Mahesh Sathiamoorthy
@madiator
Followers
13K
Following
5K
Media
595
Statuses
4K
Post training agents using RL and data curation. Co-founder @bespokelabsai. Ex-GoogleDeepMind.
Token Town
Joined February 2008
Very proud of this work and the team! . Nvidia released nemotron recently which is a great open reasoning model. The OpenThinker team worked tirelessly and heroically and curated what's arguably the best reasoning data, and got the model to be better than nemotron (and gpt4.1).
Announcing OpenThinker3-7B, the new SOTA open-data 7B reasoning model: improving over DeepSeek-R1-Distill-Qwen-7B by 33% on average over code, science, and math evals. We also release our dataset, OpenThoughts3-1.2M, which is the best open reasoning dataset across all data
1
7
79
100 is very little. Announcing my new product where you can basically read 18k books in ten minutes. Every book will be converted into an image and then at 30fps in ten minutes you can finish 18k books. Like a champ. Sign up will open tomorrow.
pro tip: you can basically read >100 books per day by asking chatgpt to summarize them for you.
1
0
8
For every first principles thinking based reasoning, you can find a contrarian reasoning that's also based on first principles thinking!. And also people seem to like to throw this term to everything and claim victory.
2
0
3
We made it guys. Now straight to IPO.
Soham Parekh 50 Index 🤣. Maybe A16Z and other VCs should start tracking the companies Soham applied to, worked at, or did trial periods with. Some big names on that list.
3
0
13
Looks like Soham applied to Bespoke as well (via a google form we had -- and his CV was uploaded). This is the new badge to carry: if Soham didn't apply you are not a serious startup. :D
13
0
69
Most of the people using Microsoft Word seem to be collaborating like it is the 2000s: they will add a comment and send the doc by email. Nobody gave them the memo?.
3
0
6
All this hype, and ultimately it's a meeting app. Is it useful? Yeah, but if you keep using it, you become useless because you have no idea what happened in a meeting. You are just a conduit for the LLM. Come to think of it, these guys just say it aloud ("cheating"), for a lot.
introducing @cluely. today is the start of a world where you never have to think again. we just killed 9 industries (thread):
6
2
15
Aditya continuing the awesome work on Matformers!.The models are open:
📢Now open, Gemma 3n weights & it is natively flexible, first of its kind, thanks to MatFormer🪆. Any model between E4B & E2B with ZERO training near Pareto -- we found a bunch!. Find a better E3B than what we released, I will send you a 🪆😉. Find the colab for extraction 🧵👇🪆
1
2
10
RT @MercatJean: We evaluated more than 1000 reasoning LLMs on 12 reasoning-focused benchmarks and made fascinating observations about cross….
0
18
0
Open research is a rising tide that lifts all the boats. This is amazing and thanks to Andy and the Laude institute for this effort!.
Today, I’m launching a deeply personal project. I’m betting $100M that we can help computer scientists create more upside impact for humanity. Built for and by researchers, including @JeffDean & @jpineau1 on the board, @LaudeInstitute catalyzes research with real-world impact.
0
1
8
If you come with a cursor for X startup idea, there is probably already a startup in YC. And as someone noted, the C in YC is Cursor :).
We’ve got Cursor for coding… but where’s the Cursor for design?.
0
0
5
Claude is the current winner for coding for web-design and where design element is important.
1
0
10
Love this discussion. Nathan's blog is great because he didn't stick to some incorrect info: he had come up with some hypothesis based on the knowledge at that time and then he posted his new learnings once new info came to light. I also think this is a good exhibit for the
1
13
125
Does anyone know how big are the Kim 1.5 models? They don't seem to mention it anywhere!
1
0
8
Has anyone researched/analyzed how much transfer learning there is in computer use? Let's say we consider only using websites: does training on 1% of the websites is enough to cover 90% or do you need to train on 10%?.(1% is already an absurdly high number).
0
0
5
In classic Google fashion, there is now this and jules. This is an example of shipping the org chart: one is from Google Cloud and the other is from Google Labs. I am sure both are great, but it's confusing to the users.
I’ll be honest with you. Our @googlecloud Gemini Code Assist was useful, but with too many rough edges. My DevRel team, along with product, engineering, and even @ThomasOrTK had enough. So we’ve spent months removing friction from it. It’s not perfect, but it’s better! A 🧵:
4
1
12