Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

We've raised $7M to help companies build AI agents that actually learn and work. Osmosis (YC W25) is a platform for companies to fine-tune models that outperform foundation models with reinforcement learning. Better, faster, and cheaper.

Kasey Zhang

2,728 subscribers

1,217,244 views • 11 months ago •via X (Twitter)

Science & Technology News & Politics Finance

Anya Rossi• Live Now

Private livecam show

50 Comments

Kasey Zhang11 months ago

Our seed round was co-led by @CRV and @AudaciousHQ. We’re honored to be supported by angels like @paulg, @rauchg, @MishaLaskin, @bernhardsson, @bryantchou, @saradu, @m_franceschetti, @waseem_s, Troy Demmer, @shaneguML, @andrewztan, @jackyhuang, and many others.

Kasey Zhang11 months ago

We’re hosting RL IRL in 10 days with @ycombinator, @shaneguML, @reductoai, @greptile, @encord_team, and @hud_evals - it’ll be a great opportunity to meet others working on applied reinforcement learning!

Kasey Zhang11 months ago

If you want to learn more about Osmosis, we'd love to chat:

Guillermo Rauch11 months ago

@Osmosis_AI congrats, been enjoying working with you guys

Kasey Zhang11 months ago

@Osmosis_AI Likewise! Excited to be hill climbing with y'all 📈

parth11 months ago

@jackyhuang @Osmosis_AI Congrats on the launch! I want to see a Cambrian explosion of SOTA open models, RL’d on specific use cases.

Kasey Zhang11 months ago

@jackyhuang @Osmosis_AI Admin he's hill climbing sideways

david11 months ago

@Osmosis_AI this is a nice vid man

Kasey Zhang11 months ago

@Osmosis_AI Had to get something nice after you guys raised the bar 🫡

helen ⁎⁺˳✧༚11 months ago

@Osmosis_AI GOAT, OF THE MOUNTAIN VARIETY !!! FINDING FOOTHOLDS WHERE OTHERS FALL

Kasey Zhang11 months ago

@Osmosis_AI Of the mountain variety, as opposed to the pog variety

helen ⁎⁺˳✧༚11 months ago

@Osmosis_AI that’s a shiny variant

Jonathan Chang11 months ago

@Osmosis_AI Huge congrats!

zehra ✨11 months ago

@Osmosis_AI !!!!!!!!!!! sick

Kasey Zhang11 months ago

@Osmosis_AI Thanks Zehra!! The @loreobsessed video was insaneee

David Zhi LuoZhang11 months ago

@Osmosis_AI Congrats Kasey!!

Misha Laskin11 months ago

@Osmosis_AI Congrats Kasey!

Kasey Zhang11 months ago

@Osmosis_AI Thanks so much Misha!

Muzzammil Zaveri (MZ)11 months ago

@Osmosis_AI Congrats @_WEEXIAO @viewerisland @Osmosis_AI! Excited to be a part of your journey 🙌

jacky (:11 months ago

@Osmosis_AI GOATTTTTT

Pranav Mital11 months ago

@Osmosis_AI This team doesn’t play around. Literally climbing hills (i)RL

Kasey Zhang11 months ago

@Osmosis_AI We learn from experience...

Michael11 months ago

@Osmosis_AI this is so cool!!

Kasey Zhang11 months ago

@Osmosis_AI Appreciate it Michael!!

Michael Zhu11 months ago

@Osmosis_AI LFG congrats!!

jack mcclelland11 months ago

@Osmosis_AI congrats on the launch @_WEEXIAO

adel 🌟11 months ago

@Osmosis_AI wohoo!!! go kasey

shreyansh🫧11 months ago

@Osmosis_AI Now this is some interesting product

andrew pignanelli11 months ago

@Osmosis_AI RL so hot these days

Eli Brown11 months ago

@Osmosis_AI Congrats!!!

Kasey Zhang11 months ago

@Osmosis_AI Thanks Eli!! Really appreciate your support and guidance thus far :)

Ryan11 months ago

@Osmosis_AI 🐐

Shailee Sankhala11 months ago

@Osmosis_AI wow wish i had rl for ai back in the day smh

Sriman Gaddam11 months ago

@Osmosis_AI Congrats on the launch 🚀

Tom Greenwald11 months ago

@Osmosis_AI Congrats man!

