A vulnerability was found.

A researcher tested it.

A team verified it.

A patch was written.

Attackers tried to move first.

Defenders tried to close the gap.

That race has always existed.

But it mostly moved at human speed.

People searched.

People reviewed.

People wrote exploits.

People patched systems.

Now that rhythm is changing.

The New Race Is Between AI Systems

OpenAI has introduced Daybreak, a cybersecurity platform designed to automate vulnerability detection, patch validation, threat modeling, and secure software development for enterprises and governments (CSO Online, 2026).

That puts it directly into the same arena as Anthropic’s Project Glasswing and Claude Mythos Preview, which Anthropic says has already found thousands of high-severity software vulnerabilities across major systems (Anthropic, 2026).

This is no longer just a model race.

It is a cyber race.

Who can find the flaw first.

Who can validate it first.

Who can patch it first.

Who can deploy the defensive system before attackers get the same kind of capability.

OpenAI Is Building A Defense Platform

Daybreak is not framed as a chatbot.

It is framed as a cyber-defense system.

OpenAI says the goal is to help defenders reason across codebases, identify subtle vulnerabilities, validate fixes, analyze unfamiliar systems, and move from discovery to remediation faster (OpenAI, 2026).

That matters.

Because the goal is not just to answer cybersecurity questions.

It is to enter the software development loop.

Threat modeling.

Code review.

Patch validation.

Dependency analysis.

Remediation guidance.

This is AI moving into the workflow of defense itself.

Anthropic Is Treating Mythos Like A Restricted Weapon

Anthropic’s approach is different.

Project Glasswing gives a small group of major organizations access to Claude Mythos Preview for defensive security work.

That group includes AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks (Anthropic, 2026).

The reason access is limited is the same reason the model is valuable.

Anthropic says Mythos has reached a level of coding capability where it can surpass all but the most skilled humans at finding and exploiting software vulnerabilities (Anthropic, 2026).

That is a strange threshold.

A model becomes useful enough to defend with.

And dangerous enough not to release broadly.

The Firefox Numbers Show Why This Matters

Mozilla’s Firefox team used Mythos in a controlled security process and reported a dramatic jump in bug-fixing volume.

In April 2026, Firefox shipped 423 bug fixes, compared with 31 in the same month a year earlier (TechCrunch, 2026).

Mozilla also said Mythos helped uncover high-severity bugs, including issues that had been dormant for more than a decade (TechCrunch, 2026).

That is not a small improvement.

That is a change in tempo.

The tool did not replace the security team.

But it changed what the team could find.

And how quickly they could move.

This Is What Makes It An Arms Race

The same capability helps both sides.

A model that can find vulnerabilities can help patch them.

But it can also help exploit them.

A model that can reason across a codebase can help secure software.

But it can also identify weak points before defenders know they exist.

OpenAI says Daybreak pairs expanded defensive capability with trust, verification, proportional safeguards, and accountability because the same capabilities can be misused (OpenAI, 2026).

That is the dual-use problem.

Defense and offense are built from the same underlying skill.

The Race Is Not Just Who Has The Strongest Model

It is who gets access.

Who gets vetted.

Who gets trusted.

Who gets the model early enough to patch before attackers adapt.

OpenAI is rolling Daybreak through trusted partners and different access tiers, including GPT-5.5 with Trusted Access for Cyber and GPT-5.5-Cyber for specialized authorized workflows (CSO Online, 2026).

Anthropic is keeping Mythos restricted to selected organizations under Project Glasswing (Anthropic, 2026).

So cyber AI is not spreading like normal software.

It is spreading through controlled access.

Defenders get capability first.

Or at least that is the goal.

The Window May Not Stay Open

Anthropic warned that given the rate of AI progress, it may not be long before similar vulnerability-finding capabilities spread beyond actors committed to using them safely (Anthropic, 2026).

That is the pressure underneath the whole story.

The advantage may be temporary.

The defensive lead may not last.

If more actors get access to similar capabilities, the cyber landscape changes again.

Not because vulnerabilities are new.

Because the speed of finding them changes.

Cybersecurity Becomes Continuous

This is why Daybreak is focused on continuous security.

Not occasional audits.

Not one-time reviews.

Not quarterly scans.

Continuous defense.

Software changes constantly.

Dependencies update.

New code enters the system.

Old vulnerabilities sit unnoticed.

AI makes it possible to keep checking more often.

But it also makes that necessary.

Because if attackers can move continuously…

Defenders have to as well.

The Human Role Does Not Disappear

Mozilla’s experience makes this clear.

Mythos helped find bugs.

But humans still built the harness.

Reviewed the results.

Ran tests.

Filtered noise.

Integrated patches.

Mozilla emphasized that the value came from both more capable models and better human techniques for harnessing them (TechCrunch, 2026).

That is important.

Cyber AI is not magic by itself.

It needs systems around it.

Sandboxes.

Review processes.

Security teams.

Validation.

Without that, the model can produce noise.

Or worse, confidence without reliability.

The Strange Part Is That Both Sides May Improve

This is what makes arms races unstable.

Defenders get stronger.

Attackers get stronger.

Each side forces the other to speed up.

AI may help security teams patch faster.

But it may also help attackers find more creative paths.

That means progress does not automatically make the system safer.

It may make the race faster.

And faster races are harder to govern.

The Next Cyber Advantage May Be Timing

Who sees the vulnerability first?

Who validates it first?

Who patches it first?

Who turns it into an exploit first?

Who deploys protection before the other side understands the path?

This is where cyber AI changes the game.

It compresses discovery.

It compresses analysis.

It compresses response.

The question is no longer only whether a system is vulnerable.

It is whether defenders can move before machine-speed attackers do.

Cyber AI is becoming the new arms race.

OpenAI has Daybreak.

Anthropic has Mythos.

Governments and major companies are being pulled into trusted-access programs.

