CodeWhisperer

@benhylak Can you share what you know ? How exactly will it change ?

multiple people who knew better and had no incentive to lie explained to me why i was wrong.

@Nithin0dha Money and stocks are ultimately notional, valuable only because we believe they are. In ancient India, professions like healthcare were driven by respect & status, not profit. Somethings are invaluable. Globalization shifted that balance, making money the primary measure of value

I recently had dinner with Dr Devi Shetty, the founder of Narayana Hospitals. For those who don't know him, he's the guy who figured out how to do open heart surgery for a few hundred dollars when the same procedure costs a bomb in the US. Narayana has 18,000 beds across India, and if you ask most middle-class people in Bangalore about it, they'll speak highly of it. There was one thing I kept thinking about over and over again after meeting him. Narayana's market cap is around ₹38,000 crore. Now compare that to pretty much any half-decent financial services business in India, and it'll be valued more than that, including Zerodha. A brokerage, worth more than a hospital chain, that has probably saved hundreds of thousands of lives. I get the arguments. If you're a fund manager/analyst, you can immediately explain it away using margins, capex, asset-light vs asset-heavy, and all that, and I'm not saying the market is wrong. But it's still a strange world we've built, where the businesses closest to money get valued the highest, and the ones doing the hard and essential things get priced like boring utilities. A hospital carries physical infrastructure, enormous liability, thin margins and the actual weight of keeping people alive. And somehow that's worth less than a platform for buying and selling stocks. I don't have a clean take on this. All of this just felt odd. Ps: Nothing here is investment advice. For that, go to @zerodhavarsity

@prompt_Tunes Can you share reason behind choosing to optimize SNAC over transformer decoding or scheduling ? SNAC is a few ms any way ?

put out a new worklog on optimizing snake1d activation kernel in triton

@null_hawk Whats your reward model ?

Its been approx 20+ hrs and finally reduced my GRPO runtime from ~12.8 hrs to ~1.5 hrs. used vLLM inference on one GPU and DDP on the other 3, applied pretty much everything I had. using 3 GPUs for DDP, hosted vLLM inference on one GPU, with 8 samples generating parallely, along with codec tokens, grad_accum=4, data sharding across ranks, micro-batching within the loss function, TF32 tensor cores, BF16 autocast for forward pass, Flash SDP + mem-efficient SDP, cuDNN benchmark mode, BF16 reduced precision reduction... works well for Hopper architecture. also turned on gradient checkpointing, manual gradient all-reduce (instead of full DDP wrapping), plus a file-based vLLM watcher that restarts the inference server with fresh merged weights at every optimizer step from a clean process tree (to avoid NCCL conflicts with torchrun), with retry logic on generation calls during server. biggest debugging rabbit holes: vLLM V1 silently ignoring stop_token_ids (had to force V0 with VLLM_USE_V1=0), and merge_and_unload() with tie_word_embeddings=True dropping the trained lm_head during save, model generates infinite codec tokens and never stops. fix: untie before merge so both embed_tokens and lm_head are saved separately. GRPO on TTS is a different beast from text!

@NayrhitB @SusquehannaVC @LightspeedIndia @BCapitalGroup @beenextVC @sparrowcapvc @2point2club deck

The exact pitch deck that helped us raise a $9M Seed Round copy whatever you want VCs that invested: → @SusquehannaVC (led) → @LightspeedIndia → @BCapitalGroup → Seaborne Capital → @beenextVC → @sparrowcapvc → @2point2club joined. fundraising is hard enough without guessing what investors want to see. so - I'm making our deck public. if you're raising right now, take it and make it yours. Reply 'deck' + follow (so I can DM it over)

@kadirnardev @yukiarimo @realmrfakename @kadirnardev can you DM me ? I will share training code if you have time to take a look. I'm using SNAC and I'm seeing a representation collapse. Model predicts all 7 tokens per frame but they are not unique. It predicts first 3 instead of using all 7.

@kadirnardev @yukiarimo @realmrfakename Wrote Jax code, Yodas around 88k. Training on TPUv4-32, half way through. 60hrs to finish 1 epoch

My new TTS architecture, trained on 10k hours of data for just $350:

@Yatharth3501 Can it handle 10+ models in a single model ?

Introducing LuxTTS: A 150x realtime high-quality voice cloning TTS model. LuxTTS's other features: - Highly expressive and realistic. - Excellent voice cloning. - Fit within 1gb vram. - 2-6x realtime on CPU. Repo: github.com/ysharma3501/Lu… Model: huggingface.co/YatharthS/LuxT…

@kadirnardev @yukiarimo @realmrfakename @kadirnardev if you have trained llm based models. How does loss curves look for them ? After how many hours of training will we start listening audio ? I’m getting random noises after 40000 hrs on yodas. Not sure if it’s because of data or should I train more or something else

@yukiarimo @realmrfakename @WhispererCode We have shared the training code for LLM-based models, but I really want to share the training or architecture for a DiT-based TTS, but these studies need to remain private :/

@BineDivyesh How did you get it ? How to apply and what’s the criteria ?

