Lightwork

Lightwork

By LightspeedTechnology
Download on the App Store

All Lightwork episodes

  • OpenAI’s GPT-6 Astra, Meta Muse & ByteDance Enters the World Model Race | Lightwork

    Claire Zau and Josh Machiz are joined by Lightspeed partner Bucky Moore for a week about how much control we are handing over to AI. OpenAI released GPT-6 Astra, its most capable model and, by OpenAI's own account, a harder one to monitor. Bucky explains why looping inference compresses the chain of thought that used to make these models readable, and why the near term answer may be tighter permissions and access controls rather than interpretability research.

    Then to the agents people are actually running. Bucky has given Instinct his personal email and credit card and let it book flights and hotels, and he argues the ChatGPT moment for personal assistants came from a startup rather than an incumbent. Meta launched Muse the same week. Claire and Josh weigh where they still want a person reviewing the work, and the group gets into what the web does when booking sites cannot tell an agent from a customer. Bloomberg also reported that ByteDance is building a real time world model, and Bucky explains why serving those models is a hard problem of its own. Reactor is a Lightspeed portfolio company. Lightspeed co-led the seed round and led the Series A.

    Chapters


    0:00 Welcome, Bucky Moore!

    1:12 OpenAI's Astra

    3:50 The New AI Model’s Architecture

    4:40 AI Safety and Guardrails

    8:01 Trusting AI Agents

    11:09 The Future of AI Models

    12:06 The AI Agent Land Grab

    17:09 AI, Trust, and Personal Data

    20:12 Improving AI Agent Capabilities

    22:11 Enterprise AI Assistants and Personal Context

    24:05 AI Agents and the Future of the Web

    25:50 ByteDance and World Models

    28:19 World Models and Consumer Content

    30:56 Reactor and World Model Infrastructure

    34:13 Real-Time Video Models and Scaling

    37 min
  • Why Mistral Just Raised €3B to Build Sovereign AI | Lightwork

    Mistral has raised €3 billion in a Series D, the largest equity round ever completed by a European tech company, just three years after the Paris AI lab launched. The round brings Samsung and EQT onto the cap table. The money will go toward compute, infrastructure, commercial growth, and international expansion, with Mistral aiming for a gigawatt of European compute by 2030.


    Chapters:

    0:00 How Did Lightspeed Get Involved With Mistral?

    2:55 How Has Mistral Changed Since Its Early Days?

    4:30 Where Will the €3B be Invested?

    6:05 Why Does Sovereign AI Matter?

    7:37 What Role Is Mistral Trying to Play in AI?

    10:29 Why Is Mistral Building European Compute?

    11:59 How Does Mistral’s Full-Stack Strategy Fit Its Vision?

    13:38 What’s Next for Mistral and European AI?

    15:25 Could Mistral Transform Physical Industries?

    17 min
  • GPT-6 Astra, Meta Muse & ByteDance Enters the World Model Race | Lightwork

    Claire Zau and Josh Machiz are joined by Lightspeed partner Bucky Moore for a week about how much control we are handing over to AI. OpenAI released GPT-6 Astra, its most capable model and, by OpenAI's own account, a harder one to monitor. Bucky explains why looping inference compresses the chain of thought that used to make these models readable, and why the near term answer is tighter permissions and access controls rather than interpretability research.

    Then to the agents people are actually running. Bucky has given Instinct his personal email and credit card and let it book flights and hotels, and he argues the ChatGPT moment for personal assistants came from a startup rather than an incumbent. Meta launched Muse the same week. Claire and Josh weigh where they still want a person reviewing the work, and the group gets into what the web does when booking sites cannot tell an agent from a customer. Bloomberg also reported that ByteDance is building a real time world model, and Bucky explains why serving those models is a hard problem of its own. Reactor is a Lightspeed portfolio company. Lightspeed co-led the seed round and led the Series A.

    Chapters

    0:00 Welcome, Bucky Moore!