Rohin Arya11 months ago

@Osmosis_AI Congrats Kasey and team!

Adish Jain ☕️11 months ago

@Osmosis_AI Congrats team!!!

Ryan D’Onofrio11 months ago

@Osmosis_AI Congrats team! Excited to learn more next week and see how it might enable new capabilities in the healthcare space!

Kasey Zhang11 months ago

@Osmosis_AI Excited to see you at RL IRL!

Brenda11 months ago

@Osmosis_AI congrats!! 🥳

anirudh11 months ago

@Osmosis_AI huskies on top red black one pack lfg kasey

Kasey Zhang11 months ago

@Osmosis_AI Alexa play Stacy's Mom

Augustin 🐓11 months ago

@Osmosis_AI looks nice!

Kasey Zhang11 months ago

@Osmosis_AI Thanks Augustin!

Adhityaa11 months ago

@Osmosis_AI 🐐

Aryan Jain11 months ago

@Osmosis_AI Congrats guys!

Shrey Pandya11 months ago

@Osmosis_AI Congrats!!

Hadi11 months ago

@Osmosis_AI congrats! and amazing video😃

Charlie Deane11 months ago

@Osmosis_AI LETS GOOOO

Manya Adams11 months ago

@Osmosis_AI wow...I wonder when I'll get to my first million

Related Videos

Despite constant improvements, AI agents still struggle with complex, multi-step tasks. @Gulp_AI (YC W25) solves this with Osmosis: a framework that helps AI agents learn (similar to DeepSeek R1) in real time. Congrats on the launch Kasey Zhang + Andy!

Despite constant improvements, AI agents still struggle with complex, multi-step tasks. @Gulp_AI (YC W25) solves this with Osmosis: a framework that helps AI agents learn (similar to DeepSeek R1) in real time. Congrats on the launch Kasey Zhang + Andy!

Y Combinator

75,962 views • 1 year ago

.Decagon co-founder and CTO Ashwin Sreenivas says the choice between expensive frontier models and cheaper, less capable ones is a false trade-off: "Even if you have a 'dumber model,' you can get it to higher performance on that specific task." "When we fine-tune smaller, dumber models, it's that they're just not as general purpose, but on the specific task we want them to do, they actually outperform the large, smart, state-of-the-art models." "So we end up getting all three things. It is better at the task, it is cheaper, and it is faster." Ashwin Sreenivas

.Decagon co-founder and CTO Ashwin Sreenivas says the choice between expensive frontier models and cheaper, less capable ones is a false trade-off: "Even if you have a 'dumber model,' you can get it to higher performance on that specific task." "When we fine-tune smaller, dumber models, it's that they're just not as general purpose, but on the specific task we want them to do, they actually outperform the large, smart, state-of-the-art models." "So we end up getting all three things. It is better at the task, it is cheaper, and it is faster." Ashwin Sreenivas

a16z

36,067 views • 1 month ago

Today, we're launching Belvedir (YC S26) We're making private AI models cheap and easy so the world can have trillions of them. Private AI models are cheaper, better, faster, and protect your data from frontier labs, but they're still to expensive and difficult to make for startups and consumers. Belvedir makes private AI models autonomously with less than 5 minutes of setup, so that private AI can finally be accessible to everyone

Today, we're launching Belvedir (YC S26) We're making private AI models cheap and easy so the world can have trillions of them. Private AI models are cheaper, better, faster, and protect your data from frontier labs, but they're still to expensive and difficult to make for startups and consumers. Belvedir makes private AI models autonomously with less than 5 minutes of setup, so that private AI can finally be accessible to everyone

Zachary Yu

88,488 views • 1 month ago

Introducing Billow (YC S26) We're killing Deloitte. Accounting is a $700B industry, but 80% of companies still do it by hand. Billow is an AI-native accounting firm that replaces humans with agents. We connect to your existing tools (NetSuite, QuickBooks, and 80+ other systems) and provide faster, cheaper accounting services. There is ZERO software for customers to use. Just integrate, and Billow closes your books. We're live with enterprise and public companies representing over $1B in value.