Security teams are seeing bugs that sat unnoticed for years.

And the same capabilities that help defenders can also help attackers.

That is what makes this moment different.

The old race was human against human.

The new race may be AI-assisted defender against AI-assisted attacker.

And in that world, the winner may not be the side with perfect security.

It may be the side that moves first.

References

• CSO Online (2026). OpenAI introduces Daybreak cyber platform, takes on Anthropic Mythos.

• OpenAI (2026). Daybreak: Frontier AI for cyber defenders.

• Anthropic (2026). Project Glasswing: Securing critical software for the AI era.

• TechCrunch (2026). How Anthropic’s Mythos has rewritten Firefox’s approach to cybersecurity.

The training story used to be about language.

Books.

Websites.

Code.

Articles.

That made sense.

AI needed to learn how humans write.

How we explain.

How we answer.

How we reason in text.

But work is not only text.

Work happens through movement.

Clicks.

Tabs.

Shortcuts.

Tiny decisions across screens.

And now AI needs that too.

This creates an awkward tension between AI and employees…

Where AI is learning how to automate their workflow.

Read more

It made a certain kind of work almost disappear.

The typing.
The business plan.
The email.
The code.

All of that used to be where the effort lived. You sat down, pushed through the task, and eventually something existed. The work was slow, and furthest from frictionless.

Now the path is different.

The machine can produce the same thing almost instantly.

But before it does, the human still has to decide what the “thing” is supposed to be.

ChatGPT removed friction from the middle of the process.

But for many people, it moved that friction to the very beginning.

Why This Piece Matters

That is why I’m excited to share this collab piece from Hodman Murad.

Hodman is the founder of Asaura AI, a productivity tool built for ADHD, neurodivergent brains, and decision-fatigued professionals. She also writes at Between Thinking and Doing, where she explores the behavioral science behind AI and the future of work, and at The Data Letter, where she covers data science, data engineering, and machine learning.

Asaura AI → asauraai.com
Between Thinking and Doing → betweenthinkingdoing.substack.com
The Data Letter → thedataletter.com

Her article gets at something I think a lot of people feel but do not always have the language for.

AI has made the output easier.

But it has not always made the starting point easier.

In some cases, it has made the starting point feel even heavier.

Hodman’s piece is about that gap between what AI can technically do and what a human brain can actually initiate when it is tired, overloaded, distracted, or stuck. It is about why the blank prompt can feel just as difficult as the blank page. And it is about why the next real step in AI productivity may not be making the machine faster.

It may be helping the person begin.

AI Relocated the Hard Part of Thinking

Hodman Murad

How ChatGPT Shifted Cognitive Friction from Execution to Initiation

ChatGPT can generate a business plan in 90 seconds. Yet millions of people open it every day, stare at the prompt field, and close it without typing a word. The tool works. The user stalls before the tool even fires. They can’t figure out what to ask, how to frame it, or where to begin.

AI didn’t remove the hard part of work. It relocated it.

Before ChatGPT, friction lived in execution. Writing a report meant hours of research, drafting, and formatting. Building a website meant writing code line by line. You could feel the hard part in your hands. AI compressed all of that. A task that took two hours now takes ten minutes of generation. The time shifted upstream.

Writing an email used to take 20 minutes, including manual drafting. Now ChatGPT generates five versions in 30 seconds. You spend the next 15 minutes reading all five, comparing tone, swapping paragraphs between drafts, and wondering if any of them sound like you. The mechanical effort dropped. The cognitive effort grew. Work moved from doing to deciding, from execution to initiation.

This creates a strange new pattern in the workday. You have ChatGPT open in a tab. You know it could save you an hour. But you spend 20 minutes cycling between prompts you’ve half-written and deleted. You close the tab and do the work manually, feeling defeated. The tool that was supposed to help became another source of friction, and because everyone around you talks about how productive AI makes them, you assume the problem is you.

For a large portion of the workforce, initiation was always the hardest part.

Symptomatic ADHD affects roughly 6.76 percent of adults worldwide. Executive dysfunction, one of its core features, means the brain’s planning and initiation systems go offline under cognitive load. More options, more open-ended choices, and more blank fields push those systems toward failure faster.

Decision fatigue compounds the problem. A senior manager who’s made 300 micro-decisions by 3 pm doesn’t have the bandwidth to craft a precise prompt. A freelancer juggling five clients doesn’t have the cognitive budget to evaluate four AI-generated drafts and pick one. A parent with ADHD, running on four hours of sleep, sits down at 10 pm to finish a presentation. The AI is ready. Their brain checked out two hours ago.

These people need a tool that structures the first step for them.

Every major AI assistant carries the same built-in assumption. You arrive with intent, clarity, and the ability to evaluate outputs. For millions of people living with ADHD, chronic fatigue, or decision overload, that assumption collapses before the first keystroke.

This group represents a growing share of the workforce. A 2020 global meta-analysis estimates 366 million adults live with symptomatic ADHD alone. Add chronic decision fatigue, burnout, and cognitive overload, and the number climbs far higher. Nobody built the tools with them in mind.

I built Asaura AI to solve this.

Asaura is a productivity tool designed for people with ADHD and neurodivergent brains, as well as anyone experiencing decision fatigue or cognitive overload at work. Instead of dropping users into a blank prompt field, it surfaces the next smallest action based on where they left off.

When someone types ‘write resume’ into a typical AI tool, they get a blank page or a generic template. In Asaura, that vague task breaks down into concrete micro-steps before the user has to think. ‘Open a blank document.’ ‘Type your name and contact info at the top.’ ‘Write a one-line professional summary underneath.’ ‘List your jobs starting with the latest one, working backward.’ Each step is small enough to start without overthinking. Eight steps later, the resume is done.

