ChatGPT-6 Astra Is Here. Will Human Existence Remain, or Will Everything Slowly Get Automated?

The model has arrived. The fear has arrived with it.

ChatGPT Astra is out, and this is the GPT-6 model, described by its makers as the most intelligent and aligned model they have ever released. Ever since it appeared there has been a huge amount of fear about AI. The talk everywhere is that AGI has started to enter, that artificial general intelligence has begun arriving, and that nobody can say with certainty how fast AI, or AGI, or robotic technology working together with it, can automate every field, every domain, every single thing human beings do. With agentic AI and robotics combined, the claim being made is that humans will not be needed for anything at all, and that human existence itself stands on the verge of ending.

Right next to that fear sit two completely opposite pieces of advice, and both come from professionals who actually work in this field. One side still says be a manual developer, meaning code by hand and do not hand your thinking over to the machine. The other side says keep AI as a co-pilot, work alongside it, use it, and move faster with it. Both things are being said at the same time by people who know the field. That contradiction is exactly why a student feels stuck, and why a working person feels stuck.

So look at what actually arrived first. Fear built on half information is always bigger than fear that survives full information.

What Astra Actually Is

Astra is not a rumour. It arrived in September 2026, first to a small set of organisations and then to users on the Plus, Pro, Business and Enterprise plans, along with the API and the major cloud platforms.

Its strengths are described as using a computer, browsing, software engineering, cybersecurity, science and professional work. The important word in that list is using. This is not a model that answers a question and waits. It operates a machine, fills forms, updates records, arranges calendars and runs software tests, and it also keeps context across sessions, which means it remembers a project the way a colleague remembers it.

That last part deserves a pause. A model that answers one question well is a tool. A model that remembers your project across days starts to behave like a junior colleague who was present yesterday. That is a different thing altogether, and it is the reason this fear feels heavier than the fear that came with the previous five models.

What Astra Is Not Allowed To Do

Here is the part the headlines did not carry, and it is the most important part of the whole release.

Astra is not allowed to do everything. It refuses the more advanced cybersecurity work, including writing a working exploit for a vulnerability. It runs under monitoring systems that watch its reasoning and its actions, and in sensitive situations it pauses and asks a human to approve the next step before it continues. In the developer environment, risky tasks simply stop instead of proceeding. It is also the first model of its kind to cross a critical cybersecurity capability threshold, which means the people who built it decided that crossing that line required new rules rather than more confidence.

The frontier has already moved past one question and into another. The old question was can it do this. The new question is who is allowed to do this, and the one after that is who signs off on it.

That is not the language of a machine that has taken authority away from people. That is the language of a very powerful instrument that still needs a responsible hand on it.

The Other Side Of The Argument

Other people say something very different, and they have history on their side. They say technical jobs will not end, that no kind of human work will end, and that this holds today as well as in the ten or fifteen years to come.

When computers came there was enormous fear about jobs, about human work, about human existence itself. Before computers, when calculators came, the fear was the same in kind and the same in certainty that the end was near. Work is still going on today.

What actually happened in those years was transformation rather than deletion. Everything got modernised and everything changed, while nothing stayed exactly as it was. The calculator took the arithmetic away and left the judgement. The spreadsheet took the manual tallying away and left the analysis. The computer took the typing away and left the thinking. What people call the death of a job is usually the death of one part of a job, while the rest of it quietly grows.

AI is more powerful than anything that came before it, and powerful still does not automatically mean total. AI can transform jobs, which is not the same thing as deleting them, and ten opportunities will go while twenty new opportunities are created. That sentence sounds like comfort, so the numbers behind it are worth checking.

Why This Fear Feels More Real

It would be dishonest to say this is just like the calculator. The fear is not stupid. It is aimed at the wrong target.

Three things are genuinely different.

Earlier automation replaced a task, whereas this wave is attempting to replace a chain. A ticket gets read, a database gets checked, a fix gets written and tested, and the result gets reported back, all in one flow. Chains are what real jobs are actually made of.

The model is also partly improving through its own output, since AI is now being used to build better AI. That curve does not behave like the old industrial curves that took decades to move, and anyone who says this will be slow because it was slow last time is selling comfort rather than analysis.

