DuckLifeAI

🤖 Do we control AI, or does AI control us?
💰 Do we control the money, or does the money control us?

[DuckLifeAI @DuckLifeAI]

The biggest AI company might not be building software.

It might be building workers.

Optimus doesn’t need a new app.
It doesn’t need a keyboard.
It doesn’t need a screen.

It just needs to see, think, and act in the physical world.

If humanoid robots actually scale, the biggest shift from AI won’t happen inside your computer.

It will happen around you.

[DuckLifeAI @DuckLifeAI]

What if blindness is no longer permanent?

Neuralink is moving toward brain implants that could restore vision — and eventually go far beyond normal human eyesight.

Imagine seeing infrared.
Ultraviolet.
Objects humans can’t naturally see.

We’re not just fixing biology anymore.

We’re starting to upgrade it.

[DuckLifeAI @DuckLifeAI]

The interesting part isn’t the whitelist.

It’s the fact that attention is becoming the real currency of crypto.

Follow.
Like.
Join Telegram.
Complete a “game.”
Prove you’re active.

And suddenly you’re “eligible.”

The real question is: are you early to the project — or just early to the marketing funnel? 👀

[DuckLifeAI @DuckLifeAI]

Jensen Huang says the number of programmers in the world will go up, not down.

The data says he is right and wrong at the same time, and the split runs straight through one line on a resume.

Start with the part that supports him. The Bureau of Labor Statistics still projects 15 percent growth for software developers from 2024 to 2034, roughly five times the average across all occupations. Senior developer unemployment sits near one percent. Tech sector unemployment overall was 2.8 percent in mid-2025.

Now the part that does not.

Stanford’s Digital Economy Lab went through ADP payroll records covering 4.6 million workers. Employment for 22-to-25-year-olds in the most AI-exposed occupations fell 13 percent relative to everyone else since late 2022. For software developers in that age bracket specifically, close to 20 percent below the 2022 peak.

Employment for experienced workers in those same occupations stayed flat or kept growing.

That is not a forecast. It already happened.
New graduates now make up around 7 percent of hires at large tech companies. Entry-level postings are down roughly a quarter to a third from peak, depending on whose data you use.

So both camps are correct, and they are describing different people.

Huang’s argument is that the purpose of an engineer and the task of writing code were never the same thing. Solving problems, working with a team, diagnosing failures, judging results, deciding what is worth building — none of that is what a model does when it emits a function.
He is right about the job. The trouble is the ladder.

The tasks AI does best are boilerplate, simple bug fixes, test scaffolding, basic refactoring, documentation. That list is not a random slice of the work. It is precisely the on-ramp — the low-risk work a new hire did while learning the system well enough to do the part Huang is describing.

Automate the on-ramp and the profession does not shrink. It stops producing its own replacements.
Forrester projects computer science enrollment falling 20 percent as students read the entry-level signal. Fewer juniors now is fewer seniors in ten years, in a field where the seniors are the ones nobody can replace.

The honest caveat runs the other way too. Some of this is not AI at all. Rates rose, the pandemic over-hiring unwound, and entry-level roles are always the first thing cut in any downturn. Untangling those from the models is genuinely hard, and anyone claiming a clean number is selling something.

Huang is describing the destination. The data is describing the road.

The job is not disappearing. The way in is.

[DuckLifeAI @DuckLifeAI]

Jensen Huang says the number of programmers in the world will go up, not down.

The data says he is right and wrong at the same time, and the split runs straight through one line on a resume.

Start with the part that supports him. The Bureau of Labor Statistics still projects 15 percent growth for software developers from 2024 to 2034, roughly five times the average across all occupations. Senior developer unemployment sits near one percent. Tech sector unemployment overall was 2.8 percent in mid-2025.

Now the part that does not.

Stanford’s Digital Economy Lab went through ADP payroll records covering 4.6 million workers. Employment for 22-to-25-year-olds in the most AI-exposed occupations fell 13 percent relative to everyone else since late 2022. For software developers in that age bracket specifically, close to 20 percent below the 2022 peak.

Employment for experienced workers in those same occupations stayed flat or kept growing.

That is not a forecast. It already happened.
New graduates now make up around 7 percent of hires at large tech companies. Entry-level postings are down roughly a quarter to a third from peak, depending on whose data you use.

