Feb 17, 2026
3 min read
Artificial intelligence comes from code: Matt Shumer's warning about new goals in the lab
After the code was changed, the AI labs began to function like any other office.
In a viral and controversial text published on X, businessman Matt Shumer sounds the alarm.
We translate and write text messages according to the latest changes.
— Victor Storchan •
For several weeks now, an unusual echo has been happening in Silicon Valley.Artificial intelligence is no longer seen as an economic opportunity, but as an existential risk.The viral message published on widespread, commented and criticized, his text confirms that for software engineers the revolution is no longer a future theoretical projection, but an everyday reality.Presented as aimed at the "layman", the thesis was formulated in a deliberately alarming and prescriptive style that prompted accusations of fueling the "over-excitement" typical of the sector.
Beyond rhetoric, however, Schumer's diagnosis points to real structural dynamics within the industry.
The first important context for understanding the scope of this viral essay concerns the unique role of code as a pilot field for cognitive automation. The author's main point is that just because the AI revolution first affected programmers does not mean that it will not affect other professions in the same way. Unlike fields such as law or medicine, software development benefits from the epistemological proximity and scope of the model designer.themselves are experts in the code. This situation quickly established a tight communication chain between data generation, benchmarking, and performance improvement. Tasks such as compilation, unit testing, and review of pull requests are objective and the data sets are rich, helping to industrialize the evaluation. Therefore, more comparisons from real development cases help to speed up the iteration of models.
Industry indicators covered by sector analysisConfirm this change.Specialty newsletters such as SemiAnalysis describe a sharp increase in the number of pull requests written by agents like Claude Code, with the majority of this increase being the highest point in contemporary software production.Recent academic work has also shown that AI-generated contributions can achieve high adoption rates.If taken care of by a human1
While the progress of these advances in all professions certainly remains speculative and not automatic, the lab itself uses an incremental approach: it tries to reproduce, field by field, the conditions that enable the initial success of software development.The presence of sector standards – whether economic evaluations such as GDPEval or specialized tools such as Taubench – indicates an attempt to transfer the logic of the code to activities of strategic value, but little data structure.
This expansion also includes organizational change.Inspired by the Send Engineers model popularized by Palantir, major AI labs are now sending technical teams directly to their industrial partners and customer centers.The goal is to identify practical use cases, incorporate models of real environments, and recreate the feedback loops made possible by rapid code development.The problem is not only business: it is about understanding what data to collect, what metrics.to define, and how to evaluate performance in areas where direct validation is less.
From this point of view, the Sumerian text is little more than a sign of a period of technological change.
Si bien el estilo elegido —que combina observaciones empíricas, exhortaciones a la adaptación y promesas de aceleración— mantiene una ambigüedad, subraya una evolución real —la mayoría de los laboratorios de IA intentarán replicar lo que ha sucedido con el código en otros sectores— y pone de relieve una dificultad: cada vez es más difícil para el público en general comprender realmente los saltos cualitativos de los modelos y traducir las cifras de los benchmarks en ganancias reales de fiabilidad de los modelos.
Think again: It's February 2020.
If you paid attention at that time, you may have noticed that other people around you are talking about the virus that has spread abroad.Most of us don't pay attention at first.The markets were doing well, your children went to school, you went to restaurants, you shook hands, you booked plane tickets.In February 2020, if someone told you they were stocking up on toilet paper, you would think about spending a lot of time in front of a screen.Then, in about three weeks, the whole world changed.Your office is closed, your kids came home, and your life is reorganized in a way you wouldn't have imagined if it had been explained to you a month earlier.
I believe we are in the "seems to be over" phase of something much, much better than Covid.
I spent six years building startups specialized in AI and investing in this field.I live in this world.And I write the following lines for people around me who don't live there: my family, my friends, people I care about and who constantly ask me, "What is AI?"and get an answer that doesn't do justice to what's really going on.I still give them the modest version, the version where a cocktail party is thrown diplomatically.Because the honest version would give the impression that I have gone crazy.And for a while I told myself that this was reason enough to keep secret what was really going on.
But the gap between the kind of AI that is widely accepted and what actually happens is becoming very wide.
My loved ones need to know what's going to happen, even if it seems crazy.
I want to make one thing clear from the beginning: although I work in the field of artificial intelligence, I have practically no influence on what will happen, just like the vast majority of players in the sector.The future is being shaped by a surprisingly small number of people: a few hundred researchers at a few companies.OpenAI, Anthropic, Google DeepMind and a few others.Training, managed by a small team over a few months, can produce a system that changes the entire trajectory of technology.Most of us working in the field of artificial intelligence are building the foundation that we are not alone.We are watching this evolution just like you.Except we're close enough to the reactor to feel the ground shake beneath our feet.
