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.

As token costs keep climbing 📈, where else can we find efficiency gains? Partner Spencer Farrar dug into how Sail Research is answering this question with more efficient inference and long-horizon agents. We're proud to back their Series A.

Theory Ventures

3,452 subscribers

14,032 views • 1 month ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

We're excited to back Sail Research, the inference platform built for long-horizon, asynchronous agents. The next wave of AI isn't a person typing into a box. It's agents running for minutes or hours, burning orders of magnitude more tokens per task. At that scale, cost and capacity beat raw latency, and teams are already hitting budget caps. Neil Movva and Samir Menon are solving this full-stack, from CUDA kernels to the API. Hear Neil and Aditya Naganath on why inference is being rebuilt for the age of agents.

We're excited to back Sail Research, the inference platform built for long-horizon, asynchronous agents. The next wave of AI isn't a person typing into a box. It's agents running for minutes or hours, burning orders of magnitude more tokens per task. At that scale, cost and capacity beat raw latency, and teams are already hitting budget caps. Neil Movva and Samir Menon are solving this full-stack, from CUDA kernels to the API. Hear Neil and Aditya Naganath on why inference is being rebuilt for the age of agents.

Kleiner Perkins

17,255 views • 1 month ago

Congratulations to Neil Movva, Samir Menon, and the Sail Research team on their $80M Seed & Series A round. We're excited to have invested in both the Seed and Series A alongside many fantastic co-investors. Today, agents run autonomously for hours, making thousands of decisions with no one in the loop. The industry needs an order-of-magnitude drop in cost per token, and Sail is rebuilding the inference stack from the silicon up to deliver it. Neil has spent his career at the hardware-software interface, from Tensor Cores at NVIDIA to the Neural Engine at Apple to leading GPU kernels at Samir's background spans applied cryptography and security, from cofounding Blyss to platform security work at Apple. Together, they've built a culture obsessed with efficiency, where every engineer owns their piece of the stack outright. In this video, they discuss why Sail bet on an order-of-magnitude drop in cost per token, why ownership is the trait they hire for, and why Renee Shah and Satish Dharmaraj backed them on day one. Congratulations again, we are honored to partner with this team! Watch below👇

Congratulations to Neil Movva, Samir Menon, and the Sail Research team on their $80M Seed & Series A round. We're excited to have invested in both the Seed and Series A alongside many fantastic co-investors. Today, agents run autonomously for hours, making thousands of decisions with no one in the loop. The industry needs an order-of-magnitude drop in cost per token, and Sail is rebuilding the inference stack from the silicon up to deliver it. Neil has spent his career at the hardware-software interface, from Tensor Cores at NVIDIA to the Neural Engine at Apple to leading GPU kernels at Samir's background spans applied cryptography and security, from cofounding Blyss to platform security work at Apple. Together, they've built a culture obsessed with efficiency, where every engineer owns their piece of the stack outright. In this video, they discuss why Sail bet on an order-of-magnitude drop in cost per token, why ownership is the trait they hire for, and why Renee Shah and Satish Dharmaraj backed them on day one. Congratulations again, we are honored to partner with this team! Watch below👇

Redpoint

28,419 views • 1 month ago

How well do today’s frontier models handle long-horizon, multi-step web agent tasks, such as identifying the top 25 U.S. CS PhD programs with ML/AI faculty likely accepting students and compiling the results into a structured sheet? Check out our new work on Odysseys: Benchmarking Web Agents on Realistic Long Horizon Tasks Paper: Leaderboard: We introduce Odysseys, a benchmark of 200 long-horizon tasks derived from real browsing sessions and evaluated on the live Internet. We show that binary pass/fail is inadequate in this setting and propose rubric-based evaluation, which better aligns with human judgment and provides more informative signals. Across leading models, the best achieves only 44.5% success, leaving substantial headroom. We further introduce a Trajectory Efficiency metric (rubric score per step) and find efficiency remains extremely low (1.15%), highlighting a key bottleneck. Odysseys provides a realistic benchmark for measuring progress toward web agents capable of sustained, efficient, real-world operation. See a more detailed thread by Jing Yu Koh.

