Skip to content
kevinferron.com

Software development philosophies · An opinion with sources and checks · My opinion

A-changin’

The first one now will later be last

Bob Dylan, The Times They Are a-Changin’, 1964 [Dylan, “The Times They Are a-Changin’”] · Dylan on my songwriters wall

The pile of small tasks used to decide who ranked first. AI is shrinking the pile. Here’s who I’d hire when it stops deciding.

The claim

My opinion

I’m optimistic about what AI does to work. It isn’t the cleverness. It’s what comes off people’s plates.

I think the ranking is changing, so Dylan’s line sits under the title. Some people ranked first because they cleared the pile fastest. They may not now. Some ranked last because the pile buried them, and they may.

Who to hire now covers what I’d hire for. Here’s why I think that’s good news. For the organisational changes I expect, there’s The Only Thing You’ll Need to Change is Everything.

This page is my opinion. The facts in it come from sources linked at the bottom, and when one of those is a preprint, a vendor, a personal post or a news report of an interview, I say so.

The lazy programmer

The Perl documentation calls laziness the first great virtue of a programmer. It’s from the glossary of Programming Perl, a book Larry Wall co-wrote: “The quality that makes you go to great effort to reduce overall energy expenditure.” [Perl glossary (laziness)]

In my words

Rigid process and filling out forms and minutiae has always been a liability for my performance, whereas I thrive in thinking in systems and design patterns, and tools that empower others to affect the software I write without requiring me to modify the software (or to allow me or others to build software additions easily).

In my words

Being overwhelmed with a lot of tasks related to the job but not the software itself. As well as always having to choose whether writing the helper scripts or build processes would impede.

The cliché here with myself and other programmers is the joke: 'I spent 10 hours automating a task that would have taken 15 minutes.'

xkcd charted the real arithmetic under the joke [xkcd 1205, Is It Worth the Time?]. My version is Figure 1, and the chore in it is invented. Make automating it cheaper. Watch which cells flip.

Invented chore: Renaming and filing the build output by hand.

25 days
10 h
Time saved over the horizon, by how often and how long. Green with a tick pays off.
Per day \ minutes1 min5 min15 min60 min
1 a day0.4 h does not pay off2.1 h does not pay off6.3 h does not pay off25.0 h pays off
5 a day2.1 h does not pay off10.4 h pays off31.3 h pays off125.0 h pays off
20 a day8.3 h does not pay off41.7 h pays off125.0 h pays off500.0 h pays off

7 of 12 cells pay off at 10 h. A 15-minute chore you do once saves 0.25 h, which does not pay for 10 h.

Figure 1. Is it worth automating? An invented chore: how often it comes up against how long it takes, over the working days you choose. A cell pays off when the time saved is at least what automating it costs. It’s the idea of the well-known xkcd chart [xkcd 1205, Is It Worth the Time?], redrawn with my own numbers, not a copy.

The smallest chores are the cells that flip, and that’s where the pile came from. A task you do once never flips, because fifteen minutes saved can’t repay half an hour. So the joke survives. I like that.

Three people who said it

On 2026-05-30 Judy Grupenhoff wrote that sequential thinking suits complicated problems, that this moment is complex, and that AI can be a translation layer for people who think in many streams at once [Grupenhoff (essay)]. Her line is “The parallel processor was never the wrong kind of mind.” [Grupenhoff (essay)] She cites no research.

The split between complicated and complex problems isn’t hers. Complexity research drew it, in Glouberman and Zimmerman’s 2002 paper for Canada’s Romanow Commission [Glouberman and Zimmerman 2002] and in Snowden and Boone’s Cynefin framework, Harvard Business Review, 2007 [Snowden and Boone 2007].

On 2026-11-03 James Camp posted on X that AI made him “10x more productive” [Camp (personal post)]. He wrote that he credits his ADHD. The small boring tasks used to pile up, he wrote, and now he says what he wants and gets the hit of finishing something. That’s one person’s account, and I treat it as one.