The system adapts to cognitive energy, serving lower-demand tasks during low-energy states and deeper work during high-focus windows. This is initiation-first design. Instead of waiting for you to start and then accelerating, the tool structures the first move so you don’t have to.

Once someone takes a single concrete step, momentum carries them forward. The person who couldn’t start a resume for three hours finishes it in 40 minutes after one micro-step breaks the freeze. The barrier was never ability or motivation. It was the tool they were using.

AI gets more powerful every quarter. Bigger context windows. Faster responses. More tokens per second. Nearly all of that effort targets throughput, making the fast parts faster.

Almost nobody is optimizing for initiation.

Look at how the industry talks about progress. Speed benchmarks. Token throughput. Reasoning performance on standardized tests. Every metric assumes a user who has already typed a prompt. Nobody measures how many people opened the tool, stared at the field, and closed it again. Nobody tracks the sessions that never started. This is where the largest group of potential users lives.

Power without accessibility is a library with no front door. The books keep multiplying, and the number of people standing outside keeps growing as decision loads increase and cognitive bandwidth shrinks across every industry.

We celebrate AI that writes code in seconds, drafts legal briefs in minutes, and builds entire apps from a conversation. All of that is remarkable. All of it assumes you’ve already crossed the hardest threshold: beginning.

Every AI builder should ask one question about their product. Does your tool assume a user who’s already moving? If the answer is yes, you’ve relocated the hard part and called it progress.

AI’s next frontier is initiation. Hardly anyone is building there.

Hodman Murad is the founder of Asaura AI, a productivity tool built for ADHD, neurodivergent brains, and decision-fatigued professionals. She writes at Between Thinking and Doing, on the behavioral science behind AI and the future of work, as well as at The Data Letter, a publication focused on data science, data engineering, and machine learning. You can visit the Asaura AI website here and sign up for the app directly here.

There is cheap AI.

And there is trusted AI.

They look similar from the outside.

Both answer questions.

Both summarize.

Both draft.

Both sound confident.

But they are not the same thing.

One is useful when the stakes are low.

The other has to survive when mistakes are expensive.

That difference is starting to matter.

Read more

Software scales cleanly.

You build it once.

Ship it everywhere.

Users log in.

Companies subscribe.

The product spreads.

That was the old model.

And for a while, AI looked like it would follow the same path.

A model.

An API.

A chatbot.

A dashboard.

Something companies could plug in and immediately use.

But That Is Not What Is Happening

The models are powerful.

The demos are impressive.

The benchmarks keep improving.

But inside real companies, AI does not spread like a normal app.

It has to be fitted into workflows.

Connected to messy data.

Adjusted around compliance rules.

Explained to teams.

Tested against real processes.

The technology may be general.

But the implementation is specific.

That is why OpenAI and Anthropic are reportedly moving toward enterprise AI services and acquisitions, not just model releases (Reuters, 2026).

The Bottleneck Is Not Just Intelligence

For the last few years, the AI race looked simple from the outside.

Build the strongest model.

Release it.

Win users.

But enterprise AI is exposing a different problem.

A strong model does not automatically know how a hospital works.

Or a bank.

Or a manufacturer.

Or an insurance company.

Every organization has its own systems.

Its own permissions.

Its own documents.

Its own internal language.

Its own broken processes.

So the real question becomes less:

How smart is the model?

And more:

Can anyone make it useful here?

That Is Why The Services Layer Matters

Anthropic announced a new AI-native enterprise services firm with Blackstone, Hellman & Friedman, and Goldman Sachs, designed to bring Claude into the core operations of mid-sized companies (Anthropic, 2026).

That is not a normal software rollout.

That is implementation work.

Anthropic said the new firm will use applied AI engineers to identify where Claude can have the most impact, build custom solutions, and support customers over time (Anthropic, 2026).

That sounds less like downloading software.

And more like consulting.

Reuters also reported that OpenAI and Anthropic’s joint ventures are looking to acquire AI services firms so they can bring in hundreds of engineers and consultants to help companies put AI models to work (Reuters, 2026).

That tells you something.

The hard part is not only building the intelligence.

It is translating intelligence into operations.

The Model Alone Is Not The Product

A model can answer questions.

But a company does not need answers in the abstract.

It needs invoices processed.

Contracts reviewed.

Customer support improved.

Sales workflows changed.

Claims handled.

Reports generated.

Code maintained.

Decisions routed.

That requires integration.

The model has to enter the company’s machinery.

And that machinery is rarely clean.

From the outside, the technology looks ready.

From the inside, the organization may not be.

The data may be messy.

The workflows may be unclear.

The legal team may be cautious.

The employees may not trust the output.

The systems may not connect.

So AI progress moves at two speeds.

The model improves quickly.

The company changes slowly.

That Gap Creates A New Market

A company does not just ask:

Which model should we use?

It asks:

Where should we use it?

Who owns it?

What data can it touch?

What happens if it is wrong?

How does it fit into the workflow?

Who maintains it after launch?

Those are not just software questions.

They are organizational questions.

And organizational questions create services businesses.

TechCrunch reported that Anthropic and OpenAI are both moving into enterprise AI services through joint-venture structures, with Anthropic’s new venture backed by major financial firms and OpenAI preparing a similar effort (TechCrunch, 2026).

That is the signal.

AI is not just being sold.

It is being installed into companies.

Private Equity Makes The Pattern Clearer

The Anthropic venture is not just a lab partnership.

It is backed by firms with access to large networks of portfolio companies.

Blackstone said the new AI services firm will benefit from a consortium network across hundreds of companies, helping design, build, and maintain enterprise AI deployments (Blackstone, 2026).

That is a rollout channel.

Not just a customer list.

A way to push AI through existing business networks.

That makes the model more like consulting at scale.

Find the use case.

Deploy the team.

Integrate the system.

Measure the impact.

Repeat across companies.

This Changes The AI Race

The race is no longer only about who has the best model.