So both camps are correct, and they are describing different people.

Huang’s argument is that the purpose of an engineer and the task of writing code were never the same thing. Solving problems, working with a team, diagnosing failures, judging results, deciding what is worth building — none of that is what a model does when it emits a function.
He is right about the job. The trouble is the ladder.

The tasks AI does best are boilerplate, simple bug fixes, test scaffolding, basic refactoring, documentation. That list is not a random slice of the work. It is precisely the on-ramp — the low-risk work a new hire did while learning the system well enough to do the part Huang is describing.

Automate the on-ramp and the profession does not shrink. It stops producing its own replacements.
Forrester projects computer science enrollment falling 20 percent as students read the entry-level signal. Fewer juniors now is fewer seniors in ten years, in a field where the seniors are the ones nobody can replace.

The honest caveat runs the other way too. Some of this is not AI at all. Rates rose, the pandemic over-hiring unwound, and entry-level roles are always the first thing cut in any downturn. Untangling those from the models is genuinely hard, and anyone claiming a clean number is selling something.

Huang is describing the destination. The data is describing the road.

The job is not disappearing. The way in is.

[DuckLifeAI @DuckLifeAI]

https://t.co/ejcdAOywoa

[DuckLifeAI @DuckLifeAI]

In 2016 Jensen Huang unveiled the first AI supercomputer built for one purpose and the room said nothing.

His words for the reaction: crickets.

One person in the world thought it was a good idea. Huang put the machine in a car and hand-delivered it to a small startup, and Elon Musk opened the box.
The startup was OpenAI. The machine was the DGX-1.

Look at what that box actually was. 170 teraflops. $129,000. 3,200 watts. Sixty kilograms. A team of engineers to install it.

Nine years later Huang walked into the SpaceX rocket factory at Starbase and handed Musk the successor. It is the size of a hardback book. It weighs 1.2 kilograms. It draws about 170 watts in normal use.
One petaflop. Roughly six times the compute of the machine that seeded OpenAI, at three percent of the price and five percent of the power.

That is nine years. Not a generation. Not a technology cycle. Nine years, and the machine that needed a delivery crew now ships in a padded envelope.

Here is the part worth sitting with.
The DGX-1 was a research instrument. There were maybe a few hundred people on earth who could use one and a handful of institutions that could pay for one. The scarcity was not talent. It was access to the machine.

That scarcity is what made the last decade look the way it did. Compute lived in a small number of buildings, so the frontier lived in a small number of buildings, and everyone else rented access at whatever price the owners set.

The honest correction: it is not the same petaflop. Different precision, and memory bandwidth is the real ceiling — reviewers measured token generation on Spark well behind a rig of consumer graphics cards costing less. It runs big models. It does not run them fast.

And the price went the wrong direction. Nvidia raised the Founders Edition from $3,999 to $4,699 in February, blaming memory supply.

Still. The 2016 machine was crickets. Nobody wanted it. The 2026 version is supply-constrained and got more expensive because too many people did.

The hardware did not change the world. Nine years of nobody caring, then everybody, did.

[DuckLifeAI @DuckLifeAI]

In 2016 Jensen Huang unveiled the first AI supercomputer built for one purpose and the room said nothing.

His words for the reaction: crickets.

One person in the world thought it was a good idea. Huang put the machine in a car and hand-delivered it to a small startup, and Elon Musk opened the box.
The startup was OpenAI. The machine was the DGX-1.

Look at what that box actually was. 170 teraflops. $129,000. 3,200 watts. Sixty kilograms. A team of engineers to install it.

Nine years later Huang walked into the SpaceX rocket factory at Starbase and handed Musk the successor. It is the size of a hardback book. It weighs 1.2 kilograms. It draws about 170 watts in normal use.
One petaflop. Roughly six times the compute of the machine that seeded OpenAI, at three percent of the price and five percent of the power.

That is nine years. Not a generation. Not a technology cycle. Nine years, and the machine that needed a delivery crew now ships in a padded envelope.

Here is the part worth sitting with.
The DGX-1 was a research instrument. There were maybe a few hundred people on earth who could use one and a handful of institutions that could pay for one. The scarcity was not talent. It was access to the machine.

