Building Praxel || Helping individuals and companies AI-Native

Joined October 2011
5 Photos and videos
Pinned Tweet
Two papers on @huggingface Daily Papers today 🇮🇳 Praxy Voice: open-source Indic TTS that beats Sarvam Bulbul on Telugu retroflex collapse (26.7% vs 33.3%), built from a frozen English base a LoRA adapter. No commercial training data PSP: Accent benchmark we used to measure it
1
2
162
First message gets me this warning 😂Fable 5's safety measures flagged this message for cybersecurity or biology topics. They may flag safe, normal content as well. These measures let us bring you Mythos-level capability in other areas sooner. Guardrailing too much? @ClaudeDevs
43
I have ~10 active Claude terminals at any given moment and almost always, all of them are not running the same version (check below image). Can't think of any other software where I ever used where this happened. Keep up with the speed and quality @ClaudeDevs @AnthropicAI
10
Amazing concept! X today has been raving Clicky's launch but this is the real deal. Kickass website and video.
this is Zero. i was going to launch @zero_university 6 months ago, but decided not to. i felt if we launched too early, the team would get distracted. the product still had areas where i wanted to make the experience incredible, not just 'good'. Zero is more like a video game or movie, where the experience matters. it's not a tool, it's immersive. how people feel when they use it matters. it needed to be fast, feel real, and actually work. so instead of launching, we went heads down for 6 months to craft the best learning experience the world has seen. it has the best applied AI models, immersive experiences that feel like the real-world (like kick-off meetings, feedback walk-throughs, and real-time collaboration), and modern workflows that train people on how to use AI tools to make them more productive, rather than use them mindlessly as a crutch. it was a tough decision because my team put a ton of work into this video and wanted the world to see it. imagine, this was made 6 months ago! i'm super proud of my team for what they made, and i'm even more proud to finally share this with the world. so yea, this is Zero. this is the future of education. to join the beta, let me know in the thread.
1
1
260
Useful update from @claudeai @ClaudeDevs @AnthropicAI Token usage transparency is becoming super important. Waiting to see it from Codex too.
29
1/ The best commercial ASR gets 0.16 Entity-Hit-Rate on Telugu. Open-source SOTA: 0.027. Neither handles phone numbers, prices, or brand names reliably. We (@PraxelHQ ) synthesized 22,000 utterances for <$50 and got 0.473. Thread on how.
1
34
6/ The ablation worth running: same LoRA recipe, same hyperparameters, no synthetic data. Result: 0.020 EHR. Fine-tuning without entity-dense data does nothing. 100% of the gain is from the data, not the recipe.
1
27
7/ Everything's open: code, synthetic data, eval metrics, model predictions. If you're working on Indic ASR or low-resource speech, it should save you some time. Paper: huggingface.co/papers/2605.0… Code: github.com/praxelhq/stt-flyw…
1
343
Is it just me, or have Codex limits reset? Thank you, @sama @OpenAI @OpenAIDevs
3
138
🧵 I gave Claude Code and Codex the same impossible goal. They took opposite approaches. One felt smarter. The other was technically correct. @AnthropicAI @OpenAI @kaggle Here's what happened:
1
1
80
My instinct was to prefer Codex. It shipped real results. But Claude was following the spec exactly. If one exploit had landed, it would've moved 10x farther than everything Codex gained. Codex adapted. Claude optimized for the stated goal. Both are defensible.
1
53
Do you want agents that chase the goal at all costs – or agents that adapt when the goal looks unreachable? What's your read? 🎯
29
Two headline social media releases this week: @X open-sourcing their recommendation algorithm @instagram launching "Instants" Tells you the direction and tg of these social media platforms.
1
1
66