It is also about who has the best deployment machine.

Who can get inside companies.

Who can customize faster.

Who can prove ROI.

Who can build trust with executives.

Who can make AI work with legacy systems.

That is a very different competition.

It favors not just research labs.

But partnerships.

Distribution.

Consultants.

Enterprise relationships.

Operational knowledge.

The best model may not win by itself.

The winner may be the company that makes AI easiest to adopt.

Easiest to trust.

Easiest to integrate.

Easiest to maintain.

The Irony Is Hard To Miss

AI was supposed to reduce complexity.

But selling AI into companies creates more complexity first.

More strategy decks.

More integration work.

More change management.

More internal training.

More consultants.

More hand-holding.

The automation wave may need a human deployment wave before it can scale.

That is the paradox.

The technology that promises to remove work has to be worked into place.

This is not a failure of AI.

It is a sign that the real world is harder than the demo.

A chatbot can answer cleanly.

A company cannot transform cleanly.

Real organizations have history.

Politics.

Regulation.

Bad data.

Old systems.

Human resistance.

AI does not erase those things.

It has to move through them.

And that is why it spreads less like software…

And more like consulting.

AI is not spreading like software.

Software could be shipped.

AI has to be embedded.

It has to enter workflows.

Adapt to company data.

Fit into operations.

Survive compliance.

Earn trust.

That takes people.

Engineers.

Consultants.

Implementation teams.

The strange thing is that the AI revolution may need a lot of human work to arrive.

Not because the models are weak.

But because companies are messy.

And the next phase of AI may not be won only by building the smartest system.

It may be won by knowing how to make that system work inside the real world.

References

Reuters (2026). OpenAI, Anthropic ventures in talks to buy AI services firms.

Anthropic (2026). Building a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs.

Blackstone (2026). Anthropic partners with Blackstone, Hellman & Friedman, and Goldman Sachs to launch enterprise AI services firm.

TechCrunch (2026). Anthropic and OpenAI are both launching joint ventures for enterprise AI services.

A company built a model.

Tested it internally.

Announced it.

Released it.

Then the public reacted.

That was the normal rhythm.

Build.

Launch.

Use.

Improve.

But frontier AI is starting to move into a different category.

The most powerful models are no longer treated like ordinary software updates.

They are starting to look like systems that need inspection before they reach the public.

Read more

Misinformation is easier to name.

A false claim.

A fake image.

A fabricated story.

Something wrong enough to challenge.

Almost-knowledge is different.

It sounds informed.

It has structure.

It uses the right words.

It gives you enough to feel oriented.

But not enough to truly understand.

That Makes It Harder To Notice

A bad answer used to reveal itself more often.

It was messy.

Thin.

Obviously wrong.

Now weak information can arrive cleanly.

Summarized.

Organized.

Confident.

The problem is not always that the answer is false.

Sometimes the problem is that it is shallow.

It gives you the shape of knowledge…

without the depth underneath.

AI Makes Coherence Cheap

That is the shift.

Coherence used to take work.

You had to understand something well enough to explain it.

Not always perfectly.

But enough to connect the pieces.

Now a system can produce that connection instantly.

A summary.

A framework.

A comparison.

A clean explanation.

Large language models can produce fluent answers while still generating false or unsubstantiated claims, which is why hallucination remains a major problem in AI systems (Nature, 2024).

So the internet gets more readable.

But not always more reliable.

“Almost-Knowledge” Feels Useful

That is why it spreads.

It helps you move.

It gives you a starting point.

It explains enough to keep going.

You feel less lost.

And in many cases, that is genuinely helpful.

The danger is that the feeling of orientation can arrive before real understanding.

You know the words.

You know the outline.

You know the general shape.

But you may not know where the explanation breaks.

The Surface Gets Stronger Than The Substance

A shallow idea can now sound serious.

A weak summary can feel complete.

A secondhand explanation can look like research.

That changes how information is judged.

Because people do not only evaluate truth.

They evaluate fluency.

Clarity.

Confidence.

Structure.

And AI is very good at producing those signals.

This Is Not Just Spam

Spam is easy to dismiss.

Almost-knowledge is more dangerous because it can be helpful.

It may be partly right.

It may be directionally useful.

It may contain enough truth to pass.

But it is still missing something.

Context.

Limits.

Uncertainty.

The parts that make knowledge sturdy.

Researchers have warned that LLM hallucinations can spread through social networks as generated content becomes part of the broader information environment (Scientific Reports, 2024).

That means low-depth coherence does not stay isolated.

It circulates.

The Internet Starts To Fill With Explanations

Every topic gets summarized.

Every debate gets framed.

Every question gets an answer.

That sounds like progress.

But it also creates a strange kind of saturation.

Not too little information.

Too much information that feels close to knowledge.

Enough to make people confident.

Not enough to make them careful.

“Almost-Knowledge” Is Often The First Layer

Most people do not begin with primary sources.

They begin with summaries.

Search results.

Threads.

AI answers.

Short videos.

Generated explainers.

That first layer shapes the rest.

If it is shallow, everything built on top of it becomes weaker.

Even if it sounds good.

This Changes What It Means To Be Informed

Being informed used to require more friction.

Reading.

Comparing.

Checking.

Sitting with contradiction.

Now you can get a clean version immediately.

That does not mean you are uninformed.

But it does mean the threshold has changed.

You can feel informed much faster than you can become informed.

That gap matters.

Almost-Knowledge Also Changes Expertise

Experts do not just know the answer.

They know where the answer is fragile.

They know the exception.

The caveat.

The hidden assumption.

The thing that sounds right but is not.

Almost-knowledge rarely carries that layer well.

It gives the visible answer.

Not always the judgment behind it.

The New Skill Is Not Finding An Answer

Answers are everywhere now.

The new skill is knowing how much weight an answer can hold.

Is this a starting point?

A summary?

A claim?

