An ode to mediocrity

The comments

Some comments were supportive, but some weren't. A principal software architect wrote: “Look, I use AI plenty... but this is 100% AI slop.” A reader asked how I knew the parts list was real, and an AI engineer at a large software company answered for me: “spoiler alert: its not.” And finally, the founder of a company that helps circuit board teams with compliance said I should look at my own repo before posting nonsense, because the circuit board had zero routes on it (routes are the copper tracks that connect parts on a circuit board to each other. It's like having all the appliances installed in a house, but no wiring). He then quoted the AI's notes back to me: “What a layout engineer still has to do: Route everything (225 connections).”

He was right. I posted a board that I claimed AI had made, and it didn't have any wires. Two other things the critics said are also true (there's no ring yet, as my project page says in bold letters, and the video in my post was generated). I thanked the first gentleman for his comment and asked what else was wrong since it's all open. His response is the best question I got all week: “If you missed something that fundamental, how are you evaluating the rest of the design?”

I'll answer his question further down in the post, but let's take a look at the board first, and who was calling it slop.

What happened to the board

Here is what actually happened. Late on the 12th of November, I told Claude the wiring was missing, and asked it to finish the job properly. At a minute past midnight the next day, it had already done an automatic pass (137 out of 225 connections were done) and got stuck with the two densest chips on the board. In the morning I got impatient, and told it to finish the job. At 10:52 the main board was fully routed, and the design rules checker (a program that checks that all the copper tracks are manufacturable) showed 0 violations and 0 unconnected items. If you want to see how much happened that day, feel free to browse through the 197 commits. As I write this, there are 17 boards fully routed in the repository, and a script that double checks that all the 133 parts on the parts list exist in the 3D model and are wired up correctly.

Not a single one of those was routed by me. I just asked for it.

Working on hardware means that the first version will often have something wrong with it. Even for very skilled professionals. A trade study found that the average number of respins for a circuit board is 2.9. Intertek, a big test house, says that around half of all products fail their first EMC test (a test to make sure the product neither emits nor is affected by radio waves). A study of 144 Kickstarter hardware projects that received funding found that at most 32% shipped on time. In other words: it's pretty normal for the first version to have something wrong with it. What's not normal is how fast it got repaired: a specialist publicly pointed it out, and within a day it was gone.

Who liked it and who didn't

I noticed this pattern with the comments on this post, but also with most things I publish: the negative comments tend to be from people who are professionals in the field, and the positive comments from people outside the field. It's tempting to say it has something to do with pride, but after doing some research I think both groups are actually seeing the truth.

There's a lot of evidence that AI helps novices more than experts. A controlled study of 758 BCG consultants found that below-average consultants improved by 43% with AI, and above-average consultants improved by 17%. A study of 5,172 customer support agents found that newer agents improved by about 30% (equivalent to turning a 2-month experienced agent into a 6-month experienced agent) and that the most experienced agents had almost no improvement. Field experiments from Microsoft, Accenture, and a third company found that junior software developers improved by 27 to 39% while senior developers only improved by 8 to 13%. My favourite study had 16 experienced open source developers work on their own projects, timed both with and without AI, in early 2025. They took 19% longer with AI tools, and thought they were 20% faster.

When a beginner tries out these tools, they give them the equivalent of months of experience instantly. When an expert tries them out, they give a small gain, slow them down a little and keep making errors in the one field where the expert is qualified to spot those errors. When both groups go on LinkedIn to talk about their experience, they're both telling the truth. The engineer who told me my ring was slop was talking about his tool, and I was talking about mine.

I should clarify that I don't think experienced workers are losing out here. It would be easy to say the experts are just afraid for their jobs, but the data so far (collected by a team of economists at Stanford) suggests that they are doing fine. The ones losing out are 22 to 25 year olds, whose employment in the most exposed occupations is about 19% lower than it otherwise would be. What experienced workers are losing is gatekeeper status. Until very recently, if you wanted a circuit board designed then you either had to be an electrical engineer or hire one. Now there's a third option.

What the word “slop” does

There's a second effect, and it has to do with the word itself. Merriam-Webster has already picked their Word of the Year for 2025, and it's slop. (They made sure to mention that it was picked by their human editors.) Merriam-Webster defines slop as “digital content of low quality that is produced usually in quantity by means of artificial intelligence”, but a lot of the time it's about who made it.

