
Scaling tokens @togethercompute. Ex: Co-founder/CEO @RefuelAI (acq), @Stanford, @primer_ai,
San Francisco, CA
Joined January 2014
- Tweets175
- Following689
- Followers423
- Likes989
Pinned Tweet
Excited to announce that Refuel is officially out of stealth today! We’re building a platform to create, clean, enrich and label datasets at the speed of thought, using LLMs. 🚀
More details below.
Rishabh Bhargava retweeted
It has been clear to many of us, and now it’s becoming clear more tangibly, that AI models will commoditize to various degrees. This is of course a difficult business reality if your core business depends on exclusivity on intelligence.
But commodity markets are not communism. They are the largest markets on earth. Oil, grain, steel, electricity, memory: trillions clear through them every year, priced by competition among thousands of suppliers. In economic terms, communism is one provider and no price. A commodity market is the precise inverse. The world that actually resembles central planning is the one Dean argues for: a set of protected incumbents, access gated by the state, agencies instructed to manufacture FUD until every regulated buyer, and transitively every tool maker upstream, backs away from cheaper competitors.
Open weights don't deter capex. They move it. When the model layer commoditizes, spend shifts to inference, data, tooling, and applications, and builds far broader industrial infrastructure rather than concentrating capital in a handful of companies. Most of our digital infrastructure today, hyperscalers included, runs on open source. The businesses built atop it keep excellent margins and compound at extraordinary rates. Open-weights intelligence will likely rank among the most important economic accelerations in history. It won't be kind to every early incumbent, Linux wasn't kind to Sun Microsystems, but it will be very good for almost everyone else. I suspect OpenAI and Anthropic, given their positions, excellent products, resources and talent density, will be just fine. They will simply hold a little less pricing power.
The security theater around Mythos continues to do damage. Of course, there is no evidence for the hysterical claims. The evidence is in fact so thin that proponents of AI's existential risks now openly recommend FUD as the strategy. That should be telling.
Some observations on Kimi:
1. It's a very good model! I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public models of Q1 2026. In my fairly limited use, it also seemed very token hungry. It's not obvious to me that this model is actually that cheap to run.
2. I am personally surprised the Chinese state continues to allow the open sourcing of models this good, given potential risks. To be clear, I *myself* might be fine with models presenting this level of marginal risk being open weight, but I am surprised that China is fine with it. I suspect the reason they are is 75% explained by strategic blindness/lack of AGI-pilledness (the CCP is very Yann Lecun-y in its views of AI). The other 25% or so is their lack of compute for customer inference (making China's open-weight strategy an unintended byproduct of US export controls) and the normal Chinese strategy of aggressive exports. For the companies, as opposed to the government, the decision to open source is partially ideological and partially because they are behind, and they know that very few people would pay for sub-frontier models from China.
3. Open-weight models are inherently decelerationist, and I'm continually surprised to see the so-called "accelerationists" so excited about open-weight models. I suspect the reason they are is that they know open-weight models are effectively ungovernable, and they simply like the overall cloak of ungovernability open-weight models create over the whole of AI. It's not a bad strategy; it reminds me of James Scott's recounting of the hill people in "the art of not being governed." Still, in the end, open-weight models deter further AI capex.
4. One probable outcome of an open-weight-model-dominant world is full AI communism, which is precisely what China proposes: rather than a market product, AI is a "public good" which will ultimately be provided by the state as a kind of "digital public infrastructure." This future strikes me as a dystopian hellscape, but I've never met an open-weight models advocate who doesn't ultimately concede this is where things end. You'd be surprised how many 'accelerationists' lobbied me, while I was in government, to support an eleven or twelve-figure federally funded data center so that startups could train models at a subsidy and then give them away for free. There was no other way for AI to progress, they said. Perhaps this is the logical end state of things. Nonetheless, I find myself surprised to see supposed accelerationists excited about such an outcome. I think many of them just don't know what they're doing. Many accelerationists do not view the creation and serving of frontier models as a legitimate business.
5. I would guess that the Trump Administration will at some point realize that their best strategy here would be to create large amounts of regulatory risk around the use of open-weight Chinese models. You don't need to "ban open source" (one of the dumber motifs of AI policy discussion). You just need to direct every agency to issue soft law that creates FUD. "A Federal Reserve Advisory Bulletin found that there may be backdoors in Chinese AI models." It needn't be that well justified. You just create enough regulatory risk that every regulated enterprise backs off. You probably don't want to create so much regulatory risk that you scare off the hyperscalers from serving Chinese models; this will just drive startups to sketchier providers. There's a happy middle ground here. I'd assume they will do some version of this.