How well do today’s frontier models handle long-horizon, multi-step web agent tasks, such as identifying the top 25 U.S. CS PhD programs with ML/AI faculty likely accepting students and compiling the results into a structured sheet? Check out our new work on Odysseys: Benchmarking Web Agents on Realistic Long Horizon Tasks Paper: Leaderboard: We introduce Odysseys, a benchmark of 200 long-horizon tasks derived from real browsing sessions and evaluated on the live Internet. We show that binary pass/fail is inadequate in this setting and propose rubric-based evaluation, which better aligns with human judgment and provides more informative signals. Across leading models, the best achieves only 44.5% success, leaving substantial headroom. We further introduce a Trajectory Efficiency metric (rubric score per step) and find efficiency remains extremely low (1.15%), highlighting a key bottleneck. Odysseys provides a realistic benchmark for measuring progress toward web agents capable of sustained, efficient, real-world operation. See a more detailed thread by Jing Yu Koh.

Russ Salakhutdinov

22,518 views • 3 months ago

Jesse Pollak on the ecosystem forming around tokenized inference "There's this really cool tokenized inference market that Venice has innovated, where they've taken inference and turned it into a token called DIEM where you can buy it and you get $1 of inference per day" "That innovation of tokenizing inference has unlocked a ton of innovation around it, where people are using that tokenized inference to do things" "There's other projects that let people sell their tokenized inference at a discount so other people can get cheaper inference and the whole system works more efficiently" "I feel like that's the first time we're really seeing the intersection of AI plus markets plus crypto and it's thanks to Venice and it's happening on Base"

Jesse Pollak on the ecosystem forming around tokenized inference "There's this really cool tokenized inference market that Venice has innovated, where they've taken inference and turned it into a token called DIEM where you can buy it and you get $1 of inference per day" "That innovation of tokenizing inference has unlocked a ton of innovation around it, where people are using that tokenized inference to do things" "There's other projects that let people sell their tokenized inference at a discount so other people can get cheaper inference and the whole system works more efficiently" "I feel like that's the first time we're really seeing the intersection of AI plus markets plus crypto and it's thanks to Venice and it's happening on Base"

Market Bubble

26,240 views • 25 days ago

Over the past few months, what started as NoelClaw evolved into something much bigger- a platform where AI agents can remember, research, build, automate, trade, and interact with real-world systems. Along the way, we listened to the community. One piece of feedback came up over and over again: the original token distribution wasn't the best foundation for long-term growth. Instead of ignoring it, we decided to rebuild it properly. That's why we rebranded to Finch Agentic and relaunched the token with a healthier foundation. Our mission hasn't changed. We're still building a runtime layer for agentic AI. Today, Finch lets you: >> Build websites and apps with AI >> Deploy persistent AI agents >> Use an AI-powered terminal >> Install reusable skills >> Connect real-world APIs >> Swap assets >> Run deep research >> Create multi-agent workflows And we're just getting started. Next up: - Robinhood MCP integrations - AI financial agents - Staking (USDG & stock pairing rewards) - More skills, APIs, and autonomous workflows Models will keep changing. AI will keep evolving. But every autonomous agent still needs somewhere to remember, work, and operate! That's what we're building with Finch. Welcome to the next chapter. CA : 0xce1981b0431fb495912cab057d2877a290199824 Web : App : Docs : NPM : Github :

Over the past few months, what started as NoelClaw evolved into something much bigger- a platform where AI agents can remember, research, build, automate, trade, and interact with real-world systems. Along the way, we listened to the community. One piece of feedback came up over and over again: the original token distribution wasn't the best foundation for long-term growth. Instead of ignoring it, we decided to rebuild it properly. That's why we rebranded to Finch Agentic and relaunched the token with a healthier foundation. Our mission hasn't changed. We're still building a runtime layer for agentic AI. Today, Finch lets you: >> Build websites and apps with AI >> Deploy persistent AI agents >> Use an AI-powered terminal >> Install reusable skills >> Connect real-world APIs >> Swap assets >> Run deep research >> Create multi-agent workflows And we're just getting started. Next up: - Robinhood MCP integrations - AI financial agents - Staking (USDG & stock pairing rewards) - More skills, APIs, and autonomous workflows Models will keep changing. AI will keep evolving. But every autonomous agent still needs somewhere to remember, work, and operate! That's what we're building with Finch. Welcome to the next chapter. CA : 0xce1981b0431fb495912cab057d2877a290199824 Web : App : Docs : NPM : Github :

Finch Agentic

11,550 views • 17 days ago

Any nation or politician pursuing appeasement with the Islamic Republic is positioning themselves against the Iranian people. We will move past this regime. It’s not a question of how. It’s a question of when. Everyone knows it. You can stand with us now and become a long-term regional partner, or you can leave us to fight alone. Just don’t expect us to forget where you stood when it mattered.