    1:12 OpenAI's Astra

    3:50 The New AI Model’s Architecture

    4:40 AI Safety and Guardrails

    8:01 Trusting AI Agents

    11:09 The Future of AI Models

    12:06 The AI Agent Land Grab

    17:09 AI, Trust, and Personal Data

    20:12 Improving AI Agent Capabilities

    22:11 Enterprise AI Assistants and Personal Context

    24:05 AI Agents and the Future of the Web

    25:50 ByteDance and World Models

    28:19 World Models and Consumer Content

    30:56 Reactor and World Model Infrastructure

    34:13 Real-Time Video Models and Scaling

    37 min
  • Tim Cook’s 15-Year Run, Meta's $18B Teen Safety Deal & 21 Banks Convene on Stablecoins

    Tim Cook's fifteen years running Apple ended this week and John Ternus becomes the company's eighth CEO, a hardware engineer who led every iPad, the last five iPhone generations, and AirPods. Claire and Sam argue Apple's decision to sit out the model building race may have been the right one, since it left the company competing on product experience while licensing models from others. Ternus fronts his first iPhone event next Wednesday.

    Meta agreed to pay eighteen billion dollars to settle state claims that Instagram and Facebook were designed to addict minors, without admitting wrongdoing, and the product terms are the part that matters. A two hour daily cap across both apps, no feeds or Reels between midnight and six, and only a parent can lift any of it. The same morning Meta ran full page ads asking TikTok and YouTube to adopt the same rules, offering to tighten its own cap to one hour per app if they follow. Then Sam explains why twenty one financial institutions want to issue a dollar stablecoin of their own, and why the banks watched transaction volume move onto rails they do not own. She reads it as a treasury tool before it is a consumer one.


    Chapter

    0:00 Tim Cook Steps Down As Apple CEO

    1:45 Tim Cook's Legacy By The Numbers

    4:00 John Ternus: Apple's 8th CEO

    5:05 Apple's AI Strategy Explained

    10:33 Meta's $18 Billion Teen Safety Settlement

    12:29 New Instagram And Facebook Teen Limits

    17:52 21 Banks Explore A Joint Stablecoin

    18:28 What Is A Stablecoin?

    22:33 Why JPMorgan Isn't Joining The Group

    28 min
  • The Company Predicting Whether You Will Return Something | Seel on Lightwork

    Zack Peng, co-founder and CEO of Seel, joins Lightwork to explain why he treated e-commerce returns as an underwriting problem rather than a logistics one. He was a philosophy and physics major who learned to code at Orbital Insight, where he watched algorithms predict events for hedge funds and insurers, and noticed that payment chargebacks drew heavy investment despite representing under 1% of GMV while returns ran 15 to 25%. Seel started by predicting returns from pre-purchase behavior using roughly 200 signals, including one Zack still finds surprising, which is whether a product photo uses a human model.

    The conversation gets into the choice to sell the liability rather than the signals, and a go-to-market that is free for brands and paid by shoppers at checkout, where an insurance-backed refund policy sits alongside features like 24/7 support and one-click resale through an agent that lists an item across seven marketplaces. Zack explains why Seel builds agencies rather than agents, arguing that an agency owns the outcome while a product can always blame the user, and why he thinks that accountability gap is what leaves many agentic support tools short of full resolution. He closes on where agentic commerce leaves the shopper, the consumer app Seel built around that, and why he expects a one-person billion dollar brand to stop sounding far-fetched.

    Chapters:

    0:00 Welcome, Zack Peng!

    1:45 What Does Seel Do for E-Commerce?

    4:21 How Zack Peng Went From Satellites to E-Commerce

    6:07 How Does Seel Predict Returns?

    7:25 How Seel’s Return Prediction Model Became Profitable

    9:05 What Did Seel Learn From Smaller Brands?

    11:17 How Does Seel’s Business Model Work?

    14:52 Who Is Accountable When Things Wrong?

    17:28 Where Does Seel Draw the Line on AI Operations?

    18:21 How Will Agentic Commerce Change E-Commerce?

    20:07 How Will AI Change the Consumer Shopping Experience?

    22:21 The Future of Seel

    24:57 Closing Thoughts

    26 min
  • Buying Insurance Agencies Instead of Just Selling Them Software | Gyde on Lightwork

    Will Johnson joins Claire Zau and Josh Machiz to explain why he buys health insurance agencies instead of just selling them software. Gyde calls itself the first AI-native insurance, wealth and health brokerage. Will's argument is that broker software did not really automate labor, it only organized data into dashboards, so he believes software alone was never going to produce the unlock he wanted. Will resists the roll up label. He calls it a transformation company. Gyde launched in January with $60 million led by Lightspeed.