6. It's probably true that open-weight models of this capability make the world a bit more dangerous, but not so much more that you'll really notice. At some point the models will be capable enough that you will notice. "A nonliving, invisible, dangerous, and infinitely self-replicating agent escaped from a Chinese lab," you say? Color me shocked.
Rishabh Bhargava retweeted
Congratulations to our portfolio companies, @togethercompute and @RefuelAI, on uniting their strengths to power the next generation of AI infrastructure!
Together AI’s AI Acceleration Cloud enables developers and enterprises to train and deploy generative AI models with speed, control, and cost-efficiency. By bringing in Refuel’s purpose-built data models and orchestration platform, they’re tackling one of the biggest bottlenecks in AI: cleaning and structuring messy data at scale.
We’re proud of @vipulved, @rish_bhargava, @nihit_desai, and both teams for building the foundation that enables applied AI to thrive.
Rishabh Bhargava retweeted
We have some big news to share today - @RefuelAI is joining @togethercompute to help accelerate the future of open source and enterprise AI! together.ai/blog/together-ai…
Rishabh Bhargava retweeted
To Reason or Not to Reason, that is the question 🤔
This quoted post is unavailable.
Rishabh Bhargava retweeted
Data intelligence too cheap to meter
RefuelLLM-2-mini (75.02%), our latest 1.5B param SLM, outperforms all comparable models including Phi-3.5 (65.3%), Qwen2.5 (67.62%), Gemma2 (64.52%), Llama3-3B (55.8%) and Llama3-1B (39.92%) across our benchmark of data processing tasks such as labeling, enrichment and structure extraction
RefuelLLM-2-mini is a Qwen2-1.5B base model, trained on a corpus of 2750+ datasets spanning tasks such as classification, reading comprehension, structured attribute extraction and entity resolution, using the same recipe as other models in the Refuel-LLM family.
It's fast!
We’re open sourcing the model weights, available on @huggingface - huggingface.co/refuelai/Qwen…
If you'd like to access models, along with fine tuning support, DM me or reach out to us: refuel.ai/get-started
Grateful to our early customers for their partnership, and the entire @RefuelAI team for their hard work 🚀
Thrilled to introduce RefuelLLM-2, our latest family of LLMs built for data labeling and enrichment tasks. RefuelLLM-2 (83.82%) outperforms GPT-4-Turbo (80.88%), Claude-3-Opus (79.19%), Llama3-70B (78.2%) and Gemini-1.5-Pro (74.59%) on a benchmark of ~30 data labeling tasks:
RefuelLLM-2-small (79.67%), aka Llama-3-Refueled, outperforms all comparable LLMs including Claude3-Sonnet (70.99%), Haiku (69.23%) and GPT-3.5-Turbo (68.13%). We’re open sourcing the model: huggingface.co/refuelai/Llam…
You can try out the models here and give us some feedback! labs.refuel.ai/playground. The code and data used for benchmarking the LLMs is available in our Autolabel library: github.com/refuel-ai/autolab…
One more thing: RefuelLLM-2 family of models output much better calibrated confidence scores - a useful lever to reject, retry or ensemble low confidence outputs.
Rishabh Bhargava retweeted
At the summit, @tomas_hk led a panel featuring @rish_bhargava, @corbtt, and @JunchenJiang
discussing "Scaling AI Operations for Diverse Workloads."
#GenAISummit
You know it's a busy week when you have the Mount Rushmore of caffeine on your desk at 4pm on a Friday.
1
15
353
Rishabh Bhargava retweeted
Thank you @databricks @DbrxMosaicAI for the keynote shoutout! Always great connecting with new and old friends at the @Data_AI_Summit
Rishabh Bhargava retweeted
Rishabh Bhargava retweeted
LLMs can label data 100x faster than humans
@RefuelAI just released RefuelLLM-2, built for data labelling. It outperforms GPT-4-Turbo, Claude-3-Opus, Llama3-70B, Gemini-1.5-Pro on this task
Here's how much time + money you can save using LLMs over humans (🧵):
Rishabh Bhargava retweeted
We're trending on @huggingface!
