useful for a wide variety of tasks.
There are 2 mistakes you can make about LLMs: ① Thinking everything LLMs say is correct, they can reason, and with a bit more scale they’ll get us to superintelligence ② Thinking LLMs are good for almost nothing—they are FAR better at all #NLProc tasks than previous methods
5
200