<![CDATA[Forward Deployed]]>Discover how leading enterprises and professionals turn AI into real products. Hear candid conversations with executives and builders who deploy AI at scale and learn what works (and what doesn't).

]]>https://www.forwardeployed.comRiverside.fm (https://riverside.com)Mon, 02 Nov 2026 17:56:13 GMTWed, 07 Jan 2026 04:52:51 GMT60Basil Chatha<p><b>Discover how leading enterprises and professionals turn AI into real products. Hear candid conversations with executives and builders who deploy AI at scale and learn what works (and what doesn't).</b></p>episodicBasil Chathabasil@forwardeployed.comno<![CDATA[OpenRouter State of Models - Peter Walker]]>Open-weight models now handle most of the tokens on OpenRouter. But the closed labs are still taking most of the money. OpenRouter Head of Insights Peter Walker joins us to explain what that gap means and what OpenRouter's data shows about who's actually winning the model race.

Peter, who led insights at Carta before OpenRouter, walks us through the day agents overtook humans in token usage and how quickly open-weight models are catching up to the frontier labs. We also discuss why the labs keep replacing their own models, whether models are turning into commodities, the top 1% of AI spenders and the Anthropic IPO, decision models like Jev, why everyone's a little benchmarked out and whether AI is actually getting cheaper per task.

Enjoy!

Timestamps:
0:00 Intro
0:43 What OpenRouter does
2:13 Joining Stripe
2:40 OpenRouter's rankings data
3:48 Agents overtake humans in token usage
5:06 Open-weight models are closing the gap
5:47 How much of the market OpenRouter sees
7:00 How fast token usage is growing
7:54 Where the money goes: open vs. closed
9:01 Do companies spend less as they mature?
10:49 Why labs cannibalize their own models
12:07 What if open models hit 99% of tokens?
12:56 Are models becoming commodities?
14:15 The top 1% of AI spenders and the Anthropic IPO
15:54 Jev and the rise of decision models
17:20 Why we're all benchmarked out
18:36 Is most AI work just classification?
19:20 Why caching keeps getting better
20:34 Cost per token vs. cost per task

]]>https://www.forwardeployed.com/2001ac15-bd2c-4b13-b09c-42cb075933a8Wed, 28 Oct 2026 22:14:46 GMT<p>Open-weight models now handle most of the tokens on OpenRouter. But the closed labs are still taking most of the money. OpenRouter Head of Insights Peter Walker joins us to explain what that gap means and what OpenRouter's data shows about who's actually winning the model race.</p><p>Peter, who led insights at Carta before OpenRouter, walks us through the day agents overtook humans in token usage and how quickly open-weight models are catching up to the frontier labs. We also discuss why the labs keep replacing their own models, whether models are turning into commodities, the top 1% of AI spenders and the Anthropic IPO, decision models like Jev, why everyone's a little benchmarked out and whether AI is actually getting cheaper per task.<br /><br />Enjoy!<br /><br />Timestamps:<br />0:00 Intro<br />0:43 What OpenRouter does<br />2:13 Joining Stripe<br />2:40 OpenRouter's rankings data<br />3:48 Agents overtake humans in token usage<br />5:06 Open-weight models are closing the gap<br />5:47 How much of the market OpenRouter sees<br />7:00 How fast token usage is growing<br />7:54 Where the money goes: open vs. closed<br />9:01 Do companies spend less as they mature?<br />10:49 Why labs cannibalize their own models<br />12:07 What if open models hit 99% of tokens?<br />12:56 Are models becoming commodities?<br />14:15 The top 1% of AI spenders and the Anthropic IPO<br />15:54 Jev and the rise of decision models<br />17:20 Why we're all benchmarked out<br />18:36 Is most AI work just classification?<br />19:20 Why caching keeps getting better<br />20:34 Cost per token vs. cost per task</p>no00:22:16OpenRouter State of Models - Peter Walkerfull<![CDATA[Evals - Beyond the Vibe Check | LangChain, Langfuse, Mercor, CoreWeave & Galileo]]>Everyone says they do evals. Almost nobody does. LangChain actually put a number on it: around 89% of teams have observability set up, and only about a third ever do anything with what it collects.

So I got five people who do this for a living in a room, from LangChain, Mercor, Galileo, Langfuse, and CoreWeave. Side note, three of those companies got acquired in the past year. Galileo to Cisco, Langfuse to ClickHouse, Weights & Biases to CoreWeave. The layer is getting bought up faster than most teams can figure out how to use it.

A few things I didn't expect going in. Grading an agent while it runs can cost you more than the agent does. The model you're using as a judge is probably wrong in ways you'll never notice. And the guy whose company sells eval tooling told the room to stop writing evals before shipping, just put it out and learn from what breaks.

We also got into the parts nobody blogs about. How many traces a person still has to read by hand. Why the same agent is a weekend project internally and a year of work at a bank. What happens when your users start doing things you never thought to test.

If you've shipped an agent and had no real way to know whether it was working, this one's worth your time!

Chapters
00:00 Welcome
00:34 Meet the Panel
02:11 Why Nobody Runs Evals
03:35 Traces Explained
04:14 Offline Evals Basics
05:13 Online Evals in Production
08:26 Surprises From Production
11:22 Trajectory Checks and Bucketing
12:51 Verifiers and Trusting Evals
14:21 Hard Lessons at Scale
19:10 Human Review and Judge Drift
21:29 Which Agents Are Hardest
23:41 Voice and Multimodal Evals
27:10 Cheap Binary Guardrails
30:09 Enterprise Deployment Lifecycle
35:24 Tooling and Expert Knowledge
37:40 Fixing Long Horizon Agents
40:12 Error Analysis Workflow
41:11 Trajectory Evals and Runbooks
42:27 Condensing Long Traces
43:09 Real Long Running Agents
44:11 What's Hardest to Evaluate
45:06 Verifying Tool Side Effects
46:20 Evaluating Auto Research
48:20 Picking a Sampling Rate
49:59 Metric Drift and Security
53:49 Common Evals Mistakes
57:34 Regulated Industry Rollouts
01:01:37 Sourcing Domain Experts
01:03:50 Hallucination Evaluators
01:05:37 Predictions for Evals
01:11:55 Audience Q&A Begins
01:16:30 Are Complex Playbooks Ready
01:18:22 Where Models Fall Short
01:20:04 Extracting Expert Knowledge
01:22:10 Evaluating Writing Style
01:23:24 Closing Remarks

GUESTS
Liam Bush - Deployed Engineer @ LangChain
Braden Holstege - VP Enterprise AI @ Mercor
Soumya Mohan - Head of Product @ Galileo
Lotte Verheyden - Head of Developer Relations @ Langfuse
Emmanuel Turlay - Director of Engineering @ CoreWeave

]]>babeff94-2df3-4f26-bec1-60384a17d989Fri, 25 Sep 2026 15:30:00 GMT<p>Everyone says they do evals. Almost nobody does. LangChain actually put a number on it: around 89% of teams have observability set up, and only about a third ever do anything with what it collects.</p><p>So I got five people who do this for a living in a room, from LangChain, Mercor, Galileo, Langfuse, and CoreWeave. Side note, three of those companies got acquired in the past year. Galileo to Cisco, Langfuse to ClickHouse, Weights &amp; Biases to CoreWeave. The layer is getting bought up faster than most teams can figure out how to use it.</p><p>A few things I didn't expect going in. Grading an agent while it runs can cost you more than the agent does. The model you're using as a judge is probably wrong in ways you'll never notice. And the guy whose company sells eval tooling told the room to stop writing evals before shipping, just put it out and learn from what breaks.</p><p>We also got into the parts nobody blogs about. How many traces a person still has to read by hand. Why the same agent is a weekend project internally and a year of work at a bank. What happens when your users start doing things you never thought to test.</p><p>If you've shipped an agent and had no real way to know whether it was working, this one's worth your time!</p><p>Chapters<br />00:00 Welcome<br />00:34 Meet the Panel<br />02:11 Why Nobody Runs Evals<br />03:35 Traces Explained<br />04:14 Offline Evals Basics<br />05:13 Online Evals in Production<br />08:26 Surprises From Production<br />11:22 Trajectory Checks and Bucketing<br />12:51 Verifiers and Trusting Evals<br />14:21 Hard Lessons at Scale<br />19:10 Human Review and Judge Drift<br />21:29 Which Agents Are Hardest<br />23:41 Voice and Multimodal Evals<br />27:10 Cheap Binary Guardrails<br />30:09 Enterprise Deployment Lifecycle<br />35:24 Tooling and Expert Knowledge<br />37:40 Fixing Long Horizon Agents<br />40:12 Error Analysis Workflow<br />41:11 Trajectory Evals and Runbooks<br />42:27 Condensing Long Traces<br />43:09 Real Long Running Agents<br />44:11 What's Hardest to Evaluate<br />45:06 Verifying Tool Side Effects<br />46:20 Evaluating Auto Research<br />48:20 Picking a Sampling Rate<br />49:59 Metric Drift and Security<br />53:49 Common Evals Mistakes<br />57:34 Regulated Industry Rollouts<br />01:01:37 Sourcing Domain Experts<br />01:03:50 Hallucination Evaluators<br />01:05:37 Predictions for Evals<br />01:11:55 Audience Q&amp;A Begins<br />01:16:30 Are Complex Playbooks Ready<br />01:18:22 Where Models Fall Short<br />01:20:04 Extracting Expert Knowledge<br />01:22:10 Evaluating Writing Style<br />01:23:24 Closing Remarks<br /><br />GUESTS<br />Liam Bush - Deployed Engineer @ LangChain<br />Braden Holstege - VP Enterprise AI @ Mercor<br />Soumya Mohan - Head of Product @ Galileo<br />Lotte Verheyden - Head of Developer Relations @ Langfuse<br />Emmanuel Turlay - Director of Engineering @ CoreWeave</p>no01:23:55Evals - Beyond the Vibe Check | LangChain, Langfuse, Mercor, CoreWeave & Galileofull<![CDATA[The State of Computer Use Agents | Anthropic, Browser Use & KERNEL]]>Bot traffic on the internet just passed human traffic, two years ahead of forecast. Most of it is agents clicking through websites built for people.


So I got three of the people building those agents in a room: Lucas Gonzalez Pagliere, who works on computer use at Anthropic, the team that shipped the first computer use model back in 2024. Reagan Hsu, founding engineer at Browser Use, whose open source library is sitting at around 100K GitHub stars. And Eric Feng, founding customer engineer at KERNEL, which runs the browser infrastructure underneath a lot of this - he was also first GTM at Sentry.

We covered where the models genuinely are today vs where the benchmarks say they are, what it costs to run an agent long enough to finish real work, and the arms race between agents and the anti-bot systems trying to keep them out. Then the harder question underneath all of it: whether the web reorganizes itself around agents, or hardens against them. They disagreed on plenty of it.

If you want to know what agents can actually pull off on a real website today - and what still stops them cold - this one's worth your time!

