<- all tokdocs

Dex Horthy on AI Coding: Why 2-3x Is the Ceiling, Not 10x

By , founder of Mavric and Apso

View on TikTok ->

Dex Horthy, founder of HumanLayer, says AI coding agents can produce code close to human quality two to three times faster, and that 10x "can't be done. Not today." A TikTok from @wordman.dev keeps the 2-3x part and cuts the ceiling, and it never names the speaker. Research on AI coding supports Horthy's warning about code quality, but no study measures his 2-3x figure.

Who is the speaker in the clip?

The clip carries no name, caption or description, so viewers would assume the speaker is the account owner. The account belongs to Jan-Niklas Wortmann, who describes himself on his site as a Solution Architect at Cursor, formerly leading AI Developer Advocacy at JetBrains, and host of the podcast The Weekly Dev's Brew.

The man on camera is his guest, Dex Horthy. The transcript matches an episode Wortmann published on 13 August 2026, "Dex Horthy - What Actually Gets You 2-3x With AI Coding." Y Combinator lists HumanLayer as founded in 2023 by Dexter Horthy, based in San Francisco, in the Fall 2024 batch. The HumanLayer "H" logo is on his shirt in the first half of the clip.

Both people sell into this market. Horthy sells agent tooling at HumanLayer, and Wortmann works at Cursor. Neither fact appears in the clip. The clip is also spliced from two recording sessions, with a change of headphones and shirt at about 26 seconds.

What did Dex Horthy actually say about 10x?

The clip ends on "but two to three times faster." The episode page gives the full quote:

"You can get 99% of human-quality code, like very good code as if you had written every character by hand, but two to three times faster. You can't get 10x. It can't be done. Not today."

The episode's first chapter marker is titled "You can't get 10x (not today)." Horthy's point was that 2-3x is the realistic ceiling. Cut where it is, the clip makes 2-3x sound like a starting point.

The episode page also quotes a line that appears in the clip: "You can give it a really good architecture doc and the model can follow it to the letter...but somehow the code is still garbage."

Is the 2-3x speedup measured?

No. Horthy gives no methodology, and the episode does not present a study. The figure is a practitioner's estimate from building coding-agent tools.

The best-known measurement points the other way. METR's randomized controlled trial, published 10 July 2025, had 16 experienced open-source developers work 246 real issues on repositories they knew well. They predicted a 24% speedup and afterward reported a 20% speedup. Measured against the clock, they took 19% longer when allowed to use AI tools.

That result does not disprove Horthy. METR tested Cursor Pro with Claude 3.5 and 3.7 Sonnet in early 2025, and METR says its result does not show that AI fails to speed up most developers or that better setup could not change the outcome. Horthy describes a late-2026 workflow built on context engineering with newer models. What METR does show is that developers misjudged their own speed by about 39 percentage points, in the optimistic direction. A self-reported 2-3x belongs in the same category until someone times it.

Does a good spec produce good AI code?

Horthy's core claim is that a model can follow an architecture document to the letter and still write bad code. SlopCodeBench measured this.

The paper, "SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks," was submitted to arXiv on 25 March 2026 by Gabriel Orlanski and colleagues and revised on 7 May 2026. The current version tests 15 coding agents on 36 problems across 196 checkpoints. No agent solved any problem end to end, and the best checkpoint pass rate was 14.8%. Structural erosion rose in 77% of runs and verbosity rose in 75.5%. Compared with maintained human repositories, agent code was 2.3x more verbose and 2.0x more eroded.

The prompt results match Horthy's point. In the first version, quality-aware prompts, including anti-slop and plan-first variants, cut starting verbosity by 33 to 35% and reduced initial erosion, but "the accumulation of issues persists regardless of prompt." Pass rates did not improve, and the cleaner start cost 30 to 48% more compute. A good spec gives a better first draft and does not stop the decline.

The two arXiv versions report different numbers. Version 1 lists 20 problems, 93 checkpoints, 11 models and a 17.2% best solve rate. Check which version a secondary source quotes.

What does industry data say about AI code quality?

Horthy opens by saying that "abandoning code quality and system quality, giving engineers permission to ship slop" is a mistake. Codebase-scale data supports that concern.

GitClear analyzed 211 million changed lines of code from January 2020 through December 2024. Refactored lines fell from 25% of changed lines in 2021 to under 10% in 2024, and copy-pasted lines rose from 8.3% to 12.3%. 2024 was the first year in which cloned code exceeded refactored code.

Google's 2025 DORA report surveyed nearly 5,000 technology professionals. 90% used AI at work, and more than 80% believed it raised their productivity. DORA found AI adoption linked to higher delivery throughput and states that "AI adoption does continue to have a negative relationship with software delivery stability." Its explanation is that faster change exposes weak testing, version control and feedback loops.

Neither source measures a 2-3x multiplier or "99% of human quality." Both measure the quality problem Horthy warns about.

Bottom line

Dex Horthy's actual claim is that AI coding tops out around 2-3x with near-human quality, and that 10x is not possible today. The TikTok cuts the limit and the speaker's name. SlopCodeBench, GitClear and DORA back his warning about code quality, while his speed figure remains an unmeasured estimate.

Sources

Published October 5, 2026. Based on a TikTok by @wordman.dev. Full video notes (markdown)