The Great Bifurcation: The Widening Split Between AI-Native and Everyone Else

The Tour That Reset Twenty Years of Pattern Matching

Vivek and Austen had traded emails for years but first met in person when Vivek, newly moved to Austin, came through Gauntlet HQ in early 2025. He watched a team demo a procurement app that had gone from a Sunday-night idea to a working product by end of day Monday, then asked offhand whether the data could be exported as a CSV — and Austen had the engineer ship it while he stood there. Vivek sat in his Jeep in the parking lot for fifteen or twenty minutes afterward. He describes it as disrupting almost twenty years of mental models built as a multi-time founder: what a great team looks like, what product-market fit looks like, what OpEx and go-to-market look like. He went home, told his wife, and then wrote his LPs an essay about software at the speed of thought.

Scoring the Predictions

That LP letter carried a list of predictions, and Vivek walked through which ones held. The bifurcation of engineering talent: correct. The 10x engineer being replaced by the 100x and soon the 1000x engineer: correct, with eight- and nine-figure comp packages as the visible edge of it. Company OpEx falling from headcount reduction but being offset by software and compute spend: correct, and now a live budgeting problem for enterprises trying to plan token spend across the whole org rather than just engineering. The one he got wrong was framing: he predicted agents would become pair-programming partners, assuming a gradual walk toward autonomy. Agents blew past those baby steps. Directionally right, wrong shape.

Where Team Size Is Actually Going

The widely predicted collapse in team size is real at the startup level and has not landed at large enterprises. Junior talent was the first casualty; middle management is now on the chopping block; senior architects are holding. Vivek’s original vision was that teams of 30 to 50 engineers become three to five AI engineers who are domain experts, coordinating in pods and reporting to an architect or the CTO — he does not think we are there yet, but believes the trend line is clear. What has already changed is the default: on the best teams he sees, a function is delegated or automated to AI first, and a hire happens only after those options are exhausted. Austen put the startup version bluntly — needing two or three million dollars to get an MVP out is no longer a fundable plan.

QA Didn’t Die, It Got Rejiggered

One prediction was that quality assurance would become irrelevant, replaced by simply reprompting the model. Both of them think that one aged badly in an interesting way: QA matters more now, not less. Vivek’s argument is that there is an enormous amount of slop in the market — in content, in code, in product — and most of it comes from human laziness rather than model weakness. What has changed is the shape of the function: not traditional QA but eval engineers, agent harnesses, and test infrastructure. When everyone can ship, taste and quality are what separate a product from the noise. A year ago you could get away with one-shotting it. Now everyone is one-shotting it, so you can’t.

Marketing and GTM Are the Laggards

Austen said the thing that surprises him most is how little marketing teams have changed compared to engineering teams, even at companies that describe themselves as AI-first. Vivek sees two shapes. The first, common at growth-stage and enterprise companies, treats AI as a productivity tool and gets 30 to 100 percent more output from the same team. The second is a single early-career, AI-native operator who works at a completely different scale — he relayed a story about one person on an ads team replicating the work of thirteen or fourteen colleagues, after which the manager fired much of the team and promoted them. Austen’s framing: one person showed up with a tractor while everyone else is hoeing the ground by hand, and twenty years of perfecting the craft of hoeing does not transfer.

What Replaces Application-Layer SaaS

Vivek’s thesis has moved down and across the stack. He now looks for full-stack enterprise AI — companies building the application layer, the data and MLOps layers, and their own models at the same time rather than sequentially. He pointed to Writer, which has trained its own Palmyra models since 2021 and sells into the C-suite with a replace-seven-vendors-with-us pitch. He has also moved from dev tools into software infrastructure, on the logic that you can ship a buggy app-layer experience but you cannot ship buggy infrastructure to a Fortune 500. Two other categories interest him: neolabs — applied AI research labs that he calls the fourth home for AI talent after startups, big tech, and academia — and differentiated AI data, the category that produced Scale AI, Mercor, and Handshake.

The New Org Chart and the Quarterly Reset

Vivek is tracking how org charts are being redrawn. The old model had domains — front end, back end, DevOps, security, design, marketing — with founders ranging across all of them and ICs moving up and down inside one. What is emerging cuts across that entirely, and he found the labels he was missing in a post from Boris Cherny, head of Claude Code: Prototyper, Builder, Sweeper, Grower, Maintainer. Prototypers generate ideas and test them fast; builders turn a prototype into production; sweepers refactor and optimize; growers do GTM engineering and hold product-market fit; maintainers scale it. The best people span two or three. His broader point applies to everyone: he thinks the best investors now update their strategy annually and will be doing it quarterly within two or three years, and that operators should evaluate their own companies and vendors on the same cadence. What has stood the test of time is short — durability, defensibility, and monotonic customer value at the company level; speed, experimentation, self-driven people, and customer centricity across the whole org, not just the customer-facing half.