Introducing Billow (YC S26) We're killing Deloitte. Accounting is a $700B industry, but 80% of companies still do it by hand. Billow is an AI-native accounting firm that replaces humans with agents. We connect to your existing tools (NetSuite, QuickBooks, and 80+ other systems) and provide faster, cheaper accounting services. There is ZERO software for customers to use. Just integrate, and Billow closes your books. We're live with enterprise and public companies representing over $1B in value.

Joanathan McIntosh

374,798 views • 1 month ago

Plexe AI (Plexe AI) is building an open-source agent to build ML models from a prompt. Their agent is already helping companies go from idea to deployed models 10x faster. Congrats on the launch, Vaibhav Dubey and Marcello De Bernardi!

Plexe AI (Plexe AI) is building an open-source agent to build ML models from a prompt. Their agent is already helping companies go from idea to deployed models 10x faster. Congrats on the launch, Vaibhav Dubey and Marcello De Bernardi!

Y Combinator

55,896 views • 1 year ago

Specialized AI agents help enterprises turn AI into systems built for their own workflows. NVIDIA Agent Toolkit brings together open Nemotron models, tools, skills and secure runtime support to help teams build agents tuned for domain-specific work. Learn more:

Specialized AI agents help enterprises turn AI into systems built for their own workflows. NVIDIA Agent Toolkit brings together open Nemotron models, tools, skills and secure runtime support to help teams build agents tuned for domain-specific work. Learn more:

NVIDIA

81,019 views • 3 months ago

⚡Faster models mean smarter AI. ⚡ With Nemotron's open-source technologies, we're enabling organizations to create AI agents that get more done in less time. Big or small, we are committed to helping companies all around the world build & deploy AI for their own needs. See how:

⚡Faster models mean smarter AI. ⚡ With Nemotron's open-source technologies, we're enabling organizations to create AI agents that get more done in less time. Big or small, we are committed to helping companies all around the world build & deploy AI for their own needs. See how:

NVIDIA

35,710 views • 1 year ago

The founder of a $4B inference company says that if you're building agents, foundational models could become your IP. According to Lin Qiao, 90% of the world's data is still private and locked inside applications and enterprises. Foundation models are trained on public internet + labeling company data, which is barely 10% of all the data. And this is why your application and foundation models are misaligned by definition. Companies building agents don't treat models as APIs. They opt for a product-model co-design. - Models continuously learn from your private data - Pick up domain-specific intelligence - Run faster - Cost less - Scale to millions of users

The founder of a $4B inference company says that if you're building agents, foundational models could become your IP. According to Lin Qiao, 90% of the world's data is still private and locked inside applications and enterprises. Foundation models are trained on public internet + labeling company data, which is barely 10% of all the data. And this is why your application and foundation models are misaligned by definition. Companies building agents don't treat models as APIs. They opt for a product-model co-design. - Models continuously learn from your private data - Pick up domain-specific intelligence - Run faster - Cost less - Scale to millions of users

Ivan Burazin

55,344 views • 5 months ago

YC W24 companies like Sync Labs have instant access to GPUs for fine tuning and training foundation models, and get $500K in credits so they don't even have to touch their balance sheet to show results like this below. You don't have to start with $100M on day one in AI

YC W24 companies like Sync Labs have instant access to GPUs for fine tuning and training foundation models, and get $500K in credits so they don't even have to touch their balance sheet to show results like this below. You don't have to start with $100M on day one in AI

Garry Tan

76,971 views • 2 years ago

1/ Announcing the new Lindy: the first platform letting you build a team of AI employees that work together to perform any task — 100x better, faster and cheaper than humans would. Watch how I build a Lindy to track competitors in 30s:

1/ Announcing the new Lindy: the first platform letting you build a team of AI employees that work together to perform any task — 100x better, faster and cheaper than humans would. Watch how I build a Lindy to track competitors in 30s:

Flo Crivello

1,085,027 views • 2 years ago

The best way to learn AI is to build with agents. To help with that, we've launched hands-on labs and a new series on Agentic Engineering. First topic: Agent Skills. Next in the pipeline: planning, context engineering, multi-agent systems, long-running agents,.. Go build!

