Programming in the Era of AI

Daniel Lemire, professor
Université du Québec (TÉLUQ)
Montréal 🇨🇦

blog: https://lemire.me

X: @lemire · GitHub: github.com/lemire

Normal vs. revolutionary science

Thomas Kuhn, The Structure of Scientific Revolutions (1962)

  • Normal science: puzzle-solving inside an accepted paradigm. Shared tools, shared methods, shared idea of what counts as a good answer.
  • Revolutionary science: anomalies pile up, the old paradigm cracks, a new one replaces it.
  • After a revolution, the old experts are not always the best guides.

Is programming in a period of normal science, or of revolution?

Where I am coming from

  • Author of high-performance libraries integrated into major browsers, runtimes and standard libraries: simdutf, fast_float, simdjson, Roaring Bitmaps.
  • Among the top 2% of scientists worldwide (Stanford/Elsevier), 100+ peer-reviewed papers.
  • Among the 1000 most-followed developers on GitHub.
  • And: I have not typed most of my code by hand in over a year.

Preface

Back in the 1990s

The singularity

« I believe that the creation of greater than human intelligence will occur during the next thirty years. (I'll be surprised if this event occurs before 2005 or after 2030.) »

Vernor Vinge, 1993

Where we actually are

A world of compute

« software is eating the world »

Andressen, 2011

A world of AI

« Software is eating the world, but AI is going to eat software. »

Jensen Huang (Nvidia)

SWE-bench Verified

  • Real GitHub issues from real open-source Python repositories.
  • The model gets the repository, the issue text, and nothing else.
  • Success = the project's own hidden tests pass afterwards.
  • 500 tasks, human-validated as solvable.

Terminal-Bench 4.0

  • The agent gets a real terminal in a sandboxed container.
  • Tasks: build, configure, debug, migrate, train... whatever the shell allows.
  • Success = automated tests check the final state of the machine.
  • 66 hard tasks, up to 8 hours each (August 2026).

Human experts scored about 50% on Terminal-Bench 2.0.
There is no human baseline for 4.0.

January 1st marked the beginning of the singularity.

—Stripe, August 19, 2026

People often talk about this concept called AGI, meaning artificial general intelligence, that is, an AI as intelligent as a person. I actually think we crossed that threshold about three months ago.

— Marc Andreessen, May 19, 2026

A summer of AI optimization

  • Six mature open-source libraries: roaring, CRoaring, ada, fast_float, simdjson, simdutf.
  • Shipped in GCC, Chromium, Node.js... used by millions.

--- c

Are we all going to be out of a job?

Jevons paradox

  • 1865: William Stanley Jevons, The Coal Question.
  • James Watt's steam engines burned far less coal per unit of work...
  • ...and Britain's coal consumption went up, not down.

When something gets cheaper to use, we use much more of it.

  • Tokens got cheaper: we consume far more tokens.
  • Code got cheaper: we write far more code, and review, test, host and maintain it.

How does it work

How an LLM works

Virtuous circle

  • AI helps us build better, faster software
  • Better, faster software helps us build better AI

How can a language model write software?

  • I write a query, it is maybe 80 characters.
  • I get back a full program with 100,000 lines.
  • Not how it works, there is more information invovled.

The inference loop

Reinforcement learning

Jagged: metacognition

Jagged: sensorimotor skill

Moravec's paradox

« It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility. »

Hans Moravec, Mind Children, 1988

Jagged: visual-spatial reasoning

Jagged: long-term memory

Human attention is limited

arXiv capped submissions at two a month

  • October 1, 2026: two submissions per person per month.
  • September 2026: over 40,000 papers, up from about 20,000 two years earlier.
  • Doubling every two years. Moderated by volunteers.
  • Google stopped accepting new bug reports in its open-source bounty program: they could not cope with the influx.

lemire.me/blog/2026/10/04/arxiv-capped-submissions-at-two-a-month/

Emerging ideas

Emerging idea: ephemeral testing

  1. Build a component (by hand or with an agent).
  2. Ask an agent to build throwaway software on top of it: an app, another layer.
  3. Have it test what it built.
  4. Judge the foundation by how well the layers above work. Then discard them.

lemire.me/blog/2026/10/05/ephemeral-testing/

Emerging idea: a layered model

  • Keep a small core that changes slowly and on purpose.
  • For the core, you read the code. You insist on tests. No vibe coding.
  • Everything else can move fast. There will be bugs, but the AI fixes them quickly.
  • The core cannot depend on the outer layers.

lemire.me/blog/2026/09/05/ai-programming-a-layered-model/

Emerging idea: AI cloning

Business challenge

  • Once your application is public, people can clone it.
  • Meaning of copyright changed. Think fashion industry, cooking.

Emerging idea: cheap custom software

  1. Build your own video software https://github.com/lemire/lemirecut
  2. Your own election-night app https://lemire.me/montmorency2026/

Business opportunity

  • Help people deploy software specific for their needs, even if just 3 users.
  • Build throw-away application (short lived): E.g., an application for one event

Emerging idea: AI-first operating software

Summary

  1. It is real: mature libraries 2–5× faster in one summer.
  2. Cheaper code means more code (Jevons), not less work.
  3. The skills are jagged: superhuman at some tasks, weak at self-doubt, touch, space and memory.
  4. Human attention is the bottleneck: review, moderation, trust.
  5. New practices are emerging: ephemeral testing, a small stable core, custom software for a few users.

Questions?

Daniel Lemire — lemire.me

🇨🇦

credit : @fromdevoid @birkenkrahe @gcomeaudotcom @JoyfulBadger @nabendu82 @MatRopert @dyaroshev @opdroid1234 @jrdi @Hassan_Abedi @NinadSachania @RonanTournier