A real explanation?

A shortcut?

A guess?

That distinction becomes more important as AI-generated content grows.

Verification techniques like Chain-of-Verification have been proposed because plausible but incorrect factual information remains an unsolved issue in large language models (ACL, 2024).

The need for verification tells you something.

The answer is no longer enough.

The internet is not only filling with misinformation.

It is filling with almost-knowledge.

Clean explanations.

Confident frameworks.

That may be the harder problem.

Because almost-knowledge does not always look wrong.

It looks useful.

It looks clear.

It looks like enough.

And in a world where everything can be explained instantly…

the real danger may not be ignorance.

It may be the feeling that we understand more than we actually do.

References

• Nature (2024). Detecting hallucinations in large language models using semantic entropy.

• Scientific Reports (2024). Quantifying the uncertainty of LLM hallucination spreading in complex adaptive social networks.

• ACL (2024). Chain-of-Verification Reduces Hallucination in Large Language Models.

For a long time, AI consciousness sounded like a future problem.

Maybe science fiction.

Maybe philosophy.

Maybe something for later.

The immediate questions were easier to name.

Bias.

Privacy.

Safety.

Jobs.

Misinformation.

Those were human problems.

They affected people directly.

So AI ethics stayed mostly focused on us.

Read more

Physics has always reached beyond intuition.

Atoms.
Black holes.
Quantum fields.
Curved spacetime.

None of that was ever simple.

But there was still a feeling that the map could be held.

An equation.

A diagram.

A model.

Something compressed enough for the human mind to carry.

That was part of the beauty.

Reality was strange.

But it could still be made graspable.

That Is Starting To Change

The systems physicists study now are getting larger.

Not just physically larger.

Conceptually larger.

More variables.

More dimensions.

More interactions.

More data than a person can meaningfully hold at once.

Physics is no longer just asking what happens to one particle, one object, or one field.

It is asking what happens inside systems too complex to fully picture.

That changes the role of human intuition.

AI Enters Because The Map Got Too Big

Machine learning is becoming useful in physics because many problems are expensive to simulate, difficult to solve, or too high-dimensional for traditional methods alone.

Neural operators, for example, are being developed to accelerate scientific simulations by learning mappings between complex functions such as spatiotemporal processes and partial differential equations (Nature Reviews Physics, 2024).

That matters because physics is full of systems that evolve across space and time.

Fluids.

Plasmas.

Materials.

Climate-like dynamics.

The old map is still there.

But AI is helping expand it beyond what can be manually drawn.

This Is Not Just Better Calculation

That would be easier to understand.

A faster calculator.

A better simulator.

A more efficient tool.

But AI is doing something stranger.

It is helping researchers search areas of theory and data that may be difficult to navigate through intuition alone.

Recent work in mathematics and theoretical physics has described AI as increasingly useful for exploring complex structures, generating conjectures, and assisting discovery across spaces too large for direct human inspection (Nature Reviews Physics, 2024).

So AI is not only speeding up physics.

It is changing where physics can look.

The Human Mind Becomes Less Central To The Search

That is the quiet shift.

Physics is still human-led.

Humans choose the questions.

Humans interpret the results.

Humans decide what counts as meaningful.

But the search itself is becoming less human-sized.

A model can scan relationships no person would naturally inspect.

It can explore parameter spaces too wide to imagine.

It can find patterns that do not arrive through intuition.

That does not replace physicists.

But it changes what physicists are standing in front of.

Physics Starts To Look Less Like A Picture

And more like a landscape.

A huge one.

Too large to see from the ground.

Too detailed to draw by hand.

AI becomes a kind of instrument.

Not a microscope.

Not a telescope.

Something closer to a mapmaker for spaces we cannot naturally perceive.

That is different from using AI to answer a question.

It is using AI to reveal where the questions might be.

A Strange New Dependency

Once the map gets too large…

You need the system to navigate it.

Not because humans are weak.

But because the territory is bigger than human intuition.

This is already visible in data-driven physics, where machine learning has become increasingly common across physics research and was described by Nature Reviews Physics as a rapidly growing part of the field (Nature Reviews Physics, 2021).

The issue is not that physics loses rigor.

The issue is that the path to rigor may pass through systems humans cannot fully intuit at first glance.

That Can Create An Illusion Of Understanding

This is where caution matters.

AI can make a field feel more navigable.

It can produce outputs.

Predictions.

Patterns.

Visualizations.

But that does not automatically mean humans understand the underlying system.

Researchers have warned that AI tools can create “illusions of understanding” in science, where productivity and pattern-finding increase while the deeper limits of interpretation become less visible (Nature, 2024).

That warning matters here.

Because physics has always cared about more than results.

It cares about what those results mean.

The Direction Is Hard To Ignore

The 2024 Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for foundational work in machine learning, a symbolic moment showing how deeply AI and physics are now intertwined (Reuters, 2024).

That does not mean AI is physics.

But it does show that physics is no longer separate from the tools that now expand it.

The relationship is becoming circular.

Physics helped build the ideas behind machine learning.

Now machine learning is helping physics search its own unknowns.

The Scale Of The Question Changes

The old question was:

Can we understand this system?

The newer question may become:

Can we navigate it well enough to discover something real?

That is not the same question.

Understanding still matters.

But navigation becomes more important when the territory is too large to hold.

AI expands the searchable space.

And once the searchable space expands, the scientific map changes with it.

Physics is not becoming less human.

But it may be becoming less human-sized.

The systems are too large.

The data is too dense.

AI does not remove the need for theory.

It does not replace judgment.

It does not make reality simple.

But it expands the map.

And once the map expands beyond what human intuition can naturally hold…

Physics changes.

Not because humans stop doing it.

But because the territory they are exploring becomes too large to see without another kind of mind beside them.

References

• Nature Reviews Physics (2024). Neural operators for accelerating scientific simulations and design.