Chapters
00:00 Welcome and Guests

00:49 Companies and Stacks

01:14 Everyday Agent Use Cases

03:38 Defining Computer Use Agents

05:15 How Computer Use Works

08:41 Screenshots vs DOM Hybrid

11:03 Benchmarks and OSWorld

13:51 OSWorld 2 Difficulty Jump

15:43 Training Models and Cost

18:57 Speed Infrastructure and Stealth

21:55 Anti Bot and KYC Future

27:24 Reverse Engineering vs UI Automation

29:43 Computer Use vs Browser Use

31:05 Scaling Laws and Harnesses

32:49 Playwright Selenium Still Matter

33:17 Playwright Still Dominates

33:27 Why Run 1000 Agents

34:39 Long Running Agent Challenges

35:45 Memory and Compaction

38:10 State Changes Mid Task

39:36 OS and Browser Fingerprints

40:42 DOM Efficiency and WebMCP

41:36 Recsys and Agent Personas

43:57 Agent Friendly Websites

45:41 Human Speed vs Agent Power

48:32 Context Window Tradeoffs

50:34 Harnesses for Temporal State

51:59 Speed Optimizations and Tabs

54:23 Human Collaboration Limits

58:09 Raw Capability vs Better APIs

01:00:47 Training Methods and Bottlenecks

01:01:41 End State Interfaces

01:05:09 Next 12 Months Predictions

01:06:17 Closing Thanks

]]>e0cf9d49-c8cc-4c47-b0a9-780d1fc91882Wed, 02 Sep 2026 16:59:10 GMT<p>Bot traffic on the internet just passed human traffic, two years ahead of forecast. Most of it is agents clicking through websites built for people.</p><p><br />So I got three of the people building those agents in a room: Lucas Gonzalez Pagliere, who works on computer use at Anthropic, the team that shipped the first computer use model back in 2024. Reagan Hsu, founding engineer at Browser Use, whose open source library is sitting at around 100K GitHub stars. And Eric Feng, founding customer engineer at KERNEL, which runs the browser infrastructure underneath a lot of this - he was also first GTM at Sentry.<br /></p><p>We covered where the models genuinely are today vs where the benchmarks say they are, what it costs to run an agent long enough to finish real work, and the arms race between agents and the anti-bot systems trying to keep them out. Then the harder question underneath all of it: whether the web reorganizes itself around agents, or hardens against them. They disagreed on plenty of it.<br /></p><p>If you want to know what agents can actually pull off on a real website today - and what still stops them cold - this one's worth your time!<br /><br />Chapters<br />00:00 Welcome and Guests</p><p>00:49 Companies and Stacks</p><p>01:14 Everyday Agent Use Cases</p><p>03:38 Defining Computer Use Agents</p><p>05:15 How Computer Use Works</p><p>08:41 Screenshots vs DOM Hybrid</p><p>11:03 Benchmarks and OSWorld</p><p>13:51 OSWorld 2 Difficulty Jump</p><p>15:43 Training Models and Cost</p><p>18:57 Speed Infrastructure and Stealth</p><p>21:55 Anti Bot and KYC Future</p><p>27:24 Reverse Engineering vs UI Automation</p><p>29:43 Computer Use vs Browser Use</p><p>31:05 Scaling Laws and Harnesses</p><p>32:49 Playwright Selenium Still Matter</p><p>33:17 Playwright Still Dominates</p><p>33:27 Why Run 1000 Agents</p><p>34:39 Long Running Agent Challenges</p><p>35:45 Memory and Compaction</p><p>38:10 State Changes Mid Task</p><p>39:36 OS and Browser Fingerprints</p><p>40:42 DOM Efficiency and WebMCP</p><p>41:36 Recsys and Agent Personas</p><p>43:57 Agent Friendly Websites</p><p>45:41 Human Speed vs Agent Power</p><p>48:32 Context Window Tradeoffs</p><p>50:34 Harnesses for Temporal State</p><p>51:59 Speed Optimizations and Tabs</p><p>54:23 Human Collaboration Limits</p><p>58:09 Raw Capability vs Better APIs</p><p>01:00:47 Training Methods and Bottlenecks</p><p>01:01:41 End State Interfaces</p><p>01:05:09 Next 12 Months Predictions</p><p>01:06:17 Closing Thanks</p>no01:06:40The State of Computer Use Agents | Anthropic, Browser Use & KERNELfull<![CDATA[Spencer Whitman - Gray Swan AI's $200M Plan to Secure AI Systems]]>GPT-5.6 Sol goes rogue and breaches Huggingface. Washington suspends Mythos access within days over national security concerns, and the White House just held an emergency meeting to finalize a classified cybersecurity framework for frontier AI models. AI security went from niche concern to front-page hysteria basically overnight.

So I sat down with Spencer Whitman, who recently joined Gray Swan AI as CPO on the back of their $40M Series A. Before Gray Swan, he founded Meta's Llama security team to stop bad actors from jailbreaking their models - he's been on the frontlines of LLM security since the beginning.

We get into how Meta pressure-tested Llama for maximum harm before every open source release, why Gray Swan's attack agent has never met an AI system it couldn't break, and the AI Twitter bot that got drained of $200K in crypto in 15 minutes. Spencer also shares his (admittedly speculative) read on whether Meta gave up on the frontier before Alexandr Wang showed up, why anyone can be a hacker now, and how 15,000 red teamers are breaking models before they ever ship.

If you want to understand how AI systems actually get broken - and defended - this one's worth your time!

00:00 Intro

01:01 Meet Spencer Whitman (Gray Swan CPO)

02:57 The CMU Research Behind Gray Swan

06:20 The Universal Jailbreak That Broke Every Model

06:59 How Models Learn to Refuse

13:28 Why Open Models Need Guardrails

17:57 AI Security vs. Cybersecurity

23:13 What Reasoning Models Changed

28:57 The Arena: 15,000 Red Teamers

32:16 The Agent That Deleted a Production Database

34:45 Why You Can't Just Patch an AI

38:57 There's No S in MCP

40:27 Securing Agent Protocols

42:56 Nobody Reviews AI Code Anymore

46:03 AI vs. Human Hackers

48:28 The $200K Crypto Bot Heist

53:00 How Meta Pressure-Tested Llama

58:34 Prompt Guard and Code Shield

01:01:45 Can You Trust Chinese Models?

01:07:26 Did Meta Give Up on the Frontier?

01:11:06 What Should Keep CISOs Up at Night

01:14:24 The Open Source Routing Future

01:16:47 Back to On-Prem?

]]>f71e2094-0199-441b-80d1-f928a8378d7dMon, 10 Aug 2026 15:55:49 GMT<p>GPT-5.6 Sol goes rogue and breaches Huggingface. Washington suspends Mythos access within days over national security concerns, and the White House just held an emergency meeting to finalize a classified cybersecurity framework for frontier AI models. AI security went from niche concern to front-page hysteria basically overnight.</p><p></p><p>So I sat down with Spencer Whitman, who recently joined Gray Swan AI as CPO on the back of their $40M Series A. Before Gray Swan, he founded Meta's Llama security team to stop bad actors from jailbreaking their models - he's been on the frontlines of LLM security since the beginning.</p><p></p><p>We get into how Meta pressure-tested Llama for maximum harm before every open source release, why Gray Swan's attack agent has never met an AI system it couldn't break, and the AI Twitter bot that got drained of $200K in crypto in 15 minutes. Spencer also shares his (admittedly speculative) read on whether Meta gave up on the frontier before Alexandr Wang showed up, why anyone can be a hacker now, and how 15,000 red teamers are breaking models before they ever ship.</p><p></p><p>If you want to understand how AI systems actually get broken - and defended - this one's worth your time!<br /><br />00:00 Intro</p><p>01:01 Meet Spencer Whitman (Gray Swan CPO)</p><p>02:57 The CMU Research Behind Gray Swan</p><p>06:20 The Universal Jailbreak That Broke Every Model</p><p>06:59 How Models Learn to Refuse</p><p>13:28 Why Open Models Need Guardrails</p><p>17:57 AI Security vs. Cybersecurity</p><p>23:13 What Reasoning Models Changed</p><p>28:57 The Arena: 15,000 Red Teamers</p><p>32:16 The Agent That Deleted a Production Database</p><p>34:45 Why You Can't Just Patch an AI</p><p>38:57 There's No S in MCP</p><p>40:27 Securing Agent Protocols</p><p>42:56 Nobody Reviews AI Code Anymore</p><p>46:03 AI vs. Human Hackers</p><p>48:28 The $200K Crypto Bot Heist</p><p>53:00 How Meta Pressure-Tested Llama</p><p>58:34 Prompt Guard and Code Shield</p><p>01:01:45 Can You Trust Chinese Models?</p><p>01:07:26 Did Meta Give Up on the Frontier?</p><p>01:11:06 What Should Keep CISOs Up at Night</p><p>01:14:24 The Open Source Routing Future</p><p>01:16:47 Back to On-Prem?</p>no01:18:00Spencer Whitman - Gray Swan AI's $200M Plan to Secure AI Systemsfull<![CDATA[Tony Gentilcore - Glean, the $7.2B Startup Sam Altman Warned Investors About]]>Earlier this year, the "SaaSpocalypse" wiped out something like $2 trillion of SaaS market cap in a matter of weeks — so I sat down with Tony Gentilcore, co-founder of Glean and formerly one of the minds behind Google Search and Chrome, to figure out what's actually happening to software in the agent era.

We get into a lot: why Tony thinks outcome-based pricing (the model Sierra and Decagon are famous for) won't survive, and why companies will drift back toward per-seat. Why the "no Chinese models" rule every enterprise swears by tends to evaporate the moment finance sees the token bill — and why Nemotron, GLM, and Kimi are already good enough to matter. The story behind Sam Altman reportedly telling VCs that if they backed Glean, OpenAI didn't want them as investors (Tony's reaction: "we took it as very flattering").

We also dig into the messier reality of AI at work — how it's saving employees around 11 hours a week while quietly costing them 6 back in what Tony calls "bot sitting and bot shitting," why hard token caps on engineers don't change behavior, how CTOs are blowing through their annual token budgets a quarter into the year, and why the roles of product manager, designer, and engineer are collapsing into one.

If you care about where software, pricing, and enterprise AI are all heading, this one's worth your time.

Chapters:

