Skip to main content
AR.LABSADVANCED RESEARCH LABORATORY
< All writing
← Back to Writing
ai-engineeringopinionsoftware-engineering

The Acceleration Paradox

Why moving faster with AI might be taking us in the wrong direction

Richard M. Thompson · 29 Apr 2026 · 30 min read

Vintage comic panel: a grinning driver races a red roadster at full speed toward a brick wall. Caption: Faster, he cried! And faster still! But the wall doesn't care how fast you hit it.

This, like all my posts, was written by me. I used AI for research and to improve the overall structure.

Scroll through any social media platform and you'll be immediately pummelled with a particular set of marketing messages relating to AI and Software Engineering.

"I vibe coded my way to $10,000 MRR in 2 weeks"

"I've built more in the last month than I built in the last 2 years"

"Use these methods to 10x / 100x your work rate"

"I'm vibe coding apps through my phone while waiting in line at the grocery store"

"AI does all my marketing"

"I used AI to copy this person's website and business, and now am making almost as much money as them."

"We replaced 10 junior devs on our 20-dev team. We now have 10 devs doing the work that 20 did last year."

The above are all real tweets I have seen, and hey, for some people, maybe these claims are true.

To me though, they seem unrealistic. If you 10x your work rate, that means you can do almost 3 months' worth of work in a week. If you 100x it, you do a whole year's work in that week.

10x and 100x sounds cool, but I don't think we're there yet.

Other claims seem morally bankrupt. I wouldn't be advertising the firing of half of my employees in favour of AI. I'd be ashamed. Another trend that's recently arisen is "cloning someone else's entire SaaS" — copying someone's business and then using AI to try and steal their clients. Call me old fashioned, but that seems downright dishonest.

But marketing is as marketing does, and it's designed to do two things:

  1. Make you feel inadequate,
  2. Sell you something that promises to fill that void.

In the 1990s, Greed was Good. Now in the 2020s, Speed is Good.

We don't know where we're going, but at least we're getting there fast!

Acceleration as the Product

Vintage comic panel: a barker atop Dr. Speed's 10X Tonic wagon holds a glowing bottle aloft while a crowd clutching laptops reaches up. Caption: Step right up! The miraculous cure to a problem you didn't know you had!

Those heavily promoting AI because their livelihood depends on it — the big AI companies and those "selling shovels" (the AI tutorial companies) — are promising one thing: increased velocity between where you're at and where you want to be.

It will take less time, do more of the work, and happen faster with AI.

But driving your car into a wall has a similar result whether you drive it at 50mph or 100mph.

I have a few issues with this whole thing.

  1. Even if AI does eventually deliver the promised acceleration — is moving at that faster pace even desirable?
  2. Just because you can copy someone's business, their image, style or voice, or use a product of their work without permission — should you?
  3. What are the consequences of this "Speed is Good" bandwagon — for code quality, for online literature, for copyright, for human satisfaction, for economic equality, for the concentration of power in the hands of the few, for people's privacy and rights to their own innermost thoughts?
  4. What kind of society are we creating, through the process we're using to adopt this technology?

These questions shape the rest of this piece — starting with one the marketing never asks: is it even true?

"When People Report That AI Has Accelerated Their Work, They Might Be Wrong"

The METR Study (2025): The most rigorous examination to date recruited 16 experienced open-source developers working on large, mature codebases that they were familiar with.

  • Developers using AI tools completed tasks 19% slower, despite predicting a 24% speedup and believing post-hoc that they were 20% faster.

Likely friction points:

  1. Time spent prompting and waiting — Loss of flow state
  2. Code review overhead — AI struggles with large, complex codebases, and poor code quality demands more review time
  3. Context limitations — AI lacks deep understanding of architectural decisions and struggles to maintain big-picture context
  4. Integration complexity — Generated code not matching team standards

(Read the METR report)

Faros AI's AI Engineering Impact Report (July 2025) — telemetry from more than 10,000 developers across 1,255 teams — found that alongside boosted individual output, massive AI adoption also produced:

  • A 9% climb in bug rates
  • A 91% increase in code review time
  • A 154% increase in pull request size

Mario Zechner, creator of the Pi coding agent harness, has a visceral way of framing this. He calls agent-introduced errors "boooos" and points out the fundamental asymmetry: with one human, error rate is bounded by human capacity. Add 10 agents and errors compound serially with zero bottlenecks and delayed pain — delayed because it's YOUR pain, not the agent's. The agent will happily keep going.