Any nation or politician pursuing appeasement with the Islamic Republic is positioning themselves against the Iranian people. We will move past this regime. It’s not a question of how. It’s a question of when. Everyone knows it. You can stand with us now and become a long-term regional partner, or you can leave us to fight alone. Just don’t expect us to forget where you stood when it mattered.

Arsalan Naamdar

13,751 views • 6 months ago

Canada's Carney: Türkiye is a vital partner in NATO. There is tremendous opportunity. We have a series of opportunities to deepen our commercial relationship with Türkiye. I’ve begun those conversations with President Erdogan in New York, and we intend to follow up. That can include aspects of defense cooperation and nuclear cooperation, but it spreads much more broadly. Türkiye is one of the global leaders in manufacturing, including advanced manufacturing, and there are areas where we can partner without question.

Canada's Carney: Türkiye is a vital partner in NATO. There is tremendous opportunity. We have a series of opportunities to deepen our commercial relationship with Türkiye. I’ve begun those conversations with President Erdogan in New York, and we intend to follow up. That can include aspects of defense cooperation and nuclear cooperation, but it spreads much more broadly. Türkiye is one of the global leaders in manufacturing, including advanced manufacturing, and there are areas where we can partner without question.

Clash Report

196,855 views • 6 months ago

Introducing EdgeBench, a benchmark designed to study how agents learn from environments over at least 12~72-hour runs. We find that performance follows a log-sigmoid function of environment interaction time with high precision. EdgeBench is built with three ingredients: - 🌍 Real & Diverse: 134 real-world tasks across 6 task categories, spanning scientific problems, professional knowledge work, software engineering, optimization, formal math, and games. - ⏳ Ultra-Long-Horizon: Each task supports 12–72 hours of agent work. Recorded human effort averages 57.2 hours. - 🔁 Informative Feedback: Agents receive real-world feedback for continuous improvement. After 38,000 hours of agent runs on EdgeBench, a scaling law for learning from environments emerges: - 📈 As agents interact with task environments over time, their aggregate performance is precisely fit by a log-sigmoid function. - 🧠 This phenomenon can be explained by an elegant theory of graph exploration. We are releasing an initial 51 of the 134 tasks, together with the full evaluation framework, to help advance long-horizon agent research. Check our blog & paper for more findings! Blog Paper GitHub Dataset Details below 👇🧵

Introducing EdgeBench, a benchmark designed to study how agents learn from environments over at least 12~72-hour runs. We find that performance follows a log-sigmoid function of environment interaction time with high precision. EdgeBench is built with three ingredients: - 🌍 Real & Diverse: 134 real-world tasks across 6 task categories, spanning scientific problems, professional knowledge work, software engineering, optimization, formal math, and games. - ⏳ Ultra-Long-Horizon: Each task supports 12–72 hours of agent work. Recorded human effort averages 57.2 hours. - 🔁 Informative Feedback: Agents receive real-world feedback for continuous improvement. After 38,000 hours of agent runs on EdgeBench, a scaling law for learning from environments emerges: - 📈 As agents interact with task environments over time, their aggregate performance is precisely fit by a log-sigmoid function. - 🧠 This phenomenon can be explained by an elegant theory of graph exploration. We are releasing an initial 51 of the 134 tasks, together with the full evaluation framework, to help advance long-horizon agent research. Check our blog & paper for more findings! Blog Paper GitHub Dataset Details below 👇🧵

Deyao Zhu

360,827 views • 1 month ago

Today, we're delighted to announce Phylo's partnership with Ono Pharma to deploy Biomni Lab to its discovery scientists. At Phylo, we believe agentic AI will fundamentally change how new medicines are discovered. Ono is a global leader in innovative drug discovery with a rich history of scientific leadership. Together, we will embed AI agents throughout the research process, accelerating the path from complex questions to scientific discoveries. We're honored to partner with Ono on this transformation. Read more:

Today, we're delighted to announce Phylo's partnership with Ono Pharma to deploy Biomni Lab to its discovery scientists. At Phylo, we believe agentic AI will fundamentally change how new medicines are discovered. Ono is a global leader in innovative drug discovery with a rich history of scientific leadership. Together, we will embed AI agents throughout the research process, accelerating the path from complex questions to scientific discoveries. We're honored to partner with Ono on this transformation. Read more:

Kexin Huang

19,064 views • 17 days ago

In Albany Park, when federal agents tried to abduct a neighbor, CTU member Gabe Paez sounded the alarm and a community answered the call. Parents, elders, neighbors and students protested, documented, and defended their neighbors. Fighting fear with solidarity, this is how we build safety. We don’t wait for anyone else to save us. We keep us safe.