In my words

ADHD fits the way the tweet I shared mentioned. Small tasks that were out of band of my work often felt overwhelming and paralyzing, but with agents I can delegate.

Palantir runs a Neurodivergent Fellowship. The posting says it isn’t a diversity initiative and asks for no diagnosis, and it names “Pattern recognition. Non-linear thinking. Hyperfocus.” [Palantir fellowship (company page)] That’s the company’s claim. It isn’t evidence that the traits matter.

Alex Karp, its CEO, has called his dyslexia the formative moment of his life (as reported by Fortune on 2025-12-03): “There is no playbook a dyslexic can master.” [Fortune on Karp, 2025 (as reported by Fortune)] He named two groups with a future in an AI economy, Fortune reported on 2026-10-25: people with vocational training, and neurodivergent people [Fortune on Karp, 2026 (as reported by Fortune)]. That’s his view, as reported. I haven’t seen the recording.

What the evidence says

Here’s what the studies say, before I lean on any of it. Mostly, not what the essays say.

  • When 453 college-educated professionals did writing tasks in an experiment, average time fell 40%, quality rose 18% and the gap between workers narrowed [Noy and Zhang 2023].
  • Less experienced and lower-skilled workers took the gains when a study of 5,172 customer-support agents found productivity up 15% on average. The most experienced saw small gains in speed and small declines in quality [Brynjolfsson, Li and Raymond 2025].
  • On tasks inside what the model does well, people using AI completed 12.2% more tasks, 25.1% more quickly, in a preregistered experiment with 758 consultants. On a task outside that line, they were 19% less likely to get a correct answer [Dell’Acqua et al. 2026].
  • For every 25% increase in AI adoption, DORA’s 2024 survey of software teams estimated delivery stability 7.2% lower and throughput 1.5% lower. Those are survey associations, not an experiment. The report itself says AI “does not appear to be a panacea.” [DORA 2024]
  • Neurodiverse users came out statistically significantly more satisfied than other users in a UK government evaluation of Microsoft 365 Copilot. The same evaluation “did not find evidence that time savings have led to improved productivity”. Most users were volunteers [UK DBT 2025 (government report)].

None of it says AI favours one kind of mind. It says AI flattens the part of the job that used to separate people. The jagged-frontier result is the catch: outside what the model does well, people did worse. So knowing where that line is, is a skill.

Then there’s how fast people think they’re going. In METR’s trial with 16 experienced open-source developers, they believed afterwards that they’d been about 20% faster. They were measured 19% slower [METR 2025 (preprint)]. A later METR update points the other way for returning developers, 18% less task time, and METR calls its own data “only very weak evidence” [METR 2026 update (preprint-level)].

METR 2025, 16 experienced developers

Believed (hollow dot)
+20%
Measured (solid dot)
−19%
Gap, in points
39
Figure 2. Felt against measured. What METR’s experienced developers believed afterwards and what was measured, as a change in speed [METR 2025 (preprint)]. METR’s later update is a label, not a bar, because METR calls its data weak [METR 2026 update (preprint-level)]. A self-reported multiple from a post isn’t plotted.

So I don’t take a self-reported multiple as a measurement. Not Camp’s, and not mine either. I also didn’t find a controlled study of AI’s effect on productivity for ADHD, dyslexic or autistic workers specifically.

The pile

Here’s how I think about it: everyone has a pile of small, sequential tasks around the real work, and for some people it costs more than it does for others. AI shrinks the pile. It doesn’t shrink the reading.

Figure 3 is an invented working day, with invented numbers. Two workers get the same day and the same amount of small work, and one of them finds it more costly. Hand off a share and the gap closes. The hour of reading stays.

0%

Worker A

Small tasks cost ×1 for this worker

Solid square: still yours. Dashed: handed off. Amber circle: reading that stays.

16 cards left, 0 handed off, 0 of reading.

Hours left for judgement: 4

Worker B

Small tasks cost ×1.5 for this worker

Solid square: still yours. Dashed: handed off. Amber circle: reading that stays.

24 cards left, 0 handed off, 0 of reading.