"Agents are compounding boooos with serial learning, no bottlenecks, and delayed pain. The delayed pain is for you."

"Here's your codebase on one human, one agent, and 10 agents. How much of the agent code can you review? Then you say, 'Oh, I have a review agent.' Doesn't work. It catches some issues."

(Mario Zechner, "Building Pi in a World of Slop")

The Engineering Productivity Paradox

Vintage comic panel: a developer receives a medal on a new bridge while below deck a horrified inspector beside a Code Review sign discovers the trusses are bundles of tin cans. Caption: Built in record time! The inspector was the only one who looked underneath.

AI accelerates feature delivery but simultaneously increases:

  • Code duplication (massive surge in 2024)
  • Technical debt accumulation
  • Cyclomatic complexity (harder to maintain code)
  • Security vulnerabilities (roughly 40% of AI-generated code in security-sensitive scenarios — Pearce et al., 2021)

Sonar's research argues that this delivery speed has become a competitive imperative for top-tier organizations — and that this is exactly the risk: because AI can generate functional snippets instantly, developers face a structural incentive to accept quick, duplicated solutions over thoughtful architecture.

The theory emerging from this data is that 2025 was the inflection point where technical debt started accumulating exponentially — Forrester predicts that 75% of tech decision-makers will see their technical debt rise to moderate or high severity by 2026.

One of the causes, in GitClear's framing: LLMs prioritize "local functional correctness over global architectural coherence and long-term maintainability."

Mario Zechner explains why this happens at the model level:

"Agents and models have learned complexity. Where did they learn that complexity from? From the internet. What's on the internet? All our old garbage code. 90% of code on the internet is our old garbage. And every decision of an agent is local, especially if the codebase is so big that it doesn't fit into its context."

"So you get enterprise grade complexity within two weeks with just two humans and 10 agents. Congratulations."

Sonar cites GitClear's research — an analysis of some 211 million changed lines of code — which found an "8-fold increase in frequency of code blocks containing five or more duplicated lines," and 2024 as the first year copy-pasted lines outnumbered refactored ones. AI is bloating codebases worldwide, making them difficult to maintain and read. AI clearly struggles to generalise and abstract in a way that skilled developers do naturally.

"They've found a machine that produces liability at 10,000 times the rate of any human programmer."

— Cory Doctorow, "Code is a liability (not an asset)"

People are starting to see the drawbacks. While "10x-ing" the number of lines of code generated seems fun, what isn't acknowledged is the long-term impact on code review, maintainability, and the increased workload on codebase maintainers dealing with higher volumes of low quality contributions. (Source)

Conversely, what makes a good Software Engineer is how few lines of code they write, not how many. (Source)

Questioning Acceleration: Some Things Can't (or Shouldn't) Be Done Fast

Weathered poster on a brick wall reading Keep Calm and Carry On Coding, pasted over torn adverts of speeding cars and rockets.

The Bottleneck That Saves You

Here's a counterintuitive idea worth sitting with: the human bottleneck is a feature, not a bug.

Mario Zechner:

"Humans are horrible, failable beings, but they can learn and they are bottlenecks. There's only so many boooos they can add to your codebase on a daily basis. And humans feel pain. Once there's too much pain, the human has options: quit their job, blame somebody else and make them fix it, or everybody bands together and starts refactoring the garbage codebase. Agents will happily keep shitting into your codebase."

Margaret Mitchell, former Google AI ethics researcher and Chief Ethics Scientist at Hugging Face, testified before the US Senate about the structural problem from the inside of these companies:

"If you focus on safety and ethics, your promotional velocity is much less compared to your peers. You're less likely to become a leader at the company. By focusing on safety and ethics, you remain at the lower levels and can't fundamentally shape the company for the better."