Check out: huggingface.co/refuelai/Llam…
Try out the model here: labs.refuel.ai/playground
Thrilled to introduce RefuelLLM-2, our latest family of LLMs built for data labeling and enrichment tasks. RefuelLLM-2 (83.82%) outperforms GPT-4-Turbo (80.88%), Claude-3-Opus (79.19%), Llama3-70B (78.2%) and Gemini-1.5-Pro (74.59%) on a benchmark of ~30 data labeling tasks:
RefuelLLM-2-small (79.67%), aka Llama-3-Refueled, outperforms all comparable LLMs including Claude3-Sonnet (70.99%), Haiku (69.23%) and GPT-3.5-Turbo (68.13%). We’re open sourcing the model: huggingface.co/refuelai/Llam…
You can try out the models here and give us some feedback! labs.refuel.ai/playground. The code and data used for benchmarking the LLMs is available in our Autolabel library: github.com/refuel-ai/autolab…
One more thing: RefuelLLM-2 family of models output much better calibrated confidence scores - a useful lever to reject, retry or ensemble low confidence outputs.
Rishabh Bhargava retweeted
We are the specialized LLM camp. Love the results from refuelLLM-2.
Thrilled to introduce RefuelLLM-2, our latest family of LLMs built for data labeling and enrichment tasks. RefuelLLM-2 (83.82%) outperforms GPT-4-Turbo (80.88%), Claude-3-Opus (79.19%), Llama3-70B (78.2%) and Gemini-1.5-Pro (74.59%) on a benchmark of ~30 data labeling tasks:
RefuelLLM-2-small (79.67%), aka Llama-3-Refueled, outperforms all comparable LLMs including Claude3-Sonnet (70.99%), Haiku (69.23%) and GPT-3.5-Turbo (68.13%). We’re open sourcing the model: huggingface.co/refuelai/Llam…
You can try out the models here and give us some feedback! labs.refuel.ai/playground. The code and data used for benchmarking the LLMs is available in our Autolabel library: github.com/refuel-ai/autolab…
One more thing: RefuelLLM-2 family of models output much better calibrated confidence scores - a useful lever to reject, retry or ensemble low confidence outputs.
Rishabh Bhargava retweeted
Thrilled to introduce RefuelLLM-2, our latest family of LLMs built for data labeling and enrichment tasks. RefuelLLM-2 (83.82%) outperforms GPT-4-Turbo (80.88%), Claude-3-Opus (79.19%), Llama3-70B (78.2%) and Gemini-1.5-Pro (74.59%) on a benchmark of ~30 data labeling tasks:
RefuelLLM-2-small (79.67%), aka Llama-3-Refueled, outperforms all comparable LLMs including Claude3-Sonnet (70.99%), Haiku (69.23%) and GPT-3.5-Turbo (68.13%). We’re open sourcing the model: huggingface.co/refuelai/Llam…
You can try out the models here and give us some feedback! labs.refuel.ai/playground. The code and data used for benchmarking the LLMs is available in our Autolabel library: github.com/refuel-ai/autolab…
One more thing: RefuelLLM-2 family of models output much better calibrated confidence scores - a useful lever to reject, retry or ensemble low confidence outputs.
Data is the fuel for every business. Check out this awesome podcast by my awesome co-founder @nihit_desai on how we use AI models at @RefuelAI to improve the quality of data and the outcomes for everything downstream of the data.
Nihit Desai of @RefuelAI joins the show with @seanfalconer to talk about the platform, and how to manage data in the current AI era.
Listen here:
softwareengineeringdaily.com…
@nihit_desai
Rishabh Bhargava retweeted
Better data = Better AI. In this episode of @software_daily I dive into why this is true, what makes it hard and how @RefuelAI is solving this at scale. Shoutout to @seanfalconer for hosting! 👇🚀
Nihit Desai of @RefuelAI joins the show with @seanfalconer to talk about the platform, and how to manage data in the current AI era.
Listen here:
softwareengineeringdaily.com…
@nihit_desai
Rishabh Bhargava retweeted
Nihit Desai of @RefuelAI joins the show with @seanfalconer to talk about the platform, and how to manage data in the current AI era.
Listen here:
softwareengineeringdaily.com…
@nihit_desai
Great case study of the challenges involved in moving from one LLM API version to another. While in aggregate the new LLM might be on-par or better, there are no guarantees for YOUR specific task.
We saw the same behavior benchmarking different gpt-3.5-turbo versions last year across a variety of datasets. refuel.ai/blog-posts/gpt-3-5…
Rishabh Bhargava retweeted
Clock speed is a team's biggest competitive advantage. Here's a fun case study of how smart ML teams like @TeachFX are leveraging Refuel to move faster and make things ⏩
refuel.ai/blog-posts/teachfx…






