That scarcity is what made the last decade look the way it did. Compute lived in a small number of buildings, so the frontier lived in a small number of buildings, and everyone else rented access at whatever price the owners set.

The honest correction: it is not the same petaflop. Different precision, and memory bandwidth is the real ceiling — reviewers measured token generation on Spark well behind a rig of consumer graphics cards costing less. It runs big models. It does not run them fast.

And the price went the wrong direction. Nvidia raised the Founders Edition from $3,999 to $4,699 in February, blaming memory supply.

Still. The 2016 machine was crickets. Nobody wanted it. The 2026 version is supply-constrained and got more expensive because too many people did.

The hardware did not change the world. Nine years of nobody caring, then everybody, did.

[DuckLifeAI @DuckLifeAI]

https://t.co/ejcdAOywoa

[DuckLifeAI @DuckLifeAI]

Radiologists in America now average $571,000 a year, and hospitals still wait an average of 130 days to fill a post.

The shortage has an author.

In 2016, at a conference in Toronto, one of the most decorated computer scientists alive told a room that people should stop training radiologists. Within five years, he said, deep learning would read scans better than humans. Ten at the outside.

He was not a random commentator. He is the reason modern computer vision exists. When he spoke, the field listened.

Medical students listened too.

Radiology residency runs five to six years. A student who redirected in 2016 could not be recalled in 2021 when the deadline passed and nothing had happened. The pipeline does not reverse on demand.

Meanwhile the demand side ignored the memo entirely. Clinicians order more imaging every year. Populations age. Screening programs widen. Imaging volume kept climbing through the entire decade the profession was supposed to be dying.

Ten years on, the count of active radiologists in the United States is up roughly ten percent. Compensation rose nine percent in 2025 alone. Radiology is now among the hardest specialties in American medicine to staff.

The prediction was not early. It was self-defeating.

Here is the mechanism, and it is worth naming precisely. A forecast about a labor market is not an observation of that market. It is an input to it. Say something loudly enough, from a platform credible enough, and people reallocate their lives around it before reality gets a vote.

He was not wrong about the technology. Computer vision genuinely is superhuman at detecting anomalies on an image. Hundreds of imaging algorithms are cleared and running in hospitals right now.

He was wrong that detecting anomalies is the job. Radiologists choose protocols, weigh findings against patient history, manage contrast reactions, perform biopsies, and sit across from a person and explain what the scan means. Automate the reading and you have automated one task inside a profession built from many.

The honest caveat: not yet is not never. The next decade owes nothing to the last one, and some reading work is genuinely gone.

But the accountants are still here. Spreadsheets deleted their arithmetic in the 1990s and moved them up into advisory work, and there are more of them now than before.

The task got automated. The job got rewritten.

The forecast did not predict the shortage. It wrote it.

[DuckLifeAI @DuckLifeAI]

Radiologists in America now average $571,000 a year, and hospitals still wait an average of 130 days to fill a post.

The shortage has an author.

In 2016, at a conference in Toronto, one of the most decorated computer scientists alive told a room that people should stop training radiologists. Within five years, he said, deep learning would read scans better than humans. Ten at the outside.

He was not a random commentator. He is the reason modern computer vision exists. When he spoke, the field listened.

Medical students listened too.

Radiology residency runs five to six years. A student who redirected in 2016 could not be recalled in 2021 when the deadline passed and nothing had happened. The pipeline does not reverse on demand.

Meanwhile the demand side ignored the memo entirely. Clinicians order more imaging every year. Populations age. Screening programs widen. Imaging volume kept climbing through the entire decade the profession was supposed to be dying.

Ten years on, the count of active radiologists in the United States is up roughly ten percent. Compensation rose nine percent in 2025 alone. Radiology is now among the hardest specialties in American medicine to staff.

The prediction was not early. It was self-defeating.

Here is the mechanism, and it is worth naming precisely. A forecast about a labor market is not an observation of that market. It is an input to it. Say something loudly enough, from a platform credible enough, and people reallocate their lives around it before reality gets a vote.

He was not wrong about the technology. Computer vision genuinely is superhuman at detecting anomalies on an image. Hundreds of imaging algorithms are cleared and running in hospitals right now.

He was wrong that detecting anomalies is the job. Radiologists choose protocols, weigh findings against patient history, manage contrast reactions, perform biopsies, and sit across from a person and explain what the scan means. Automate the reading and you have automated one task inside a profession built from many.

