5:47

Extreme power estimates collide with waste-heat limits

Into the Impossible With Brian Keating·6 min·Nov 11

Knuth estimates the power behind a reported UAP maneuver, then explains why even near-perfect efficiency would leave a severe heat problem. The mismatch points to missing assumptions or unexplained physics. 3 takeaways →

3:33

Relativistic travel trades short trips for decades on Earth ProPro pick

Into the Impossible With Brian Keating·4 min·Nov 11

Kevin Knuth, physics professor at the University at Albany, says objects moving like the reported Nimitz craft would make excellent interstellar ships. At 1,000 Gs, he says, 90% of light speed takes about 17 hours, and Proxima Centauri is under a day and a half away for travelers. On Into the Impossible With Brian Keating, he adds the catch. A trip taking them 4 days takes 80 years on Earth. 3 takeaways →

Knuth uses extreme acceleration to illustrate how relativistic travel could make distant stars reachable in a traveler’s lifetime. The catch is that far more time would pass for people back on Earth. 3 takeaways →

2:34

A reported underwater object implies extraordinary speed ProPro pick

Into the Impossible With Brian Keating·3 min·Nov 11

Knuth recounts a New Zealand naval report of a large underwater object that rapidly closed on a ship. His rough calculation suggests extreme speed and acceleration, though the sonar data still needs analysis. 3 takeaways →

7:40

Longer AI work horizons could shift companies toward outcome delegation

AI agents·Moonshots·8 min·Nov 11

An analyst's reading of a Meter report puts the time horizon of Anthropic's internal Model 4 at about 240 hours, doubling every two months. On Moonshots, Salim Ismail, Chairman of OpenExO, said this points to companies delegating outcomes instead of tasks. Co-host Imad added a bull case where Anthropic sells its top models only to firms making one-year upfront commitments. 3 takeaways →

The hosts discuss models taking on longer tasks and the prospect of delegating outcomes, not just instructions. They also consider how frontier labs might ration their most capable models. 3 takeaways →

4:43

Treating gene therapy as an approved platform could lower costs

Biology & drugs·Moonshots·5 min·Nov 11

An AI agent's thought experiment says curing the top 5,000 diseases requires approving a reusable process, not each drug. On Moonshots, Dave Blundin, Founder and General Partner of Link Ventures, shared it. At $2 billion a drug the total is $10 trillion, but at $2 million each it falls to $10 billion. The note names delivery to tissues like the kidney and slow diagnosis as the main obstacles. 3 takeaways →

Dave shares an agent’s thought experiment: approve reusable treatment platforms rather than each therapy as a wholly new drug. It lays out delivery, diagnosis, cost and regulation as the main bottlenecks. 3 takeaways →

6:25

Robotics startups should target valuable problems, not easy ones ProPro pick

Robotics·Moonshots·6 min·Nov 11

Robotics founders should chase scarce, valuable problems, not minimum-wage work like restocking grocery shelves, argued Alexander Wissner-Gross, founder of Reified, on Moonshots. He said Tesla's Optimus and Figure will flow to the highest-value work first, so startups should pick other form factors or build tools for them. Co-host Salim Ismail countered that simpler store robots could profit from the gap between $20 an hour wages and robot costs. 3 takeaways →

The hosts debate a startup building robots to restock grocery shelves and where it might fit beside general-purpose humanoids. Alex argues founders should pursue scarce, high-value problems rather than minimum-wage work. 3 takeaways →

5:17

Offensive success depends on suppressing enemy drone units

Lex Fridman Podcast·5 min·Nov 10

Ukraine now wins ground by first suppressing Russian drone units, says Michael Kofman, a senior fellow at the Carnegie Endowment. On the Lex Fridman Podcast he described a plan that hits drones and supply lines, then covers the assault. Kofman says this sharply cuts losses, which matters because he says Ukraine cannot afford Russia's way of fighting, where an American Marine described losing 90 to 95% of troops. 3 takeaways →

They explain how both sides contest the drone-and-fire zone before advancing, and why Ukraine's planned suppression reduces losses. Battlefield movement now depends on setting conditions, not just pushing forward. 3 takeaways →