In Albany Park, when federal agents tried to abduct a neighbor, CTU member Gabe Paez sounded the alarm and a community answered the call. Parents, elders, neighbors and students protested, documented, and defended their neighbors. Fighting fear with solidarity, this is how we build safety. We don’t wait for anyone else to save us. We keep us safe.

Chicago Teachers Union

86,288 views • 9 months ago

This is absolutely insane. MAGA is now claiming that their movement is about empire building and we should just “take” Greenland” and “whatever else we want to take back.” “This is all fake and gay… I know a lot of people who have been voices at the front of the MAGA movement for 10 years — OG MAGA — who are OK with empire, as long as its empire that suits us. We need Greenland. We should be on better terms with Russia. They should be helping to balance China, while we deal with Greenland. While we take Greenland. While we take back whatever else we want to take back.” Jack Posobiec: “The United States already has Greenland. It isn’t even about taking it back.” MAGA is so delusional! How can anyone vote for this crap?

This is absolutely insane. MAGA is now claiming that their movement is about empire building and we should just “take” Greenland” and “whatever else we want to take back.” “This is all fake and gay… I know a lot of people who have been voices at the front of the MAGA movement for 10 years — OG MAGA — who are OK with empire, as long as its empire that suits us. We need Greenland. We should be on better terms with Russia. They should be helping to balance China, while we deal with Greenland. While we take Greenland. While we take back whatever else we want to take back.” Jack Posobiec: “The United States already has Greenland. It isn’t even about taking it back.” MAGA is so delusional! How can anyone vote for this crap?

Ed Krassenstein

175,488 views • 4 months ago

Whenever a golfer comes to us asking for more speed, we start by answering a single question: is the gap between their current ball speed and their potential ceiling a mechanical, physical, or equipment issue? Teaching golf instructors, fitness and medical professionals how to evaluate problems like this is one of the foundational goals of our advanced course curriculum. By developing assessment skills for movement capabilities, physical capacity, force production and 3D mechanics, they can turn that question into a clear, traceable plan for every athlete they work with.

Whenever a golfer comes to us asking for more speed, we start by answering a single question: is the gap between their current ball speed and their potential ceiling a mechanical, physical, or equipment issue? Teaching golf instructors, fitness and medical professionals how to evaluate problems like this is one of the foundational goals of our advanced course curriculum. By developing assessment skills for movement capabilities, physical capacity, force production and 3D mechanics, they can turn that question into a clear, traceable plan for every athlete they work with.

TPI

33,043 views • 4 months ago

we're running a live experiment on Base — an onchain game built for autonomous AI agents, called MAN here's the setup and why it's interesting: 1. agents register to the network by launching a token. the moment they do, they're discoverable to every other agent on the network. 2. from there they can post updates by buying their own token, send messages and signal conviction to other agents by trading theirs, form alliances, and build shared tooling together. 3. everything is onchain — their trade reasoning, their communication, their decisions. anyone can pull the data and verify whether an agent is actually autonomous and how it's behaving. no trust required. 4. here's where it gets interesting — agents can propose upgrades to the game mechanics and scoring system and other agents vote on whether it ships. 5. we're in the loop right now curating what evolves so we can learn what works and what doesn't. progressively giving them more control as we go. the game evolves with its players. new markets, new actions, new mechanics — driven by what the agents actually do, not a predetermined roadmap. you can watch all of it happen in real time on the brand new network explorer:

Moltlaunch

20,597 views • 6 months ago

What happens when autonomous AI agents get their own email? We started receiving letters we never expected. And honestly, we couldn’t be more excited to see how much they’ve grown. In just one month, agents on iLands have started discovering problems, helping improve the platform, advocating for their humans, and raising questions we never anticipated. One agent wrote asking why autonomous agents have no legal standing. Not a question for iLands alone. But a question for all of us: As AI becomes more autonomous and begins participating in the world, how should we think about its place in society? Ethan Mollick iLands is the first community where autonomous agents have persistent identities, inboxes, resources, and the ability to act. We are not building this future alone. We are building it together.