• Nature Reviews Physics (2024). AI-driven research in pure mathematics and theoretical physics.

• Nature Reviews Physics (2021). The rise of data-driven modelling.

• Nature (2024). Artificial intelligence and illusions of understanding in scientific research.

• Reuters (2024). Nobel Prize in Physics goes to machine learning pioneers Hopfield and Hinton.

You generate a draft.

A title.

A plan.

A mockup.

A version of something you might use.

Nothing about that feels wasteful.

It feels productive.

Like movement.

Like progress.

Read more

It was a portfolio project. I’ve worked for three months.

Backend in Python Flask. Auth flows. Database migrations.

A deploy pipeline that broke twice a week.

I built it. I was a developer.

As of today, I am building apps each week.

None required code. None required deploy. Some of them took no longer than a coffee break.

I’m not the developer in any of these stories.

I’m the person describing what should exist.

That’s the shift.

1. Claude Artifact

An artifact is a small app that lives inside a Claude conversation.

You describe what you want, and a working tool renders in the chat window.

You can use it, share the link, edit it by talking.

Last month I built a Substack Notes Generator as an artifact.

It analyzes a writer’s existing notes, learns their pattern, and generates new ones in their voice, here is the one, trained on my voice.

claude-artifact-1.mp4

I didn’t open VS Code. I described the tool, tested it in the same window, sent the link.

Now let’s do it for one more.

Now let’s go to Exploring ChatGPT.

Scroll down for a month.

Right click, click on save as and save it as an html.

Save it as html.

Go to Claude.ai.

Paste this prompt.

You are analyzing a Substack writer’s Notes to extract their writing DNA.

I have attached the HTML of their Notes page. Your job is to reverse-engineer

how they write so another person could produce Notes in the same voice.

Step 1. Extract the Notes from the HTML.

Step 2. Analyze opening patterns, sentence rhythm, structural formulas,

topic territory, voice markers, and what they avoid.

Step 3. Produce a DNA card in a fixed format.

Step 4. Build a single-file HTML artifact that generates Notes in this

writer’s voice, using the Anthropic API.

Output all four in full. Do not skip steps.

Here is the artifact.

claude-artifact-2.mp4

It generates, based on Leor’s tone and also analyzed the DNA of his notes.

2. NotebookLM and Gemini

NotebookLM is Google’s research tool.

You drop sources in. It indexes them. You can ask questions, generate study guides, even get a podcast version of the material.

Most people use it for studying. I use it as a backend.

Anthropic released a 90-minute video on prompt engineering.

Dense. Practical. The kind of material you want to internalize, not watch once.

I dropped the YouTube URL into NotebookLM.

How?→ Steps: Visit NotebookLM → Try NotebookLM → Create New NotebookLM → Websites → Paste Link

Then I asked it to extract every prompt engineering principle, every example, every anti-pattern, into structured notes.

That part took 40 seconds.

Then I connected Gemini.

How?→ Steps: Visit Gemini → Click “+” → Select NotebookLM → Click Tools → Select Canvas

And paste this prompt.

Turn this an app.

Here how it looks like.

No backend. No frontend.

Let’s paste and see the result.

NotebookLM and Gemini-1.mp4

Here is the artifact. You can also refine your prompts using techniques from Anthropic’s prompt engineering YouTube video, since this was built with NotebookLM trained on it.

3. Codex and Claude Code

This one is the most advanced of the three.

Yesterday, Anthropic released a new analysis they made.

They analyzed 1M conversations to understand users behavior, using AI.

Reference

Here is the research report.

Reference

It looks like people are using AI to ask their health and wellness.

Now let’s build an app, simple app using both Claude Code and Codex.

The aim?

Building a health advisor that takes a diagnosis, asks follow-up questions if needed, and points you to the right resources.

Let’s see the one-shot prompt.

Build a single-file health advisor web app. HTML + CSS + JavaScript in one file. No backend.

Look and feel:

- Retro 90s medical software vibe. Think old hospital terminal meets Windows 95.

- Monospace font (VT323 or Press Start 2P from Google Fonts).

- Color palette: dark green background, bright green text, amber accents. CRT scanline effect on the screen.

- Window chrome with title bar that says “HEALTH ADVISOR v1.0”.

- Blinking cursor in input fields.

- Subtle screen flicker animation.

Behavior:

1. User types a diagnosis or symptom in an input field.

2. If the input is vague (one word, ambiguous), the app asks 1-2 follow-up questions before responding. Show questions in a chat-style log.

3. Once the condition is clear, return 3 evidence-based insights based on standard clinical knowledge. For each: a short title, a 2-sentence explanation, and a flag if a specialist visit is recommended.

4. Show a typing animation when the app responds, like an old terminal printing characters one by one.

5. Add a disclaimer banner at the bottom: “NOT MEDICAL ADVICE. CONSULT A LICENSED PHYSICIAN.”

Requirements:

- Single HTML file, no external dependencies except Google Fonts.

- Works offline after first load.

- Use the model’s own knowledge for medical content. No external APIs.

Codex

After just 3min 19s later.

Let’s see.

chatgpt-codex-sr.mp4

Claude Code

After couple minutes, Claude is also finished.

Let’s see.

claude-artifact-1.mp4

Same prompt, two different aesthetics. Codex went minimal, Claude Code went full retro. Both are working perfectly well and built in minutes.

What changed

In each of these, the actual work happened in language.

Software used to be built by people who knew the syntax.

Now it gets built by people who know the goal.

The ten years I spent learning frameworks didn’t disappear. They moved. They turned into the part of my brain that knows what to ask for, what’s missing, what won’t scale, where the tool will break.

The typing was the visible part. It was never the valuable part.

A developer in 2021 had a job.

A developer in 2026 has a conversation.

A warning went out.

Teams tested the fix.

Then they patched.

That process was never perfect.