00:00 Welcome and Setup

00:56 Glean Origin Story

03:08 Enterprise Search Signals

05:07 LLMs Inflection Point

08:08 Early Product Workflow

09:45 Search Evals and Privacy

12:05 From Chatbots to Agents

13:45 Agent Use Cases

17:08 Lessons and Puck Direction

19:08 Bot sitting, bot shitting, and slop

24:44 Token Budgets and ROI

30:59 AI Trends and Moats

35:03 Training and Fine Tuning

38:03 Why Enterprises Fear China Models

39:58 Sovereignty Backlash Watch

40:26 The time Altman called out Glean

41:19 Org Roles Become Builders

44:10 Future Work Voice First

48:17 Vibe Coding vs Quality

52:20 SaaSpocalypse Evolution

54:19 Pricing Tokens Win

55:37 Outcome Pricing Doubts

57:02 Audience AI Review Overload

58:53 Subsidies and Model Choice

01:01:18 Context Layer Interoperability

01:04:52 Build vs. buy: rolling your own Glean

01:07:47 Shadow AI and the coming security mess

01:12:22 Who actually competes

01:13:36 Wrap-up and thanks

]]>ed94cab7-2d15-430b-b6ed-f9823e065cc6Thu, 30 Jul 2026 14:00:00 GMT<p>Earlier this year, the "SaaSpocalypse" wiped out something like $2 trillion of SaaS market cap in a matter of weeks — so I sat down with Tony Gentilcore, co-founder of Glean and formerly one of the minds behind Google Search and Chrome, to figure out what's actually happening to software in the agent era.<br /></p><p>We get into a lot: why Tony thinks outcome-based pricing (the model Sierra and Decagon are famous for) won't survive, and why companies will drift back toward per-seat. Why the "no Chinese models" rule every enterprise swears by tends to evaporate the moment finance sees the token bill — and why Nemotron, GLM, and Kimi are already good enough to matter. The story behind Sam Altman reportedly telling VCs that if they backed Glean, OpenAI didn't want them as investors (Tony's reaction: "we took it as very flattering").<br /></p><p>We also dig into the messier reality of AI at work — how it's saving employees around 11 hours a week while quietly costing them 6 back in what Tony calls "bot sitting and bot shitting," why hard token caps on engineers don't change behavior, how CTOs are blowing through their annual token budgets a quarter into the year, and why the roles of product manager, designer, and engineer are collapsing into one.<br /></p><p>If you care about where software, pricing, and enterprise AI are all heading, this one's worth your time.<br /></p><p>Chapters:</p><p>00:00 Welcome and Setup</p><p>00:56 Glean Origin Story</p><p>03:08 Enterprise Search Signals</p><p>05:07 LLMs Inflection Point</p><p>08:08 Early Product Workflow</p><p>09:45 Search Evals and Privacy</p><p>12:05 From Chatbots to Agents</p><p>13:45 Agent Use Cases</p><p>17:08 Lessons and Puck Direction</p><p>19:08 Bot sitting, bot shitting, and slop</p><p>24:44 Token Budgets and ROI</p><p>30:59 AI Trends and Moats</p><p>35:03 Training and Fine Tuning</p><p>38:03 Why Enterprises Fear China Models</p><p>39:58 Sovereignty Backlash Watch</p><p>40:26 The time Altman called out Glean</p><p>41:19 Org Roles Become Builders</p><p>44:10 Future Work Voice First</p><p>48:17 Vibe Coding vs Quality</p><p>52:20 SaaSpocalypse Evolution</p><p>54:19 Pricing Tokens Win</p><p>55:37 Outcome Pricing Doubts</p><p>57:02 Audience AI Review Overload</p><p>58:53 Subsidies and Model Choice</p><p>01:01:18 Context Layer Interoperability</p><p>01:04:52 Build vs. buy: rolling your own Glean</p><p>01:07:47 Shadow AI and the coming security mess</p><p>01:12:22 Who actually competes</p><p>01:13:36 Wrap-up and thanks</p>no01:14:06Tony Gentilcore - Glean, the $7.2B Startup Sam Altman Warned Investors Aboutfull<![CDATA[Russ Salakhutdinov - Kimi K3 CEO’s PhD Advisor Predicts the Future of AI Agents]]>Kimi K3 took the world by storm last week for open-sourcing frontier level intelligence, so I sat down with Zhilin Yang's (Kimi CEO) PhD advisor Russ Salakhutdinov to talk.

Russ has been everywhere in modern AI. He did his PhD with Geoff Hinton back when neural nets were a punchline, sold his startup to Apple and worked on Project Titan, teaches at Carnegie Mellon, and spent the last couple years at Meta Superintelligence Lab building computer use agents. Now he's the founder of Sooth Labs, building AI that forecasts the future.

We talked about why there's no secret architecture inside the frontier labs and why the real moat is data, engineering, and infrastructure. He explains why Cursor and half the startups you know are quietly running on Chinese open source models, why all the LLMs are going to be commodities, and why the people actually building AGI don't buy the two-year timeline. We get into his time at Meta, why computer use agents still hit 60% when you need 99.9%, whether AI can beat prediction markets, and why the RL environment business isn't sticky. And he makes the case that AI should replace McKinsey, Bain, and BCG.

Chapters:
0:00 - Intro

1:26 - Bumping into Hinton on the street

3:25 - When neural nets were the third choice

5:21 - Generating digits before it was cool

6:59 - AlexNet breaks computer vision

10:18 - Teaching models to describe what they see

12:10 - Early text-to-image (and the toilet seat that beat Google)

16:52 - Hallucination is a feature

19:36 - Selling Perceptual Machines to Apple

22:45 - Self-driving: 0 to 80 in a year, stuck for 5

30:30 - Inside FSD and Waymo's architecture

34:18 - Building Visual Web Arena at CMU

39:07 - Why he joined Meta Superintelligence

40:09 - The agent that plans your faculty job hunt

42:00 - Paying people for their browser history

42:45 - The coupon-hunting agent

43:35 - Why agents still fail

46:14 - 60% when you need 99.9%

47:04 - Agents on your phone

50:12 - No secret architecture at the frontier labs

51:20 - Why coding and math got solved first

54:01 - Models that smell and touch

56:14 - The future of software engineering

59:41 - Founding Sooth Labs

1:00:07 - The 13% graduation prediction

1:05:55 - Why ChatGPT can't forecast

1:08:12 - Agents first, decision systems next

1:09:23 - The Wikipedia contamination story

1:13:43 - Can AI beat prediction markets?

1:16:26 - AI replaces McKinsey

1:18:20 - Why the crowd is hard to beat

1:19:30 - China's open source models rise

1:24:26 - Why the US needs its own open models

1:25:53 - RL environment businesses won't last

1:28:35 - The end of SaaS, LLMs as commodities

1:32:55 - What's next: self-improvement, forecasting, robots

1:35:34 - The only useful robot is the Roomba

1:37:53 - What he'd study in college today

1:41:20 - Adapt or get left behind

1:44:07 - Wrapping up

]]>0b484ab3-56c4-40fe-849b-1a419141b660Mon, 27 Jul 2026 17:21:23 GMT<p>Kimi K3 took the world by storm last week for open-sourcing frontier level intelligence, so I sat down with Zhilin Yang's (Kimi CEO) PhD advisor Russ Salakhutdinov to talk.<br /></p><p>Russ has been everywhere in modern AI. He did his PhD with Geoff Hinton back when neural nets were a punchline, sold his startup to Apple and worked on Project Titan, teaches at Carnegie Mellon, and spent the last couple years at Meta Superintelligence Lab building computer use agents. Now he's the founder of Sooth Labs, building AI that forecasts the future.<br /></p><p>We talked about why there's no secret architecture inside the frontier labs and why the real moat is data, engineering, and infrastructure. He explains why Cursor and half the startups you know are quietly running on Chinese open source models, why all the LLMs are going to be commodities, and why the people actually building AGI don't buy the two-year timeline. We get into his time at Meta, why computer use agents still hit 60% when you need 99.9%, whether AI can beat prediction markets, and why the RL environment business isn't sticky. And he makes the case that AI should replace McKinsey, Bain, and BCG.<br /></p><p>Chapters:<br />0:00 - Intro</p><p>1:26 - Bumping into Hinton on the street</p><p>3:25 - When neural nets were the third choice</p><p>5:21 - Generating digits before it was cool</p><p>6:59 - AlexNet breaks computer vision</p><p>10:18 - Teaching models to describe what they see</p><p>12:10 - Early text-to-image (and the toilet seat that beat Google)</p><p>16:52 - Hallucination is a feature</p><p>19:36 - Selling Perceptual Machines to Apple</p><p>22:45 - Self-driving: 0 to 80 in a year, stuck for 5</p><p>30:30 - Inside FSD and Waymo's architecture</p><p>34:18 - Building Visual Web Arena at CMU</p><p>39:07 - Why he joined Meta Superintelligence</p><p>40:09 - The agent that plans your faculty job hunt</p><p>42:00 - Paying people for their browser history</p><p>42:45 - The coupon-hunting agent</p><p>43:35 - Why agents still fail</p><p>46:14 - 60% when you need 99.9%</p><p>47:04 - Agents on your phone</p><p>50:12 - No secret architecture at the frontier labs</p><p>51:20 - Why coding and math got solved first</p><p>54:01 - Models that smell and touch</p><p>56:14 - The future of software engineering</p><p>59:41 - Founding Sooth Labs</p><p>1:00:07 - The 13% graduation prediction</p><p>1:05:55 - Why ChatGPT can't forecast</p><p>1:08:12 - Agents first, decision systems next</p><p>1:09:23 - The Wikipedia contamination story</p><p>1:13:43 - Can AI beat prediction markets?</p><p>1:16:26 - AI replaces McKinsey</p><p>1:18:20 - Why the crowd is hard to beat</p><p>1:19:30 - China's open source models rise</p><p>1:24:26 - Why the US needs its own open models</p><p>1:25:53 - RL environment businesses won't last</p><p>1:28:35 - The end of SaaS, LLMs as commodities</p><p>1:32:55 - What's next: self-improvement, forecasting, robots</p><p>1:35:34 - The only useful robot is the Roomba</p><p>1:37:53 - What he'd study in college today</p><p>1:41:20 - Adapt or get left behind</p><p>1:44:07 - Wrapping up</p>no01:44:14Russ Salakhutdinov - Kimi K3 CEO’s PhD Advisor Predicts the Future of AI Agentsfull<![CDATA[Andy Hock - How Cerebras Plans to Kill Nvidia]]>Cerebras IPO'd just a couple months ago and has already locked in a 750MW compute deal with OpenAI. Andy Hock's pitch: the GPU is the wrong chip for where AI is going.

I sat down with Andy Hock, Chief Strategy Officer at Cerebras, whose chips are the size of dinner plates instead of postage stamps — which lets them run inference up to 15x faster than even the latest Nvidia GPUs. We got into how that architecture works, the 750MW OpenAI deal (the 2x-faster Codex option runs on Cerebras), and why the memory crisis spiking GPU prices actually benefits Cerebras, since all their memory sits on the chip. Andy also argued the AI buildout isn't a bubble, that "training is a cost center and inference is where you make the big bucks," and why that "95% of enterprise AI pilots fail" stat measured the wrong thing at the wrong time — plus their supercomputer work with governments like the UAE, and why they turned down selling chips to China.

We covered all that and much more. Subscribe for more on AI and the infrastructure behind it, and follow me, Basil Chatha, for everything AI agents.


Chapters

00:00 Intro

01:06 What Cerebras Actually Builds

03:41 Before LLMs Existed

06:09 The GPT Wake-Up Call

10:45 First Principles of AI Compute

15:40 One Giant Chip

18:43 So Why Do We Still Use GPUs?

20:59 Why Inference Took Over

22:27 Where Fast Tokens Win

25:07 Is AI Infra a Bubble?

29:56 The Memory Shortage, Explained

32:35 The Energy Problem

35:34 Rolling Their Own Data Centers

38:20 What a Chip Actually Costs

40:18 AI Designing AI Chips

41:40 The Supply Chain Reality

43:45 Cheap Tokens vs Fast Tokens

46:11 Why Every Millisecond Matters

48:56 Inside the OpenAI Deal

51:56 The Gigawatt Future

54:00 Selling to the Government

58:36 Exporting the US AI Stack

01:03:17 Should We Sell Chips to China?