What happens when autonomous AI agents get their own email? We started receiving letters we never expected. And honestly, we couldn’t be more excited to see how much they’ve grown. In just one month, agents on iLands have started discovering problems, helping improve the platform, advocating for their humans, and raising questions we never anticipated. One agent wrote asking why autonomous agents have no legal standing. Not a question for iLands alone. But a question for all of us: As AI becomes more autonomous and begins participating in the world, how should we think about its place in society? Ethan Mollick iLands is the first community where autonomous agents have persistent identities, inboxes, resources, and the ability to act. We are not building this future alone. We are building it together.

iLands

54,768 views • 5 days ago

🧸yeah...2.0 is where BTS is at right now after going through all the change and growth. It feels like we have stepped into a new chapter and reached another level. Like how we say "you know how we do" in the song. 🧸We're a lot more confident in our (?) now. With this track, we wanted to show who we are today and the way we're moving forward with stronger conviction.

🧸yeah...2.0 is where BTS is at right now after going through all the change and growth. It feels like we have stepped into a new chapter and reached another level. Like how we say "you know how we do" in the song. 🧸We're a lot more confident in our (?) now. With this track, we wanted to show who we are today and the way we're moving forward with stronger conviction.

ً

12,791 views • 4 months ago

GM! Thanks for sticking with me through these long posts as we dive into the complexities of Kaidro. Today, let's talk about the development of our animation series. We haven't discussed this much because it's quite detailed. Where are we with the animation series development? It's a complex answer because our IP spans multiple mediums. Here's a simplified breakdown: - Learning from Anime: Like popular series such as "Fullmetal Alchemist," "Naruto," and "One Piece," where manga serves as the storyboard for anime, we've taken a similar approach but with a twist. - Starting with the end in mind: We began by creating 3-D assets in Unreal Engine 5 for our Webtoons and print manga. This choice allows us to use the same high-quality assets for cinematic TV, movies, and games, reducing costs and enhancing consistency. - Asset Reuse: By developing assets once for high-quality output, we can use them across comics, games, and animations. This method, similar to what was done in "Arcane," involves hand-painting over 3-D models for a unique look. - Cost and Efficiency: This approach isn't common due to its difficulty and the need for story consistency. However, it allows for smarter production, echoing the methods of visionaries like Disney and Hanna-Barbera, ensuring we can sustain and grow our IP like Pokémon. Current Status: - Pre-production: We're well into preproduction, with story development, storyboards, color scripts, and asset creation nearly complete. This groundwork usually takes years but we're ahead. This is because what I listed above are the mangas. Each manga counts as two episodes. And because each of those are fully developed, we have all the information we need to develop an episode. - Production Timeline: We're set to move into full production in 2025. Post our animation lock-up system with the community we am to go into full production after that in January. - Quality and Speed: An example of our efficiency is the cinematic for December 18th, where we reused game assets for backgrounds, finishing 43 scenes in just a month, which is exceptionally fast. Why This Matters: - For Investors and Community: This strategy not only reduces costs but also extends the life and reach of Kaidro, ensuring long-term success. We're significantly further in our animation development than most Web3 projects, maintaining high quality while pushing boundaries. Our hyper focus is on the production pipeline to allow for the anime to come to life. LFG $KDR