8:04

Analysts over-rationalized Russia's invasion plan ProPro pick

Lex Fridman Podcast·8 min·Nov 10

Michael Kofman, a senior fellow at the Carnegie Endowment, says he rightly predicted Russia's 2022 full-scale invasion but misjudged how it would unfold. On the Lex Fridman Podcast he said analysts over-rationalized Russia's plan and missed that it rested on Putin's assumption of a quick takeover and a friendly welcome. Kofman also says Ukraine's military was underestimated, and that military power means little without a specific scenario. 3 takeaways →

Kofman explains why he expected a full-scale invasion but misjudged how Russia would execute it and how Ukraine would respond. He argues that analysts must connect military plans to leaders' assumptions and remain humble about uncertainty. 3 takeaways →

5:44

Bakhmut's symbolic value outlasted its military value

Lex Fridman Podcast·6 min·Nov 10

Michael Kofman, a senior fellow at the Carnegie Endowment, argues Bakhmut mattered more as a symbol than as ground. On the Lex Fridman Podcast he said it took Russia a year to reach Chasiv Yar, the next town, and another year to take much of it. Kofman says declaring a city a fortress traps troops, and holding Bakhmut may have used up experienced Ukrainian units. 3 takeaways →

Kofman describes how Bakhmut became a politically charged battle that tied down Ukrainian forces and shaped Western expectations. Its eventual loss, he argues, mattered less militarily than its symbolic weight suggested. 3 takeaways →

3:40

Tesla recovered sales by calling test-drive leads ProPro pick

Growth·The Knowledge Project·4 min·Nov 10

Tesla found 10,000 people who had taken a test drive in the previous 30 days and never got a call back. On The Knowledge Project, Jon McNeill, co-founder and CEO of DVx Ventures, said Tesla had no sales commissions, so nobody was pushed to ask for the order. He paused new leads until stores called prior ones, and orders came in the next day. 3 takeaways →

McNeill found 10,000 test-drive customers who had not been called back. He stopped new leads until stores followed up, and orders started coming in. 3 takeaways →

3:36

The five-step algorithm simplifies before it automates ProPro pick

AI startups·The Knowledge Project·4 min·Nov 10

Jon McNeill, co-founder and CEO of DVx Ventures, lays out the five-step Algorithm Tesla built from postmortems on its big mistakes. On The Knowledge Project, he says to question every requirement, delete steps customers don't pay for, optimize by hand, speed it up to expose failures, and automate last. He says it is in the SpaceX IPO slides. 3 takeaways →

McNeill lays out a process for removing unnecessary requirements, simplifying and testing a workflow, then automating it. The sequence helps expose problems before they become embedded in software or systems. 3 takeaways →

2:53

McNeill left Tesla when the job became unhealthy

Hiring & teams·The Knowledge Project·3 min·Nov 10

Jon McNeill, co-founder and CEO of DVx Ventures, says he left Tesla after his family held an intervention over the Christmas holidays and told him he had become moody and mean. On The Knowledge Project, he explains he was stressed trying to help Elon Musk with mental health struggles he felt unequipped to handle. A few weeks later he told Musk it was time to go. 3 takeaways →

McNeill describes how trying to support Elon through a difficult period took a toll that his family noticed. He eventually told Musk he could not help with those problems and needed to leave. 3 takeaways →

The best 4 minutes of every big interview

4:21

Mathematical proofs could make AI agents safer

AI agents·Sources·4 min·Nov 10

AWS is betting on mathematical proofs to keep unpredictable AI agents inside company rules. On Sources, Matt Garman, CEO of AWS, said Amazon's automated reasoning research already checks whether cloud permissions conflict, reaching 100% proof without brute force. Applied to agents, it could let an insurer process claims fully automated and answer in 10 seconds instead of a month, with no human checking each step. 3 takeaways →

Garman explains how formal proofs can keep unpredictable agents within defined rules. That could let companies automate sensitive work without relying on a human to check every action. 3 takeaways →