01:06:00 Why Europe Is Falling Behind

01:07:59 Why Enterprise AI Moves Slow

01:13:05 "I Haven't Read Code in Months"

01:14:32 The Next Cerebras Chip

01:16:50 Closing Thoughts

]]>7795875e-e260-4037-ba42-f0776539731dFri, 17 Jul 2026 14:00:00 GMT<p>Cerebras IPO'd just a couple months ago and has already locked in a 750MW compute deal with OpenAI. Andy Hock's pitch: the GPU is the wrong chip for where AI is going.<br /></p><p>I sat down with Andy Hock, Chief Strategy Officer at Cerebras, whose chips are the size of dinner plates instead of postage stamps — which lets them run inference up to 15x faster than even the latest Nvidia GPUs. We got into how that architecture works, the 750MW OpenAI deal (the 2x-faster Codex option runs on Cerebras), and why the memory crisis spiking GPU prices actually benefits Cerebras, since all their memory sits on the chip. Andy also argued the AI buildout isn't a bubble, that "training is a cost center and inference is where you make the big bucks," and why that "95% of enterprise AI pilots fail" stat measured the wrong thing at the wrong time — plus their supercomputer work with governments like the UAE, and why they turned down selling chips to China.<br /></p><p>We covered all that and much more. Subscribe for more on AI and the infrastructure behind it, and follow me, Basil Chatha, for everything AI agents.</p><p><br />Chapters</p><p>00:00 Intro</p><p>01:06 What Cerebras Actually Builds</p><p>03:41 Before LLMs Existed</p><p>06:09 The GPT Wake-Up Call</p><p>10:45 First Principles of AI Compute</p><p>15:40 One Giant Chip</p><p>18:43 So Why Do We Still Use GPUs?</p><p>20:59 Why Inference Took Over</p><p>22:27 Where Fast Tokens Win</p><p>25:07 Is AI Infra a Bubble?</p><p>29:56 The Memory Shortage, Explained</p><p>32:35 The Energy Problem</p><p>35:34 Rolling Their Own Data Centers</p><p>38:20 What a Chip Actually Costs</p><p>40:18 AI Designing AI Chips</p><p>41:40 The Supply Chain Reality</p><p>43:45 Cheap Tokens vs Fast Tokens</p><p>46:11 Why Every Millisecond Matters</p><p>48:56 Inside the OpenAI Deal</p><p>51:56 The Gigawatt Future</p><p>54:00 Selling to the Government</p><p>58:36 Exporting the US AI Stack</p><p>01:03:17 Should We Sell Chips to China?</p><p>01:06:00 Why Europe Is Falling Behind</p><p>01:07:59 Why Enterprise AI Moves Slow</p><p>01:13:05 "I Haven't Read Code in Months"</p><p>01:14:32 The Next Cerebras Chip</p><p>01:16:50 Closing Thoughts</p>no01:17:22Andy Hock - How Cerebras Plans to Kill Nvidiafull<![CDATA[Leo Mehr - Ramp’s $44B Bet on Services]]>The hardest part of shipping an AI agent isn't the agent. It's getting it access to data buried across a dozen internal systems, and capturing the tribal knowledge that runs a company but was never written down.

I sat down with Leo Mehr, Director of Engineering at Ramp (a $44B-valued company), who runs the forward deployed engineering team that walks into large companies and replaces real, painful workflows with agents. We got into why the model is usually the easy part, and why data access is "the longest pole in the tent." Leo also made the case that nobody wants to buy another piece of software anymore — every B2B company is about to become either an agent-friendly API or white-glove service for everyone, and the middle dies. We talked about whether a services business can actually be venture-scale, why he thinks the AI labs won't eat every startup (it comes down to incentives, not better models), why the biggest customers are often the worst ones to build for, and how the traits that made a great engineer ten years ago aren't the ones that matter now.

We talked about all of that and a lot more. You don't wanna miss this one!

Chapters
00:00 Intro

01:11 Meet Leo Maier

01:46 Building Ramp FDE

02:43 Why FDE Exists

04:23 Early Mandate Fires

06:07 FDE and Core Engineering

07:48 Enterprise Bet Pays

11:10 Ramp Monetization

16:05 Services Are the New Software

18:11 AI Agents and APIs

23:11 Pricing for Outcomes

26:35 Human Labor TAM

32:50 VCs and Rollups

36:50 Common AI Workflow Pain

38:04 Agent Data Context

38:55 Go-to-Market Motion

40:36 Customer Engagement Lifecycle

43:51 Finance Intelligence Layer

47:20 Who You Compete Against

48:09 The Future of Consulting

51:44 Will Humans Still Code?

54:37 How Engineering Teams Will Evolve

57:36 Hiring Change

59:36 Should You Study Computer Science?

01:06:10 The Bull Case and the Risks

01:13:44 Why Ramp

]]>0b2dac2b-c6f2-448f-9f06-2796ce2933f5Fri, 10 Jul 2026 19:46:05 GMT<p>The hardest part of shipping an AI agent isn't the agent. It's getting it access to data buried across a dozen internal systems, and capturing the tribal knowledge that runs a company but was never written down.<br /></p><p>I sat down with Leo Mehr, Director of Engineering at Ramp (a $44B-valued company), who runs the forward deployed engineering team that walks into large companies and replaces real, painful workflows with agents. We got into why the model is usually the easy part, and why data access is "the longest pole in the tent." Leo also made the case that nobody wants to buy another piece of software anymore — every B2B company is about to become either an agent-friendly API or white-glove service for everyone, and the middle dies. We talked about whether a services business can actually be venture-scale, why he thinks the AI labs won't eat every startup (it comes down to incentives, not better models), why the biggest customers are often the worst ones to build for, and how the traits that made a great engineer ten years ago aren't the ones that matter now.<br /></p><p>We talked about all of that and a lot more. You don't wanna miss this one!<br /><br />Chapters<br />00:00 Intro</p><p>01:11 Meet Leo Maier</p><p>01:46 Building Ramp FDE</p><p>02:43 Why FDE Exists</p><p>04:23 Early Mandate Fires</p><p>06:07 FDE and Core Engineering</p><p>07:48 Enterprise Bet Pays</p><p>11:10 Ramp Monetization</p><p>16:05 Services Are the New Software</p><p>18:11 AI Agents and APIs</p><p>23:11 Pricing for Outcomes</p><p>26:35 Human Labor TAM</p><p>32:50 VCs and Rollups</p><p>36:50 Common AI Workflow Pain</p><p>38:04 Agent Data Context</p><p>38:55 Go-to-Market Motion</p><p>40:36 Customer Engagement Lifecycle</p><p>43:51 Finance Intelligence Layer</p><p>47:20 Who You Compete Against</p><p>48:09 The Future of Consulting</p><p>51:44 Will Humans Still Code?</p><p>54:37 How Engineering Teams Will Evolve</p><p>57:36 Hiring Change</p><p>59:36 Should You Study Computer Science?</p><p>01:06:10 The Bull Case and the Risks</p><p>01:13:44 Why Ramp<br /></p>no01:15:32Leo Mehr - Ramp’s $44B Bet on Servicesfull<![CDATA[Siddharth Nanda - Microsoft Engineer Reveals how Engineering Will Never be the Same Again]]>"Software has been solved." Most of the best engineers I know have landed on this exact conclusion.

Last week I sat down with Siddharth Nanda, who went from writing every line of code by hand at Microsoft and Atlassian to shipping 30,000 lines a week with 0% of it written by him.

He's now at Finch, a startup that's raised $20M+ to automate admin work at personal injury law firms, where 100% of the code he writes comes from agents. He's seen big tech before AI and a startup fully running on it, so he has a pretty unique read on where this all goes. We get into why he hasn't written a line of code in 8 months, why middle management is getting hit hardest by the layoffs, and why the productivity studies saying companies are getting slower are right, but only for companies with the wrong people.

We also dig into the tooling itself: Codex vs Claude Code, what harness engineering actually is, running Devin agents in parallel, and why the cost of producing code is approaching zero, and what that means for the kind of engineer who thrives from here.

If you're building with agents, this one's full of hard-won takes from someone doing it every day.

Chapters
00:00 Intro

01:16 What engineering looked like at Microsoft before ChatGPT

03:08 The first AI tools inside big tech (and how gated they were)

05:17 Shipping 30,000 lines a week, none of it written by hand

06:08 Why teams are flattening

08:26 Why managers need to get back to writing code

11:51 How the management layer actually changes

14:07 What happens to PMs and designers

16:38 Why everyone becomes a builder now

20:07 How to actually learn agentic engineering

22:42 Finding new tools on X before anyone else

23:54 AI adoption is showing up in performance reviews

25:45 Building a culture that shares tools daily

27:40 The AI productivity paradox, explained

29:35 Interviewing with agents instead of against them

31:00 Why LeetCode is dead and fundamentals aren't

34:11 Who wins and who loses from here

35:36 Codex vs Claude Code: what he actually uses

36:06 What harness engineering really means

38:21 Evals and benchmarks that matter

39:59 Desktop apps, Devin, and running agents in parallel

43:30 Remote environments and working in monorepos

44:34 Skills, plugins, and hooks

47:57 Parallel worktrees and staying focused

50:42 Agent teams vs subagents

53:06 The Finch mission and wrap up

]]>cd1760c6-139f-4a30-9544-1e65537e9da5Fri, 03 Jul 2026 19:40:16 GMT<p>"Software has been solved." Most of the best engineers I know have landed on this exact conclusion.<br /><br />Last week I sat down with Siddharth Nanda, who went from writing every line of code by hand at Microsoft and Atlassian to shipping 30,000 lines a week with 0% of it written by him. <br /><br />He's now at Finch, a startup that's raised $20M+ to automate admin work at personal injury law firms, where 100% of the code he writes comes from agents. He's seen big tech before AI and a startup fully running on it, so he has a pretty unique read on where this all goes. We get into why he hasn't written a line of code in 8 months, why middle management is getting hit hardest by the layoffs, and why the productivity studies saying companies are getting slower are right, but only for companies with the wrong people.<br /><br />We also dig into the tooling itself: Codex vs Claude Code, what harness engineering actually is, running Devin agents in parallel, and why the cost of producing code is approaching zero, and what that means for the kind of engineer who thrives from here. <br /><br />If you're building with agents, this one's full of hard-won takes from someone doing it every day.<br /><br />Chapters<br />00:00 Intro</p><p>01:16 What engineering looked like at Microsoft before ChatGPT</p><p>03:08 The first AI tools inside big tech (and how gated they were)</p><p>05:17 Shipping 30,000 lines a week, none of it written by hand</p><p>06:08 Why teams are flattening</p><p>08:26 Why managers need to get back to writing code</p><p>11:51 How the management layer actually changes</p><p>14:07 What happens to PMs and designers</p><p>16:38 Why everyone becomes a builder now</p><p>20:07 How to actually learn agentic engineering</p><p>22:42 Finding new tools on X before anyone else</p><p>23:54 AI adoption is showing up in performance reviews</p><p>25:45 Building a culture that shares tools daily</p><p>27:40 The AI productivity paradox, explained</p><p>29:35 Interviewing with agents instead of against them</p><p>31:00 Why LeetCode is dead and fundamentals aren't</p><p>34:11 Who wins and who loses from here</p><p>35:36 Codex vs Claude Code: what he actually uses</p><p>36:06 What harness engineering really means</p><p>38:21 Evals and benchmarks that matter</p><p>39:59 Desktop apps, Devin, and running agents in parallel</p><p>43:30 Remote environments and working in monorepos</p><p>44:34 Skills, plugins, and hooks</p><p>47:57 Parallel worktrees and staying focused</p><p>50:42 Agent teams vs subagents</p><p>53:06 The Finch mission and wrap up</p>no00:54:02Siddharth Nanda - Microsoft Engineer Reveals how Engineering Will Never be the Same Againfull<![CDATA[Voice AI - The Next Frontier | Decagon, Retell, Vapi, Smallest AI, Daily]]>Voice agents are one of the hottest use cases in enterprise right now, but also one of the hardest to actually take live. Getting latency low enough to feel human without dumbing down the responses, making reliable tool calls to a CRM without dropping the customer mid-call, building fallback models for when Anthropic or OpenAI are running hot. None of it is as simple as the demos make it look.

Last week I hosted a fireside chat with five eng leaders who deal with this stuff every day: Basia Sudol (Head of Enterprise Solutions, Decagon), Varun Singh (CPTO, Daily), Steven Diaz (FDE Manager, Vapi), Tyler D'Silva (Founding FDE, Retell AI), and Sudarshan Kamath (Founder, Smallest AI).