Hours left for judgement: 2

The gap between them: 2 h. The reading stays: 0 h.

Figure 3. The pile. An invented working day. Each card is a slice of small work, and a slice handed to an agent leaves a card of reading behind. The numbers are mine, to show the shape, not a finding.

That’s the version where it goes well. Handing off isn’t free. What came back still needs reading, as if someone else wrote it, and Who to hire now puts that second among the skills. Attention is the bottleneck is about why that hour is the scarce one.

Trying paths before committing

In my words

Agent handover allows trying out and experimenting with multiple paths quickly before being committed. The old ways, sometimes the core framework being written would prove itself off target, but it would be too late to re-architect because things had already been built with it.

Most software engineers I know often wish they could refactor their software with ease, and that really was what Scrum and XP programming was supposed to be about, but it got lost in the project management culture as 'tech debt' instead of system iteration to make better software. All of those gates are lifted in this new world.

There’s a pattern for trying pieces of a design in parallel, the fan-out. Designing an agent graph walks through it.

In 1992, on a portfolio system, Ward Cunningham coined the debt metaphor: “Shipping first time code is like going into debt.” [Cunningham 1992] In a 2009 video he drew the line: people had confused it with writing code poorly on purpose, and he’s “never in favor of writing code poorly” [Cunningham 2009].

On design, the Agile Manifesto’s principles land in the same place: “Continuous attention to technical excellence and good design enhances agility.” [Agile Manifesto principles] For Kent Beck, in a 2000 interview, refactoring sat next to testing: “The two most important practices are testing and refactoring.” [Beck, Overload interview 2000]

So changing the design as you learn isn’t new. It was meant to be how the work went, and what got in the way was cost: rebuilding was expensive, so the first architecture got defended. I think that cost is what dropped. I wrote about the same shift in the process around the code in The 50-year storm is here. Are you surfing yet?

The best counter I know is DORA’s. In their 2024 survey, more AI adoption went with lower delivery stability, and they suspect changes got bigger [DORA 2024]. Trying more paths quickly isn’t the same as keeping the batches small, and More people, more agents, same plan makes the related point that work has to be partitioned before more hands help. Fast without gates is just fast, which I argued in Fast or good was never a law, and The end of software engineering lasted six days looks at a benchmark where earlier mistakes stay in the code.

Who I’d hire, then

Something else has to sort people if the pile stops doing it. I’d sort on judgement. What to build, what not to build, and where the model is wrong.

Figure 4 is the subtitle, literally. It has four invented candidates with no names, each scored out of ten for throughput and for judgement. Weight judgement more and the order changes. With AI on, the throughput gap between them shrinks.

0%
  1. 1Candidate Athroughput 9, judgement 4, score 9
  2. 2Candidate Bthroughput 7, judgement 7, score 7
  3. 4Candidate Cthroughput 3, judgement 9, score 3
  4. 3Candidate Dthroughput 5, judgement 5, score 5

Candidate A is first. Throughput spread: 6 points with AI off, 3 with AI on.

Figure 4. The leaderboard flip. Four invented candidates, scored by a weighted mix of throughput and judgement. Turning AI on lifts each throughput part of the way to the top, so the spread shrinks. The numbers are mine, not data.

A hiring manager at Anthropic, the company that makes Claude, wrote on 2026-10-27 that he stopped giving the traditional coding interview. He says the model beats candidates at those small challenges and he now looks for small prototypes that show taste, empathy and principled thinking; execution, he argues, is no longer the expensive part [Rieseberg (essay, by a hiring manager at Anthropic, which makes Claude)]. He also describes becoming the old-timer himself, the subtitle from the other side: “Somewhere in that year, I had become the old-timer.” [Rieseberg (essay, by a hiring manager at Anthropic, which makes Claude)] That’s one hiring manager’s essay, and a vendor-adjacent one.

That isn’t hiring a neurotype. It changes what the screen measures, so the people the pile filtered out get through. On how to think about it, the best source I found is Nancy Doyle’s review: neurodivergent people often have a spiky profile, and the fit between person and job matters [Doyle 2020].