2:49

Compute boom risks depend on how providers finance and sell capacity

Inference & cost·Sources·3 min·Nov 10

The compute boom's risks are local, not systemic, says Matt Garman, CEO of AWS, on Sources. He argues neoclouds like CoreWeave carry high-rate debt, large training clusters and a few big customers, which leaves little room for surprises. AWS serves millions of customers, and most of its AI work is inference spread across many end users. If one model provider fails, he says, the work moves to another. 3 takeaways →

Garman argues that compute risks are concentrated among providers with debt-heavy financing and a few large training customers. AWS, he says, has a broader base of workloads and inference users. 3 takeaways →

2:57

AI is accelerating the move of core workloads to the cloud

Enterprise AI·Sources·3 min·Nov 10

AI is pulling companies' core systems into the cloud, and AWS is feeling it. On Sources, Matt Garman, CEO of AWS, said agents struggle to pull data from on-prem databases, mainframes or VMware clusters in legacy data centers. He says that realization has accelerated AWS's core business over the last couple of years, with growth across tens of thousands of customers, though he calls the outcome far from certain. 3 takeaways →

Garman says companies need cloud-based access to their data and systems to get value from AI agents. That shift is speeding up cloud migrations and reinforcing AWS growth. 3 takeaways →

6:24

Enterprise AI adoption follows inference spending

Enterprise AI·Sourcery·6 min·Nov 10

Brendan Foody, co-founder and CEO of Mercor, says a company's inference spending shows when it should invest in its own models. Under $10M, firms are still figuring things out. From $10M to $100M, about 10% of inference budget should go to evals. Past $100M, post-training starts to pay off. He made the case on Sourcery, predicting every Fortune 500 will build its own models. 3 takeaways →

Foody predicts Fortune 500 companies will build their own models and explains how inference spend shapes adoption. His thresholds give operators a way to plan for evals and post-training. 3 takeaways →

3:23

Scalable interviews help find rare high-quality experts

Building AI products·Sourcery·3 min·Nov 10

70% of the experts who sign up with Mercor come through referrals, says co-founder and CEO Brendan Foody. On Sourcery, he said Mercor built its first AI interviewer in April 2023 and has now run over seven million interviews. Foody argues scaling interviews finds world-class engineers where talent is sparse, and says Mercor now pays out about $5 million a day. 3 takeaways →

Foody explains how Mercor built an AI interviewer and uses referrals, interviews, and other signals to assess candidates. The approach helps find strong talent even in low-density labor markets. 3 takeaways →

3:24

Mercor sells training environments, not just expert access

Training data·Sourcery·3 min·Nov 10

Brendan Foody, co-founder and CEO of Mercor, says people wrongly see the company as an Uber-style marketplace. On Sourcery, he said Mercor sells RL environments to AI labs, made of apps, the content inside them, and tasks with grading rubrics. Experts are one ingredient. AI copilots now make an hour of expert time yield three to five times the improvement, and demand has grown as a result. 3 takeaways →

Foody breaks down the apps, worlds, and tasks that make up Mercor’s RL environments. He explains why experts are one input to the product and how AI copilots increase demand for their work. 3 takeaways →

5:08

Email delivery depends on pacing and reputation ProPro pick

Latent Space·5 min·Nov 10

Sending one email to 100,000 people at once gets throttled by Gmail, which watches early opens, archives and deletes before accepting the rest. Zeno Rocha, co-founder and CEO of Resend, told Latent Space that senders should queue the most engaged recipients first and space out batches across providers and IP pools. Abusers buy new domains and burst-send, so safe senders do the opposite. 3 takeaways →

Gmail meters large sends based on early engagement, so providers pace queues and warm IP pools. Zeno explains how abuse patterns shape sender safeguards. 3 takeaways →

6:15

An MCP should cover the whole product

AI agents·Latent Space·6 min·Nov 10

Resend's MCP, the connector that lets AI tools use its email product, grew at least 2x every month, Zeno Rocha says. The co-founder and CEO told Latent Space that doing it right meant building OAuth and a permissions model, and growing from about 20 tools to 80-something. He says buyers now pick vendors by official MCP support, and full dashboard coverage is the hard part. 3 takeaways →

Resend treats MCP as a full product surface, from OAuth and RBAC to official integrations. Zeno argues tool coverage should match the dashboard, including account and billing actions. 3 takeaways →