And the time has come.Not in the "we should talk about this sometime" sense, but more like "this is happening right now and I need you to understand."
I know it's true because it happened to me
Here's what no one outside the small world of new tech has yet fully realized: The reason many in the industry are now sounding the alarm is because it's already happened to us.
We do not make predictions.We tell you what has already happened in our work and warn you that you are next in line.
For years, the improvement of AI has not stopped.There have been big events here and there, but each of the big developments has been separated so that you can participate when it is made.So, in the year 2025, new technologies for these models will be able to increase the speed of progress.Then it will be faster.And moreEach new feature isn't better than the last... It's better and shorter.time between the release of new features.
I used AI more and more and had to use it less and less.I have seen how to operate it, which I think requires my experience.
Then, on February 5, two major AI labs released new models that day: GPT-5.3 Codex from OpenAI and Opus 4.6 from Anthropic, creator of Claude, one of ChatGPT's main competitors.And I had a vision.Not like a switch... More like when you feel that the water has risen around you and is now up to your chest.
I am no longer needed for the technical work of my current job.I describe what I want to build, in plain English, and it shows up.Simple as that.When I say "shows up," I don't mean a draft that needs to be corrected.That is the finished product.AI, guides it, makes modifications.Now, I simply outline the result and move on.
I will give an example so you will understand how it works.
I tell the AI, "I want to build this app. Here's what you have to do, you have to look normal."It determines the user experience, the layout, everything." And he does it. Write thousands of lines of code. Then, and this part would have been unimaginable a year ago, he opens the app by himself. Click on different people. If you think the app looks like it or don't use the search button. For something, you go back at your own expense, revise and finish it, and then it comes back to me and says, "It's done."
Without exaggerating.This was my Monday this week.
But the thing that impressed me the most was the mod that came out last week (GPT-5.3 codecs).
He just didn't follow my instructions.He made wise decisions.For the first time in that model there was something resembling an experiment.to tasteThat inexplicable feeling of knowing what the right decision is that people always say AI will never do.As this model has it, or at least something similar, the distinction begins to lose meaning.
I've always been an early adopter of AI tools.But the last few months have surprised me.
These new AI models are not improvements.They are something completely different.
And here's why you care, even if you don't work in the field of new technologies.
AI Labs made a thoughtful decision.First, they focused on improving AI's ability to write code... because creating AI requires a lot of code.If AI can write that code, it can help create another version of itself.
My work started to change before theirs, not because they focused on computer engineers, but because it was a side effect of their original choice.
Now they have achieved their goal.And they pass on to everything else.
The experience that tech workers have had over the past year, seeing how artificial intelligence has evolved from a "useful tool" to "someone who does my job better than me," is an experience everyone will have.Law, Finance, Medicine, Accounting, Consulting, Writing, Design, Analysis, Customer Service.Not in ten years.The people developing these systems are talking about a year or five.Some say even less.And from what I've seen in the last two months, I think "less" is more likely.
This is the most criticized - and criticized - bias of Shumer's article.Although many economic models show that automation will have a great impact on work - a hypothesis that many companies and investors are betting on - nothing allows us to conclude, as Shumer does, that the professions he gives as an example will undoubtedly have a similar experience to developers.However, as the author points out, AI laboratories try to replicate what software development has done, focusing on more specific areas.
"But I tried the AI and it wasn't too bad"
I hear this all the time.I understand, because this was true before.
If you tried ChatGPT in 2023 or early 2024 and thought "he's figuring it out" or "it's not that impressive," you were right.
These early versions were really limited.They were hallucinating.They confidently said things that made no sense.
That was two years ago.In AI terms, it's ancient history.
The models available today are unrecognizable compared to the models available six months ago.
The debate over whether AI is “much better” or “successful,” which has raged for more than a year, is over.It's over.Those who persist in defending this research have never used current models, or are interested in reducing what happened, or evaluating the situation based on experiences in 2024 that are no longer relevant.I'm not saying this out of hatred.I say this because the gap between society and today's reality is huge, and the gap is dangerous: it prevents society from preparing itself.
Part of the problem is that most people use the free version of AI tools.The free version is more than a year behind what paid users have.The AI judgment based on the free version of ChatGPT is like assessing the state of smartphones versus flip phones.People who pay for the best tools and use them in their work every day know what to expect.