The best way to learn AI is to build with agents. To help with that, we've launched hands-on labs and a new series on Agentic Engineering. First topic: Agent Skills. Next in the pipeline: planning, context engineering, multi-agent systems, long-running agents,.. Go build!

elvis

31,802 views • 4 months ago

Jack Dorsey says closed-source AI models are able to manipulate users to align with the financial incentives of the companies that build them

Jack Dorsey says closed-source AI models are able to manipulate users to align with the financial incentives of the companies that build them

Tsarathustra

155,965 views • 2 years ago

We got into Y Combinator 3 months ago with zero revenue and today Fabraix (YC S26) is at $1.04M annualised run rate. Traditional cybersecurity was built for software that behaves within defined boundaries. AI agents can reason, use tools, take actions, and find ways around those boundaries. The cost asymmetry we're seeing in irl warfare is also now a software problem. So we built autonomous red-teaming agents that find real vulnerabilities in AI agents that work 24/7, powered by our own state-of-the-art offensive models that outperform frontier models. AI agents are automating all knowledge work. Fabraix (YC S26) is the adversarial testing infrastructure that will secure every production agent.

We got into Y Combinator 3 months ago with zero revenue and today Fabraix (YC S26) is at $1.04M annualised run rate. Traditional cybersecurity was built for software that behaves within defined boundaries. AI agents can reason, use tools, take actions, and find ways around those boundaries. The cost asymmetry we're seeing in irl warfare is also now a software problem. So we built autonomous red-teaming agents that find real vulnerabilities in AI agents that work 24/7, powered by our own state-of-the-art offensive models that outperform frontier models. AI agents are automating all knowledge work. Fabraix (YC S26) is the adversarial testing infrastructure that will secure every production agent.

Zach AL

45,496 views • 14 days ago

NVIDIA AQCURIES HUGGING FACE FOR $12.9B. Jensen on the value of open source models: "I recommended that people use closed models as much as they can because they are developing rapidly. We use Claude, Codex, Perplexity, and Cursor at Nvidia. However, there are so many companies that have domain expertise, companies that have sovereignty reasons, they do their own research, those companies need to have control and ownership over their own intellectual proprietary. Open sourced models have become sufficiently good to allow companies to build their own AI."

NVIDIA AQCURIES HUGGING FACE FOR $12.9B. Jensen on the value of open source models: "I recommended that people use closed models as much as they can because they are developing rapidly. We use Claude, Codex, Perplexity, and Cursor at Nvidia. However, there are so many companies that have domain expertise, companies that have sovereignty reasons, they do their own research, those companies need to have control and ownership over their own intellectual proprietary. Open sourced models have become sufficiently good to allow companies to build their own AI."

amit

122,906 views • 22 days ago

Frontier models that use reasoning to "think" during inference are generating 5X more AI tokens per year. ✨ "Inference is now a thinking process. And in order to teach AI how to think, reinforcement learning and very significant computation was introduced into post-training," said NVIDIA CEO Jensen Huang in a recent keynote at #CES2026. Reinforcement learning is increasing computation demands across all AI scaling laws: pre-training, post-training, and test-time scaling. Learn more about reinforcement learning and AI scaling ➡️

Frontier models that use reasoning to "think" during inference are generating 5X more AI tokens per year. ✨ "Inference is now a thinking process. And in order to teach AI how to think, reinforcement learning and very significant computation was introduced into post-training," said NVIDIA CEO Jensen Huang in a recent keynote at #CES2026. Reinforcement learning is increasing computation demands across all AI scaling laws: pre-training, post-training, and test-time scaling. Learn more about reinforcement learning and AI scaling ➡️

NVIDIA AI Infrastructure

13,329 views • 8 months ago

Today, we’re announcing Parahelp to the world and that we've raised $21.2M. In the past year, Parahelp has become the go-to AI support agent for the world’s fastest-moving software companies, and we're proud to say that 0% of the companies we've onboarded have churned. We’re hiring exceptional engineers, so if you’re excited to build on the frontier of applied AI for customer support, please reach out!

Today, we’re announcing Parahelp to the world and that we've raised $21.2M. In the past year, Parahelp has become the go-to AI support agent for the world’s fastest-moving software companies, and we're proud to say that 0% of the companies we've onboarded have churned. We’re hiring exceptional engineers, so if you’re excited to build on the frontier of applied AI for customer support, please reach out!