Doyle advises asking individuals what fits rather than assuming, and her review calls the “superpower” story weakly supported [Doyle 2020]. I wouldn’t build a rule on a diagnosis.

Limits and counter-evidence

Counter-evidence

What pushes against me:

  • A review of 31 studies found more divergent thinking with high subclinical ADHD traits, but not in people with the clinical disorder [Hoogman et al. 2020]. So the strengths story is thinner than the essays suggest.
  • ADHD goes with starting ventures, not with succeeding at them. One study of 9,869 people says it can’t speak to venture performance [Lerner, Verheul and Thurik 2019]. Another found inattention linked to less entrepreneurship [Wiklund et al. 2017].
  • Research on neurodiversity at work is thin. In a 2023 focal article, the authors say surprisingly little research bridges neurodiversity and organizational diversity [LeFevre-Levy et al. 2023].
  • AI may be cutting entry-level hiring in exposed jobs. A working paper puts employment of workers aged 22 to 25 in AI-exposed occupations 19% below where it would be, and says most of that is less hiring. The authors call it early and descriptive, not causal [Brynjolfsson, Chandar and Chen (working paper, preprint)].
  • Optimism cuts both ways. It helps people keep engaging with problems [Carver, Scheier and Segerstrom 2010]. Moderate optimists choose prudently, extreme ones don’t [Puri and Robinson 2007]. Optimistic plans run late, but the bias shrinks when people connect past experience to the forecast [Buehler, Griffin and Ross 1994].

I can’t wave away the entry-level one. The hopeful answer I found comes from an engineering manager, not a study: juniors’ fresh eyes are an advantage, he says, and the expensive part of the job has moved [Rieseberg (essay, by a hiring manager at Anthropic, which makes Claude)]. It answers the worry. It doesn’t cancel the paper.

So the optimism I’m claiming comes with checks. Don’t trust the feeling of speed: measure it. Traces, checkpoints and receipts shows what proof that a check passed looks like, and What gets cheap, and what doesn’t says why I think receipts matter.

My view

My view

I’m optimistic, on the same conditions as before: say what done looks like, read what came back as if someone else wrote it, and write it down. Some of the people the pile held back are good at the first and the third. Hire for that. The loop behind the first is in One loop, three models, where the review step reads the whole diff as if someone else wrote it, and the third has its own chapter, Writing back.

The pile shrank. That isn’t the whole job. It was never the job.

In my words

For hiring managers, I think the main takeaway is that their pattern recognition needs to be rewritten to identify a completely different set of patterns.

Related on this site

  • Who to hire nowThe engineer worth having on a team has changed because of LLMs, agent graphs and second brains: my points from other Lab pages, gathered and linked.
  • Attention is the bottleneckThe scarce thing in a team that builds with agents is a person’s attention: a vendor’s design for an always-on agent, taken apart with my own Lab pages.
  • What gets cheap, and what doesn’tWhen execution gets cheap, where the margin moves: a model’s answer taken apart with my own Lab pages, and my claim about guarantees and service contracts.