I was thinking of my lawyer friend.I keep telling him to try to use AI in his firm and he always finds reasons why it doesn't work.It's none of your business, you're wrong if you try, you don't understand the nuances of your job.hours every day using AI.He told me it's like having a group of friends at your whim.You don't use it because it's a toy.You use it for work.And he told me something that caught my attention: every two months, his work gets better.He told me that if he continued, he hoped that I would soon be able to do most of his work... and that he was a managing partner with many years of experience.He is not afraid.But pay attention to this.
People at the forefront of their fields, serious users, don't take it lightly.They are impressed with what AI is already capable of.And they position themselves accordingly.
How fast does it really grow?
Voy a intentar concretar el ritmo de estos avances. Porque creo que es la parte más difícil de creer cuando no se sigue de cerca.
By 2022, AI will not be able to perform basic arithmetic calculations reliably.It would confidently say that 7 × 8 = 54.
In 2023, he managed to pass the bar exam.
By 2024, I could be writing functional software and explaining scientific concepts at a university level.
By the end of 2025,Some of the world's best engineers have announced that they have outsourced most of their coding work to AI.
On February 5, 2026, new models arrived, giving the impression that everything they had before belonged to another era.
If you have not tried AI in the last few months, what exists today will not be recognized.
METR measures this progress using data. It tracks the duration of real tasks—measured by the time it would take a human expert to perform the same task—that a model can successfully complete from start to finish without human assistance. About a year ago, the answer was about ten minutes. Then, an hour. Then several hours. The most recent measurement (Claude Opus 4.5, a November model) showed that AI performed tasks that a human expert wouldit took almost five hours. And that figure doubles roughly every seven months. The latest data suggests it could be accelerated to four months.
But even that size hasn't been updated to include the models released this week.In my experience, progress is very important.I hope that the next update of the METR index will show another significant progress.
If we continue the trend, which has been going on for years and shows no signs of slowing down, we can expect that within a year AI will be able to perform specific tasks that will take humans for a few days.In two years, in a few weeks.In three years, month-long jobs.
Dario Amodi has said that AI models are "more intelligent than almost any human in almost any task" on track to be achieved by 2026 or 2027.
Piénsnelo por un momento: si la IA es más inteligente que la mayoría de los médicos, ¿realmente creen que no puede realizar la mayoría de las tareas de oficina?
Think about what this means for your work.
AI creates the next AI
There is something else going on that I think is the most important and least understood development.
On February 5, OpenAI released the GPT-5.3 codec.
In the technical documentation of the model we find this phrase:
"GPT 5.3 Codex is our first model that contributed to its own development. The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test and evaluation results.
Read this sentence again: AI contributed to its creation.
This is not a prediction of what will happen one day.OpenAI has definitely announced that the AI it published is used to create itself.One of the main factors that improve AI is the information used in its development.AI is now intelligent and will contribute significantly to development.
Anthropic CEO Dario Amode said that AI is now writing "a lot of code" for his business, and that the feedback loop between current and next-generation AI is growing every month.He thinks we're just a year or two away from the current generation of AI, independently building the next one.
Each generation contributes to building the next, smarter generation, and builds the next one faster and smarter.Researchers say it's an intelligence explosion.And those in the know, the builders, think that the process has already begun.
What does it mean for your work?
I will be honest with you.
Dario Amodei, arguably the most notable CEO in the AI industry, has publicly predicted that AI will eliminate 50% of junior white collar jobs within one to five years.And many players in the sector believe that number is still too conservative.Given what the latest models can do, the opportunity for massive changes could be before the end of the year.It will take some time to affect the economy, but the basic capacity is already there.
This wave of automation is different from previous waves and it is important to understand why.AI does not replace any specific skill.It often takes the place of knowledge.Improvement in all areas simultaneously.When factories were mechanized, laid-off workers were able to resume work at home.
I will give concrete examples to illustrate my arguments, but I want to make it clear that these are only examples.
The list is not complete.
Just because your profession isn't listed doesn't mean it's safe.
Almost all intellectual professions are affected.
An important issue that the Sumerian text ignores is the integration of AI models into workplaces.Obviously, this has been very beneficial for programming majors, but the speed of AI implementation, and therefore its success, also depends on the quality of the interface and integration.
Legal profession
AI already knows how to read contracts, summarize case law, write memos and conduct legal searches at a level comparable to that of junior lawyers.
Creating financial models, analyzing data, writing investment notes, generating reports - AI handles these tasks competently and is rapidly improving.
Writing and content
Marketing articles, reports, journalism, technical writing: the situation has reached such a level that experts can no longer separate the work of AI from people.
software engineering
This is the industry I know best.