I found a clean experiment about this effect on poetry. Two researchers had 1,634 people each classify ten poems as human or AI. Five of the poems were from a famous poet (Byron, Sylvia Plath, etc.) and five were written by an AI in the style of the poet. People correctly identified the AI-written poems 46.6% of the time, which is slightly worse than random. A second set of 696 people rated the poems according to 14 qualities. Actual AI-written poems scored higher than actual human-written poems. But poems that participants were told were AI-written scored lower than poems they were told were human-written, regardless of which they actually were. This effect held for 13 of the 14 qualities.

This effect isn't limited to creative works. In a set of four experiments with 4,439 participants, employees who used AI to complete work tasks were seen as less competent and motivated. In a set of 13 experiments, people who admitted to using AI were trusted less than people who didn't. So when people realised that my ring was AI-generated, they probably lowered their opinion of the repo. The issues in the board were real, and I'm glad they got found, but I think the leap from “two faults” to “100% slop” was not.

We've been here before

I went down a small rabbit hole and found my new favourite insult. In 1592 a pamphlet came out in London under the name of Robert Greene, a writer who signed some of his books “Master of Arts in both Universities” (Cambridge and Oxford, in case you wondered). It warned fellow playwrights about a newcomer, “an upstart Crow, beautified with our feathers”, who thinks he can write verse as well as the best of them, and who is “an absolute Iohannes fac totum”. Johannes Factotum is Johnny do everything. In other words, a jack of all trades. The guy was also an actor and the son of a glove maker. Greene (or perhaps his editor, nobody knows) even calls him “the onely Shake-scene in a countrey”. To be clear, this is about a guy with two university degrees calling someone a jack of all trades, who happened to be William Shakespeare. Ok, perhaps they were talking about someone else, but very few people think so.

In 1991, a student in Helsinki announced he was building a free operating system and said “just a hobby, won't be big and professional like gnu”. Today, Linux powers all of the world's 500 fastest supercomputers. It has done so since 2017. In 2005, Nature had a group of experts compare 42 articles on scientific subjects in Wikipedia and the Encyclopaedia Britannica. They found Wikipedia had an average of four inaccuracies per article, while Britannica had three. Britannica declared the study fatally flawed and demanded a retraction. In 2012, Britannica ceased printing.

Both Linux and Wikipedia started out as a worse product. However, since they were open, all the faults people found were faults people fixed. Programmers call this Linus's Law: “Given enough eyeballs, all bugs are shallow.” Even my circuit board got some eyeballs. This happened in the comment thread on LinkedIn. Since the people criticising me did exactly what Linus's Law suggests, I'd like to say thank you. And yes, I mean it.

How I evaluate what I can't judge

Okay, back to the best question of the week. How could I do any evaluation if I've missed something so fundamental?

It turns out there are a lot of other people judging me, and it's by design. For instance I can't judge board layouts so I've made a series of arrangements to make sure that I don't have to. Everything is published (that's how the missing routes were found in the first place) and the work can check itself. All the boards have rule checks. The software has 732 tests which all pass when run on a simulated board. And other AI have checked my board wiring 5 times over. In fact the 5th time they found the board wasn't good enough and failed it, and the failure (two ribbon cables couldn't be fitted in the ring body as drawn) is written near the top of the project page, right next to “No ring has been built yet.”

But ultimately it comes down to judgement: what to build, how high to set the bar, when the answer is wrong. It's very important, and I think undervalued. Take a study where AI was used to advise 640 small business owners in Kenya. The better business owners improved their results by about 15%. The worse business owners made their results about 8% worse. Both sets of people tended to ask similar questions and get similar answers, but acted on different parts of the answers. AI doesn't make everyone better. It lifts people who can tell a good answer from a bad one, including in fields they've never worked in.

Mediocre at everything

I want to talk about the title. I looked up the etymology of “mediocre”, and it's better than I expected. It comes from the Latin mediocris, which itself comes from medius (middle) and probably ocris (a rugged mountain). So mediocre means something like “halfway up the mountain”. There's even an ode to mediocrity by Horace from around 23 BC: “Auream quisquis mediocritatem diligit”, whoever loves the golden middle. His argument was that the middle is the safest, since it's the mountain tops that get struck by lightning.