The Microsoft sponsorship is coming in clutch. 🫡

@pshishodiaa Nice work @pshishodiaa ! Would you be able to share any insights into model architecture and how you managed to train it in $1000 ? Did you build your own codec and any other optimizations ?

this is just the beginning. The lightning strike of bitter lesson for speech models is not very far.

@AP_CRDANews @apsrtc @ncbn @AndhraPradeshCM @naralokesh @RamuduM_RCT @AndhraNexus @Rajahmundry4U @introvert123_ @sainad777 @GreaterBezawada Add Kadapa to it

To @apsrtc please start "Airport Shuttle" services to nearby Urban Centers in #AndhraPradesh Rajahmundry Airport - #Rajahmundry #Kakinada Vijayawada Airport - #Vijayawada #Guntur #Eluru Tirupati Airport - #Tirupati #Srikalahasthi Kurnool Airport - #Kurnool #Nandyal

@apsrtc Pls expand vennela, starliner and Amaravati from Proddatur to Bangalore, Hyderabad, Chennai

APSRTC ensures your safety and comfort on every journey 🚍 With thoughtfully designed emergency exits in every gangway, our VENNELA A/C and STAR LINER Non-A/C Sleeper Buses redefine safe and peaceful night travel Book your tickets online@ apsrtconline.in

@balajis In low income democracies leaders end up serving least common denominator. They can’t think beyond that as they are the biggest vote bank. And this vote bank is not far sighted

RECLAIMING DEMOCRACY I disagree with the idea that democracy must limit economic growth. Moreover, I think we actually need to reclaim democracy just as we reclaimed free speech, because above all democracy means the consent of the governed. But let me first defend democracy's economic track record. (1) First: democratic India posted the highest economic growth rate in the world over the last ten years, proving that democracy need not hold back an economy: (2) Second: India is now arguably better at democracy than many Western jurisdictions, as half a billion votes were counted in one day. That shows democracy is feasible even at billion-person scale in the modern age: (3) Third, perhaps obviously, democratic America was the richest and most successful country in the world. It was able to build for 200+ years, in part because it combined democracy with capitalism and the frontier spirit. (4) Fourth, as another example of how democracy can actually catalyze economic growth, the democratic vote in the late USSR publicly repudiated the Soviet far left, and led to the economic growth of Eastern Europe and the Baltics. (5) Fifth, what can happen economically when we lack democracy? Well, in California, once the blues succeeded in destroying democracy and turning it into a one party state where the Party always won, the looting of the public trough truly began in earnest. The abolition of competitive multiparty elections after 2010 is why Californians got $100B nonexistent trains: (6) Sixth, even in those contexts where the voting is more with feet than ballot, the essential principle of consent is preserved. For example, tech companies do have CEOs, and so there is top-down leadership. But they also have bottom-up consent, as every single person who's there has consented to be there and can leave at any time. This consent is what drives the performance of tech. (7) Seventh, the recent 97%+ vote in Starbase shows that we can manifest this type of consent in the physical world as an official vote, by combining voting with your feet (moving to Starbase), voting with your wallet (building Starbase), and then voting with your ballot (incorporating Starbase). (8) Eighth, just like crony capitalism isn’t a good reason to implement communism, so too democratic corruption isn’t a good reason to implement dictatorship. (9) Basically: I do understand the right’s critique of democracy, just as I understand the left’s critique of capitalism, but at the end of the day one must earn legitimacy through consensual votes just as one builds wealth through mutually beneficial transactions. Otherwise the right-wing anti-democrat is just the inverse of the left-wing anti-capitalist. The anti-capitalist wants wealth without earning it, while the anti-democrat wants legitimacy without building it. But in reality there are no shortcuts. You need the consent of free people, freely given, to build something truly great. And that’s what democracy really represents: the consent of the governed. Anyway, I'll write more on this, but we should reclaim democracy as we reclaimed free speech, by strengthening it for the Internet age. Because the alternative to democratic capitalism is communist dictatorship, and that just isn’t an acceptable alternative.

@arpit_bhayani Customer support in banking commerce

Not enough people care about Indic languages and transitively Indic LLMs. Tell me the last time we wrote and typed in our regional language. Most people who hold any kind of purchasing capability are not even comfortable reading an Indic script. I don't think there is enough incentive to do Indic (anything) except hopium - gambling, astrology, mandir apps, etc. and brainrot content. I'd be glad to be proven wrong. But I don't see that happening.

@deedydas @TheOneDit Pls show

@TheOneDit Sarvam raised $41M and I can definitely point to maybe 10 companies that raised less and delivered more.

India's biggest AI startup, $1B Sarvam, just launched its flagship LLM. It's a 24B Mistral small post trained on Indic data with a mere 23 downloads 2 days after launch. In contrast, 2 Korean college trained an open-source model that did ~200k last month. Embarrassing.

@deedydas Sarvam is man in the arena doing what they are doing with limited ecosystem and support. Don’t diss them, in fact don’t diss anyone doing something. Praise in public, criticise in private.

@deedydas We shouldn’t stop building models, may be we will imitate others initially eventually we will get to the top but if we stop building we will always stay at the bottom.