Sources

  • Dylan, “The Times They Are a-Changin’” · The Times They Are A-Changin’, official song page; the subtitle is its one line on this page. Bob Dylan, bobdylan.com (the album of 1964); the official lyrics page; read 2026-11-04.
  • Perl glossary (laziness) · perlglossary, the entries “laziness”, “impatience” and “hubris”, derived from the glossary of Programming Perl, 4th edition. Christiansen, foy, Wall and Orwant (the book); the Perl documentation, perldoc.perl.org; the official Perl documentation; read 2026-11-04. The printed book was not read.
  • xkcd 1205, Is It Worth the Time? · Randall Munroe, xkcd 1205: how long you can spend making a routine task more efficient before you lose time overall (linked, not copied). Randall Munroe, xkcd; the comic; read 2026-11-04.
  • Grupenhoff (essay) · Judy Grupenhoff, “If You Think In Systems and See Patterns that Others Don’t, This Is Your Moment”, 2026-05-30. An opinion essay that cites no research. Judy Grupenhoff, Human Transition (Substack); the essay in full; read 2026-11-04.
  • Glouberman and Zimmerman 2002 · Glouberman and Zimmerman, “Complicated and Complex Systems: What Would Successful Reform of Medicare Look Like?”, Romanow Commission discussion paper 8 (credited for the distinction only). Sholom Glouberman, Brenda Zimmerman, Commission on the Future of Health Care in Canada; the catalogue record only; the PDF text was not read; read 2026-11-04.
  • Snowden and Boone 2007 · Snowden and Boone, “A Leader’s Framework for Decision Making”, Harvard Business Review, November 2007 (credited for the Cynefin framework only). David J. Snowden, Mary E. Boone, Harvard Business Review; the venue page only; the article is paywalled and was not read; read 2026-11-04.
  • Camp (personal post) · James Camp (@JamesonCamp), X post of 2026-11-03: his own account of AI and his ADHD. One person’s self-report. James Camp, X; the post in full, read in a browser; read 2026-11-04.
  • Palantir fellowship (company page) · Palantir Technologies, “Neurodivergent Fellowship”, a job posting. The company’s own claims; its pay range is not used. Palantir Technologies, Palantir’s careers page; the posting in full; read 2026-11-04.
  • Fortune on Karp, 2025 (as reported by Fortune) · Fortune, “Alex Karp credits his dyslexia for Palantir’s success”, 2025-12-03, reporting his remarks at the DealBook Summit; the video was not watched. Lily Mae Lazarus, Fortune; the article; read 2026-11-04.
  • Fortune on Karp, 2026 (as reported by Fortune) · Fortune, on Karp’s remarks on TBPN about neurodivergent people and AI, 2026-10-25; the recording was not found. Marco Quiroz-Gutierrez, Fortune; the article; read 2026-11-04.
  • Noy and Zhang 2023 · Noy and Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence”, Science 381 (2023). Shakked Noy, Whitney Zhang, Science; the abstract and the working paper’s introduction; read 2026-11-04.
  • Brynjolfsson, Li and Raymond 2025 · Brynjolfsson, Li and Raymond, “Generative AI at Work”, Quarterly Journal of Economics (2025); the journal’s figures, not the working paper’s. Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond, Quarterly Journal of Economics; the published abstract; read 2026-11-04.
  • Dell’Acqua et al. 2026 · Dell’Acqua and others, “Navigating the Jagged Technological Frontier”, Organization Science 37(2) (2026). Fabrizio Dell’Acqua and eight others, Organization Science; the published abstract (through OpenAlex); read 2026-11-04.
  • DORA 2024 · DORA (Google Cloud), Accelerate State of DevOps Report 2024: survey associations, not controlled experiments. DORA, dora.dev; the report’s chapter on AI’s downstream impact; read 2026-11-04.
  • METR 2025 (preprint) · Becker, Rush, Barnes and Rein (METR), “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”, arXiv:2507.09089 (preprint). Joel Becker, Nate Rush, Beth Barnes, David Rein, arXiv; the abstract and METR’s post; read 2026-11-04.
  • METR 2026 update (preprint-level) · METR, uplift update of 2026-02-24: METR’s own post, not peer reviewed. METR, metr.org; the post in full; read 2026-11-04.