    Then the conversation turns to open enrollment and to jobs. Will says this could be one of the most difficult fourth quarters brokers have seen. Renewals AI is Gyde's answer, with Gia running the outreach, the needs assessment, and the scheduling while the human broker delivers the advice. On jobs he says not one person has been removed because of AI and that Gyde has been hiring into acquired teams. He also puts the workforce math at six brokers retiring for every one who enters, which by Gyde's own analysis makes the problem capacity rather than replacement. He also describes speaking with more than a hundred agency owners who had already sold their firms, and ninety nine of them being deeply unhappy with how it went.

    56 min
  • Robots Outrun Usain Bolt, Skild's ChatGPT Moment & Google Bids $10M for Spirit Data | Lightwork

    Claire Zau and Josh Machiz break down a week about robot bodies, robot brains, and the data underneath both. Humanoid robots broke three human world records at the World Humanoid Robot Games in Beijing, where more than 600 teams and 2,000 robots competed across 51 events, up from roughly 280 teams and 26 events last year. One robot beat Usain Bolt's 9.58 second 100 meter record, and last year's winning time was 21.50. The hosts argue the sprints are the fanfare and the scenario events are the harder test, since a robot that outruns Bolt still cannot open a jar, fold socks, or stack cardboard. Unitree, which the hosts compare to Boston Dynamics, also went public in Shanghai.

    Skild AI launched S1, a model the company says lets a robot watch one video of a person doing a task and then do it, with no teleoperation and no retraining. Teaching a robot a new task used to mean 50 to 100 hours of puppeteering data. Skild says ingesting the video takes about 11 minutes, and pancake flipping was not in the roughly one million hours behind the model. Both hosts call it the ChatGPT moment for robotics. Lightspeed participated in Skild's Series A. The episode closes on Google's roughly $10 million bid for Spirit Airlines' digital estate in bankruptcy, a package holding about 100 million emails, 500 million Teams messages, 30 million recorded service calls, and 7.5 billion passenger transaction records. Google says a third party will scrub personal information first. Mercor bid $7.5 million, micro1 came in higher after the auction closed, existing employees have not consented to their workplace communications becoming training data, and the sale still needs judicial sign off.


    Chapters:

    0:00 Welcome to Lightwork

    0:59 Robots Beat Human World Records

    4:01 What Happened at the Robot Games?

    9:14 UniTree Goes Public in Shanghai

    12:29 Skild AI’s S1: The ChatGPT Moment for Robotics

    16:20 The ChatGPT Moment For Robotics

    18:40 In-Context Learning, But For Robots

    21:04 Google Wants Spirit Airlines’ Data

    23:35 What's Valuable Inside A Digital Estate

    28:58 Closing Thoughts


    Stay In Touch:

    Website: www.lsvp.com

    X: https://x.com/lightspeedvp

    LinkedIn: https://www.linkedin.com/company/lightspeed-venture-partners/

    Instagram: https://www.instagram.com/lightspeed/


    The content here does not constitute tax, legal, business or investment advice or an offer to provide such advice, should not be construed as advocating the purchase or sale of any security or investment or a recommendation of any company, and is not an offer, or solicitation of an offer, for the purchase or sale of any security or investment product. The views expressed by our guests do not necessarily represent the views of Lightspeed. For more details please see lsvp.com/legal.


    Robots Outrun Usain Bolt, Skild's ChatGPT Moment & Google Bids $10M for Spirit Data | Lightwork

    https://youtu.be/RZPgoLNTOOE

    31 min
  • Stripe Buys OpenRouter, Grok Bot Arrives, ChatGPT for Teens & AI Designs Proteins | Lightwork

    On this episode of Lightwork, Claire Zau and Josh Machiz break down four stories. Stripe acquired OpenRouter, the model gateway that lets a developer integrate once and reach hundreds of models, routing each request by budget, latency, or task, with automatic failover when a model goes down. The hosts get into why a payments company wants the layer that decides which model handles a request, whether Stripe is neutral enough to own it, and what happens as the labs pull routing into their own products and app builders consider building it themselves.