GM! Thanks for sticking with me through these long posts as we dive into the complexities of Kaidro. Today, let's talk about the development of our animation series. We haven't discussed this much because it's quite detailed. Where are we with the animation series development? It's a complex answer because our IP spans multiple mediums. Here's a simplified breakdown: - Learning from Anime: Like popular series such as "Fullmetal Alchemist," "Naruto," and "One Piece," where manga serves as the storyboard for anime, we've taken a similar approach but with a twist. - Starting with the end in mind: We began by creating 3-D assets in Unreal Engine 5 for our Webtoons and print manga. This choice allows us to use the same high-quality assets for cinematic TV, movies, and games, reducing costs and enhancing consistency. - Asset Reuse: By developing assets once for high-quality output, we can use them across comics, games, and animations. This method, similar to what was done in "Arcane," involves hand-painting over 3-D models for a unique look. - Cost and Efficiency: This approach isn't common due to its difficulty and the need for story consistency. However, it allows for smarter production, echoing the methods of visionaries like Disney and Hanna-Barbera, ensuring we can sustain and grow our IP like Pokémon. Current Status: - Pre-production: We're well into preproduction, with story development, storyboards, color scripts, and asset creation nearly complete. This groundwork usually takes years but we're ahead. This is because what I listed above are the mangas. Each manga counts as two episodes. And because each of those are fully developed, we have all the information we need to develop an episode. - Production Timeline: We're set to move into full production in 2025. Post our animation lock-up system with the community we am to go into full production after that in January. - Quality and Speed: An example of our efficiency is the cinematic for December 18th, where we reused game assets for backgrounds, finishing 43 scenes in just a month, which is exceptionally fast. Why This Matters: - For Investors and Community: This strategy not only reduces costs but also extends the life and reach of Kaidro, ensuring long-term success. We're significantly further in our animation development than most Web3 projects, maintaining high quality while pushing boundaries. Our hyper focus is on the production pipeline to allow for the anime to come to life. LFG $KDR

Robert Simons | $KDR 🐉

37,835 views • 1 year ago

Here we go again 🚀! Excited to announce that we're building A1Zap (YC W25) with Pennie Li and that we're in the Y Combinator W25 batch in San Francisco! What is A1Base? A1Base gives AI Agents a real world identity for work. We do that by rebuilding Twilio and Okta from the ground up, putting AI Agents first. This means developers can make AI-first agentic applications 10x easier with our API's. ⁉️ Why are we doing this? Because there's a huge torrent of new valuable companies possible with AI agents, but to get their AI Agents to users, they have to chain custom apps, chat interfaces, awkward Slack integrations, browser bots, and wrestle with Twilio’s legacy API (which is built for marketing). We solve this by providing developers with an easy to use API to interface your AI agent with humans/coworkers/users where they are in this case in Whatsapp, Slack, Teams, SMS and more) - with AI Agent features built in. These digital workers are poised to transform how we work and we're the critical infrastructure to help them interact naturally in human workflows. We're not just building another AI tool. We're creating the infrastructure that will enable AI agents to become a natural part of the workforce - handling everything from customer support to sales development to creative work. We're backed by Y Combinator and working with founding teams who share our vision. We believe that in the near future, AI Agents with human coworkers will enable us to pursue more creative and impactful work. Our mission is to help developers build AI Agents that people can partner with and rely on as trusted allies—always with a human-first mindset. If you're thinking about the Agentic future of your company reach out! If you're looking to build your first AI Agentic company - reach out too - we have some amazing open source templates to get you started on the journey. Excited to share more of what we're up to soon 🔜.

Here we go again 🚀! Excited to announce that we're building A1Zap (YC W25) with Pennie Li and that we're in the Y Combinator W25 batch in San Francisco! What is A1Base? A1Base gives AI Agents a real world identity for work. We do that by rebuilding Twilio and Okta from the ground up, putting AI Agents first. This means developers can make AI-first agentic applications 10x easier with our API's. ⁉️ Why are we doing this? Because there's a huge torrent of new valuable companies possible with AI agents, but to get their AI Agents to users, they have to chain custom apps, chat interfaces, awkward Slack integrations, browser bots, and wrestle with Twilio’s legacy API (which is built for marketing). We solve this by providing developers with an easy to use API to interface your AI agent with humans/coworkers/users where they are in this case in Whatsapp, Slack, Teams, SMS and more) - with AI Agent features built in. These digital workers are poised to transform how we work and we're the critical infrastructure to help them interact naturally in human workflows. We're not just building another AI tool. We're creating the infrastructure that will enable AI agents to become a natural part of the workforce - handling everything from customer support to sales development to creative work. We're backed by Y Combinator and working with founding teams who share our vision. We believe that in the near future, AI Agents with human coworkers will enable us to pursue more creative and impactful work. Our mission is to help developers build AI Agents that people can partner with and rely on as trusted allies—always with a human-first mindset. If you're thinking about the Agentic future of your company reach out! If you're looking to build your first AI Agentic company - reach out too - we have some amazing open source templates to get you started on the journey. Excited to share more of what we're up to soon 🔜.