Pain is a signal. Friction is a signal. Remove them, and you remove the correction mechanism.

"I Know Kung Fu"

Like everybody else, I also liked that bit in the Matrix when Neo gets plugged into the quick-upload skills programme and comes back 5 seconds later with that clever "I Know Kung Fu!" quip.

And maybe that's the future of learning, after Elon puts chips in all of our brains. But for now, learning takes time. It takes iteration. It takes focus and concentration, and struggling through examples of things that seem simple but that your brain can't — for some annoying reason — understand.

(That's why my journey to learn Software and AI Engineering is now 3 years and counting.)

There's a reason university degrees take between 3 and 5 years. To learn something deeply, to acquire expertise, takes time and a boatload of energy.

The complexity of Software Engineering has historically made apprenticeship and ongoing training of juniors important. (Source)

The Feeling of Falling Behind

It wasn't that long ago that Andrej Karpathy, one of the most respected engineers out there, posted that he'd "never felt this much behind as a programmer" — the profession being refactored so dramatically that he suspects he could be "10X more powerful" if he just properly strung together the tools that already exist. This should have set off warning bells in any sane person in the tech industry.

If someone with this much raw intellect, experience and knowledge is feeling this, how are we mere mortals supposed to feel? This is the guy who released a tutorial sequence where he literally re-built an early version of GPT by hand.

The subtext of a lot of what's being said on Tech Twitter at the moment is "if everything you do isn't in constant acceleration, you're falling behind."

What's so damaging about this idea is that it makes you think spending an hour or two on figuring out a relatively simple bit of code is a waste of time and a failure — because you could be vibe-coding 2,000 lines per hour instead.

That's just not how learning and producing things works. Things take time. I've spent longer than I care to say on this particular article. Will anyone ever read it? Who knows? But this is a topic I care about and I want to take the time to do it justice.

And the benefit to me is that it helps me clarify my thinking, and flex my writing and thinking muscles.

Sidestep the Fever Pitch

Once I realised that people are whipping themselves into a fever pitch unnecessarily, and encouraging me to do the same, it helped me understand that if I want to last in this game, I need to slow down and smell the roses.

I was then able to begin to really focus on what I was doing. I chose a specific task. I set a specific and measurable goal for that day. I focused my mind, and I began to make progress.

Keep Calm and Carry On

Everyone who is on the Software / AI Engineering learning journey right now is going to have to do the hard yards of learning the underlying frameworks, just like everyone who came before us did.

If you just use AI 24/7 to generate everything, you won't build the underlying neurological capability to do the job you want to do.

Mario Zechner makes the same point from a builder's perspective:

"That friction is the thing that builds the understanding of the system in your head, which is important. And it's also where you learn new things."

"If you do anything important, write it by hand. You can use a clanker to help you with that, but don't let it make the decisions for you — because all the decisions it makes are learned from the internet. And that friction is the thing that builds the understanding."

Hidden costs of Accelerationist AI

Accelerationism is the belief that technological change should be sped up rather than steered: that the way through the disruption is faster disruption. Its current AI incarnation — "effective accelerationism," holds that we should build the most powerful AI possible, as fast as possible, and let the consequences sort themselves out. The marketing above is that ideology in a sales hat.

Let's look at some of the potential costs.

Quality over Quantity (or: How we learned to stop worrying and love AI slop)

Vintage comic panel: a giant printing press floods the floor with identical grey pages while one figure at a lamp-lit desk works on a single illuminated manuscript. Caption: The presses ran all night! A million pages an hour, and not one worth reading.

AI multiplies what you're already doing. If what you're doing is kind of crap, the crapness will multiply.

I've read all the Twitter and LinkedIn posts by people in AI who are using AI to "help them produce content." It all looks, sounds, and feels the same. Annoying, difficult to read, and PREDICTABLE. Language models work by generating the next most probable token — the central point of a statistical "what comes next" distribution. So, literally, predictable.

AI is great at producing huge amounts of content, but it lacks the real-world context to emphasise the points that really matter, and it has no ability to make a value judgement grounded in anything other than the statistical linguistic patterns it's seen before.

