Forward Deployed

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

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

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

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?

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

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

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

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

"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

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