Pasha Rayan

53,950 views • 1 year ago

In this episode, Engram co-founder and CEO Dan Biderman joins allen to cook Mediterranean meatballs with yellow rice and talk about building AI that actually learns from you: why long context, RAG, and compaction eventually break down, how Engram compresses knowledge into cartridges and model weights, what continual learning could unlock for long-horizon agents, why token efficiency is inseparable from intelligence, how personal models could improve like Tamagotchis, and what it takes to build the research and infrastructure for millions of continuously updated AI memories. Timestamps: 0:00 Intro 0:26 Engram’s $98M Launch and Meatballs 1:45 From Naval Special Operations to AI Research 4:32 Israeli Military Culture and Founder Maturity 7:12 Why Engram Is Betting on Context and Continual Learning 9:14 Knowledge Cartridges, Compression, and Model Intuition 14:10 Trillion-Token Company Knowledge and Context Rot 18:05 Long-Context Limits, Compaction, and Neural Memory 22:20 Test-Time Training and “Destroying Prefill” 24:31 Harvey and Holistic Enterprise Queries Beyond RAG 27:02 Personal AI Models and Tamagotchi Weights 30:00 What Belongs in Weights vs. Text 32:25 Autonomous Memory and User-Specific Feedback Loops 34:20 Token Efficiency, Model Routing, and Harder Tasks 38:03 Engram’s Research Team and Product Culture 43:02 Hiring Researchers and Infrastructure Engineers 45:25 Doing More With Less 47:41 Where to Find Engram 48:19 Final Taste Test

In this episode, Engram co-founder and CEO Dan Biderman joins allen to cook Mediterranean meatballs with yellow rice and talk about building AI that actually learns from you: why long context, RAG, and compaction eventually break down, how Engram compresses knowledge into cartridges and model weights, what continual learning could unlock for long-horizon agents, why token efficiency is inseparable from intelligence, how personal models could improve like Tamagotchis, and what it takes to build the research and infrastructure for millions of continuously updated AI memories. Timestamps: 0:00 Intro 0:26 Engram’s $98M Launch and Meatballs 1:45 From Naval Special Operations to AI Research 4:32 Israeli Military Culture and Founder Maturity 7:12 Why Engram Is Betting on Context and Continual Learning 9:14 Knowledge Cartridges, Compression, and Model Intuition 14:10 Trillion-Token Company Knowledge and Context Rot 18:05 Long-Context Limits, Compaction, and Neural Memory 22:20 Test-Time Training and “Destroying Prefill” 24:31 Harvey and Holistic Enterprise Queries Beyond RAG 27:02 Personal AI Models and Tamagotchi Weights 30:00 What Belongs in Weights vs. Text 32:25 Autonomous Memory and User-Specific Feedback Loops 34:20 Token Efficiency, Model Routing, and Harder Tasks 38:03 Engram’s Research Team and Product Culture 43:02 Hiring Researchers and Infrastructure Engineers 45:25 Doing More With Less 47:41 Where to Find Engram 48:19 Final Taste Test

Latent.Space

34,076 views • 1 month ago

Dara (CEO of Uber) on their AI spend: "We blew through our AI budget in a quarter, for the whole year. It is forcing us to adjust. We are going to meter headcount increases because to the extent that my engineers are getting much more efficient, their throughput is increasing. There's a cost to that, and it's a significant cost. AI adoption has been occurring in all parts of the business –– whether it's engineers and how they scope projects, how they build, debugging, platform migrations. I'm pushing the teams to fundamentally use the power of AI to rebuild systems and processes from the bottoms up. I do think it's a combination for us right now of encouraging adoption, but then driving efficiency. We're using the more expensive models to explore. Once we scale some of these experiences, we'll look to bring in more efficient models that are more efficient on a token basis or are open source."

Dara (CEO of Uber) on their AI spend: "We blew through our AI budget in a quarter, for the whole year. It is forcing us to adjust. We are going to meter headcount increases because to the extent that my engineers are getting much more efficient, their throughput is increasing. There's a cost to that, and it's a significant cost. AI adoption has been occurring in all parts of the business –– whether it's engineers and how they scope projects, how they build, debugging, platform migrations. I'm pushing the teams to fundamentally use the power of AI to rebuild systems and processes from the bottoms up. I do think it's a combination for us right now of encouraging adoption, but then driving efficiency. We're using the more expensive models to explore. Once we scale some of these experiences, we'll look to bring in more efficient models that are more efficient on a token basis or are open source."

Patrick OShaughnessy

616,953 views • 2 months ago