Sharvil Nanavati

🚨 Hey folks I got something to announce: I am launching a new investment fund brand: Prototype Capital 🦾 A globally-active first-check investment fund. We focus on cool stuff that's technical hard to achieve, but also still new enough to be a bit weird – globally. Think Robotics in Eastern Europe, AI in SF, Software remotely, OpenSource SaaS, Space-tech in India, Reindustrialization of Europe, AI Agents in Brazil, among many others… - $100k-$200k - we say we invest early and we actually mean it - with every 5th investment we are the first to commit - only 6% of all ivnestments were after seed - the brand will be a larger umbrella for all kinds of stuff from our own prototypes 😬, to streams, to communities, to investinpublic transparency efforts - this is legally speaking a rebrand of my previous solo GP fund so we got track record to proof - we do about 15-20 investments per year Link in thread! 🔥 Please RT if you got a sec – Appreciate your launch support 🙏❤️

@OfficialLoganK What's the quota on Gemini Pro Exp-0801? The experimental model is sooooooo goooooood!!

We just removed the RPD limit for Gemini 1.5 Pro (paid tier), giving developers the ability to go from 10,000 RPD to now 518,400 RPD 📈 Happy building : )

@OfficialLoganK The experimental 0801 model is a huge leap forward – what are the quotas on that? I keep getting rate limited and don't know which limit I'm hitting.

Gemini 1.5 Pro free tier comes with: - 2 RPM (requests per minute) - 32,000 TPM (tokens per minute) - 50 RPD (requests per day) More modest, but shows what our higher intelligence models are capable of. (3/4)

We are giving developers 1,500,000,000 tokens for free everyday in the Gemini API There is no stronger developer value proposition out there 🧵 (1/4)

Whoa did @GoogleDeepMind silently drop their new Gemini model? It's already topping the lmsys arena! chat.lmsys.org #LLMs #ai

Chatted with @cerebral_valley folks. You know the drill. We're hiring, so hmu!

@unilightwf Why is point 1 not surprising? I assume it's because of limited labeled data, but it seems that in principle it should perform better... what am I missing?

VoxSim: A perceptual voice similarity dataset mm.kaist.ac.kr/pubs/pdfs/ahn2… Two takes: 1. No surprise: similarity predictor (two samples in -> sim score out) << spk emb cos similarity 2. Cos similarity only has <0.8 correlation -- so why the hell did we use it so much in TTS papers?

@unilightwf Agreed, seems unlikely even in the case of masked autoencoding models since the masks don't always span semantic boundaries (and even if they did, you'd still have an alignment problem?)

@snrrrub IMHO most existing works are just exploring the possibility, but it's hard to believe there would really be semantic info in these encodings. For instance, wav2vec is based on CPC, which essentially learns slow features, and I think that's basically phonetic info in speech.

arxiv.org/abs/2406.08619 I thought this is a known fact…

@rezmeram Which kinds of personalities would you want him to chat with? I presume it wouldn't be a generic assistant.

Can you accelerate deployment of the next generation of voice please? I care for an elderly person who often feels lonely. It would be nice if I could drop a GPT on his phone to help him ease the loneliness he feels during the last days of his journey. Thanks. p.s. I need to make a living, can't spend all my time with him.

GPT-4o as an assistant for helping doctors screen and treat cancer patients:

@erogol Any experience with Schrodinger Bridges?

I'm playing with flow-matching models and I started to hate the time I wasted with GANs. FM trains way faster..

@fchollet Given the long and expensive road to AGI, performing economically valuable tasks along the way may be a commercial prereq to building it.

I am not convinced that general intelligence is "the ability to perform most economically valuable tasks." My 3-year old can perform *no* economically valuable task, but he's one of the smartest guys I've ever interacted with. Meanwhile, control theory has automated millions of highly valuable industrial jobs, but no one would call a PID controller intelligent.

@MrCatid Sometimes you have to slow down to speed up...

@snrrrub Will look nice on top of all the other books I don't have time to read because AI is way cooler to work on

Richard Hamming (of Hamming distance fame) wrote one of the best books on critical thinking and problem solving I've come across. And it has a whole section dedicated to AI (in 1997!). Highly recommended! goodreads.com/book/show/5304…

@Marktechpost @MSFTResearch They highlight one of the key reasons I work on speech synthesis: "The advantages of this work could contribute to valuable endeavors, such as generating speech for individuals with aphasia or people with amyotrophic lateral sclerosis."

Microsoft Researchers Introduce VALL-E 2: A Language Modeling Approach that Achieves Human Parity Zero-Shot Text-to-Speech Synthesis (TTS) Quick read: marktechpost.com/2024/06/13/mic… Paper: arxiv.org/abs/2406.05370 @MSFTResearch

@dorienherremans @ArxivSound @Women_MIR Nice! Do you know of other accent datasets that can help your technique scale up?

Generate speech in your desired accent using a tts with Multi-Level VAE with Adversarial Training. 🔥🔥🔥 Paper: arxiv.org/abs/2406.01018 Github: github.com/AMAAI-Lab/Acce… Demo: amaai-lab.github.io/Accented-TTS-M… #tts #AI #womeninscience @ArxivSound @Women_MIR

@aipin_io Pretty sweet, what are you using for TTS?

Experience the charm of 🌲 Elmwood 🌲 through our Text to Speech feature! From the serene village ambience to the symphony of nocturnal sounds 🌙🎶, let our TTS bring the story to life. Check out the demo below to see it in 📽️ action! #AI #AIPIN #TextToSpeech #Storytelling #TTS

@Project0dyssey Love this! Are you hosting an in-person event?

We are excited to bring on @elevenlabs to elevate the audio capabilities of the Project Odyssey AI Film competition through their Text-To-Speech, Speech-To-Speech, Dubbing, and SFX tooling! Sign up now at: projectodyssey.ai #projectodyssey #elevenlabs

@ithinkimrishi Pretty much all configuration-based languages. Ansible was one of the more recent ones.

@snrrrub which domain specific language are you referring to?

``CLaM-TTS: Improving Neural Codec Language Model for Zero-Shot Text-to-Speech,'' Jaehyeon Kim, Keon Lee, Seungjun Chung, Jaewoong Cho, ift.tt/8oVgxDY

@MrCatid @knuth is a legend. He wrote about surreal numbers in the 70s!

@snrrrub Is this an April Fool's day joke? Because I've almost escaped unscathed.

Surreal number is a real thing.