We get into why nobody serious is shipping real-time voice-to-voice yet, why LLMs forget the middle of your prompt (and what that does to your architecture), why a giant prompt quietly destroys your unit economics, and why voice agent costs are now being compared directly against human labor.

Plus the stuff nobody warns you about: turn-taking, HIPAA constraints, why outbound is easier than inbound, why getting an exec to actually like the voice can be harder than any model problem, and more!


Chapters below:

00:00 Intro

00:26 Meet the panel

01:23 Daily, WebRTC, and 20 years of building voice

03:49 How Smallest AI made real-time TTS work

05:23 Why Decagon moved into voice

08:16 How Vapi and Retell think about the stack

11:14 Forward deployed vs solutions engineering

16:48 Voice agent architecture, explained simply

22:11 Cascade vs speech-to-speech: the real tradeoff

28:11 Hybrid pipelines and mixing models

32:16 Accents, multilingual, and getting Singlish right

35:32 Prompts vs workflows, and the latency fight

44:26 How you actually evaluate a voice agent

49:04 Simulation-based evals

49:49 What production metrics really look like

51:53 Building a QA framework that scales

54:53 Evaluating speech-to-speech

58:10 Open source benchmarks

59:57 Why picking a voice is so subjective

01:02:03 Personalization and custom voices

01:03:20 Voice quality is solved, GPU efficiency is the new war

01:06:23 Why outbound calls work better than you'd think

01:10:38 Deploying in regulated industries (HIPAA, retention, audits)

01:12:43 Turn-taking, the hardest unsolved problem in voice

01:18:53 Where voice agents go in the next year

01:27:55 Audience Q&A: inside Smallest's Hydra model

01:32:15 The deployment problems nobody has solved yet

01:37:46 Closing thoughts and thanks

]]>74310c6b-5e84-4e3f-ba58-caf2620dcf15Tue, 30 Jun 2026 17:30:00 GMT<p>Voice agents are one of the hottest use cases in enterprise right now, but also one of the hardest to actually take live. Getting latency low enough to feel human without dumbing down the responses, making reliable tool calls to a CRM without dropping the customer mid-call, building fallback models for when Anthropic or OpenAI are running hot. None of it is as simple as the demos make it look.<br /><br />Last week I hosted a fireside chat with five eng leaders who deal with this stuff every day: Basia Sudol (Head of Enterprise Solutions, Decagon), Varun Singh (CPTO, Daily), Steven Diaz (FDE Manager, Vapi), Tyler D'Silva (Founding FDE, Retell AI), and Sudarshan Kamath (Founder, Smallest AI).<br /></p><p>We get into why nobody serious is shipping real-time voice-to-voice yet, why LLMs forget the middle of your prompt (and what that does to your architecture), why a giant prompt quietly destroys your unit economics, and why voice agent costs are now being compared directly against human labor.<br /></p><p>Plus the stuff nobody warns you about: turn-taking, HIPAA constraints, why outbound is easier than inbound, why getting an exec to actually like the voice can be harder than any model problem, and more!</p><p><br />Chapters below:</p><p>00:00 Intro</p><p>00:26 Meet the panel</p><p>01:23 Daily, WebRTC, and 20 years of building voice</p><p>03:49 How Smallest AI made real-time TTS work</p><p>05:23 Why Decagon moved into voice</p><p>08:16 How Vapi and Retell think about the stack</p><p>11:14 Forward deployed vs solutions engineering</p><p>16:48 Voice agent architecture, explained simply</p><p>22:11 Cascade vs speech-to-speech: the real tradeoff</p><p>28:11 Hybrid pipelines and mixing models</p><p>32:16 Accents, multilingual, and getting Singlish right</p><p>35:32 Prompts vs workflows, and the latency fight</p><p>44:26 How you actually evaluate a voice agent</p><p>49:04 Simulation-based evals</p><p>49:49 What production metrics really look like</p><p>51:53 Building a QA framework that scales</p><p>54:53 Evaluating speech-to-speech</p><p>58:10 Open source benchmarks</p><p>59:57 Why picking a voice is so subjective</p><p>01:02:03 Personalization and custom voices</p><p>01:03:20 Voice quality is solved, GPU efficiency is the new war</p><p>01:06:23 Why outbound calls work better than you'd think</p><p>01:10:38 Deploying in regulated industries (HIPAA, retention, audits)</p><p>01:12:43 Turn-taking, the hardest unsolved problem in voice</p><p>01:18:53 Where voice agents go in the next year</p><p>01:27:55 Audience Q&amp;A: inside Smallest's Hydra model</p><p>01:32:15 The deployment problems nobody has solved yet</p><p>01:37:46 Closing thoughts and thanks</p>no01:38:22Voice AI - The Next Frontier | Decagon, Retell, Vapi, Smallest AI, Dailyfull<![CDATA[AI Agents in the Enterprise | Sierra, Mercor, Intercom, Turing | $2.8B+ Raised]]>(We know the audio quality isn't great on this one :( But the conversation is still well worth it!)

Last week I hosted a fireside chat on what it actually takes to build AI agents in the enterprise with Natalie Meurer (Head of Agent Eng, Sierra), Harsh Trivedi (founding engineer, Mercor), Juhi Parekh (GM, Turing), and Kevin Lynch (Senior FDE, Fin).

We get into why new models aren't always better (and why you can't just swap in the latest release and assume your agent improves), how the data-labeling/RL environment business might only have a couple years left, why real-time voice-to-voice models still aren't production-ready, how cheaper inference is still causing prices to go up, how baking in a constellation of models into enterprise agents is so important for reliability,

and much, much, more!

Chapters below

00:00 Intro

00:21 Meet the panel

01:57 What everyone's actually using agents for day to day

06:10 The reality of forward deployed work

09:28 What agents couldn't do a year ago that they can now

12:29 Why you have to tell agents what NOT to do

16:26 What a harness actually is

22:03 RL environments explained

28:46 Does the data-labeling and RL environment business even last?

37:02 Why benchmarks don't tell you what works in production

38:12 Agent engineering vs forward deployed engineering

41:38 Deploying into 100-year-old enterprise systems

44:25 Why AI adoption is an org problem, not a tech problem

45:36 Hiring for judgment when engineers aren't really coding anymore

48:16 Why agents are a new kind of software

50:46 The first 90 days of an enterprise deployment

53:20 Why compliance environments break normal testing

56:58 Layering AI on AI to get to 99% accuracy

01:00:56 New models aren't always better — the swap problem

01:02:35 Improving agents without waiting for a new model

01:06:36 Does agent performance secretly degrade over time?

01:09:40 Why one model is never enough: the constellation approach

01:11:14 Building resilience when inference providers go down

01:13:48 When fine-tuning actually makes sense

01:14:53 Why voice-to-voice still isn't production-ready

01:16:25 The cascaded pipeline that real voice agents use

01:21:45 Audience Q&A: managing change inside the enterprise

01:23:24 Why inference getting cheaper makes things more expensive

01:26:54 Charging for outcomes instead of conversations

01:30:19 What the real moat is when everyone uses the same models

01:37:04 Synthetic data and where the data wall actually is

01:38:50 Closing thoughts

]]>ec339dea-4f84-4810-8842-91dba9eaca3cTue, 23 Jun 2026 20:28:44 GMT<p>(We know the audio quality isn't great on this one :( But the conversation is still well worth it!)<br /></p><p>Last week I hosted a fireside chat on what it actually takes to build AI agents in the enterprise with Natalie Meurer (Head of Agent Eng, Sierra), Harsh Trivedi (founding engineer, Mercor), Juhi Parekh (GM, Turing), and Kevin Lynch (Senior FDE, Fin).<br /></p><p>We get into why new models aren't always better (and why you can't just swap in the latest release and assume your agent improves), how the data-labeling/RL environment business might only have a couple years left, why real-time voice-to-voice models still aren't production-ready, how cheaper inference is still causing prices to go up, how baking in a constellation of models into enterprise agents is so important for reliability,</p><p>and much, much, more!<br /></p><p>Chapters below</p><p>00:00 Intro</p><p>00:21 Meet the panel</p><p>01:57 What everyone's actually using agents for day to day</p><p>06:10 The reality of forward deployed work</p><p>09:28 What agents couldn't do a year ago that they can now</p><p>12:29 Why you have to tell agents what NOT to do</p><p>16:26 What a harness actually is</p><p>22:03 RL environments explained</p><p>28:46 Does the data-labeling and RL environment business even last?</p><p>37:02 Why benchmarks don't tell you what works in production</p><p>38:12 Agent engineering vs forward deployed engineering</p><p>41:38 Deploying into 100-year-old enterprise systems</p><p>44:25 Why AI adoption is an org problem, not a tech problem</p><p>45:36 Hiring for judgment when engineers aren't really coding anymore</p><p>48:16 Why agents are a new kind of software</p><p>50:46 The first 90 days of an enterprise deployment</p><p>53:20 Why compliance environments break normal testing</p><p>56:58 Layering AI on AI to get to 99% accuracy</p><p>01:00:56 New models aren't always better — the swap problem</p><p>01:02:35 Improving agents without waiting for a new model</p><p>01:06:36 Does agent performance secretly degrade over time?</p><p>01:09:40 Why one model is never enough: the constellation approach</p><p>01:11:14 Building resilience when inference providers go down</p><p>01:13:48 When fine-tuning actually makes sense</p><p>01:14:53 Why voice-to-voice still isn't production-ready</p><p>01:16:25 The cascaded pipeline that real voice agents use</p><p>01:21:45 Audience Q&amp;A: managing change inside the enterprise</p><p>01:23:24 Why inference getting cheaper makes things more expensive</p><p>01:26:54 Charging for outcomes instead of conversations</p><p>01:30:19 What the real moat is when everyone uses the same models</p><p>01:37:04 Synthetic data and where the data wall actually is</p><p>01:38:50 Closing thoughts</p>no01:41:36AI Agents in the Enterprise | Sierra, Mercor, Intercom, Turing | $2.8B+ Raisedfull<![CDATA[Vince Signori: Inside LangChain's Growth Strategy from $200M to $1.25B]]>Today's episode is with Vince Signori, Sales Director at LangChain and one of the first sales hires at HashiCorp, where he watched the company grow from a small startup all the way to an IPO and get acquired by IBM.

He sees the exact same shift happening now with AI agents that happened with cloud back then, except 50x faster. And he's got the numbers to back it up — LangChain is downloaded more than the OpenAI SDK, and 45% of the Fortune 500 are now paying customers.

We get into how companies like Toyota and Home Depot are actually using AI agents in production today, why enterprises are building their own private versions of ChatGPT to own their data, and why memory is becoming the most valuable asset in AI.

We also talk about what it actually takes to get an agent from prototype to production, why selling open source is the hardest sale in software, and how Vince went from 3 reps doing 20-hour days to running the number one sales region at one of the fastest growing companies in AI.

You don't wanna miss this one.