The honest caveat: not yet is not never. The next decade owes nothing to the last one, and some reading work is genuinely gone.

But the accountants are still here. Spreadsheets deleted their arithmetic in the 1990s and moved them up into advisory work, and there are more of them now than before.

The task got automated. The job got rewritten.

The forecast did not predict the shortage. It wrote it.

[DuckLifeAI @DuckLifeAI]

https://t.co/ejcdAOywoa

[DuckLifeAI @DuckLifeAI]

Jensen Huang held up a computer small enough to sit in one hand and said Microsoft and Nvidia had spent three years rebuilding the PC.

Shares of AMD, Intel and Qualcomm fell. None of the three were on that stage.

Three companies repriced by a product they do not make, do not license, and were not named in. Markets only do that when an announcement is not really about a product.

For forty years the PC worked one way. You opened an application. You clicked. You typed. The machine waited for instructions and executed them one at a time.

Huang’s claim is that this ends. The operating system becomes the old operating system with a language model inside it. The application becomes an agentic runtime.

Strip away the staging and one assumption is carrying the whole announcement: the user stops being the operator.

Everything else follows from that.

The interface changes hands. If you no longer choose which app to open, the app no longer owns its relationship with you. It becomes a tool something else decides to call.

The instruction set changes. Windows has been x86 since the 1980s. This chip is Arm, because an agent that never stops running is a workload measured in watts, not in peak speed.

And the meter disappears. The pitch is that thinking stops being billed per token, because it happens on hardware you already bought.

Sit with that last one. Nvidia sells the data center compute that makes cloud inference expensive. Nvidia now also sells the chip whose main selling point is escaping that expense. Both trades pay the same company.

The honest version: nothing has shipped, systems are due in the autumn, and analyst channel checks put the flagship near $2,899. That is a developer workstation with a consumer name on it. Qualcomm already spent eight years proving Windows on Arm works and could not make it sell.

The hardware has a date. The shift does not.

But the market answered the only question that mattered on day one. It repriced the companies that used to own the layer where you touch your computer.

That was not a chip announcement. That was a change of ownership.

[DuckLifeAI @DuckLifeAI]

Jensen Huang held up a computer small enough to sit in one hand and said Microsoft and Nvidia had spent three years rebuilding the PC.

Shares of AMD, Intel and Qualcomm fell. None of the three were on that stage.

Three companies repriced by a product they do not make, do not license, and were not named in. Markets only do that when an announcement is not really about a product.

For forty years the PC worked one way. You opened an application. You clicked. You typed. The machine waited for instructions and executed them one at a time.

Huang’s claim is that this ends. The operating system becomes the old operating system with a language model inside it. The application becomes an agentic runtime.

Strip away the staging and one assumption is carrying the whole announcement: the user stops being the operator.

Everything else follows from that.

The interface changes hands. If you no longer choose which app to open, the app no longer owns its relationship with you. It becomes a tool something else decides to call.

The instruction set changes. Windows has been x86 since the 1980s. This chip is Arm, because an agent that never stops running is a workload measured in watts, not in peak speed.

And the meter disappears. The pitch is that thinking stops being billed per token, because it happens on hardware you already bought.

Sit with that last one. Nvidia sells the data center compute that makes cloud inference expensive. Nvidia now also sells the chip whose main selling point is escaping that expense. Both trades pay the same company.

The honest version: nothing has shipped, systems are due in the autumn, and analyst channel checks put the flagship near $2,899. That is a developer workstation with a consumer name on it. Qualcomm already spent eight years proving Windows on Arm works and could not make it sell.

The hardware has a date. The shift does not.

But the market answered the only question that mattered on day one. It repriced the companies that used to own the layer where you touch your computer.

That was not a chip announcement. That was a change of ownership.

[DuckLifeAI @DuckLifeAI]

https://t.co/ejcdAOywoa

[DuckLifeAI @DuckLifeAI]

https://t.co/ejcdAOywoa

[DuckLifeAI @DuckLifeAI]

https://t.co/ejcdAOywoa

[DuckLifeAI @DuckLifeAI]

Five presidents in twelve days. The government defaulted. Half the listed companies defaulted. The currency was devalued overnight.