Then robotics entered the room. Software used to only advise, and now it can operate a physical machine. Hiding inside knowledge work has stopped being a safe strategy.

The anxiety is not wrong, and nobody should be ashamed of feeling it. The anxiety is simply aimed at the wrong thing.

The AGI Question, Without Drama

Most discussion either frightens people or comforts them, and both habits are worth dropping.

AGI has no single agreed meaning. The working definition used by most forecasters is fairly simple, which is that most purely mental work becomes automatable at better quality, speed and cost than humans can manage.

Now look at what the serious forecasters actually believe. Their middle estimates for that milestone stretch from the late 2020s all the way into the 2040s, which is not a consensus of next Tuesday but a spread wide enough to hold two completely different careers inside it.

Their answers also keep moving in both directions. Through 2025 several prominent forecasters pushed their timelines further away after deciding AGI was harder than they had thought. In early 2026 the movement reversed and timelines came closer again. That is what an unsettled question looks like from the inside, and if the best informed people in the world keep revising in both directions inside eighteen months, then anyone who states the date with total confidence is describing what they are selling rather than what they know.

The honest position is that nobody knows the date. Treat confident claims in either direction as salesmanship.

The Bill Nobody Talks About

Now comes the part the panic skips, and this is the part that actually decides your future.

Intelligence is not free. It runs on electricity, on water and on land, and the bill is arriving now. AI companies already face a shortage of data centres, a shortage of water and a shortage of electricity, because everything they need is something they are short of.

The numbers show it. Data centres used around four hundred and fifteen terawatt-hours in 2024, which is roughly one and a half per cent of the world’s electricity, and that figure had already been growing at about twelve per cent a year for five years.

The base case projection nearly doubles it. Consumption reaches about nine hundred and forty-five terawatt-hours by 2030, close to three per cent of global electricity, growing at roughly fifteen per cent a year, which is more than four times faster than demand from everything else on the planet.

The AI slice is smaller than the shouting suggests. AI focused data centres were about half a per cent of global electricity in 2025, around a third of all data centre consumption, and by 2030 they are projected to consume roughly as much as every other kind of data centre combined. The AI share is the part that explodes.

Water comes in through cooling. A single hundred-megawatt AI data centre is estimated to use between one and a half and three million cubic metres of water a year for evaporative cooling alone. Add rare materials, scarce land, grid connection queues measured in years, and a supply chain that moves in quarters rather than weeks.

So why does this matter to your career? Because it puts a price on automation, and anything with a price can be afforded or not afforded.

Cheaper Price, Bigger Bill

Here is something worth understanding, because it destroys the simple story that AI gets cheaper every year and therefore humans get replaced every year.

The price of a single token, which is the unit of text a model processes, has fallen by roughly ninety-nine and a half per cent since the early GPT-3 era. That is not a typing mistake. The unit cost of intelligence has genuinely collapsed.

So AI bills should be falling. They are not.

In company after company the total bill has tripled while the price per unit fell to almost nothing. How does that happen? Usage grows faster than price falls.

When something becomes cheaper people do not buy the same amount of it. They buy far more. They stop asking one question and start running a thousand tasks, and they stop summarising one document and start processing every document the company owns. The unit got cheap, the volume exploded, and the total went up anyway.

This is the most under reported fact in the AI economy. Cheaper intelligence does not mean less spending. It means more automation at a price that keeps climbing.

Hold that next to the energy numbers. Every task handed to a machine carries a compute bill, an electricity bill and a water bill attached to it. That does not stop automation. It meters it.

Where The Money Is Going

Look at the money, because the numbers are not small.

The biggest technology companies plan to spend between six hundred and sixty and six hundred and ninety billion dollars on AI infrastructure in 2026 alone, and global AI investment is forecast to cross one trillion dollars in the same year.

Here is the part that matters. This spending is not coming out of spare profit.

Capital expenditure is expected to rise by roughly five hundred and thirty-four billion dollars against additional cash flow of around three hundred and forty billion, which works out to about one dollar and fifty-seven cents of new investment for every single dollar of new cash coming in. People say this casually as two rupees spent to earn one rupee, except here it sits in the actual ledgers.