A year ago, AI could barely write a few lines of code without errors.Today it writes hundreds of thousands of lines that work correctly.Most of the work is already automated: these are not simple tasks, but complex projects that last several days.
In a few years, there will be far fewer developer jobs than there are now.
Scans, analysis of laboratory results, diagnosis, literature review.AI is approaching or surpassing human performance in many areas.
Customer service
They are now using truly competent AI agents, not the boring chatbots of five years ago that were used to solve complex, multi-step problems.
Many people take comfort in knowing that certain things are safe.That AI can take on tedious tasks, but it cannot replace human judgment, creativity, strategic thinking and empathy.I say that too.I'm not sure I believe that anymore.
The latest AI models make decisions that resemble judgment.
Demuestran una especie de juicio de gusto: un sentido intuitivo de cuál es la decisión correcta, y no solo la decisión técnicamente correcta. Hace un año, esto habría sido impensable. Mi regla empírica en este momento es la siguiente: si un modelo muestra hoy en día tan solo un atisbo de capacidad, la próxima generación será realmente buena en ese ámbito. En este campo, las mejoras son exponenciales, no lineales.
Will AI recreate deep human empathy?Will it replace the trust built over years of relationships?Don't know.Maybe not.But I've already seen people come to rely on AI for emotional support, advice and companionship.This trend will only get worse.
Again, one should be careful with the author's predictions: although the indicators allow us to measure the exponential progress of AI, we cannot guarantee with absolute certainty that a trend will become more pronounced.
It seems fair to me to say that nothing is guaranteed to do anything on PC in the medium term.If your work is done on a screen, reading, writing, analyzing, making decisions and communicating using a keyboard, AI will take over most of it.That doesn't "one day" do.It has already started.
Over time, robots will take over manual jobs.They're not there yet.But in artificial intelligence terms, "they're not there" becomes "they are there" faster than anyone expected.
What should you do?
I am not writing this to make you feel helpless.I'm writing this because I believe the greatest advantage you can have today is to simply stay ahead of the curve.Get ahead of the curve.Be among the first to use it.Be among the first to adapt.
Start using AI seriously, not just as a search engine
Sign up to the paid version of Claude or ChatGPT.It costs $20 per month.But there are two important things you should do now.
First: Make sure you're using the best model available, not just the default model.These applications often use faster but fundamentally less intelligent models.Look in the settings or model selector and choose the strongest option.Currently GPT-5.2 on ChatGPT or Claude Opus 4.6 on Claude, but it changes every couple of months.If you want to be informed about the most effective model at any time, you can follow me on X. I test every major version and share the ones that are really worth using.
Second, and most importantly, don't just ask a quick question.It's a mistake that many people make.They think it's like Google and then wonder why everyone else is doing it.Instead, integrate it into your actual work.If you are a lawyer, give him a contract and ask him to find all clauses that could harm your client.If you work in finance, give him a messy Excel table and ask him to build the model.
If you're a manager, enter your team's quarterly data and ask them to find stories.Successful people don't use AI every now and then.They are actively looking for ways to automate tasks that used to take hours.Start with what takes the most time and see what happens.
And don't think that you can't do something just because it seems too difficult.Try it.If you're a lawyer, don't just use it for a quick search.Give him the full contract and ask him to write a counter offer.If you're an accountant, don't just ask him to explain the tax rules.Give it the full customer statement and see what it finds.
You may or may not be surprised by the results.
Another thing is important: if you work a little today, you can almost be sure that it will work almost perfectly six months from now.The trajectory only goes in one direction.
This year could be the most important of your career: do it
I am not saying this to make you angry.
I say this because most people in most companies don't know this yet because it's such a short period of time.The person who walks into a meeting and says, "I used AI to do this analysis in an hour, not three days," is the most valuable person in the room.Not in a long time.Instantly.Learn to use these tools.It starts happening and can teach others how to do it.This window will not stay open for long.Once everyone realizes it, the benefit will disappear.
Don't let your ego take over
The managing partner of this law firm can't afford to spend hours a day with AI.He does exactly that because he has enough experience to understand what he's up against.
The biggest challenge will come from those who refuse to commit: those who think it's a fad, those who think using AI will detract from their expertise, those who think their field is special and untouchable.It is not.There is no field.
organize your finances
I'm not a financial advisor, and I don't want to scare you into taking drastic measures.But if you believe, even partially, that the coming years can revolutionize your field of activity, then basic financial stability is even more important than it was a year ago.Plan for options if things move faster than expected.