    Then SpaceX closed its acquisition of Cursor and shipped Grok Bot, agents that each get their own cloud computer and keep working after you close your laptop, with a teach a task feature that turns one demonstration into a repeatable skill. OpenAI launched ChatGPT for Teens, a separate experience for 13 to 17 year olds that redirects homework shortcuts into study mode and nudges a break after 90 minutes, arriving alongside the GUARD Act and California's SB 243. And an Anthropic experiment had Claude run a protein binder design campaign start to finish, with outside labs building and testing the results and binders produced against 14 of 15 targets.

    0:00 Welcome to Lightwork

    1:03 Stripe Buys OpenRouter: The AI Tollbooth

    5:40 OpenRouter’s Business Model

    8:54 Is Stripe Neutral Enough?

    12:01 SpaceX Closes the Cursor Acquisition

    18:04 What is Grok Bot?

    22:17 Open AI Launches ChatGPT for Teens

    23:55 What’s Different About Teen Mode?

    27:48 AI Regulation Is Coming

    30:57 Claude Designs Protein Binders

    35:44 Neko Health Announced NYC Clinic

    36:18 Castelion Raises $1B Series C

    39 min
  • Eon CEO: “The Cloud Is Just Someone Else's Computer” | Lightwork

    Ofir Ehrlich, CEO and co-founder of Eon, joins Claire Zau to explain why the cloud data enterprises pay to store and rarely open turned out to be the thing their AI needs most. Ofir walks through the four companies he has started, including CloudEndure, which he sold to Amazon before leading AWS's cloud migration and disaster recovery business. He assumed backup was a solved problem until a customer he had worked closely with was hit by ransomware and lost access to a significant portion of their infrastructure, which sent him looking at what breaks in the cloud that never broke on premises.

    They discuss how enterprise data sits locked across business units that cannot say what they hold, what happens when an auditor asks for records from five years ago and the database no longer runs, and why backup can rack up a company's cloud spend while staying something nobody trusts. As models and compute platforms become easy to swap, Ofir argues, a company's accumulated data becomes what actually differentiates it. He also talks about building Eon with a best friend of 26 years and a colleague he hired at AWS, and why he would not start a company without friends in it.

    Chapters:00:00 Welcome, Ofir Ehrlich!05:22 Why Companies Moved to the Cloud07:08 Learning Through Angel Investing08:41 The Founder Mindset: Imposter Syndrome vs. Conviction10:23 Building a Company Around What You Love13:11 How AI Is Changing Enterprise Data17:05 The New Cloud Infrastructure Organization19:09 Why Recovering Old Data Can Be So Difficult20:25 Why AI Teams Need Enterprise Data22:33 What Eon Actually Does24:40 How Eon Helps Make Data Usable26:47 Choosing Eon's Co-Founders30:29 What's Exciting in AI Right Now?31:37 Final Thoughts

    33 min
  • AI Creates New Viruses, Building AI for Healthcare & Moats in Healthcare IT | Lightwork

    Claire Zau is joined by Dr. Brenton Fargnoli, a physician and partner on Lightspeed's healthcare team, to unpack the AI-designed bacteriophage story making headlines this week. Stanford and Arc Institute researchers used a genome language model to design nearly 300 candidate phage genomes and landed on 16 that worked, a cocktail able to overcome resistance the natural phage couldn't beat on its own. The phages only infect E. coli, not human cells, and a scientist still has to synthesize and test every design in a lab.

    From there, Brenton and Claire get into what they believe it takes to build a vertical AI company in healthcare right now, from finding the right wedge to surviving the moment an incumbent like Epic decides to build the same feature natively. They cover why founders are seeing faster adoption than past software waves, what separates a real data or workflow moat from a feature a platform can copy, and what Brenton looks for in a team before backing them.


    Chapters:

    0:00 Welcome to Lightwork

    1:11 What Actually Happened With the 16 AI-Designed Viruses

    5:54 Could AI-Designed Viruses Infect Humans?

    6:32 AI, Antibiotic Resistance & Medical Applications

    10:47 Building Vertical AI in Healthcare

    13:39 What’s Making Healthcare AI Ripe for Startups?

    15:38 Top-Down vs. Bottom-Up Healthcare AI

    16:19 How Founders Can Compete With Big Tech

    20:13 What Makes a Great Healthcare AI Founder?

    21:44 Building Great Healthcare Tech Teams

    23:06 What Will Epic Announce Next?

    24:21 What Healthcare AI Opportunities Is Dr. Brenton Most Excited About?

    27:28 Final Thoughts

    29 min