Think about what a company does next. It has spent that much and the return has not arrived, which leaves only three doors. It raises prices, or it cuts costs, or it rations the service.

All three doors eventually reach the customer, and the customer is an ordinary person with an ordinary budget. When the ordinary person can no longer afford AI, the human becomes affordable again.

That is the loop nobody advertises. More AI use means more cost, more cost means more expensive tools, and more expensive tools mean the human is needed back. This is not a theory about the distant future either, because the rehiring wave has already started and it is described further down.

Why The Co-Pilot Advice Holds

This is why the co-pilot advice makes sense.

AI is not weak, and AI is not harmless. The economics simply do not allow AI to replace everything, everywhere, at once, forever.

Working alongside AI is the position that survives both directions of this argument. The people who say use the tool are right, and the people who say do not lose your own thinking are also right. The mistake is choosing one and throwing the other away.

The Big Entrepreneurs Look At The Sky

The debate continues on the ground while some professionals point somewhere else entirely. Big stalwart entrepreneurs like Elon Musk are preparing to build AI data centres in space, along with robots, on other planets.

That tells you something on its own. If land and power and cooling on Earth were not a real limit, nobody would propose the most expensive real estate on or off this planet.

The details show how real the limit is. Prototype satellites are expected in early 2027, and the first operational AI satellites are tentatively planned for the fourth quarter of that year.

Space is cold, and space is also a vacuum, which means there is no air to carry heat away. Heat simply builds up, and removing it needs large radiators and liquid cooling loops. To match a single hundred-megawatt data centre on the ground, an orbital facility would need to be something like five hundred to a thousand times the size of the International Space Station. Maintenance on Earth is a person walking in with a replacement chip, and in orbit that becomes an orbital problem instead.

Then there is money. Launching into orbit costs around a thousand dollars per kilogram today, and the assessment from the companies studying this is that the price must fall to roughly two hundred dollars per kilogram before space based data centres begin to make sense.

The people with the most resources in the world are looking at other planets for power and land. That means the limits on Earth are real. AI has not escaped every limit. AI has hit them hard enough to look at the sky.

The Layoffs That Came Back

Now comes the story that matters most, and it is also the story reported the least, because it is not dramatic enough for a headline.

Companies did replace people with AI. Between January 2025 and June 2026 roughly one hundred and twenty-six thousand employees lost their jobs across forty-seven American companies due to AI related factors.

The reversal started after that.

Fifty-five per cent of employers who cut staff because of AI already regret those cuts, and thirty-two per cent of hiring managers who removed positions have since added them back. By 2027 half of the companies that blamed AI for job cuts are predicted to rehire for similar functions, though the titles will be different. The work returned anyway.

Ford is the clearest case. The company brought back around three hundred veteran quality inspectors after AI did not meet quality expectations. Its leadership admitted that they had overestimated what the technology could do, that they assumed feeding design requirements into AI would produce a high quality result, and that they did not pay enough attention to the experience of their most knowledgeable engineers.

The reasons behind these reversals repeat across companies, and that is what makes them worth reading. They describe exactly where human workers are still needed.

Software can write, summarise and repeat faster than any human. What it cannot do is carry institutional knowledge or make nuanced decisions about technology. Remove those people too early and problems arrive months later, when the software is already in production and somebody has to handle the edge cases, the security questions and the human factors that were never in the specification. Reviewing and correcting the work of AI turns out to require far more expertise than companies first assumed, not less.

So read the pattern. The first wave replaced people because those companies believed AI could do the whole job. The second wave brought them back because AI could do the visible part and not the invisible part, and the invisible part is the judgement calls that were never written down anywhere.

The Bad News Nobody Says Out Loud

A blog that only tells the comfortable half is not worth reading, so here is the uncomfortable half.

The layoffs that came back were mostly the experienced people, and the people who did not get hired at all were the beginners.

Entry level technology hiring fell by roughly a quarter in a single year, and research into the labour market found something sharper still, which is that young workers in roles most exposed to AI saw employment decline noticeably more than older colleagues in the same fields. The door that used to open for a fresh graduate has narrowed, and that boring first job which taught everything does not open the same way now. It narrowed before the rehiring wave arrived to help.

