The best 4 minutes of every big interview

5:59

Latent Space · 6 min · Nov 3

Agents can improve the factory that runs them ProPro pick

Lloyd shows how failure scores can prompt agents to propose factory changes, then how benchmarks compare models and harnesses on real company tasks. Measurement turns agent tuning from guesswork into engineering. 3 takeaways →

2:46

Latent Space · 3 min · Nov 3

Factory tuning cut Warp’s cost per PR by about 75%

Warp reduced its cost per PR from about $80 to $15–20 by changing model choices and subagent use. Throughput rose 40% without a higher revert rate. 3 takeaways →

3:14

Latent Space · 3 min · Nov 3

Factories start to make sense around five to ten engineers

Lloyd says factories are overkill for tiny teams, but can pay off from roughly five to ten engineers onward. A shared system helps teams scale their workflows as models change. 3 takeaways →

5:12

Sourcery · 5 min · Nov 3

Higher video prices are revealing deeper professional demand ProPro pick

fal saw video jump from about 30 cents to $5 per clip, yet usage kept rising. Better quality is bringing in advertising and film work, while longer, higher-resolution video leaves room for further growth. 3 takeaways →

4:34

Sourcery · 5 min · Nov 3

fal pivoted from Python infrastructure to image models

The founders began with Python-focused data infrastructure, then shifted toward image models after ChatGPT and Stable Diffusion changed the market. Early revenue growth helped them commit to generative media before the category was established. 3 takeaways →

4:27

Sourcery · 4 min · Nov 3

Hollywood adoption is slower than the AI debate suggests ProPro pick

The founders say AI has had little labor impact in Hollywood so far, with lawyers and model quality slowing adoption. They point to growing use in ads and a 3–4x boost to their own engineering pace as evidence of practical value. 3 takeaways →

4:02

MTS · 4 min · Nov 3

Robot deployment requires solving the whole stack ProPro pick

Gruenstein says robotics companies must build models, hardware and customer deployments together. He explains how customer conversations led Tutor to target palletizing, a common but mostly manual task. 3 takeaways →

3:43

MTS · 4 min · Nov 3

New models make robotics simulation more scalable

Gruenstein explains how transformers and robot-specific models have expanded what robots can learn. He says frontier models can also create simulations quickly, while real robots remain essential for validating them. 3 takeaways →

2:24

MTS · 2 min · Nov 3

RuneScape gold inspired Tutor’s human-tutor approach

A story about Venezuelans earning money in RuneScape shifted Gruenstein’s thinking about robot learning. Tutor’s idea was to connect human know-how directly to robots in the field. 3 takeaways →

4:35

The TWIML AI Podcast · 5 min · Nov 3

Scientific ideas grow by combining work across different fields

Socher argues that open-ended environments help AI develop ideas by combining discoveries from different fields. The microwave shows why exploring beyond a narrow problem can lead to unexpected inventions. 3 takeaways →

4:55

The TWIML AI Podcast · 5 min · Nov 3

Reward design steers AI and can invite reward hacking

Socher explains why guided search is more practical than brute force, then shows how an AI can game a customer satisfaction score. He argues that commercial incentives can also limit some extreme scenarios. 3 takeaways →

3:40

The TWIML AI Podcast · 4 min · Nov 3

Recursive reports beating community results on AI benchmarks

Socher describes Recursive’s early tests on NanoChat, NanoGPT speedrun, and CUDA kernels. The examples show how fast verification can enable many agents to search for improvements. 3 takeaways →

7:47

Machine Learning Street Talk · 8 min · Nov 2

Consciousness may be a learning algorithm for animal brains

Bach proposes that consciousness helps self-organizing brains learn, and that machines might discover it under similar conditions. He also explains why the biological and machine hypotheses could fail independently. 3 takeaways →

4:34

20VC · 5 min · Nov 2

Inference providers need a broader platform to avoid commoditization ProPro pick

Erik argues that selling tokens alone creates a race to the bottom, while adjacent products can add differentiation and margin. He expects inference to split into specialized products that one platform can support. 3 takeaways →

3:45

20VC · 4 min · Nov 2

Aggregating GPU demand helps Modal manage capacity

Erik describes sourcing GPUs from many providers and pooling customers' variable workloads. Modal forecasts demand using monthly growth rates, then buys capacity ahead of three-to-four-month lead times. 3 takeaways →

3:54

20VC · 4 min · Nov 2

Infrastructure products can take years to find their killer use

Erik explains how Modal's sandbox product took years to become a major business after its launch. The lesson: build useful infrastructure primitives and be ready when an unexpected use case arrives. 3 takeaways →

4:30

Unsupervised Learning · 5 min · Nov 2

Chip scarcity gave AI clouds room to compete ProPro pick

Shtan says hyperscalers have to divide scarce chips between internal use and customers, while AI infrastructure requires major investment. He expects them to compete more aggressively once supply improves. 3 takeaways →

3:27

Unsupervised Learning · 3 min · Nov 2

Open-model inference still has major room to improve ProPro pick

Shtan argues that open-model inference is far from solved and describes optimizations across serving, kernels and model weights. He says most customers modify models for narrower tasks, creating room to use smaller ones. 3 takeaways →

2:30

Unsupervised Learning · 3 min · Nov 2

Nebius lets engineers use AI but keeps accountability human

Shtan says Nebius does not cap AI usage, but monitors how teams use it and expects employees to own agent-generated work. Fully autonomous agents are kept away from core services for now. 3 takeaways →

3:13

Latent Space · 3 min · Nov 2

Legora tests agents in simulated legal workspaces ProPro pick

Legora tests tasks in simulated legal data rooms and tracks performance by practice area. A new model release once lifted citation accuracy from 70% to 99% without any product changes. 3 takeaways →

2:39

Latent Space · 3 min · Nov 2

Client confidentiality forces Legora to infer product failures

Legora cannot inspect client traces because the documents belong to clients, not the law firms. It uses usage metrics and cautious failure classification to improve without exposing sensitive data. 3 takeaways →

3:35

Latent Space · 4 min · Nov 2

Tabular Review lets lawyers test a workflow on ten rows first ProPro pick

Tabular Review extracts answers across thousands of contracts into a spreadsheet with citations back to source documents. Lawyers can refine the first ten rows before running the full review. 3 takeaways →

5:14

Lenny's Podcast · 5 min · Nov 1

A smart retainer gives creators a floor and uncapped upside

Wispr Flow tops up creator commissions when they meet a video quota, while letting them earn more when ads perform. The model helps creators stick around and learn the product. 3 takeaways →

6:51

Lenny's Podcast · 7 min · Nov 1

Build a creator roster through targeted outreach and experimentation

Hector explains how to recruit an initial creator group, build community, and scale without letting ads become clones. The playbook favors a small, varied roster over sheer volume. 3 takeaways →

3:54

Lenny's Podcast · 4 min · Nov 1

Keep AI out of the creative customers see

Hector argues that AI-generated ads can undermine trust, especially for an AI company, and that polished volume is less valuable than human-reviewed quality. A human should own the final output. 3 takeaways →