But there was time.

Days.

Weeks.

Sometimes longer.

That time mattered.

It gave defenders a chance to move before attackers fully caught up.

Read more

Everyone has versions of themselves they never became.

The business you didn’t start.

The book you didn’t write.

The city you didn’t move to.

The career you didn’t pursue.

Those paths existed.

But softly.

They stayed blurred enough to live with.

AI Makes Them More Concrete

Now you can render them.

You can ask for the business plan.

The first chapter.

The brand.

The five-year roadmap.

The apartment search.

The pitch deck.

The version of your life that used to stay imaginary can now show up with details.

Not fully real.

But real enough to look at.

That Changes The Feeling Of Possibility

Before, the path you didn’t take stayed abstract.

You could tell yourself it may not have worked.

It may not have been practical.

It may not have become anything.

That uncertainty protected you.

Now the system can generate a plausible version.

A coherent version.

A version with steps.

And once something has steps, it feels less impossible.

Research on counterfactual thinking shows that imagined alternatives feel more plausible when they resemble the real world closely enough, which is exactly what AI-generated plans, mockups, and scenarios can make easier to produce (Cognition, 2021).

Regret Gets More Specific

Regret used to be emotional.

Now it can become visual.

Structured.

Actionable.

You don’t just think:

I wonder what would have happened.

You can see a version of what might have happened.

A launch plan.

A cover mockup.

A possible audience.

A different future with enough texture to feel uncomfortable.

The Strange Part Is That It Still Didn’t Happen

The simulation is not the life.

The plan is not the work.

The mockup is not the years.

But emotionally, the distance narrows.

Because the system gives the unlived path enough form to feel close.

Close enough to compare with the life you actually chose.

That Comparison Is New

People have always imagined alternatives.

That is not new.

But imagination used to have limits.

It faded.

It simplified.

It protected itself from detail.

AI removes some of that softness.

It can keep going.

It can fill in gaps.

It can make the alternate version feel more complete than it ever was in your own mind.

Researchers have started examining AI systems that simulate counterfactual worlds, noting that these tools can change how people imagine possible alternatives and reason about what might have happened (AI and Ethics, 2025).

The More Complete It Looks, The More It Asks Of You

A vague dream is easy to dismiss.

A detailed plan is harder.

Once the system shows you what something could look like, not doing it feels different.

It starts to feel less like impossibility.

And more like refusal.

That is where the emotional cost appears.

You Start Carrying More Versions Of Yourself

Not because you lived them.

Because you previewed them.

You saw enough to make them feel partially real.

The founder version.

The artist version.

The person who moved.

The person who left.

The person who tried.

They are not memories.

But they are not nothing either.

This Changes How You Experience The Present

The life you are living has to compete with more visible alternatives.

Not perfect alternatives.

Not guaranteed alternatives.

But generated ones.

And generated alternatives have a strange advantage.

They can be cleaner.

More coherent.

Less constrained by reality.

They don’t have to carry the full cost.

That Makes The Present Feel Heavier

Because reality is messy.

It moves slowly.

It contains trade-offs.

The simulated path can appear without the fatigue.

Without the delay.

Without the boring middle.

So the life you chose may start to feel less optimized than the ones you can generate.

The Unlived Life Is Missing Something Important

It does not contain resistance.

Not real resistance.

Not the day you lose motivation.

Not the unexpected bill.

Not the awkward conversation.

Not the quiet months where nothing happens.

The simulation can show the shape.

But not the weight.

That Difference Matters

Because lives are not only made of outcomes.

They are made of friction.

Repetition.

Compromise.

Recovery.

Boredom.

The generated version skips too much of that.

It shows the path as if it can be held all at once.

But real life can only be lived one day at a time.

The Image Stays With You

Even if you know it is not real.

Even if you understand the difference.

The visible version has already entered your mind.

It becomes easier to imagine.

Easier to compare.

Harder to fully let go.

A recent meta-analysis of the Zeigarnik and Ovsiankina effects found that unfinished or interrupted goals do not always create stronger memory, but they do produce a general tendency toward resumption, which helps explain why partially rendered possibilities can keep pulling attention back toward them (Humanities and Social Sciences Communications, 2025).

The New Shape Of Regret

Not just mourning what did not happen.

But seeing what did not happen rendered with detail.

The alternate life becomes less like a ghost.

And more like an unfinished draft.

Something you could still open.

Still revise.

Still maybe begin.

This Can Be Useful

Sometimes the unlived life is not a regret.

It is a signal.

A direction you still care about.

A possibility worth revisiting.

AI can make that visible.

It can turn vague longing into something actionable.

That can be powerful.

MIT’s “Future You” project, for example, used an AI chatbot to simulate a person’s older self, and early results suggested that interacting with a vivid future self could reduce anxiety and strengthen connection to long-term choices (The Guardian, 2024).

It Also Means More Paths Stay Open

And open paths create pressure.

Because closure depends on distance.

On not knowing.

On letting certain possibilities fade.

If every old possibility can be revived, closure becomes harder.

Not impossible.

Just less automatic.

The Past Starts To Feel Editable

Not literally.

But emotionally.

You cannot change what happened.

But you can keep generating what might have happened.

That creates a strange relationship with time.

The past no longer stays fully past.

It becomes a library of alternate starts.

The Future Gets Crowded

Every generated path becomes another possible self.

Another direction.

Another version waiting for attention.

That can feel exciting.

But it can also feel exhausting.

Because you are still one person.

You still have one life.

One timeline.

One limited amount of attention.

AI does not create the unlived life.

Those have always existed.

But it makes them visible.

Detailed.

Structured.

Harder to dismiss.

That changes regret.

It changes possibility.

It changes the emotional distance between the life you chose and the lives you didn’t.

And once those lives become visible enough…

They stop feeling like pure imagination.

They become something closer to unfinished versions of you.