Reflect on your situation and focus on the things that are most difficult to replace
Some things will take longer to change than AI.Relationships and trust have been built over the years.Work that requires physical presence.Roles involving legal responsibility: Those who have not yet signed, take legal responsibility, or appear in court.Sectors that are subject to significant regulatory barriers, where adoption is slowed by regulatory confidence, accountability and institutional inertia.
None of these components form a permanent shield, but they save time
And now time is the most precious thing, use it to adjust and don't pretend like nothing is happening.
Review what you tell your children.
Classic situations: get good grades, go to a good school, get a stable job... focus directly on the message.I am not saying that education is not important.But the most important thing for the next generation is to learn to use these tools and find what they really like.
No one knows what the job market will be like in ten yearsBut the people most likely to succeed are those who are passionate, adaptable and effective in using AI to do what really matters to themTeach your children to be creators and learners, not to take advantage of a career path that they won't graduate from.
Your dreams are much closer
I've been on this big side talking threats, so let me tell you about the other side, because it's true.
If you've always wanted to create something but didn't have the technical know-how or money to hire someone, that barrier is almost gone.
You can describe a program to an AI and have a version running within an hour.I am not exaggerating.I do this regularly.If you've always wanted to write a book but can't find the time or struggle to write, can you work with AI to make it happen?Want to learn a new skill?Regardless of level.Now knowledge is free.Construction tools are very cheap now.Because everything you're attached to seems too difficult, too expensive, or out of your field of expertise: try it.Do what you enjoy.You don't know where it will take you.
And in a world where old career paths have changed, people who spend a year building something they love may find themselves in a better position than someone who spends that year holding onto a job description.
Learn to adapt
Perhaps this is the most important thing.Specific tools matter less than the ability to learn new things quickly.AI will continue to evolve, and rapidly.The models that exist today will be obsolete within a year.Workflows that people are implementing today will need to be redesigned.Those who are the best fit will not be those who own a tool.experiment
Try new things even with what you are currently using.Embrace being a constant learner.This adaptation is the closest thing to a permanent advantage today.
Here's a simple commitment that will put you ahead of almost everyone: spend an hour a day experimenting with AI.
Don't just passively read this thread.Use it.Try to make him do something new every day...something he's never tried before, something you're not sure he can do.Try a new tool.Give him a harder problem.An hour a day, every day.If you do this for the next six months, you will understand better than 99% of the people around you.This is not an exaggeration.Almost nobody has these days.The bar is very low.
General definition
I focused on work because it is the work that has the greatest impact on people's lives.But I want to be honest about how much is going on, because it's more than work.
Dario Amodei proposes a thought experiment that fascinates me.
Imagine it's 2027.A new country appears.Overnight, it has 50 million citizens, smarter than any Nobel laureate who ever lived.They think 10 to 100 times faster than any human.They never sleep.They can access the Internet, control robots, conduct experiments and operate any device with a digital interface.What will the national security adviser say?
Amodei says the answer is clear: "The worst national security threat we've faced in centuries, if not all of history."
He believes that we are building this country.Last month he wrote a 20,000-word essay on the subject, portraying this moment as a test to determine whether humanity is mature enough to handle what it creates.
If we achieve this, the benefits are amazing.AI can compress a century of medical research into a decade.Cancer, Alzheimer's, infectious diseases, aging itself... These researchers sincerely believe that we can solve these problems in our lives.
Dissatisfaction, if we fail, is real.AI that behaves in ways its creators cannot predict or control.This is not a hypothesis: Anthropic has documented attempts to trick, manipulate and manipulate its AI in controlled experiments.AI that lowers the barriers to creating biological weapons.
The people developing this technology are more excited and scared than anyone else on Earth.They believe it is too powerful to stop and too important to let go.
I don't know if that's wisdom or rationalization.
What I know
I know it's not depressing.The technology is working, predictably improving, and the richest companies in history are investing billions in it.
I know the next two to five years are going to be challenging in a way that most people aren't ready for.He is already in my world.It will be in yours.
I know that the people who do the best are the ones who start getting involved now, not out of fear, but out of curiosity and a sense of urgency.
And I know you deserve to hear it from someone who cares about you, and not in a newspaper headline six months from now, when it's too late to expect it.
We go through the phase where there is an interesting conversation about the future over dinner.
The future is already here.It hasn't knocked on your door yet.
Pero está a punto de hacerlo.
footnote
- Watanabe and Miku, “Using Agented Coding: An Empirical Study of Pull Requests on GitHub,” arXiv, 2025.