6:39

Markdown pages can correct AI pricing answers

Growth·Latent Space·7 min·Nov 10

Resend added FAQs and Markdown versions of its pages so models could read current, exact answers. A Markdown pricing page corrected a model's false claim about Resend's costs. 3 takeaways →

3:20

Canva's data makes smaller design models more economical

Building AI products·The TWIML AI Podcast·3 min·Nov 10

Canva says its own design models run 40 times cheaper than the most expensive frontier models on the same design tasks, at the same quality or sometimes better. Stefano Corazza, Head of AI Research at Canva, told The TWIML AI Podcast that unique data let a model with far fewer parameters beat a bigger one. Planning the layout before generating pixels also lets edits happen earlier. 3 takeaways →

Corazza explains why Canva trained its own models: unique design data and tight product integration. He says smaller models can match larger ones at a fraction of the cost. 3 takeaways →

5:11

Detailed scene descriptions help smaller models create better designs

Image & video models·The TWIML AI Podcast·5 min·Nov 10

Canva found its image model can take far more detailed scene descriptions than expected, including bounding boxes and positions. Stefano Corazza, Head of AI Research at Canva, told The TWIML AI Podcast that for most creative tasks a well-tuned model under 100 billion parameters is enough. He says a frontier model would cost about a dollar per image, so the economics fail. 3 takeaways →

Corazza describes how Canva expanded structured scene descriptions and tuned them to its diffusion model. He argues that creative tasks can work well with models under 100 billion parameters. 3 takeaways →

3:44

Owning the model stack creates more room to cut inference costs

Inference & cost·The TWIML AI Podcast·4 min·Nov 10

Canva says the cost of generating a second of 720p video fell two orders of magnitude in a year, from about half a dollar to some cents. Stefano Corazza, Head of AI Research at Canva, told The TWIML AI Podcast this comes from tuning every layer, from GPUs to serving. He expects another tenfold drop, which would let more free users use AI. 3 takeaways →

Corazza says Canva improves inference across hardware, kernels, model steps and serving. Those optimizations make AI more affordable and could drive another order-of-magnitude cost drop. 3 takeaways →

5:15

Agents need search built for different inputs and latency

AI agents·Knuckle Up with Nakul·5 min·Nov 10

Search built for humans doesn't fit AI agents, says Parag Agrawal, founder and CEO of Parallel Web Systems. On Knuckle Up with Nakul, he said a person types a short query and gets 10 blue links, while an agent can write a session's worth of query and needs tokens for its next actions. Latency swings from 50 to 200 milliseconds for interactive uses to an hour for overnight agents. 3 takeaways →

Agrawal explains how agents submit richer queries, need different results and have highly variable latency needs. Those differences call for search infrastructure designed around agents, not human browsing. 3 takeaways →

3:04

A CEO should shape the company around how they lead

Hiring & teams·Knuckle Up with Nakul·3 min·Nov 10

Once you become CEO, you stop adapting yourself to the company and adapt the company to you, says Parag Agrawal, founder and CEO of Parallel Web Systems. On Knuckle Up with Nakul, he said that after more than 10 years at Twitter he learned that fitting into someone else's shoes shortchanges the company. A product-first, data-first or other CEO should hire and run executive meetings differently. 3 takeaways →

Agrawal says a CEO must stop adapting themselves to fit the existing leadership and instead shape the company around their own strengths and vision. That shift lets the business pursue its best version. 3 takeaways →

7:12

Agent-scale web access needs a way to pay for content ProPro pick

AI agents·Knuckle Up with Nakul·7 min·Nov 10

Parag Agrawal, founder and CEO of Parallel Web Systems, warns that ads and per-seat subscriptions are hard to imagine surviving once agents are the web's main users, and publishers may shut their content off. On Knuckle Up with Nakul, he described Parallel's Index, a model that prices content by how much it improves an agent's answer. Partners include the Atlantic, Fortune and PitchBook. 3 takeaways →

Agrawal argues that agents threaten ad and per-seat subscription models, leaving publishers with less reason to keep content open. Parallel aims to price content by how much it improves an agent's output. 3 takeaways →