  • UK DBT 2025 (government report) · UK Department for Business and Trade, evaluation of the Microsoft 365 Copilot pilot, 2025-08-28; not peer reviewed, no counterfactual. Department for Business and Trade, GOV.UK; the full text; read 2026-11-04.
  • Brynjolfsson, Chandar and Chen (working paper, preprint) · Brynjolfsson, Chandar and Chen, “Canaries in the Coal Mine?”, Stanford Digital Economy Lab working paper, the version dated August 2026 (early, descriptive indicators). Erik Brynjolfsson, Bharat Chandar, Ruyu Chen, Stanford Digital Economy Lab; the abstract and introduction; read 2026-11-04.
  • Rieseberg (essay, by a hiring manager at Anthropic, which makes Claude) · Felix Rieseberg, “Hunger beats tenure: Advice for junior engineers”, 2026-10-27. An opinion essay by a hiring manager at Anthropic, the company that makes Claude, so vendor-adjacent. Felix Rieseberg, felixrieseberg.com; the essay in full; read 2026-11-04. His own linked sources were not read and are not cited.
  • Doyle 2020 · Doyle, “Neurodiversity at work: a biopsychosocial model and the impact on working adults”, British Medical Bulletin 135 (2020). Nancy Doyle, British Medical Bulletin; the abstract and the full text on PubMed Central; read 2026-11-04.
  • Hoogman et al. 2020 · Hoogman and others, “Creativity and ADHD: A review of behavioral studies, the effect of psychostimulants and neural underpinnings”, Neuroscience and Biobehavioral Reviews 119 (2020). Martine Hoogman, Marije Stolte, Matthijs Baas, Evelyn Kroesbergen, Neuroscience and Biobehavioral Reviews; the abstract (through Europe PMC); read 2026-11-04.
  • Lerner, Verheul and Thurik 2019 · Lerner, Verheul and Thurik, “Entrepreneurship and attention deficit/hyperactivity disorder: a large-scale study involving the clinical condition of ADHD”, Small Business Economics 53 (2019). Dina Lerner, Ingrid Verheul, Roy Thurik, Small Business Economics; the full text; read 2026-11-04.
  • Wiklund et al. 2017 · Wiklund, Yu, Tucker and Marino, “ADHD, impulsivity and entrepreneurship”, Journal of Business Venturing 32(6) (2017). Johan Wiklund and three others, Journal of Business Venturing; the abstract (through EconPapers); read 2026-11-04.
  • LeFevre-Levy et al. 2023 · LeFevre-Levy and others, “Neurodiversity in the workplace: Considering neuroatypicality as a form of diversity”, Industrial and Organizational Psychology 16(1) (2023). Rose LeFevre-Levy and four others, Industrial and Organizational Psychology; the abstract; read 2026-11-04.
  • Carver, Scheier and Segerstrom 2010 · Carver, Scheier and Segerstrom, “Optimism”, Clinical Psychology Review 30(7) (2010). Charles Carver, Michael Scheier, Suzanne Segerstrom, Clinical Psychology Review; the abstract (through Europe PMC); read 2026-11-04.
  • Puri and Robinson 2007 · Puri and Robinson, “Optimism and economic choice”, Journal of Financial Economics 86(1) (2007). Manju Puri, David Robinson, Journal of Financial Economics; the NBER abstract and Duke’s release; the journal’s abstract was not reachable; read 2026-11-04.
  • Buehler, Griffin and Ross 1994 · Buehler, Griffin and Ross, “Exploring the ‘planning fallacy’”, Journal of Personality and Social Psychology 67(3) (1994). Roger Buehler, Dale Griffin, Michael Ross, Journal of Personality and Social Psychology; the abstract; read 2026-11-04.
  • Cunningham 1992 · Ward Cunningham, “The WyCash Portfolio Management System”, OOPSLA ’92 experience report, on his own site. Ward Cunningham, OOPSLA ’92; Cunningham’s own copy; the ACM page was not read; read 2026-11-04.
  • Cunningham 2009 · Ward Cunningham, “Debt Metaphor” (video, 2009), as transcribed on his wiki under CC BY 3.0. Ward Cunningham, c2 wiki (transcript of a YouTube video); the transcript; the video page was not read; read 2026-11-04.
  • Agile Manifesto principles · Principles behind the Agile Manifesto, agilemanifesto.org (2001). Beck, Cunningham, Fowler and fourteen others, agilemanifesto.org; the principles page; read 2026-11-04.
  • Beck, Overload interview 2000 · Nicolai Josuttis, “eXtreme Programming: An interview with Kent Beck”, Overload 8(35), January 2000 (Beck’s book was not read). Kent Beck, interviewed by Nicolai Josuttis, Overload (ACCU); the interview; read 2026-11-04.

Sources as read on 2026-11-04.