Did an Intern Using Claude Code Beat a Senior Engineer? The Evidence
By Matt Cullerton, founder of Mavric and Apso
A TikTok from Graham Weaver (@grahamcweaver) tells the story of a software CEO who gave the same rewrite to a senior engineer and a summer intern, both using Claude Code, and the intern finished faster with better code. No company, person, date or metric is named, and no primary account of the story exists in public sources. The best-documented rewrite of this kind was done by the engineer who wrote the original code, and the published research on whether AI levels the playing field splits by field.
What the video claims
Weaver says he was talking to "a CEO of about a $75 million revenue software business" whose product was "about a third of their revenue." The CEO gave a rewrite to one senior engineer, who "finished the task by himself in three months." The CEO, "unbeknownst to the software engineer, also gave the same task, exact same task, to a summer intern." The intern "finished the project quicker, and his code was significantly better than that of the senior engineer."
Weaver's explanation is prompt posture: "the senior engineer told AI how to design this code and what he wanted to do with it. Whereas the summer intern asked AI how to build it." He closes with "You don't have to have 10 years of coding experience to be really effective using AI just like that summer intern showed."
Is the story verified?
No. I searched for a CEO's own post, a company blog, a case study or first-party reporting, and found only the TikTok and pages restating its description. The $75 million revenue figure, the three-month timeline and the "significantly better" judgment are unsourced claims relayed secondhand.
The video also contradicts its own description. The spoken script says only the engineer was unaware of the second assignment. The post description says the CEO gave the rewrite "to two people without telling either of them." An intern who knows he is racing a senior engineer is in a different experiment from one who does not, and the video does not say which happened.
The timeline is tight as well. A summer intern "still in college" works a summer term, which in US programs typically runs ten to twelve weeks, and the engineer took three months. The two efforts must have run close to the same time, though the story implies the intern's assignment came second.
Who is Graham Weaver?
His credentials check out. Weaver signs Alpine Investors' Fund IX letter, dated August 16, 2025, as "Founder and CEO, Alpine Investors." The letter states $15 billion in total assets under management and a Fund IX close of $4.5 billion. Stanford Graduate School of Business lists him as a Lecturer in Management teaching GSBGEN 542 (Engineering a Remarkable Life) and STRAMGT 355 (Managing Growing Enterprises). Alpine invests in software companies, so the conversation is plausible, though that does not confirm the details.
What is the SaaSpocalypse?
Weaver links the story to "the SAASpocalypse" and "how easy it is to rewrite software code." The term has a specific origin. Long Angle's Q1 2026 software analysis says a Jefferies trader coined "SaaSpocalypse" in February 2026 as markets priced in the effect of new AI products. A software selloff on February 3, 2026 erased roughly $285 billion in market cap, and the SEG SaaS Index fell 25.7% year to date through March 31, 2026. Long Angle cautions that a broader market move also drove the decline.
The term describes a repricing of SaaS stocks on the fear that AI agents will replace per-seat licenses. An index drawdown says nothing about how long a rewrite takes, and the video does not connect the two.
What does a documented AI rewrite look like?
The clearest first-party account is Jarred Sumner's Rewriting Bun in Rust, published July 8, 2026. It covers a port of the Bun JavaScript runtime from Zig to Rust between May 3 and May 14, 2026. The numbers:
- 11 days of work
- a 535,496-line Zig codebase, excluding comments, producing +1,009,272 lines of Rust across 1,448 files
- 6,502 commits (6,778 including merges)
- 5.9 billion uncached input tokens, 690 million output tokens and 72 billion cached reads, about $165,000 at API pricing
Sumner ran about 50 workflows in Claude Code, four at a time in separate worktrees with 16 Claude instances each, roughly 64 at once, on a pre-release Claude Fable 5. Anthropic acquired Bun in December 2025.
This is strong evidence that one person with agents can compress a team-year of work. That person wrote the original Zig code himself over about a year, and the post says the existing TypeScript test suite served as the objective check that made automated verification possible. Weaver's story includes neither condition.
Anthropic's LG CNS case study is an enterprise example. LG CNS modernized a 20-year-old project management platform for a Korean construction company, converting 2,888 of 2,913 APIs (99.1%) and migrating 1,340 screens from MiPlatform to React. The work ran from November 2025 to June 2026 at roughly 50% of the cost of a conventional rebuild, with a team and a vendor.
Does AI level the playing field for less experienced workers?
In customer support, yes. Generative AI at Work by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in the Quarterly Journal of Economics, studied about 5,172 support agents given staggered access to an AI assistant. Issues resolved per hour rose 15% on average. Less experienced and lower-skilled workers improved in speed and quality, while the most skilled saw small speed gains and small quality declines.
In coding, the one randomized trial points the other way for experienced developers. METR's study, published July 10, 2025, had 16 experienced open-source developers work 246 real issues in repositories they knew, mostly with Cursor Pro and Claude 3.5 and 3.7 Sonnet. With AI they took 19% longer, and afterward they estimated AI had sped them up by 20%.
METR studied developers in codebases they knew well, where their own context competes with the model's output and review time eats the gains. That could produce an outcome like Weaver's, but it is a different mechanism from the one he names. No study cited here tests how the question to the AI was phrased.
Bottom line
The intern story is an unverified secondhand anecdote, and its script and description disagree on who knew about the race. The largest documented AI rewrite was run by the engineer with the deepest knowledge of the code. Evidence that AI helps less experienced workers most comes from customer support, while the coding trial found experienced developers slower with AI.
Sources
- Rewriting Bun in Rust - Jarred Sumner's account of the Zig to Rust port
- METR developer productivity study - randomized trial of experienced developers using AI tools
- Generative AI at Work - Brynjolfsson, Li and Raymond on customer-support agents
- Alpine Investors Fund IX letter - Weaver's title and Alpine's fund figures
- Graham Weaver at Stanford GSB - his faculty title and courses
- Long Angle on the SaaSpocalypse - origin of the term and the market figures
- LG CNS case study - Anthropic's account of an enterprise platform modernization
Published October 1, 2026. Based on a TikTok by @grahamcweaver. Full video notes (markdown)