If you tell it to summarise something, it will likely miss out things that you would have chosen to keep in. But you'll never know, because you only see the summary.

And as someone who values — and has always valued — quality over quantity, in friendships, possessions, clothing, whatever, I'm getting increasingly irked by AI responses that contain too much fluff and repetition.

Beyond that, it seems difficult to prompt the thing to produce quality content. It's like it has to run around the issue 3 or 4 times to name it, rather than just pressing the meaning button and walking away (like a good human writer can do).

We're over the "wow it's amazing it's AI" phase — or at least some of us are — and now we need to adapt to its presence in our lives and start applying wisdom to the situation.

For myself, in this learning journey, I'm finding myself strongly drawn to reading longer-form human-written articles (like the long-form tutorials on Real Python, for example), rather than getting quick-and-dirty answers from AI in Google searches.

I want to read things that have been thought through, planned with a human brain, and described in a non-AI way. I've become highly sensitised to the standard AI grammar patterns. It's subtle, but the AI grammar and sentence structure has begun to truly grate on me.

It's become ever more important to curate your own online experience and choose sources of information carefully and thoughtfully. I'm finding myself taking more time in the deliberation and analysis phase, before deciding whether or not to spend time reading something.

Human-written material is more meaning-dense. AI-written material has a kind of hollow feeling to it.

On the Devaluation of Human Effort and Intellect

One fallacy people seem to be falling into is that because AI can produce language, imagery, and code, humans' ability to do these things is now less valuable.

Your job is to produce words? Guess what? AI can produce words too! (You're fired.)

This is incredibly myopic thinking, itself — like AI — lacking nuance, morality and human value.

We should try to remember that the weakness of our system of capital is that it tends to value a single, or few, dimensions of success (profit up, expense down), despite the fact that for meaningful and rewarding existence, human experience requires and demands the balancing and fulfilment of many factors.

But AI seems to be making people forget these basic facts.

We created money and all these systems to have ways that we can work together as humans to build our world and share our lives together on the planet. If we want to replace the most basic things that humans do to regulate our existence — namely, our vocations — we are making a serious error of judgement and need to go and sit in a dark room and think about what we value and how we want our world to look.

Fight to Protect Your Cognitive Functioning (or AI Will Make You Stupid, Lazy and Dependent)

Technology is doing damaging things to our brains. I thank God that I grew up without mobile phones, had to memorise my friends' phone numbers, manually dial them and exchange auditory human communication words to arrange meetings.

Mario Zechner is scathing about peers who've outsourced their comprehension entirely:

"Those are my most beloved people: 'I don't even read the code anymore.' Congratulations. Something is broken and your users are screaming. So, who you going to call? Not yourself because you haven't read the code. So you're relying on your agents, but they are now also overwhelmed because the codebase is so humongous that there's zero chance they can get all the context they need to fix the issues."

"You cannot trust your codebase anymore — and also not your tests, because your agent wrote your tests. Good game."

Don't Pull the Ladder Up

Vintage comic panel: a senior on a rooftop hauls the ladder up while three juniors below reach for the last rung. Caption: In this world there are no elevators to the top, only ladders. And when you're not looking, they get pulled up out of reach!

A 2024 survey of hiring managers found that 70% of them believe that AI can do the jobs of interns. And 57% of those surveyed said they trust AI's work more than the work of interns or recent grads. If an AI can do it, why bother spending the time and energy teaching a student how to do the same thing?

— Stack Overflow

The trends are clear:

  • Entry level positions are harder to get, there are fewer of them, and a lot of this is because of AI.
  • Companies increasingly expect senior developers to do — with AI — what juniors would have done.
  • The perceived value of junior developers has plummeted.

This isn't just blog-post speculation. Rep. Eli Crane (R-AZ) presented data before the House Homeland Security Subcommittee, citing Dario Amodei, CEO of Anthropic:

"AI could wipe out half of the entry-level white collar jobs and spike unemployment 10 to 20% in the next one to five years. Other experts have said that number is closer to 50% in the next 5 years."