This is the real danger, and it is not the danger people shout about. The danger is not that AI takes a senior engineer’s job tomorrow. The danger is that the bottom rung of the ladder got removed, and a ladder without a bottom rung cannot be climbed by anybody.

One staffing professional made a point that is easy to forget. AI makes mistakes in coding, and if nobody reviews it the cost arrives when that code reaches production. If nobody hires junior developers now there will be nobody to hire in ten years.

The honest reading of the whole picture is that the middle of the ladder is being squeezed, the top is not, and the bottom is being removed. The pressure is highest exactly where beginners stand.

That is bad news and it should be said out loud. It is also specific, and specific problems have answers.

Which Layer Of The Job Moves

Here is the mental model that makes this manageable, and almost nobody teaches it, because panic is easier to sell than structure.

A job is not one thing. A job is a bundle of layers stacked on top of each other, and automation does not delete the bundle. It climbs the stack from the bottom and stops wherever it runs out of grip.

Picture four layers, from the bottom up.

The bottom layer is typing and syntax, which means knowing which bracket goes where, remembering the library name, and getting the loop right. This layer is already mostly gone, and that is fine because it was never the job.

The next layer is solving a known problem with a known cause. A bug with a clear error message, a form that needs validating, a report that needs generating. This layer is being automated now and quickly, so if most of your day sits here the ground is genuinely moving under you.

The layer above that is deciding what should exist at all. Which problem is worth solving, what the system should do when it fails, what the trade-off is between speed and safety, and what the customer actually meant when they said something else. This layer is barely touched, because it requires knowing things that were never written down.

The top layer is taking responsibility. Signing off. Being the one whose name is on it when it breaks at two in the morning.

Machines cannot be held accountable, because there is nobody to hold. That is not a technical limitation. It is structural, and it is permanent.

So the useful question changes now. Stop asking whether AI will take your job, and start asking which layer you are standing on. Then ask how to climb one step up.

That question has an answer. The answer is uncomfortable, and it is actionable.

Two Opposite Pieces Of Advice, And Why Both Are Right

Working professionals give both of these constantly, and the contradiction confuses most beginners. Separate them properly.

One side says keep coding by hand. Do not let AI finish your projects, struggle with the errors, read the documentation, and sit with the problem. The other side says stop resisting, use AI as a co-pilot, and accept that nobody will pay for slowness.

Both are right at different levels, and the confusion comes from mixing the levels together.

Coding by hand builds the mental model. When a bug is fixed by hand, the developer learns what the bug felt like, the shape of the error, the question that should have been asked earlier, and the assumption that was quietly wrong. That internal model is what later judges whether a machine’s answer is correct. Skip this stage and a person can produce working code while being unable to tell good output from confident nonsense.

Using AI as a co-pilot is how the work is done now, and refusing it out of pride is its own kind of slowness. Used well it removes the dull parts, including boilerplate, syntax hunting and finding the right library call, and it leaves the parts that are actually human. Deciding what should be built, deciding why, and deciding what happens when it breaks.

The skill sitting above both is judgement. Knowing what is right without being told has never been commoditised, and automation raises the value of that skill rather than lowering it. Ten thousand plausible answers are cheap now, which makes the ability to identify the correct one expensive.

What Grows, And What Quietly Dies

Some skills get more valuable the longer a person holds them, while others get cheaper every year and the person holding them does not notice until the day they need them. This distinction is worth more than any list of tools.

Skills dying right now are memorising syntax, remembering commands, knowing where a setting lives in a menu, and being fast at typing boilerplate. All of these were once worth money, and all of them are being absorbed.

Skills growing right now are reading a system and understanding why it was built that way, debugging something where the cause is not yet known, judging whether an answer is correct before anyone confirms it, writing clearly enough that a human and a machine both understand the instruction, and knowing the domain itself, which means the industry, the regulations, the customers and the way things actually work as opposed to how the manual says they work.

Look at the second list and notice that none of it is about tools. All of it is about judgement, and judgement only grows through use.

Here is a practical test that takes ten seconds. Take any skill in any field, and ask one question. If a machine got ten times better at this tomorrow, would my skill become worthless or more useful? Syntax becomes worthless, while judgement becomes more useful, because there is now ten times more output to judge.