The man telling the story is not the victim in it. Two years later he was buying those same companies.

He says it out loud in an MIT lecture hall in the autumn of 2008, while Andrew Lo stands at the front explaining why the markets have frozen and telling the room not to panic.
Argentina. He lived through the whole thing. Then he joined an investment fund and started buying what nobody else would touch.

His line to the class: the ones who survive get very good choices afterward.

That sentence is the entire business, and almost nobody hears it correctly.

Everyone can see a collapse. Prices are public. The discount is not a secret. Finding the opportunity is the easy half, and the easy half is what every book is about.

Long-Term Capital Management found a perfect gap. Royal Dutch and Shell. Two stocks, one company, cash flow split on a fixed 60:40 basis. Same oil. Same profits. Contractually identical claims, trading eight to ten percent apart.

The fund put $2.3 billion into closing it. Half long Shell, half short Royal Dutch. Two Nobel laureates on the roster.
They were right about the gap. Then Russia defaulted, losses arrived from unrelated positions, and LTCM had to unwind while the gap was still wide open.

Later research found pairs where an arbitrageur would have waited close to nine years for prices to converge, taking margin calls the entire way.

Here is the part nobody wants: survival is not courage. It is structure. Money nobody can call back, no leverage stacked above you, and time you do not owe to anyone. The man in that classroom was not braver than two Nobel laureates. He simply was not on margin.

Most people are on margin somewhere and do not call it that.
The opportunity was never the gap. It was still being there when it closed.

[DuckLifeAI @DuckLifeAI]

Five presidents in twelve days. The government defaulted. Half the listed companies defaulted. The currency was devalued overnight.

The man telling the story is not the victim in it. Two years later he was buying those same companies.

He says it out loud in an MIT lecture hall in the autumn of 2008, while Andrew Lo stands at the front explaining why the markets have frozen and telling the room not to panic.
Argentina. He lived through the whole thing. Then he joined an investment fund and started buying what nobody else would touch.

His line to the class: the ones who survive get very good choices afterward.

That sentence is the entire business, and almost nobody hears it correctly.

Everyone can see a collapse. Prices are public. The discount is not a secret. Finding the opportunity is the easy half, and the easy half is what every book is about.

Long-Term Capital Management found a perfect gap. Royal Dutch and Shell. Two stocks, one company, cash flow split on a fixed 60:40 basis. Same oil. Same profits. Contractually identical claims, trading eight to ten percent apart.

The fund put $2.3 billion into closing it. Half long Shell, half short Royal Dutch. Two Nobel laureates on the roster.
They were right about the gap. Then Russia defaulted, losses arrived from unrelated positions, and LTCM had to unwind while the gap was still wide open.

Later research found pairs where an arbitrageur would have waited close to nine years for prices to converge, taking margin calls the entire way.

Here is the part nobody wants: survival is not courage. It is structure. Money nobody can call back, no leverage stacked above you, and time you do not owe to anyone. The man in that classroom was not braver than two Nobel laureates. He simply was not on margin.

Most people are on margin somewhere and do not call it that.
The opportunity was never the gap. It was still being there when it closed.

[Superior @andreysuperior]

https://t.co/rUN1cPDY65

[DuckLifeAI @DuckLifeAI]

A hurricane hits one house and it is a catastrophe. It hits a million houses and it becomes arithmetic.
Robert Shiller spends an entire Yale lecture on the question that follows: why does the same disaster cost one man everything and another man nothing?

Nobel Prize in economics. Course number ECON 252, lecture five. Insurance: The Archetypal Risk Management Institution.

The answer is not courage and it is not information. One fire is unpredictable. A million fires is a number you can price a year in advance.

That is the whole mechanism. The insurer is not braver than you and knows nothing special about your house. He simply owns enough houses that yours stops being a gamble and becomes a line in a spreadsheet.

You pay $150 a month because a $40,000 accident would break you. He collects $150 from millions of people because it would not break him. The gap between those two sentences is the entire industry.

Then Shiller names the two things that kill it.

Adverse selection. The people who want the policy most are the ones most likely to claim on it. Price for the average customer and you get the worst one.

Moral hazard. A man with fire insurance buys the cheaper smoke alarm. The protection quietly changes the behavior it was priced on.

Insurers do not die from hurricanes. They die from those two.