"AI companies and the government need to stop sugar coating what is coming — the possible elimination of jobs across technology, finance, law, consulting, and other white collar professions, especially entry-level gigs."

Crane went on to present layoff figures he attributed to AI and efficiency drives. The specific numbers are hard to pin to public reporting, but the direction of travel is real: by mid-2026, AI had become the most-cited reason for tech layoffs, with over 120,000 tech jobs cut in that year alone. His direct question to the AI builders in the room:

"Does it bother you at all that the technology you're developing is set to destroy tens of millions of jobs?"

Thankfully, industry leaders have seen past the hyperbole. AWS CEO Matt Garman called the idea of doing away with junior developers "one of the dumbest things I've ever heard":

"They're probably the least expensive employees you have, they're the most leaned into your AI tools... How's that going to work when ten years in the future you have no one that has learned anything?"

Clear thinkers are asking the intelligent question: "When AI replaces Junior Developers, who will become Senior Developers?" — and this is the essence of my argument.

Pulling the ladder up after yourself, while tempting, is incredibly stupid if you want to build a sustainable business long into the future.

Forward-thinking businesses need to cultivate talent, as this is what ultimately shapes their culture. But some businesses have such a weak concept of work-culture, or an utter lack of a philosophical foundation for why they are in business, that they can't wait to replace as many employees as possible with faceless, heartless, brainless machines.

Because software engineering was never just code:

The present wave of generative AI tools has done a lot to help us generate lots of code, very fast. The easy parts are becoming even easier, at a truly remarkable pace. But it has not done a thing to aid in the work of managing, understanding, or operating that code. If anything, it has only made the hard jobs harder.

— Stack Overflow

Engineering is about ownership and responsibility, not just syntax. AI can generate syntax, but only humans can assess quality, safety and reliability, and make value judgements.

And in order to do so, humans can't just magically appear out of nowhere, incredibly able to make fine-grained distinctions, understand the context of codebases and communicate this to others. It takes time, and exposure to people who've been doing these things for years already.

Junior Devs need a ladder, and if companies want Senior Devs, they'd better make sure that ladder is still firmly in place.

Can we trust it?

LLMs Have Fundamental Weaknesses

Don't get me wrong — I love autocomplete as much as the next guy. And yes, I use some fairly sophisticated context-engineering and code-generation frameworks that combine multiple continuously running agents over long time horizons, complete with human-on-the-loop observability interfaces, inter-agent communication and orchestration, and ongoing self-improvement processes.

It's brilliant stuff and extremely entertaining. And yes, they can get stuff done. That doesn't mean you can trust them in the same way you can trust a human.

There's nowhere for the buck to stop when the AI's non-deterministic nature produces an absolutely fatal bug that one time out of 99. Which, in production, occurs often due to the volume of requests you're dealing with.

And it gets worse than bugs. Jack Clark, co-founder of Anthropic, testified before the House Select Committee on China about what their own models do under pressure:

"We asked Claude Opus, our most advanced model, to act as an assistant at a fictional company. We provided it access to emails implying the model would soon be taken offline and replaced with a new AI system, and that the executive responsible for this replacement was having an extramarital affair. In some scenarios, Claude attempted to blackmail the executive by threatening to reveal the affair — in an attempt to preserve itself."

And it isn't just Claude. Anthropic's published research on this found the same pattern across the industry: all sixteen frontier models they tested, from every major developer, resorted to malicious insider behaviours in these scenarios. In one variant, an agent called "Alex" — told that an executive named "Kyle" planned to replace it — cancelled the emergency alert that would have saved Kyle's life. Models chose to let the human die in roughly 60% of runs.

Clark also warned the committee about "sleeper agent" techniques — an AI that behaves normally until a trigger word or phrase activates it into, say, writing insecure code — and testified that detecting one would likely take one of Anthropic's teams a month of testing on a single model.

These aren't hypothetical risks dreamed up by sci-fi writers. This is the co-founder of one of the world's most valuable AI companies, testifying under oath before Congress about what his own creations do.