The Verification Economy

One more idea, and it changes how the whole picture reads.

When production becomes cheap, verification becomes the bottleneck, and whatever is the bottleneck is what gets paid.

This has happened before. Manufacturing made goods cheap and quality inspection became a profession. Publishing made text cheap and editing became a profession. Search made information cheap and the ability to tell good sources from bad became valuable.

Now output is cheap. Code, drafts, designs and reports can all be produced in seconds, so the scarce thing is no longer production. The scarce thing is the ability to check, and to be trusted when you say something is correct.

That is exactly what the Ford story shows. Three hundred experienced inspectors were brought back because the machine could produce and could not verify. The company had confused production with correctness, and the gap between them cost more than the salaries it saved.

The safest place to stand in any field is not the person who makes the thing. It is the person whose word decides whether the thing is right, and that role is growing rather than shrinking. It is being created right now under new titles, in companies that have already learned this lesson the hard way.

The Other Side Of The Hype

A one-sided piece is propaganda, so here is the other side.

The stories about agentic AI and robots automating everything run into something unglamorous, which is that the physical world does not behave like a demo.

A robot works in simulation and then meets reality, where it finds different friction, different lighting, different materials and different weight. The gap between simulation and the real world is one of the hardest problems in robotics, and it does not close because a press release says so. Systems that reason and then act in the physical world are still fragile. They work beautifully in the recorded test and behave unpredictably in the factory.

There is also the demo to production gap, and it applies everywhere. Early testing looks promising because the test case is narrow and clean, while the real world brings more edge cases, more security questions and more human factors than anyone expected. That is exactly how so many companies ended up rehiring the people they had just let go.

None of this means automation stops. It means automation is slower, messier and more expensive than the announcement implies. That gap is where careers are made, by people willing to look at the messy part instead of the demo.

What This Changes For Blind And Low-Vision Readers

Most articles about AI and jobs leave this part out, and it deserves its own place here.

For readers who are blind or have low vision, this wave of tools is not only a threat to employment. It is also a genuine leveller and, in some fields, a real opening.

A tool that operates a computer through instruction changes the situation directly. It reads a screen, fills a form, checks a record and reports back, which is not simply an accessibility feature bolted on at the end. It removes a dependency.

Work that once needed a sighted colleague beside you can now be driven by your own voice, at your own pace, at the hour you choose. Independence stops being a favour granted by someone else.

Notice which layer of the job that reaches. The lower layers are the mechanical, visual and repetitive parts, and those were the hardest layers to reach without sight. Automation is eating the bottom of the stack, which for blind and low-vision professionals means it is eating the part that was most in the way.

This does not mean every AI tool is accessible. Plenty are not, with unlabelled controls, visual only output, drag and drop interfaces that have no keyboard path, and feedback that exists only on screen. The habit worth building is to test a tool the way it will actually be used, with a screen reader on, with a keyboard alone, and with no sighted help in the room. If it works in that room it works, and if it only works with someone watching it does not count yet.

The practical point is simple. The same shift that makes some routine work cheaper is making some previously closed doors openable, and that is worth using rather than fearing.

Who This Is Written For

One group gets forgotten in the middle of all this, and that group is today’s youth.

Programmers who want to become developers, students who are learning coding right now, and anyone doing any kind of job in any field. All of them are scared, because AI has started entering their field and the risk of losing their job is visible to them. They can see it coming.

That fear is not imaginary, and it deserves a real answer rather than a slogan. So this is written for all of them, for staying safe and secure, for keeping the mind creative, and for what the actual solutions can be.

The Solutions

Learn the work by hand first, and then use AI on top of it. These two pieces of advice are not enemies. Coding manually builds the mental model, and that model is the only thing that lets a person judge whether the machine’s answer is correct. Skip it and a person can produce working code while being unable to tell good output from confident nonsense.

Use AI as the co-pilot without pride getting in the way. It removes the dull parts and leaves the parts that are actually human. Deciding what should be built, deciding why, and deciding what happens when it breaks. Refusing the tool is not discipline. It is just slowness.