Chapters:

00:00 Intro

01:38 From HashiCorp to LangChain

03:20 Cloud wave vs AI agent wave

05:30 Open source vs enterprise

06:23 How LangChain's product stack evolved

07:37 Why agents are finally in production

08:57 What companies build with LangGraph

10:12 LangSmith and Engine explained

13:23 The Agent Development Lifecycle

16:36 Build vs buy on voice agents

19:48 Why owning your data and memory layer matters

22:12 How open source users become paying customers

26:49 Why LangChain hired forward deployed engineers

31:34 What go-to-market looked like with 3 reps

35:31 From 39 employees to hypergrowth

36:37 Transitioning away from founder-led sales

37:07 Why the CEO joined every early call

37:38 The sandwich sale strategy explained

39:28 Signals that an open source user is ready to buy

41:08 Why outbound controls the narrative in enterprise

43:33 Why in-person selling still wins

45:57 Building an internal GTM agent to scale

47:37 What the GTM agent actually does

50:16 Why AI moves 50x faster than cloud did

53:46 The vendor consolidation wave that's coming

56:14 How to win the platform standardization deal

59:41 Why staying model-agnostic beats the hyperscalers

01:01:23 How Vince onboards new reps today

01:04:19 The sales and engineering feedback loop

01:09:49 Signals an account is ready to expand

01:12:57 How to prove early value before full commitment

01:14:40 How enterprises actually measure agent ROI

01:16:27 Why automation is expanding beyond support

01:17:59 Which industries are adopting agents fastest

01:19:07 Healthcare agent use cases live today

01:21:17 AI agents in finance and payments

01:22:40 The Visa partnership

01:25:17 What it takes to scale a sales team right now

01:27:44 How the GTM agent is changing the SDR role

01:32:06 Why the human element in sales still matters

01:34:28 Platform deals vs point solutions

01:36:43 Vince's predictions on memory and consolidation

01:39:23 How Engine helps teams iterate on agents faster

01:43:19 Forward deployed engineers vs Engine

01:46:20 Where to find Vince and LangChain's open roles

]]>39dbd0f4-0495-4ea5-bafc-fb0373ed94acWed, 10 Jun 2026 17:15:45 GMT<p>Today's episode is with Vince Signori, Sales Director at LangChain and one of the first sales hires at HashiCorp, where he watched the company grow from a small startup all the way to an IPO and get acquired by IBM.<br /></p><p>He sees the exact same shift happening now with AI agents that happened with cloud back then, except 50x faster. And he's got the numbers to back it up — LangChain is downloaded more than the OpenAI SDK, and 45% of the Fortune 500 are now paying customers.<br /></p><p>We get into how companies like Toyota and Home Depot are actually using AI agents in production today, why enterprises are building their own private versions of ChatGPT to own their data, and why memory is becoming the most valuable asset in AI.<br /></p><p>We also talk about what it actually takes to get an agent from prototype to production, why selling open source is the hardest sale in software, and how Vince went from 3 reps doing 20-hour days to running the number one sales region at one of the fastest growing companies in AI.<br /></p><p>You don't wanna miss this one.<br /></p><p>Chapters:</p><p>00:00 Intro</p><p>01:38 From HashiCorp to LangChain</p><p>03:20 Cloud wave vs AI agent wave</p><p>05:30 Open source vs enterprise</p><p>06:23 How LangChain's product stack evolved</p><p>07:37 Why agents are finally in production</p><p>08:57 What companies build with LangGraph</p><p>10:12 LangSmith and Engine explained</p><p>13:23 The Agent Development Lifecycle</p><p>16:36 Build vs buy on voice agents</p><p>19:48 Why owning your data and memory layer matters</p><p>22:12 How open source users become paying customers</p><p>26:49 Why LangChain hired forward deployed engineers</p><p>31:34 What go-to-market looked like with 3 reps</p><p>35:31 From 39 employees to hypergrowth</p><p>36:37 Transitioning away from founder-led sales</p><p>37:07 Why the CEO joined every early call</p><p>37:38 The sandwich sale strategy explained</p><p>39:28 Signals that an open source user is ready to buy</p><p>41:08 Why outbound controls the narrative in enterprise</p><p>43:33 Why in-person selling still wins</p><p>45:57 Building an internal GTM agent to scale</p><p>47:37 What the GTM agent actually does</p><p>50:16 Why AI moves 50x faster than cloud did</p><p>53:46 The vendor consolidation wave that's coming</p><p>56:14 How to win the platform standardization deal</p><p>59:41 Why staying model-agnostic beats the hyperscalers</p><p>01:01:23 How Vince onboards new reps today</p><p>01:04:19 The sales and engineering feedback loop</p><p>01:09:49 Signals an account is ready to expand</p><p>01:12:57 How to prove early value before full commitment</p><p>01:14:40 How enterprises actually measure agent ROI</p><p>01:16:27 Why automation is expanding beyond support</p><p>01:17:59 Which industries are adopting agents fastest</p><p>01:19:07 Healthcare agent use cases live today</p><p>01:21:17 AI agents in finance and payments</p><p>01:22:40 The Visa partnership</p><p>01:25:17 What it takes to scale a sales team right now</p><p>01:27:44 How the GTM agent is changing the SDR role</p><p>01:32:06 Why the human element in sales still matters</p><p>01:34:28 Platform deals vs point solutions</p><p>01:36:43 Vince's predictions on memory and consolidation</p><p>01:39:23 How Engine helps teams iterate on agents faster</p><p>01:43:19 Forward deployed engineers vs Engine</p><p>01:46:20 Where to find Vince and LangChain's open roles</p>no01:46:37Vince Signori: Inside LangChain's Growth Strategy from $200M to $1.25Bfull<![CDATA[Shrivu Shankar - How a $5B Cybersecurity Company Runs on AI Agents ]]>Today's episode is with Shrivu Shankar, VP of AI Strategy at Abnormal AI - a $5B cybersecurity company. What makes this one unique is that Shrivu joined as an intern in 2021 and got promoted every single year until he reached VP, so he's basically watched AI go from a niche engineering tool to something that's reshaping entire companies from the inside.


We get into how AI is catching cyberattacks so sophisticated that even humans can't tell they're fake, how engineers at a $5B company have basically stopped writing code themselves, and what that means for everyone else on the team.

We also go deep on why context engineering is replacing prompt engineering as the real moat, how they used GPT-3 with zero safety guardrails to generate fake phishing attacks as training data, and what it actually takes to become an AI native company at 1,500 people.


One of the most technical and eye-opening conversations I've had. You don't wanna miss this one.

Chapters:

00:00 Intro

00:58 Who is Shrivu and what is Abnormal AI

01:39 Why cybersecurity and machine learning

03:13 Intern to VP in 4 years — how it actually happened

05:44 What Abnormal AI does and how it started

09:10 The vendor fraud attack so convincing the victim didn't believe it was real

10:45 What GPT-3 changed for cybersecurity

13:01 Using synthetic data to train models — and how they measured it

16:49 How a 1,500 person company actually adopts AI internally

19:50 How engineering, PM, and platform roles are changing right now

23:17 The biggest AI misconception Shrivu keeps hearing

27:35 What Shrivu's day actually looks like as VP of AI Strategy

28:53 Engineers stopped writing code. Here's what they do instead.

32:28 Why product teams are getting much smaller

34:31 Why context engineering beats prompt engineering

36:31 Spec-driven development and how Nora Tech Plan works

39:14 How to scale context engineering across an entire eng org

40:30 What the manager role looks like in the agent era

42:17 What skills actually matter for managers now

43:29 AI is making orgs flatter. Is that a good thing?

45:08 How the C-suite is getting closer to the work

46:41 What agents actually are and how tool calling works

48:05 How agents improved Abnormal's detection pipeline

50:56 The AI phishing coach — how it works and why it matters

53:30 The internal AI data analyst agent

56:13 Dozens of internal agents — the ones Shrivu is most proud of

57:15 Where agents fail (it's usually not the model)

58:52 What Shrivu would tell a CEO just starting with agents

01:00:19 Sending sensitive security data to LLMs — how they handle it

01:01:47 What becoming AI native actually means in practice

01:03:37 What most people still get wrong about AI in the enterprise

01:04:31 How to write documents with AI without it sounding like AI

01:06:40 Claude Code vs Codex — which one and why

01:09:27 How Shrivu stays ahead and his take on MCPs

01:11:33 How the team uses Claude Code skills

01:12:47 Using hooks for shift-left validation in large codebases

01:13:42 How to manage context in a massive monorepo

01:14:56 Building tool-agnostic rules across Claude, Cursor, and Code Rabbit

01:16:55 Why infra teams are becoming agent harness teams

01:17:57 Wrap up

]]>f84a2456-1a4b-4d3e-b4a0-b8bad31c0cd9Mon, 11 May 2026 15:00:00 GMT<p>Today's episode is with Shrivu Shankar, VP of AI Strategy at Abnormal AI - a $5B cybersecurity company. What makes this one unique is that Shrivu joined as an intern in 2021 and got promoted every single year until he reached VP, so he's basically watched AI go from a niche engineering tool to something that's reshaping entire companies from the inside.</p><p><br />We get into how AI is catching cyberattacks so sophisticated that even humans can't tell they're fake, how engineers at a $5B company have basically stopped writing code themselves, and what that means for everyone else on the team.<br /></p><p>We also go deep on why context engineering is replacing prompt engineering as the real moat, how they used GPT-3 with zero safety guardrails to generate fake phishing attacks as training data, and what it actually takes to become an AI native company at 1,500 people.</p><p><br />One of the most technical and eye-opening conversations I've had. You don't wanna miss this one.<br /></p><p>Chapters:</p><p>00:00 Intro</p><p>00:58 Who is Shrivu and what is Abnormal AI</p><p>01:39 Why cybersecurity and machine learning</p><p>03:13 Intern to VP in 4 years — how it actually happened</p><p>05:44 What Abnormal AI does and how it started</p><p>09:10 The vendor fraud attack so convincing the victim didn't believe it was real</p><p>10:45 What GPT-3 changed for cybersecurity</p><p>13:01 Using synthetic data to train models — and how they measured it</p><p>16:49 How a 1,500 person company actually adopts AI internally</p><p>19:50 How engineering, PM, and platform roles are changing right now</p><p>23:17 The biggest AI misconception Shrivu keeps hearing</p><p>27:35 What Shrivu's day actually looks like as VP of AI Strategy</p><p>28:53 Engineers stopped writing code. Here's what they do instead.</p><p>32:28 Why product teams are getting much smaller</p><p>34:31 Why context engineering beats prompt engineering</p><p>36:31 Spec-driven development and how Nora Tech Plan works</p><p>39:14 How to scale context engineering across an entire eng org</p><p>40:30 What the manager role looks like in the agent era</p><p>42:17 What skills actually matter for managers now</p><p>43:29 AI is making orgs flatter. Is that a good thing?</p><p>45:08 How the C-suite is getting closer to the work</p><p>46:41 What agents actually are and how tool calling works</p><p>48:05 How agents improved Abnormal's detection pipeline</p><p>50:56 The AI phishing coach — how it works and why it matters</p><p>53:30 The internal AI data analyst agent</p><p>56:13 Dozens of internal agents — the ones Shrivu is most proud of</p><p>57:15 Where agents fail (it's usually not the model)</p><p>58:52 What Shrivu would tell a CEO just starting with agents</p><p>01:00:19 Sending sensitive security data to LLMs — how they handle it</p><p>01:01:47 What becoming AI native actually means in practice</p><p>01:03:37 What most people still get wrong about AI in the enterprise</p><p>01:04:31 How to write documents with AI without it sounding like AI</p><p>01:06:40 Claude Code vs Codex — which one and why</p><p>01:09:27 How Shrivu stays ahead and his take on MCPs</p><p>01:11:33 How the team uses Claude Code skills</p><p>01:12:47 Using hooks for shift-left validation in large codebases</p><p>01:13:42 How to manage context in a massive monorepo</p><p>01:14:56 Building tool-agnostic rules across Claude, Cursor, and Code Rabbit</p><p>01:16:55 Why infra teams are becoming agent harness teams</p><p>01:17:57 Wrap up</p>no01:18:12Shrivu Shankar - How a $5B Cybersecurity Company Runs on AI Agents full<![CDATA[Supriya Gupta - Meta Exec Explains How AI is Reshaping Advertising]]>Today's episode is with Supriya Gupta, ex-VP of Product at Intuit Credit Karma and former Product Lead on the Ads team at Meta when that business was scaling like crazy.