If I had produced this ring design ten years ago, with all the mistakes I made, I would have been considered a mediocre hardware engineer. A real hardware engineer, with a degree and experience. But today, the same output is just AI slop. But you know what? Fine, I'll take mediocre. Because imagine having the superpower of being mediocre at anything you can imagine. Being halfway up the mountain of circuit design, firmware development, 3D modelling, video production, contract law, financial stuff, whatever, whenever you want. By any definition I can think of, someone who is somewhat OK at anything and everything is a genius. We used to have a word for these people: polymaths. One study found that Nobel prize winners are at least 22 times more likely than other scientists to be an amateur actor, dancer or magician. But polymaths were going extinct. A recent analysis of 2.3 million scholars showed that the fraction of scientists working in multiple fields had decreased dramatically over the last century. My guess is that all the fields had become too deep, and one person couldn't explore more than one field deeply anymore. But AI changes that.

By the way, the mountain keeps getting taller. A group called METR has been measuring the longest task AI models can complete, in units of skilled human time. In 2019, this was 3 seconds. In 2023, 4 minutes. In February 2026, it was 12 hours (with large error bars). This number has doubled roughly every 7 months since 2019. The recent slowdown of the 16 expert developers didn't last. When the same lab did the study a year later, they found that the developers looked faster with AI, although they said the evidence wasn't very strong. So “halfway up the mountain” has gone from 3 seconds to a day and a half in seven years.

The critics are right. Compared to the best in their trade, my ring is mediocre. But when I compare it to where I was on 9th November (pretty much at the bottom of a mountain I never even thought of climbing) then it's still incredible. So yes, AI makes you mediocre. But mediocre at everything.

If you want to take a look at the repo yourself, it is still open here: github.com/NexusGPT/openring. Go ahead and tell me what's wrong with it.

Sources

The ring: OpenRing public project page, state of 15 Nov 2026, https://github.com/NexusGPT/openring, and the project's own log and requirements.

Word of the Year: Merriam-Webster, “2025 Word of the Year: Slop”, https://www.merriam-webster.com/wordplay/word-of-the-year

Hardware: Lifecycle Insights study as cited in I-Connect007, “Cutting Respins: Journey to the Single-spin PCB”; Intertek figure as cited in Military and Aerospace Electronics, 27 Aug 2026; Jensen and Özkil, Design Science 4 (2018), https://doi.org/10.1017/dsj.2018.14

Beginners and experts: Dell'Acqua et al., “Navigating the Jagged Technological Frontier”, Harvard Business School Working Paper 24-013 (2023); Brynjolfsson, Li and Raymond, “Generative AI at Work”, https://arxiv.org/abs/2304.11771; Cui et al. as reported by MIT Sloan, 4 Nov 2024; Becker et al., METR (2025), https://arxiv.org/abs/2507.09089, and the METR update of 24 Feb 2026, https://metr.org/blog/2026-02-24-uplift-update/

Jobs: Stanford Digital Economy Lab, August 2026 update of Brynjolfsson, Chandar and Chen, “Canaries in the Coal Mine”, https://digitaleconomy.stanford.edu/news/canariesaug26/

The label: Porter and Machery, Scientific Reports 14, 26133 (2024), https://www.nature.com/articles/s41598-024-76900-1; Reif, Larrick and Soll, PNAS (2025), https://doi.org/10.1073/pnas.2426766122; Schilke and Reimann, Organizational Behavior and Human Decision Processes 188 (2025), https://doi.org/10.1016/j.obhdp.2025.104405

History: Greene's Groatsworth of Wit (1592); Torvalds, comp.os.minix, 25 Aug 1991; TOP500 operating system statistics, https://top500.org/statistics/details/osfam/1/; Giles, Nature 438 (15 Dec 2005); Encyclopaedia Britannica, “Fatally Flawed” (March 2006); Raymond, The Cathedral and the Bazaar.

Judgment: Otis et al., “The Uneven Impact of Generative AI on Entrepreneurial Performance” (2024), https://escholarship.org/uc/item/40s41244

The ode: Horace, Odes 2.10, https://www.thelatinlibrary.com/horace/carm2.shtml (my rendering); https://en.wiktionary.org/wiki/mediocris; Root-Bernstein et al., Journal of Psychology of Science and Technology 1(2) (2008); Risha, Lin, Leahey and Wu, “The Death of Renaissance Scientist”, https://arxiv.org/abs/2304.14518; METR, “Task-Completion Time Horizons of Frontier AI Models”, https://metr.org/time-horizons/