Move up the chain instead of competing inside it. Writing code is becoming the floor rather than the ceiling, and understanding the system is where a person becomes hard to remove. The architecture, the trade-offs, the cost, the failure modes, and the reason a particular database was chosen.

Become the one who verifies, not only the one who produces. Reviewing AI output is a real professional skill now rather than a temporary inconvenience, and catching the mistake before it reaches production is a job title being created in real time.

Find the invisible part of any job. Institutional memory, knowing which complaint actually matters, knowing which client will forgive a delay and which one will not, and knowing how the process really runs as opposed to how it is documented. That knowledge is exactly what automation cannot reach, because nobody ever wrote it down. Write some of it down and you become the one the tool needs in order to be deployed at all.

Keep the mind creative. Creativity here does not mean art. It means refusing to accept that the first obvious framing of a problem is the right one. Automation handles the expected case beautifully and the unexpected case poorly, so living in the unexpected case is training for the part of the market that does not shrink.

For beginners, do not wait for the perfect entry level job. The bottom rung is narrow right now, so build the ladder yourself with open source contributions, small real projects for real people, and a portfolio that shows work rather than certificates. The formal door is narrow, while the informal doors are still open, and they open by showing something that works.

For students, retire the wrong question. Not which course should be done next, but what can be built, understood and proved over the next six to twelve months that cannot be done today. Three projects you can defend in detail beat fifteen that were copied. When a project breaks the instinct is to watch another tutorial, and the answer instead is to find the reason, read the documentation, get stuck and then get unstuck. That uncomfortable process is where information turns into ability.

Watch the economics rather than the headlines. The cost of AI is climbing even as its unit price falls, and the shortages of data centres, power and water are the speed limit. Anyone planning a career should plan for a world where intelligence is metered rather than free.

Will Human Existence Remain?

These are the real questions, and they deserve a straight answer without flattery.

Human existence will remain. That sentence alone is too easy, and cheap reassurance is worse than no reassurance.

Machines now beat humans at many narrow things and will take over many more. Specific jobs will suffer, specific teams will suffer, and specific people will suffer. Pretending otherwise insults those who lose something real. The survival of the species and the security of a livelihood are two different questions, and mixing them up is how people end up either panicking or complacent.

What the evidence of this year suggests is narrower and also more useful. The attempt to replace whole jobs ran into a wall built from judgement, from accountability, from context and from physical limits. Work came back. Not the same work, not under the same titles, and not for everyone equally, but it came back.

The physical wall is real rather than sentimental. Automating everything would need enormous amounts of electricity, water, land and money, and even on the best projections data centres reach only about three per cent of global electricity by 2030. The orbital data centre plan exists precisely because the ground based supply is straining. That number is the speed limit.

Effort spent doing what a machine already does better will keep losing value. That is not cruelty. It is arithmetic. Effort spent deciding, judging, taking responsibility and understanding context will keep gaining value, along with handling the case nobody anticipated, because those things remain scarce. Scarcity is what price is made of.

So here is the answer to the question of whether everything will slowly get automated. The routine part of everything, yes, and slowly. The whole of anything, no, and not while judgement, responsibility and the bill for electricity stay where they are. The cost is rising, the limits are real, and the humans get called back.

There is one more thing, which is less economic and more human. The value of effort was never only its market price. A person who spends two years becoming genuinely good at something difficult changes as a person, and no tool can deliver that change on their behalf. That change never appears in a salary band, and it is not up for auction.

The Question Worth Keeping

The fear is not a prophecy. It is a signal that something real is shifting, and the signal has been misread.

The machine is not arriving to replace the human being. It is arriving to replace the routine part of every human job and to make the rest more valuable. The decisions, the responsibility, the judgement, and the parts of the work that were never written down because everyone assumed they could not be taught.

That leaves two questions, and only one of them deserves your energy.

The first is what if it takes everything. That question cannot be answered and cannot be planned around, and worrying about it produces nothing except lost months.

The second is what part of the work is invisible, and how does a person become the one who holds it. That question has an answer and it has a plan, and it rewards acting this month rather than next year.

Most people will spend the next five years arguing about the first question. A few will quietly build the answer to the second, and they will be the ones still needed when the dust settles.