Day 4/120.
Started CS336 and The SpeechLLM Playbook today.
Already knew most of the CS336 concepts, so I'm mostly here for the experiments.
Got a surprise for y'all tomorrow. Hopefully I'll finish this course and build a few similar things by Aug 20. Let's see.
finally allowed to talk about Gemini 3.5 Pro
tested it privately for 2 weeks. it changes everything
> better than Opus 5 in every way
> real multi-tool use
> SOTA frontend
> destroys 5.6 sol and fable 5 on long coding tasks
> beats every frontier model with /teamwork-preview
> understands subagents
> understands what you really want
> keeps working until the job is done
> way cheaper than every frontier model
can’t wait to test it tomorrow 🙃
drop your prompts below, i’ll test them 👇
finally allowed to talk about Gemini 3.5 Pro
tested it privately for 2 weeks. it changes everything
> better than Opus 5 in every way
> real multi-tool use
> SOTA frontend
> destroys 5.6 sol and fable 5 on long coding tasks
> beats every frontier model with /teamwork-preview
> understands subagents
> understands what you really want
> keeps working until the job is done
> way cheaper than every frontier model
can’t wait to test it tomorrow 🙃
drop your prompts below, i’ll test them 👇
Day 3/120
Compounding a little every day.
Feels like my brain is slowly de-fragmenting after relying too much on AI. Started diving into speech and audio today, took a few meetings, and the product we're building as a team made solid progress.
I'll share what we're building, the books I'm studying, and everything I'm learning soon. :)
Day 1/120.
Started slow today. Took me some time to get back into the zone, but once I did it felt good.
Wrote some PyTorch again after a long break. Feels like I'm finally learning again.
Will start posting more detailed updates from tomorrow.
Almost forgot to mention this.
About two months ago, I started taking fitness seriously and have been showing up at the gym consistently. The progress is small, but it's finally visible, and that feels good.
I'll be documenting this journey too. Nothing fancy just trying to get compound a lil everyday! 🥂
Along the way, I'd love to connect with people doing great work in inference engineering, computational biology, and growth. Always happy to learn from builders
Day 0/120
It's been a while since I truly learned something deeply or built something from scratch.
So here's a reset.
For the next 120 days, I'll be documenting my journey as I revisit deep learning, go deep into speech systems, distributed training, and synthetic biology, while building something genuinely useful.
I'll also be learning distribution and growth by helping scale the companies I'm working with.
See you on Day 1.
The Indian AI story nobody's talking about 🇮🇳
• Maya-2-Native is now #1 on Voice Arena's real-time Hindi leaderboard.
• Ranked using blind human evaluations (Bradley-Terry Elo).
• Built by @mayaresearch_ai .
How did this fly under everyone's radar?
You are the best judge of a voice model’s quality.
Voice is human. The leap between today’s voice models and Maya Research’s new models is something no leaderboard can capture. You have to hear it for yourself.
• Yu Deng: Connected microscopic particle motion to the equations governing gases and fluids.
• John Pardon: Developed groundbreaking methods in geometry and knot theory.
• Jacob Tsimerman: Solved major problems in arithmetic geometry, including the André Oort conjecture.
• Hong Wang: Solved the 3D Kakeya problem, a landmark result in harmonic analysis.
There are a few moments you only understand looking back.
You trace where you are now, and you can see exactly where it started.
For me, it was the SPC Founder Fellowship.
Somewhere in the fellowship, it hit me: what I was actually obsessed with was bringing technology closer to the people technology usually skips.
That became Maya’s global maxima: building the voice interface for the next five billion.
Honestly, I don’t think I would have found that clarity without the SPC team.
Go Find these people. Find your global maxima.
@spc@spc_india