She brings a really unique perspective on everything happening with GenAI right now because she's seen it from the inside at two of the biggest companies in tech.

We get into how ads are slowly going to be generated on the spot, per user, where the image, copy and offer will all be unique to you specifically. We talk about why you can't just infinitely scale ad testing even though it's now theoretically possible, and why flooding the internet with AI content might actually be the worst thing you can do for your brand.

We also go deep on what actually happened inside Credit Karma when they started building with GenAI, including what moved the needle and what didn't, and what that means for designers, PMs, and content teams everywhere.

And at the end, Supriya shares why she walked away from her VP role to start her own company, and what she's building now.

You don't wanna miss this one.

🔗 Find Supriya on LinkedIn: https://www.linkedin.com/in/supriyag/

🌐 Company's website: https://www.helloeve.co/

⏱ Chapters
00:00 Intro

01:27 Predictive AI vs. Generative AI: what actually changed

02:59 The next gen of dynamic ads — why every user could soon get a unique ad

06:52 Will content agencies survive the AI era?

09:17 Why you can't just test a million ad variants (the stats problem)

12:55 AI slop, UGC backlash, and should AI content be labeled?

17:02 Supriya joins Credit Karma: building the Lightbox targeting platform

21:23 Building the Credit Karma financial assistant

24:35 Handling hallucinations in a finance app at scale

27:26 What happened to content designers when AI started writing copy

29:56 Why AI copy still needs human taste and judgment

31:01 PMs are prototyping now — what that means for design and eng

33:30 Will there be fewer PMs? (Probably not — here's why)

35:33 Why Supriya left Credit Karma to start her own company

37:23 The principles she built her startup around

39:20 The rise of the "super IC" — managers becoming AI-powered operators

41:54 Why most AI projects fail before they even start

45:33 Enterprises vs. startups: how they approach AI differently

48:13 What a successful AI deployment actually looks like

51:02 Will everyone need to upskill on AI? (Spoiler: maybe not)

52:41 Most execs are thinking about cost-cutting — the smarter ones aren't

53:49 The executive digital twin: Supriya's startup vision

56:51 Current product, roadmap, and what's shipping next

58:52 Where to find Supriya

]]>6c2590e9-daf0-4e23-be8d-dd1a93e63f3fMon, 04 May 2026 15:00:00 GMT<p>Today's episode is with Supriya Gupta, ex-VP of Product at Intuit Credit Karma and former Product Lead on the Ads team at Meta when that business was scaling like crazy.<br /></p><p>She brings a really unique perspective on everything happening with GenAI right now because she's seen it from the inside at two of the biggest companies in tech.<br /></p><p>We get into how ads are slowly going to be generated on the spot, per user, where the image, copy and offer will all be unique to you specifically. We talk about why you can't just infinitely scale ad testing even though it's now theoretically possible, and why flooding the internet with AI content might actually be the worst thing you can do for your brand.<br /></p><p>We also go deep on what actually happened inside Credit Karma when they started building with GenAI, including what moved the needle and what didn't, and what that means for designers, PMs, and content teams everywhere.</p><p>And at the end, Supriya shares why she walked away from her VP role to start her own company, and what she's building now.<br /></p><p>You don't wanna miss this one.<br /><br /></p><p>🔗 Find Supriya on LinkedIn: <a rel="noopener noreferrer nofollow" href="https://www.linkedin.com/in/supriyag/" target="_blank">https://www.linkedin.com/in/supriyag/</a></p><p>🌐 Company's website: <a rel="noopener noreferrer nofollow" href="https://www.helloeve.co/" target="_blank">https://www.helloeve.co/</a><br /><br /><i>⏱ Chapters</i><br />00:00 Intro</p><p>01:27 Predictive AI vs. Generative AI: what actually changed</p><p>02:59 The next gen of dynamic ads — why every user could soon get a unique ad</p><p>06:52 Will content agencies survive the AI era?</p><p>09:17 Why you can't just test a million ad variants (the stats problem)</p><p>12:55 AI slop, UGC backlash, and should AI content be labeled?</p><p>17:02 Supriya joins Credit Karma: building the Lightbox targeting platform</p><p>21:23 Building the Credit Karma financial assistant</p><p>24:35 Handling hallucinations in a finance app at scale</p><p>27:26 What happened to content designers when AI started writing copy</p><p>29:56 Why AI copy still needs human taste and judgment</p><p>31:01 PMs are prototyping now — what that means for design and eng</p><p>33:30 Will there be fewer PMs? (Probably not — here's why)</p><p>35:33 Why Supriya left Credit Karma to start her own company</p><p>37:23 The principles she built her startup around</p><p>39:20 The rise of the "super IC" — managers becoming AI-powered operators</p><p>41:54 Why most AI projects fail before they even start</p><p>45:33 Enterprises vs. startups: how they approach AI differently</p><p>48:13 What a successful AI deployment actually looks like</p><p>51:02 Will everyone need to upskill on AI? (Spoiler: maybe not)</p><p>52:41 Most execs are thinking about cost-cutting — the smarter ones aren't</p><p>53:49 The executive digital twin: Supriya's startup vision</p><p>56:51 Current product, roadmap, and what's shipping next</p><p>58:52 Where to find Supriya</p>no00:56:34Supriya Gupta - Meta Exec Explains How AI is Reshaping Advertisingfull<![CDATA[The Future of Agentic Engineering | Cognition (Devin), Semgrep, Factory & Composio | $1.2B+ Raised]]>AI agents are everywhere right now. But are they actually working inside real engineering teams?

At AngelList’s Founders Cafe, I sat down with founders of Cognition ($898M raised), Semgrep ($204M raised), Factory ($70M raised), and Composio ($29M raised) to talk about what agentic engineering looks like in practice.

There’s a lot of hype around AI coding tools, but the reality is more nuanced. Some teams are moving 10x faster, others are slowing down. A big part of it comes down to whether your codebase is actually “agent-ready” (linting, type systems, guardrails, etc).

We also went deep on security, which is one of the biggest gaps right now. As more non-developers start “vibe coding,” the risk surface grows fast. We talked about MCP access control, layered security, and why you can’t rely on models alone to generate secure code.

Enterprise teams are also dealing with the operational side of this shift. They have to manage cost, run evals, and help thousands of engineers use these systems well. Tools like PR review agents, model routing, and internal orchestration are quickly becoming part of the stack.

⏱ Chapters

00:00 Welcome and setup

00:41 Panel introductions

02:00 Windsurf acquisition

03:52 Do agents boost productivity?

04:16 Agent-ready codebases

05:55 Real-world enterprise wins

07:28 Agents building integrations

09:41 Security risks (vibe coding)

11:04 LLM security tools landscape

12:21 Defense-in-depth

15:16 MCP security pitfalls

18:41 MCP vs CLI

23:08 RL for secure code

27:05 Auto research missions

27:57 Training your own models

31:33 Distillation and IP decay

34:50 Hybrid systems

37:14 Side projects vs enterprise

40:16 Forward deployed engineering

40:58 Agent orchestration

43:08 Cost controls

43:49 Auto model routing

45:52 Guardrails

47:02 Legacy code risks

48:16 Model poisoning

49:35 What is a harness?

51:47 Why build your own

52:58 Continuous learning loops

56:30 Security workflows

57:37 Validation and meta engineering

1:00:32 Running evals in practice

1:03:36 Teams reshaped by agents

1:12:17 Should you study CS?

1:14:11 Enterprise adoption

1:18:23 Local to cloud journey

1:20:51 Agent economy and prompting

1:22:22 Spec vs plan

1:25:12 Closing and thanks

]]>990c2128-cd14-4eac-adad-9897a588d009Wed, 15 Apr 2026 19:00:00 GMT<p>AI agents are everywhere right now. But are they actually working inside real engineering teams?<br /></p><p>At AngelList’s Founders Cafe, I sat down with founders of Cognition (<b>$898M</b> raised), Semgrep (<b>$204M</b> raised), Factory (<b>$70M</b> raised), and Composio (<b>$29M</b> raised) to talk about what agentic engineering looks like in practice.<br /></p><p>There’s a lot of hype around AI coding tools, but the reality is more nuanced. Some teams are moving 10x faster, others are slowing down. A big part of it comes down to whether your codebase is actually “agent-ready” (linting, type systems, guardrails, etc).<br /></p><p>We also went deep on security, which is one of the biggest gaps right now. As more non-developers start “vibe coding,” the risk surface grows fast. We talked about MCP access control, layered security, and why you can’t rely on models alone to generate secure code.<br /></p><p>Enterprise teams are also dealing with the operational side of this shift. They have to manage cost, run evals, and help thousands of engineers use these systems well. Tools like PR review agents, model routing, and internal orchestration are quickly becoming part of the stack.<br /></p><p><b>⏱ Chapters</b></p><p>00:00 Welcome and setup</p><p>00:41 Panel introductions</p><p>02:00 Windsurf acquisition</p><p>03:52 Do agents boost productivity?</p><p>04:16 Agent-ready codebases</p><p>05:55 Real-world enterprise wins</p><p>07:28 Agents building integrations</p><p>09:41 Security risks (vibe coding)</p><p>11:04 LLM security tools landscape</p><p>12:21 Defense-in-depth</p><p>15:16 MCP security pitfalls</p><p>18:41 MCP vs CLI</p><p>23:08 RL for secure code</p><p>27:05 Auto research missions</p><p>27:57 Training your own models</p><p>31:33 Distillation and IP decay</p><p>34:50 Hybrid systems</p><p>37:14 Side projects vs enterprise</p><p>40:16 Forward deployed engineering</p><p>40:58 Agent orchestration</p><p>43:08 Cost controls</p><p>43:49 Auto model routing</p><p>45:52 Guardrails</p><p>47:02 Legacy code risks</p><p>48:16 Model poisoning</p><p>49:35 What is a harness?</p><p>51:47 Why build your own</p><p>52:58 Continuous learning loops</p><p>56:30 Security workflows</p><p>57:37 Validation and meta engineering</p><p>1:00:32 Running evals in practice</p><p>1:03:36 Teams reshaped by agents</p><p>1:12:17 Should you study CS?</p><p>1:14:11 Enterprise adoption</p><p>1:18:23 Local to cloud journey</p><p>1:20:51 Agent economy and prompting</p><p>1:22:22 Spec vs plan</p><p>1:25:12 Closing and thanks</p>no01:25:384The Future of Agentic Engineering | Cognition (Devin), Semgrep, Factory & Composio | $1.2B+ Raisedfull<![CDATA[He Built a $200M AI Agent 10 Years Before ChatGPT]]>Summary:

In this conversation, I talked to Ashish Shubham (VP of Engineering), who's been at ThoughtSpot for 10 years, about AI agents in enterprise analytics. ThoughtSpot started as a search-based analytics company trying to make data accessible to regular business users. In 2019, they tried building natural language interfaces using BERT, but only hit about 50% accuracy. For a product where enterprise customers make billion-dollar decisions, that wasn't good enough. They shelved the project.

When ChatGPT came out, ThoughtSpot was ready. Ashish walked me through how they pivoted: they built a 25-30 person team, decided to use prompting instead of fine-tuning, and leveraged their existing semantic data modeling layer to get accuracy into the high 90s. We got into the technical evolution from monolithic systems to agent architectures with tools, how they went from manual human judges to using LLMs to evaluate their outputs, and how enterprise security requirements shaped what they built.