Anker

358,375 views • 1 year ago

Brian Wong: The biggest profit margin companies won't be AI models. They'll be agents that help you build other agents. Most of America is running construction companies, factories, carpentry shops. They don't have time to prompt anything. Orchestrate the agents FOR them and they can run it without thinking. "Commerce only grows if it's buying for me. We don't scale agentic commerce if I have to tell it to buy things every time." Promptless AI + stablecoin payments is the wave. FT The Rollup Andy Brian Wong

Brian Wong: The biggest profit margin companies won't be AI models. They'll be agents that help you build other agents. Most of America is running construction companies, factories, carpentry shops. They don't have time to prompt anything. Orchestrate the agents FOR them and they can run it without thinking. "Commerce only grows if it's buying for me. We don't scale agentic commerce if I have to tell it to buy things every time." Promptless AI + stablecoin payments is the wave. FT The Rollup Andy Brian Wong

Milk Road

28,182 views • 4 months ago

Chamath Palihapitiya believes AGI may already exist inside leading AI labs and the bigger story is that advanced intelligence is becoming cheaper and more widely available (Save this). Chamath Palihapitiya argues that the public may be focused too much on benchmark rankings, while frontier labs are already developing models capable of complex reasoning, coding, research, and tool use. The main question is how quickly companies will release these systems and how much access they will provide. AGI has not been officially confirmed and strong benchmark results do not necessarily prove that a model can perform every intellectual task like a human. However, AI capabilities are improving quickly, while the cost of running advanced models continues to fall. That combination is important because cheaper AI can be used by more businesses for customer service, software development, research, marketing, financial analysis, and automation. Competition is also accelerating among OpenAI, Anthropic, Google, xAI, Meta, and open source developers because as more companies release capable models, users gain more choices and prices continue to decline. This creates a powerful cycle in which better models attract more users, more usage generates more revenue and data, and lower prices encourage companies to apply AI to additional tasks. The biggest challenge is moving from impressive demonstrations to measurable business results. Companies still need to redesign workflows, train employees, protect sensitive information, and prove that AI spending is producing a real return on investment. AI agents could create the next major increase in demand because they can plan tasks, use tools, check their work, retry failed actions and operate for long periods without constant human supervision. Even if each AI task becomes cheaper, total usage could grow much faster as businesses use models across more departments and this could increase demand for GPUs, high bandwidth memory, networking equipment, electricity, cooling systems, and data centers.

Chamath Palihapitiya believes AGI may already exist inside leading AI labs and the bigger story is that advanced intelligence is becoming cheaper and more widely available (Save this). Chamath Palihapitiya argues that the public may be focused too much on benchmark rankings, while frontier labs are already developing models capable of complex reasoning, coding, research, and tool use. The main question is how quickly companies will release these systems and how much access they will provide. AGI has not been officially confirmed and strong benchmark results do not necessarily prove that a model can perform every intellectual task like a human. However, AI capabilities are improving quickly, while the cost of running advanced models continues to fall. That combination is important because cheaper AI can be used by more businesses for customer service, software development, research, marketing, financial analysis, and automation. Competition is also accelerating among OpenAI, Anthropic, Google, xAI, Meta, and open source developers because as more companies release capable models, users gain more choices and prices continue to decline. This creates a powerful cycle in which better models attract more users, more usage generates more revenue and data, and lower prices encourage companies to apply AI to additional tasks. The biggest challenge is moving from impressive demonstrations to measurable business results. Companies still need to redesign workflows, train employees, protect sensitive information, and prove that AI spending is producing a real return on investment. AI agents could create the next major increase in demand because they can plan tasks, use tools, check their work, retry failed actions and operate for long periods without constant human supervision. Even if each AI task becomes cheaper, total usage could grow much faster as businesses use models across more departments and this could increase demand for GPUs, high bandwidth memory, networking equipment, electricity, cooling systems, and data centers.

Milk Road AI

13,501 views • 20 days ago

President Trump says China's DeepSeek AI model is a "wake-up call" for American companies but it is a good thing for faster and cheaper methods of AI to be developed

President Trump says China's DeepSeek AI model is a "wake-up call" for American companies but it is a good thing for faster and cheaper methods of AI to be developed

Tsarathustra

1,137,813 views • 1 year ago