TLDR Founders 2026-09-02

The biggest false dichotomy in AI: a costly frontier model for everything, or a cheap model everywhere. (Only to restart tuning with each new model release.)

Finest finds the model your work needs:

👉Finest API serves a lower-cost configuration only when it clears your task's quality bar. Otherwise, the model you named serves.

👉Finest Code brings the same evidence-gated approach to Claude Code, with Claude as the default and fallback.

All you do is change two lines of code. You keep your prompts, model IDs, tools, and workflow.

Performance or price? Choose both →

Investment committees now pause weekly on AI companies with great revenue curves to ask what is truly unique, since every viral product has copycats. As one founder-investor put it, AI made it 10x easier to build a $10M ARR company but did not make it 10x easier to build a $1B one. A seed and Series A investor at FPV Ventures calls the stall zone the new valley, $5M to $25M in ARR. Carta found 15.4% of Q1 2022 seed companies reached a Series A within two years, down from 30.6% for 2018. The four ways out are talent M&A and seed-strapping, raising small, getting profitable fast, and keeping 80% of the company.

There are three ways to make money in the AI industry: models, the harness, and everything else the model and harness need to stay healthy at scale. Founders picking a category need to make sure they're using the right playbook. The categories are obvious, but the strategies are close to inverted. This article looks at each category to see what works.

AI is changing hiring faster than many organizations can adapt. Companies are changing how they think about who they hire, when they hire, and why. Competitive advantage is increasingly shifting toward judgment as AI lowers the cost of execution. Traditional markers of pedigree still matter, but they are becoming less predictive than evidence that someone can learn, adapt, and operate in unfamiliar requirements.

Deterministic systems should be used to establish context. Large language models should be used to reason over that context. Models shouldn't be used to establish whether something has happened or what the current state of a system is. Systems should be used to assemble that information into context so that models can interpret, connect, reason, and synthesize.

The discovery layer and the infrastructure layer have never been equally measurable. The industry puts all of its attention and spend into the layer where the numbers are contested, produced by interested parties, and unstable across instruments. The layer that determines whether agents can transact with them at all returns a clean pass or fail on every check. Both citation and infrastructure matter, but infrastructure is what gets you transacted with.

This is the person on the other side of your renewal itemizing how he decides to cancel. A working CFO lists what he is cutting and why, including a workflow tool that quoted him double what he pays for Claude and a $ 50K-a-year point solution whose function a larger vendor now bundles for free. His own playbook is opening renewal conversations four or more months early, and he reckons 30 to 60 minutes of demoing alternatives buys enough negotiating power. Backdrop is the Bessemer cloud index up 55% in six months, its highest since early 2022.

Teams using coding agents open 6x more pull requests than traditional teams and 3x more than they did in 2024, according to Linear's report on AI usage in software teams. Review load is climbing faster than anyone's model of the job, which is the likeliest way your first engineering manager hire burns out. Managers already see only about 40% of what a report does in a day. The squeeze lands hardest on engineering managers and tech leads: the people expected to write code and manage at once.

Figma has introduced generative plugins and shaders to the canvas to give users a way to design the tools they need, when they need them. Users can publish plugins and shaders to the Figma community or privately within organizations on Organization or Enterprise plans. The agent can build motion and interactivity into any effect, and users can view and download the code behind any shader or plugin. The Figma MCP server can be used to view, change, and export generative plugins and shaders.

WorkOS Relay proxies third-party agent API calls and injects the credentials at the boundary so prompt injection attacks are unable to steal tokens. It lets apps call third-party APIs on behalf of users without directly handling secrets. Requests are proxied through WorkOS, where credentials are swapped in based on the user and organization context. Relay is still in early access, and access requires contacting support.

Adaptive Intelligence is a new bot detection engine from Cloudflare that makes getting through so slow and costly that attacks stop being worth running. Not every attack is obvious to spot, and the economics of deterministic detection are always in the attacker's favor. Adaptive Intelligence makes sure that nothing attackers learn stays true, so each new attempt costs more than the last. It recognizes bots without visibly reacting to them and treats detection as a statistical judgment, weighing many signals at once so there's no single piece of logic for an attacker to isolate and beat.

"Enshittification" isn't real. Social media platforms change to meet evolving user needs, not to worsen user experiences. Despite complaints, user satisfaction and engagement with platforms like Instagram continue to rise. Critics mistake platform evolution and user dislocation for deterioration, but the data shows that people use these platforms more, not less.

Agent scaling is about how cleanly a system allows agents to work in isolation without stepping on each other all the time. This requires expertise in the problem domain, a keen eye for good architectural decisions, and good measurement. Coding is the easy part of writing software - it's everything else that takes most of the time. There's only so much you can streamline before you start losing a significant amount of quality.

Zoom Phone adds a business line, team extensions, and AI call summaries, right inside the Zoom app you already use, from $10.50/user/month. See plans →

Shopify users can now use a Stripe account for setup and billing across all providers.

An open-source coding agent shipped a release carrying 16,000+ merged pull requests from 933 contributors, about half of every PR ever merged into the project, after a seven-week pause to rebuild foundations that followed 106 releases in 230 days.

Lying with data occurs when decisions lack proper retrospective analysis and disregard the Null Hypothesis.