Mahesh Sathiamoorthy

4.5K posts

Mahesh Sathiamoorthy banner
Mahesh Sathiamoorthy

Mahesh Sathiamoorthy

@madiator

RL Environment Curation. Data Curation (OpenThoughts). Post-training. CEO @bespokelabsai. Ex-GoogleDeepMind.

Inside a RL Environment Katılım Şubat 2008
1.4K Takip Edilen14.5K Takipçiler
Sabitlenmiş Tweet
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
We are announcing Open Thoughts, our large-scale open-source effort to curate the best open reasoning datasets! DeepSeek-R1 is amazing but we still don't have access to high-quality open reasoning datasets. These datasets are crucial if you want to build your reasoning models! Bespoke Labs released a 17k reasoning dataset last Wednesday, and the reception has been phenomenal (it's trending on HF). So we are joining forces with the Datacomp community to launch Open Thoughts --- an open data, open model, and open code initiative for creating the best open reasoning datasets and the associated models. Along with this, we release OpenThoughts-114k reasoning dataset and the associated OpenThinker-7B model. Links to the code, model, and data are below in 🧵.
Mahesh Sathiamoorthy tweet media
English
45
286
1.8K
226.8K
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
@mjamei that's just the interface. what about adding helper functions for calling llm-as-a-judge, storing to yaml, reading from yaml etc.
English
1
0
0
22
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
What's the library people use for defining/loading/processing rubrics?
English
6
1
11
2.9K
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
@AashaySachdeva Standardization, reuse etc. Also, I asked opus and it gave me this. Do you like this way of representing rubrics?
Mahesh Sathiamoorthy tweet media
English
1
0
0
195
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
Weekend project: getting back to some fixes in Curator..
English
0
0
4
723
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
@NandoDF They make more money since I use Claude periodically to keep my version of Claude code in sync with theirs.
English
0
0
0
154
Nando de Freitas
Nando de Freitas@NandoDF·
What happens to Anthropic when anyone can use Claude Code to generate Claude Code?
English
50
4
102
26.2K
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
Claude code high-five'ing itself about how well it explored the code :)
Mahesh Sathiamoorthy tweet media
English
2
0
18
2.1K
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
I will be talking at the Agentic Evals workshop organized by HuggingFace. It's streamed live, so mark your calendars!
Ben Burtenshaw@ben_burtenshaw

LIVE WORKSHOP: The State Agentic Evals! Agents are capable of multi-step reasoning, tool use, and real-world task completio, so evaluation needs to keep up. We will discuss topics related to questions such as: Where is the state of the art in evaluating agentic systems? Why are agentics systems’ performance on benchmarks not reflected in usage? How can we evaluate the agentic systems and language models that we use? This workshop brings together diverse perspectives from academia, industry, and policy to explore the frontier of agentic evaluation.

English
0
2
5
697
Mahesh Sathiamoorthy retweetledi
Daanish Khazi
Daanish Khazi@bertgodel·
We’re announcing Kos-1 Lite, a medical model that achieves SOTA on HealthBench Hard at 46.6%. As a medium sized language model (~100B), it achieves these results at a fraction of the serving cost of frontier trillion-parameter models.
Daanish Khazi tweet media
English
40
59
318
24.6K
Mahesh Sathiamoorthy retweetledi
rohan anil
rohan anil@_arohan_·
I feel a bit responsible for hyping agentic coding in December as I was having and still having too much fun doing best technical work. However I heard some gossip about certain big tech hiring fewer junior eng. so I wanted to make a point. If you want your engineering output to actually compound, hire ambitious junior engineers, give them exceptional tools, and pair them tightly with senior engineers who are great communicators and genuinely care about teaching. Juniors move fast and explore multiple approaches, while seniors spend their time framing the hard problems and raising the bar for everyone around them. This will avoid endless debates and death by committees.
English
20
39
472
33.7K
Mahesh Sathiamoorthy retweetledi
alex fazio
alex fazio@alxfazio·
you should be headless claude maxxing, so here’s an article that explains it better than the anthropic docs
English
20
29
822
195.6K
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
Bought a Tesla model Y in 2021 for 55k (or actually probably slightly more). I owe 15k now and it's value in the market is like 19k. So all these years, I will get 4k out of it if I had to sell it. I probably went through a time before where i was underwater..
English
0
0
13
3.9K
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
Slack is not here and makes sense that it's not here. I have seen various people say that they can just vibe code slack. But the main selling point of slack is that I can interact with other organizations via external slack connect. So it has a nice moat based on network effects. Now, can someone please vibe code a standard so that vibe coded slacks can talk to each other please?
Tenobrus@tenobrus

gigafucked: - grammarly - calendly - miro - retool - webflow - langchain - writer - harvey - glean - expedia - monday fucked: - accenture - intuit - notion - jasper - canva - alphasense - postman - airtable - talkdesk - sierra - zapier - replit - solace probably fucked: - cursor - pilot - clay - mercor naively seems fucked but so competent / plugged in they seem to be figuring it out on the fly anyway: - linear

English
0
1
11
3.7K
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
Growing spinach in the backyard. Also have radish, cilantro, mint, rosemary. Eggplant plant survived the winter so it should be producing nice yield this summer..
Mahesh Sathiamoorthy tweet media
English
1
0
21
1.2K
Mahesh Sathiamoorthy
Mahesh Sathiamoorthy@madiator·
This benchmark helps making progress on RL training of small models faster. Congrats on the release and happy to have played a role here with Bespoke.
Richard Zhuang@RichardZ412

Terminal-Bench is a leading benchmark for agents. Unfortunately it’s hard: most small coding agents get very low scores on TB2, so training/system ablations look flat - you can't tell what's working. Announcing OpenThoughts-TBLite - 100 curated TB2-style tasks, difficulty-calibrated so even 8B models can make progress. It's designed to give researchers measurable signal during development, providing faster feedback for experimental iteration while closely tracking true TB2 performance🧵

English
1
1
11
2K

Keşfet

@mjamei @AashaySachdeva @NandoDF @BobbySamuels @bespokelabsai @datologyai @elonmusk @BarackObama