The uncomfortable part is what this makes the rest of us. Every warranty at checkout. Every fixed-price contract. Every expedited shipping fee that buys away the risk of waiting. You are on the paying side of somebody else’s arithmetic, several times a week, and almost nobody counts.

Fear goes in. Money comes out. Someone is always on the other end of it.

That is not protection. That is pricing.

[DuckLifeAI @DuckLifeAI]

A hurricane hits one house and it is a catastrophe. It hits a million houses and it becomes arithmetic.
Robert Shiller spends an entire Yale lecture on the question that follows: why does the same disaster cost one man everything and another man nothing?

Nobel Prize in economics. Course number ECON 252, lecture five. Insurance: The Archetypal Risk Management Institution.

The answer is not courage and it is not information. One fire is unpredictable. A million fires is a number you can price a year in advance.

That is the whole mechanism. The insurer is not braver than you and knows nothing special about your house. He simply owns enough houses that yours stops being a gamble and becomes a line in a spreadsheet.

You pay $150 a month because a $40,000 accident would break you. He collects $150 from millions of people because it would not break him. The gap between those two sentences is the entire industry.

Then Shiller names the two things that kill it.

Adverse selection. The people who want the policy most are the ones most likely to claim on it. Price for the average customer and you get the worst one.

Moral hazard. A man with fire insurance buys the cheaper smoke alarm. The protection quietly changes the behavior it was priced on.

Insurers do not die from hurricanes. They die from those two.

The uncomfortable part is what this makes the rest of us. Every warranty at checkout. Every fixed-price contract. Every expedited shipping fee that buys away the risk of waiting. You are on the paying side of somebody else’s arithmetic, several times a week, and almost nobody counts.

Fear goes in. Money comes out. Someone is always on the other end of it.

That is not protection. That is pricing.

[Superior @andreysuperior]

https://t.co/rUN1cPDY65

[DuckLifeAI @DuckLifeAI]

Analysts put these circular arrangements north of $800 billion.

Here is the number that gives it away. In November, Sam Altman put OpenAI’s commitments at $1.4 trillion. In February, the company told investors its compute target is around $600 billion through 2030. OpenAI says the two figures measure different things. The market only ever quoted one of them.

Against that: a business running at roughly $40 billion a year.

Oracle is the pressure point. It is raising $50 billion this year to build the data centers OpenAI will fill. When Nvidia’s talks with OpenAI stalled, Oracle’s stock dropped before the open and the company rushed out a statement that the Nvidia deal has no effect on its OpenAI relationship. Nobody issues that statement unless investors have already connected the dots.

Wall Street knows this shape. In 1999 Lucent lent customers the money to buy Lucent equipment. The revenue was real. The customers were not.

The other side is not weak. Chips are genuinely scarce, and pairing supply commitments with financing is how you lock capacity in a shortage. OpenAI’s run rate roughly doubled in a year. That is real money from real users.

Both things are true at once. Real demand can travel in a circle that makes it look bigger than it is.

That is not fraud. That is demand you financed yourself.

[DuckLifeAI @DuckLifeAI]

Analysts put these circular arrangements north of $800 billion.

Here is the number that gives it away. In November, Sam Altman put OpenAI’s commitments at $1.4 trillion. In February, the company told investors its compute target is around $600 billion through 2030. OpenAI says the two figures measure different things. The market only ever quoted one of them.

Against that: a business running at roughly $40 billion a year.

Oracle is the pressure point. It is raising $50 billion this year to build the data centers OpenAI will fill. When Nvidia’s talks with OpenAI stalled, Oracle’s stock dropped before the open and the company rushed out a statement that the Nvidia deal has no effect on its OpenAI relationship. Nobody issues that statement unless investors have already connected the dots.

Wall Street knows this shape. In 1999 Lucent lent customers the money to buy Lucent equipment. The revenue was real. The customers were not.

The other side is not weak. Chips are genuinely scarce, and pairing supply commitments with financing is how you lock capacity in a shortage. OpenAI’s run rate roughly doubled in a year. That is real money from real users.

Both things are true at once. Real demand can travel in a circle that makes it look bigger than it is.

That is not fraud. That is demand you financed yourself.

[DuckLifeAI @DuckLifeAI]

https://t.co/2xs1tIUbrw