6 min read · 12 sources

ICONIQ’s new AI benchmarks cover a hand-picked group of winners, not the top quartile - the real bar is far higher.

ICONIQ's AI Revenue Numbers Are Garbage - And the Bar Is Higher Than You Think

ICONIQ published new benchmarks for AI-native revenue, and the numbers look absurd. That’s because they are - but not in the way you’d hope. The report draws from the top quartile of hand-selected AI winners at a single top VC firm, which effectively makes the numbers top-decile or higher.

General Catalyst’s Hemant Taneja now argues the traditional T2D3 growth standard (triple, triple, double, double, double) is no longer sufficient. Engineers should care because these benchmarks set the bar investors compare every startup against, and that bar just moved up. Before you chase these numbers, check the company size and cash burn behind them - the chart may show a pace you can’t afford.

Charts of the Week: So Many Apps, So Little Time

The number of apps being built has exploded, driven by AI code generation tools. The catch: for the most part, no one is really using these new apps. Time spent per app is flat or falling across every category.

The one exception is Productivity, where time-spent is growing fastest - driven by ChatGPT, Claude, Gemini, and Grok. The lesson for engineers: generating code is easy. Building compelling, successful apps still requires more than just AI output. The tooling lowered the cost of creation, not the bar for distribution.

The New Battleground in GTM: It's Not the Product, It's the Data

Someone who loved your product moves to a new company. You can send the same congratulations-and-pitch email they’ll get from everyone else, or you can bring up what they actually accomplished with your product and ask if they’re tackling something similar in the new role.

The moat around distribution - “GTM alpha” - comes from combining third-party data with first-party data you genuinely own: champion relationships, usage data, the works. One workflow waits a month before making that approach. Another reopens lost deals when the original obstacle changes. The reason to reach out comes from what happened in the account, not how long it’s been since the last touch. Build systems that leverage both data types and you get a durable advantage.

PostHog's 6-7 Loops a Day to Make Itself Self-Driving

PostHog’s self-driving team is eating its own dog food hard. They run 6-7 loops daily to automate their own operations, using a “scout” that watches important metrics and files structured feedback headlessly when it hits friction, via the PostHog MCP.

The loop routes each theme to its owner - MCP, product area, docs, or signals team - and the scout validates that fixes held before closing a theme. In one example, an investigation report arrived within five minutes, followed by a proposed code change eight minutes later. Other workflows investigate error spikes and inspect session replays. Each step passes its findings to the next, reducing how much a human has to reconstruct before they can start fixing.

Google Ax: Kubernetes for AI Agents

Google open-sourced Ax, a framework for building and orchestrating AI agents. It runs on top of Agent Substrate, and if you’ve used Kubernetes, it will feel familiar - same declarative model, same operator pattern, same “desired state vs. current state” reconciliation loop.

The pitch: developers bring the API, and the framework handles the agent, the browser, and the orchestration. For teams running agents in production, this is the missing control plane. If you’ve been duct-taping together agent workflows, Ax is worth a look this week.

MSL's Muse Voice Transcribe: SOTA in Real-Time Speech

MSL released Muse Voice Transcribe, its first real-time audio perception model, and it’s now SOTA in streaming speech-to-text. The model manages hour-long sessions with 20+ speakers, handles mid-sentence code-switching between languages, and can be biased toward names and domain terms.

This is a significant advancement in real-time, multilingual audio processing with built-in speaker separation. For anyone building meeting transcription, live captioning, or voice agents, this raises the floor on what users will expect from your product by default.

Who Controls Your Startup? Governance Matters from Day One

You can own a large share of your company and still lose your job in one board meeting. The Zenefits case is the cautionary tale: the board removed CEO Parker Conrad in a single meeting without a shareholder vote.

The composite teaching case “SoftMet” makes it concrete - founders raise Series A, B, and C, and suddenly own less than half the company on a fully diluted basis. The board has seven members, four appointed by investors, plus two observers. Governance structures negotiated early in fundraising determine who controls the company, regardless of share ownership. Pay attention to who appoints each director, who can replace them, and what happens to those rights in the next round.

People First Equity Plans: Giving Away 20% on Purpose

Athyna gave away 20% of the company to employees, and the founder says they’d do it again. This “People First” approach to equity contrasts with traditional vesting structures - think outright grants employees actually own rather than options they must purchase.

The model uses time-based vesting with a liquidity event before shares transfer, and walks employees through what they own, what happens if they leave, and how future fundraising changes their stake. If you’re considering equity comp structures that prioritize broad ownership and retention, this is a concrete data point.

The Manager's Path in the Age of AI

Leaders have the power to make things better or worse based on their choices, behaviors, and how they show up. AI didn’t change that - it amplified it. Your team watches how you handle the AI-generated code, the sloppy PRs, the “just ship it” pressure. They learn from it.

SaaStr AI: Doubling Sponsorship Revenue with Agents

SaaStr’s agent stack helped the company double sponsorship revenue in the last 12 months. The teardown covers how they run inbound, renewals, and outbound on agents. If you’re building GTM agents, this is a working example of the economics, not a slide deck.

How I Avoid Throwing Slop Grenades

One engineer generated hundreds of AI-assisted pull requests during a migration - 60 in a single day - and avoided flooding the team with slop. The key was extensive pre-work: talking to each team first, baking their input into the changes, then iterating on the workflow based on what went wrong.

Generating a code change in five minutes doesn’t help a teammate who has to review it. AI slop is a symptom of a broken SDLC, and the fix is process improvement, not better prompting.

Build Your AI Second Brain for GTM

The framework for an “AI second brain” for GTM teams distinguishes between two types of knowledge: context (what is true today) and history (why things changed). The author suggests using event sourcing - Type 1 vs. Type 2 data, borrowed from engineering and accounting - to separate the stores.

This gives AI agents a practical structure for knowledge, enabling better judgment and pattern recognition in GTM workflows. It’s a genuinely useful pattern for anyone feeding agents with company knowledge.