Citrini Memo Reactions, Kim K Enters Energy Drinks, Jane Street Sued | Patrick & John Collison, Bill Gurley, James Cadwallader, Scott Wu, Ivan Zhao, Stefano Ermon, Rune Kvist, Reiner Pope, Devansh Pandey
TBPN
00:00
Diffusion Models: Speed, Use Cases, and Distillation
Discussion on diffusion LLM throughput, coding, voice agents, retrieval, and distillation's IP implications.
Play episode from 01:40:42
Transcript
Transcript
Episode notes
Sign up for TBPN’s daily newsletter at TBPN.com
- (01:52) - Citrini Memo Reactions
- (09:35) - Is Doordash Cooked?
- (12:51) - 𝕏 Timeline Reactions
- (20:59) - Kim K Enters Energy Drinks
- (22:23) - 𝕏 Timeline Reactions
- (29:01) - Jane Street Sued
- (40:17) - Patrick Collison is the co-founder and CEO of Stripe, the global payments and financial infrastructure company powering millions of businesses. A longtime advocate for scientific and technological progress, he also co-founded the Arc Institute and writes frequently about economic growth, research, and innovation. John Collison is the co-founder and President of Stripe, where he leads strategy and product across payments, banking-as-a-service, and financial tooling. He previously co-founded Auctomatic, which was acquired in 2008, and focuses on expanding Stripe’s global reach and financial infrastructure capabilities.
- (57:46) - Bill Gurley, a general partner at Benchmark and author of "Runnin' Down a Dream," discusses his transition from a conventional tech job to venture capital, emphasizing the importance of pursuing one's passion to avoid career regret. He highlights six principles for a fulfilling career: chasing curiosity, honing one's craft, developing mentors, embracing peers, going where the action is, and giving back. Gurley also addresses the rapid public consciousness of AI advancements, noting its unprecedented speed compared to previous tech waves, and underscores the necessity for individuals to be hyper-curious and continuously learning to thrive in evolving industries.
- (01:23:18) - 𝕏 Timeline Reactions
- (01:29:16) - Ivan Zhao, co-founder and CEO of Notion, is a Chinese-Canadian entrepreneur with a background in cognitive science from the University of British Columbia. In the conversation, he discusses the launch of Notion's Custom Agents, autonomous AI teammates designed to handle repetitive tasks across various platforms, enhancing productivity and collaboration.
- (01:45:09) - 𝕏 Timeline Reactions
- (01:51:15) - Stefano Ermon, co-founder and CEO of Inception Labs, discusses his background in generative AI research at Stanford, including co-inventing diffusion models. He explains how Inception's diffusion-based language models generate text by refining entire sequences simultaneously, resulting in speeds over 1,000 tokens per second on NVIDIA GPUs. Ermon highlights the models' efficiency and scalability, making them ideal for latency-sensitive applications like coding autocomplete and voice agents.
- (02:01:07) - James Cadwallader, co-founder and CEO of Profound, discusses the company's recent $96 million Series C funding round, which values the company at $1 billion. He highlights the launch of Profound Agents, customizable autonomous tools that enhance marketing efficiency by enabling brands to monitor and influence their representation across AI platforms. Cadwallader emphasizes the transformative impact of AI on brand discovery and the necessity for marketers to adapt to this evolving landscape.
- (02:13:59) - Scott Wu, co-founder and CEO of Cognition AI, discusses the company's significant growth, noting that enterprise usage has more than doubled in the past six weeks, driven by the adoption of AI agents capable of handling end-to-end tasks. He highlights the latest launch's focus on enhancing user experience by addressing known frictions, introducing features like automated testing and improved integrations. Wu also emphasizes the importance of optimizing various aspects of the software engineering workflow, including testing and review processes, to further improve efficiency.
- (02:30:39) - Rune Kvist, co-founder and CEO of the Artificial Intelligence Underwriting Company (AIUC), discusses the company's mission to underwrite superintelligence by developing standards and insurance products for AI agents. He highlights the challenges in insuring AI systems due to unpredictable risks and emphasizes the importance of creating a standardized framework to manage these uncertainties. Kvist also mentions AIUC's recent launch of the world's first insurance policy for AI agents, in collaboration with Eleven Labs, aiming to address concerns like hallucinations leading to financial losses and data leakage.
- (02:39:02) - Reiner Pope, CEO and co-founder of MaddX, discusses his company's development of high-throughput chips tailored for large language models, emphasizing the need for a from-scratch design to achieve optimal performance. He highlights the constraints in the current market, particularly the limited silicon wafer supply, and how MaddX's approach aims to maximize throughput per dollar and per watt. Pope also addresses the challenges posed by existing technologies like Nvidia's CUDA, noting that while CUDA offers backward compatibility, it restricts hardware innovation, whereas MaddX's specialized design offers greater efficiency for frontier labs willing to adapt their software.
- (02:53:49) - Devansh Pandey, co-founder of Standard Intelligence, discusses the company's approach to pre-training computer use models by capturing 30fps video of user interactions, including screen recordings, key presses, and mouse movements, to create a comprehensive dataset for training general models capable of performing diverse computer tasks. He highlights the potential applications of these models, such as automating repetitive tasks like form filling and enhancing CAD design processes, and emphasizes the advantages of video-based training over text-based methods, noting that graphical user interfaces are designed for human interaction and that video data inherently captures temporal aspects of user behavior. Additionally, Pandey shares insights into the company's data collection methods, including the use of an application that records user screens and inputs, and discusses the potential for their models to generalize to various computer-based tasks, including applications in robotics and self-driving technology.
TBPN.com is made possible by:
Ramp - https://Ramp.com
AppLovin - https://axon.ai
Cisco - https://www.cisco.com
Cognition - https://cognition.ai
Console - https://console.com
CrowdStrike - https://crowdstrike.com
ElevenLabs - https://elevenlabs.io
Figma - https://figma.com
Fin - https://fin.ai
Gemini - https://gemini.google.com
Graphite - https://graphite.com
Gusto - https://gusto.com/tbpn
Kalshi - https://kalshi.com
Labelbox - https://labelbox.com
Lambda - https://lambda.ai
Linear - https://linear.app
MongoDB - https://mongodb.com
NYSE - https://nyse.com
Okta - https://www.okta.com
Phantom - https://phantom.com/cash
Plaid - https://plaid.com
The AI-powered Podcast Player
Save insights by tapping your headphones, chat with episodes, discover the best highlights - and more!