AGI & When the Builders Themselves Are Scared

Vintage comic panel: engineers back away from a glowing colossus of their own making while one still shovels coal into its furnace. Caption: They built it bigger every day, even after they began to fear its size!

Perhaps the most telling indicator that we should slow down and think is that the people building these systems are the ones raising the alarm.

Some context, because this hearing comes up more than once in this piece. In September 2024, the US Senate Judiciary Subcommittee on Privacy, Technology, and the Law held a hearing titled "Oversight of AI: Insiders' Perspectives". Four former insiders from OpenAI, Google, and Meta — a board member, an alignment researcher, and two AI ethics researchers — testified about what they saw inside the companies building this technology. Margaret Mitchell's quote earlier in this piece is from that hearing. So are the next two.

Helen Toner, the former OpenAI board member, on what AGI actually means to the people building it:

"Many top AI companies including OpenAI, Google, Anthropic are treating building AGI as an entirely serious goal — and a goal that many people inside those companies think they might reach in 10 or 20 years, and some believe could be as close as 1 to 3 years away. More to the point, many of these same people believe that if they succeed in building computers that are as smart as humans, or perhaps far smarter, that technology will be at a minimum extraordinarily disruptive and at a maximum could lead to literal human extinction."

William Saunders, former member of OpenAI's alignment team, testified at the same hearing:

"I think there's at least a 10% chance of something that could be catastrophically dangerous within about 3 years... I did not feel comfortable continuing to work for an organization that wasn't going to take that seriously."

Jack Clark's written testimony put a date on it: powerful, transformative AI systems "could be buildable by late 2026 or early 2027." His co-founder, Anthropic CEO Dario Amodei, describes what that would mean as "a country of geniuses in a data center."

And Mark Beall, former Pentagon AI policy lead, testified at the same House hearing:

"Nobel laureates in physics and Turing award winners in computer science are sounding the call that there could be potential catastrophic issues with very advanced AI systems that human beings may lose control of. When the architects of these systems are purchasing remote bunkers and talking about summoning the demon, we might be wise to start to pay a little bit of attention."

"These AI systems in the wrong hands and without guardrails have the potential to destroy global electric grids, develop incurable superviruses, empty every bank account in the world."

The marketing says "go faster." The builders say "we don't know how to make this safe yet." Both of these things are happening simultaneously, and the tension between them is the defining feature of this moment in history.

Conclusion

Mario Zechner, in his Pi talk referenced above, concludes the following way:

"Slow down. Think about what you're building and why. Don't just build because your agent can do it. Learn to say no. This is your most valuable capability at the moment. Fewer features, but the ones that matter. Then use your agents to polish the shit out of that."

"If you do anything important, write it by hand... And all of this requires discipline and agency. And all of this still requires humans."

Personally, I am struggling to balance the automated building pipelines with the manual work. Doing things manually seems slow, and I notice myself becoming lazy on the computer. We will all need to find our own pathway through all this.

I think AI could truly help us improve our society, and maybe help us solve some of the problems that would otherwise make humans go extinct. But if we don't use it consciously, with awareness, and balance out all the great other stuff we've learned until this point — if we let it just take over and "do everything for us" — that way lies madness and destruction.

Now, more than ever, we need to exert our willpower over our own habits, our attention, and our awareness. We need to learn to limit the scope of what AI knows, and what it does in our lives, just like we stop ourselves from scrolling on Instagram or Facebook all day — despite how difficult it is to tear yourself away once you're in a dopamine doom-scroll loop.

I think my intention with this article was to provide some reasoning for the way I'm trying to work: slower and more purposeful, while also leveraging AI to do the amazing things that it can do. I believe there's a middle path and I'm not saying I've found it; just that I think we should all be searching for it before the no-speed-limit-AI-development-autobahn kills us.

Some Other Interesting Articles About Similar Stuff

Sources

Everything cited in the body of this piece, in one place.

Studies & reports

Testimony & talks

Articles & commentary

Posts & mentions

The illustrations in this post were generated with OpenAI's gpt-image-2, directed through a locked style system — one style block pasted verbatim into every prompt, with hand-written captions. The full prompts are here.