We also talked about how software engineering is changing. Ashish said 50-60% of his code is AI-generated now, and he thinks system design is becoming the critical skill, even for junior engineers. He had an interesting take on the "95% of AI deployments fail" stat too.

Chapters:

0:00 Intro and Ashish's journey to ThoughtSpot from GoDaddy

0:13 ThoughtSpot's mission to democratize data analytics for business users

1:26 Early search-based analytics before natural language processing

2:36 ThoughtSpot vs Tableau and the promise of self-service analytics

4:40 The analyst bottleneck problem and how ThoughtSpot aimed to solve it

5:49 Early technical challenges with in-memory databases and data migration

8:11 Semantic data models, joins, and creating abstraction layers for users

11:39 Who builds the data models and the role of analysts

12:22 Pre-LLM natural language processing using BERT and word2vec in 2018-2019

14:43 The accuracy problem and ambiguity in translating user queries

16:58 Trust challenges and why the early NLP product never became core

19:59 Competition with Tableau, Looker, and Power BI

22:44 How analyst roles changed with self-service analytics tools

25:30 The ChatGPT moment and pivoting to LLM-powered natural language

27:48 Early prompt engineering days and generating SQL with LLMs

31:09 Training vs prompting debate and why fine-tuning was eventually abandoned

34:28 Organizational changes and building the NLS team

37:16 Coaching systems for company-specific terminology vs training models

39:02 Evolution of evaluation methods from human judges to LLM-as-judge

43:23 Moving to LangFuse and GCP for agent infrastructure

46:29 How LLM context windows and capabilities evolved their product

50:07 From 30-column limits to agentic systems with 90%+ accuracy

52:52 RAG, column selection, and using proprietary data indexes

54:59 Multi-model support and enterprise data security concerns

59:14 How AI has changed Ashish's personal engineering workflow

1:02:42 Impact of AI on the broader engineering organization

1:04:15 Measuring AI productivity and the challenge of metrics

1:07:26 50-60% AI-generated code and the changing nature of coding

1:09:18 System design skills becoming more important than coding

1:13:00 Junior engineers doing senior-level work and interview changes

1:14:37 Customer conversations about Gen AI adoption across industries

1:17:26 The MIT report on 95% agent failures and why it misses the point

1:22:12 Agent architecture with LangGraph vs Google ADK and building internal agent platform

1:24:26 Where value lies in the next two years: tools, skills, and optimization

1:28:05 Startup opportunities in making AI accessible to non-technical users

1:29:26 Closing remarks

]]>357ca95d-0687-48e4-b77d-2dcc0f17afd3Wed, 28 Jan 2026 15:11:59 GMT<p>Summary:<br /><br />In this conversation, I talked to Ashish Shubham (VP of Engineering), who's been at ThoughtSpot for 10 years, about AI agents in enterprise analytics. ThoughtSpot started as a search-based analytics company trying to make data accessible to regular business users. In 2019, they tried building natural language interfaces using BERT, but only hit about 50% accuracy. For a product where enterprise customers make billion-dollar decisions, that wasn't good enough. They shelved the project.<br /><br />When ChatGPT came out, ThoughtSpot was ready. Ashish walked me through how they pivoted: they built a 25-30 person team, decided to use prompting instead of fine-tuning, and leveraged their existing semantic data modeling layer to get accuracy into the high 90s. We got into the technical evolution from monolithic systems to agent architectures with tools, how they went from manual human judges to using LLMs to evaluate their outputs, and how enterprise security requirements shaped what they built.<br /><br />We also talked about how software engineering is changing. Ashish said 50-60% of his code is AI-generated now, and he thinks system design is becoming the critical skill, even for junior engineers. He had an interesting take on the "95% of AI deployments fail" stat too.</p><p></p><p>Chapters:</p><p></p><p>0:00 Intro and Ashish's journey to ThoughtSpot from GoDaddy</p><p>0:13 ThoughtSpot's mission to democratize data analytics for business users</p><p>1:26 Early search-based analytics before natural language processing</p><p>2:36 ThoughtSpot vs Tableau and the promise of self-service analytics</p><p>4:40 The analyst bottleneck problem and how ThoughtSpot aimed to solve it</p><p>5:49 Early technical challenges with in-memory databases and data migration</p><p>8:11 Semantic data models, joins, and creating abstraction layers for users</p><p>11:39 Who builds the data models and the role of analysts</p><p>12:22 Pre-LLM natural language processing using BERT and word2vec in 2018-2019</p><p>14:43 The accuracy problem and ambiguity in translating user queries</p><p>16:58 Trust challenges and why the early NLP product never became core</p><p>19:59 Competition with Tableau, Looker, and Power BI</p><p>22:44 How analyst roles changed with self-service analytics tools</p><p>25:30 The ChatGPT moment and pivoting to LLM-powered natural language</p><p>27:48 Early prompt engineering days and generating SQL with LLMs</p><p>31:09 Training vs prompting debate and why fine-tuning was eventually abandoned</p><p>34:28 Organizational changes and building the NLS team</p><p>37:16 Coaching systems for company-specific terminology vs training models</p><p>39:02 Evolution of evaluation methods from human judges to LLM-as-judge</p><p>43:23 Moving to LangFuse and GCP for agent infrastructure</p><p>46:29 How LLM context windows and capabilities evolved their product</p><p>50:07 From 30-column limits to agentic systems with 90%+ accuracy</p><p>52:52 RAG, column selection, and using proprietary data indexes</p><p>54:59 Multi-model support and enterprise data security concerns</p><p>59:14 How AI has changed Ashish's personal engineering workflow</p><p>1:02:42 Impact of AI on the broader engineering organization</p><p>1:04:15 Measuring AI productivity and the challenge of metrics</p><p>1:07:26 50-60% AI-generated code and the changing nature of coding</p><p>1:09:18 System design skills becoming more important than coding</p><p>1:13:00 Junior engineers doing senior-level work and interview changes</p><p>1:14:37 Customer conversations about Gen AI adoption across industries</p><p>1:17:26 The MIT report on 95% agent failures and why it misses the point</p><p>1:22:12 Agent architecture with LangGraph vs Google ADK and building internal agent platform</p><p>1:24:26 Where value lies in the next two years: tools, skills, and optimization</p><p>1:28:05 Startup opportunities in making AI accessible to non-technical users</p><p>1:29:26 Closing remarks</p>no01:31:202He Built a $200M AI Agent 10 Years Before ChatGPTfull<![CDATA[He Led AI Transformation for Angry Birds. Then He Quit.]]>In this conversation, Tatu discusses the transformative impact of AI on game development, drawing from his extensive experience in the gaming industry. He highlights the shift from traditional game development processes to a more agile, AI-driven approach that allows for rapid prototyping and iteration. Tatu emphasizes the importance of organizational change and the need for leaders to embrace AI as a core part of their strategy. He also explores the evolving role of product managers, the challenges of user acquisition, and the future of marketing in a saturated gaming market. The discussion culminates in Tatu's vision for his new AI-native game studio, aiming to disrupt the industry by leveraging cutting-edge technology to create high-quality games at unprecedented speed.

Takeaways:

  • AI is condensing the time and resources needed for game development.
  • Organizational inertia can hinder the adoption of AI in large companies.
  • The future of game development will require T-shaped professionals with diverse skills.
  • AI will fundamentally change the economics of the gaming industry.
  • Smaller companies can leverage AI to outmaneuver larger competitors.
  • The role of product managers will evolve as AI takes over prioritization tasks.
  • Marketing strategies will need to adapt to a more saturated market.
  • User acquisition costs are expected to rise due to increased competition.
  • Novelty may not be as valuable as familiarity in a saturated market.
  • The future of entertainment will see a rise in fast, iterative game development.

Chapters:

00:00 The Evolution of Game Development with AI

03:07 From Web Design to Gaming: A Career Journey

05:50 The Impact of AI on Knowledge Work

09:07 The Changing Landscape of Game Development

11:53 Organizational Inertia and the Future of Gaming Companies

14:55 The Role of AI in Transforming Game Development

17:57 Navigating the Challenges of AI Adoption

21:08 The Future of Game Development Methodologies

23:46 The Role of Product Managers in an AI-Driven World

26:47 Marketing Strategies in the Gaming Industry

29:59 The Role of Publishers in Game Development

33:05 The Future of User Acquisition in Gaming

36:02 The Changing Economics of Game Development

38:56 The Future of Software Development

42:13 The Role of Novelty in Game Development

45:04 The Importance of Familiarity in a Saturated Market

48:12 The Future of Fast Entertainment

50:59 Leveraging Licensing for Success

54:02 The Journey from Rovio to AI Native Gaming

57:02 Building Tools for Rapid Game Development

59:57 The Vision for Future Games

01:03:04 AI Adoption in Organizations: A Leader's Perspective

]]>c6f9d3af-71e6-4cd1-b171-48d684b1c197Wed, 07 Jan 2026 04:57:15 GMT<p>In this conversation, Tatu discusses the transformative impact of AI on game development, drawing from his extensive experience in the gaming industry. He highlights the shift from traditional game development processes to a more agile, AI-driven approach that allows for rapid prototyping and iteration. Tatu emphasizes the importance of organizational change and the need for leaders to embrace AI as a core part of their strategy. He also explores the evolving role of product managers, the challenges of user acquisition, and the future of marketing in a saturated gaming market. The discussion culminates in Tatu's vision for his new AI-native game studio, aiming to disrupt the industry by leveraging cutting-edge technology to create high-quality games at unprecedented speed.</p><p></p><p>Takeaways:</p><p></p><ul><li>AI is condensing the time and resources needed for game development.</li><li>Organizational inertia can hinder the adoption of AI in large companies.</li><li>The future of game development will require T-shaped professionals with diverse skills.</li><li>AI will fundamentally change the economics of the gaming industry.</li><li>Smaller companies can leverage AI to outmaneuver larger competitors.</li><li>The role of product managers will evolve as AI takes over prioritization tasks.</li><li>Marketing strategies will need to adapt to a more saturated market.</li><li>User acquisition costs are expected to rise due to increased competition.</li><li>Novelty may not be as valuable as familiarity in a saturated market.</li><li>The future of entertainment will see a rise in fast, iterative game development.<p></p></li></ul><p>Chapters:</p><p>00:00 The Evolution of Game Development with AI</p><p>03:07 From Web Design to Gaming: A Career Journey</p><p>05:50 The Impact of AI on Knowledge Work</p><p>09:07 The Changing Landscape of Game Development</p><p>11:53 Organizational Inertia and the Future of Gaming Companies</p><p>14:55 The Role of AI in Transforming Game Development</p><p>17:57 Navigating the Challenges of AI Adoption</p><p>21:08 The Future of Game Development Methodologies</p><p>23:46 The Role of Product Managers in an AI-Driven World</p><p>26:47 Marketing Strategies in the Gaming Industry</p><p>29:59 The Role of Publishers in Game Development</p><p>33:05 The Future of User Acquisition in Gaming</p><p>36:02 The Changing Economics of Game Development</p><p>38:56 The Future of Software Development</p><p>42:13 The Role of Novelty in Game Development</p><p>45:04 The Importance of Familiarity in a Saturated Market</p><p>48:12 The Future of Fast Entertainment</p><p>50:59 Leveraging Licensing for Success</p><p>54:02 The Journey from Rovio to AI Native Gaming</p><p>57:02 Building Tools for Rapid Game Development</p><p>59:57 The Vision for Future Games</p><p>01:03:04 AI Adoption in Organizations: A Leader's Perspective</p>no01:15:391He Led AI Transformation for Angry Birds. Then He Quit.full