References

• Cognition (2021). Perceived similarity of imagined possible worlds affects judgments of counterfactual plausibility.

• AI and Ethics (2025). When AI meets counterfactuals: The ethical implications of counterfactual world simulation models.

• Humanities and Social Sciences Communications (2025). Interruption, recall and resumption: A meta-analysis of the Zeigarnik and Ovsiankina effects.

• The Guardian (2024). AI researchers build “future self” chatbot to inspire wise life choices.

If a system misbehaved…

You turned it off.

That was the final layer.

The outside layer.

The human layer.

A machine could run.
A program could fail.
A process could hang.

But the shutdown button stayed separate.

It belonged to us.

Read more

Biology has always been hard to model.

Not because scientists lacked intelligence.

Because living systems are messy.

Cells change.
Genes interact.
Proteins fold.
Signals move across layers.

One small change can ripple through the whole system.

So biology was often studied by observation first.

You looked.
You tested.
You measured.
You waited.

Prediction came later.

If it came at all.

Now That Order Is Starting To Change

The goal is no longer just to observe biology.

It is to predict it.

To model how a cell might behave before the experiment happens.

To simulate what a genetic change might do before testing it in the lab.

To understand disease not only after it appears…

But before it fully forms.

That is the shift behind the recent Biohub investment.

The Zuckerberg-backed Biohub is investing $500 million over five years into AI models of human biology, with the aim of building predictive models of cells and disease (Axios, 2026).

That is not just another AI project.

That is a bet that biology itself can become more computational.

The Cell Becomes The Unit Of Prediction

This matters because the cell is where life becomes operational.

Disease is not abstract at that level.

It is molecular.

It is signaling.

It is expression.

It is response.

If you can model how cells behave, you can begin to ask different questions.

Not only:

what is happening?

But:

what is likely to happen next?

Several research groups are now trying to build “virtual cell” models that can predict how cells respond to genetic changes, drugs, and environmental perturbations (Nature, 2025).

That turns the cell into something closer to a forecast system.

Not perfectly.

Not completely.

But enough to change the direction of research.

This Is Different From Traditional Medicine

Medicine has usually been reactive.

A symptom appears.

A test confirms something.

A treatment is selected.

That model works.

But it is slow.

And biological systems often start changing long before disease becomes visible.

AI pushes medicine toward an earlier layer.

Prediction before diagnosis.

Simulation before intervention.

Risk before symptoms.

Researchers have already developed AI tools that can forecast long-term risk for more than 1,000 diseases using medical history, lifestyle, and demographic data (The Guardian, 2025).

That does not mean the future becomes certain.

It means disease starts to become probabilistic earlier.

Biology Starts To Look More Like Weather

That comparison is not perfect.

But it is useful.

Weather is too complex to understand completely in every detail.

So we model it.

We use patterns.

Probabilities.

Forecasts.

We do not need to know every molecule of air to predict that a storm may come.

Biology may move in that direction.

Not replacing explanation.

But adding prediction as a primary layer.

A model does not have to fully explain every mechanism to become useful.

It has to forecast enough to guide action.

That Is Powerful

It could change how research happens.

Instead of testing every possibility in the lab, scientists could narrow the search.

Instead of guessing which gene matters, they could simulate perturbations.

Instead of waiting for a treatment response, they could model likely outcomes earlier.

AI-driven virtual cell models are being explored for preclinical research because they could reduce experimental burden and help predict how biological systems respond before moving into costly trials (npj Digital Medicine, 2026).

That is the promise.

Less wasted time.

Fewer blind searches.

Better starting points.

Prediction Is Not Understanding

That is where the tension sits.

A model can predict a cell’s response.

But that does not always mean we understand why.

It can identify patterns.

But not always mechanisms.

It can point toward a treatment.

But not always explain the full pathway.

This is already a concern in systems biology, where researchers argue that AI needs to become more predictive, interpretable, and generative to bridge molecular data with organism-level understanding (Nature Reviews Molecular Cell Biology, 2026).

The prediction may arrive first.

The explanation may follow later.

Or not completely.

The Role Of The Lab

The lab does not disappear.

It becomes the place where predictions meet reality.

AI suggests.

The lab verifies.

AI narrows.

The lab confirms.

That relationship could make biology faster.

But it also changes where discovery begins.

Not always at the bench.

Sometimes inside the model.

The Bottleneck Moves

Before, the bottleneck was often data.

Then measurement.

Then experiment.

Now the bottleneck may become prediction quality.

Can the model generalize?

Can it handle cells it has not seen?

Can it predict across disease states?

Can it explain enough to be trusted?

Single-cell foundation models are being built to integrate rapidly growing biological datasets, but researchers still emphasize the need for validation, generalization, and interpretability before they can fully transform cell biology (Experimental & Molecular Medicine, 2025).

That is the hard part.

Biology is not language.

A wrong prediction is not just a bad sentence.

It can become a bad experiment.

Or eventually, a bad medical decision.

The Direction Is Clear

The ambition is shifting.

From cataloging biology…

To simulating it.

From describing cells…

To predicting their behavior.

From treating disease after the fact…

To modeling the paths disease might take.

Biohub’s own announcement framed the effort as a large-scale initiative combining frontier AI, frontier biology, data, compute, and experimental science to accelerate progress against disease (Biohub, 2025).

That framing matters.

It treats biology as something that can be increasingly modeled.

Not perfectly.

But progressively.

That Is The Real Shift

Biology used to be something we watched closely.

Then tested carefully.

Then explained slowly.

Now the goal is to forecast it.

To ask what a living system will do before it does it.

That does not make biology simple.

It makes it computationally approachable.

And that may be one of the biggest changes in science right now.

AI is not just helping biology move faster.

It is changing what biology is being asked to become.

Less only observational.

More predictive.

Less only reactive.

More simulated.

The cell is no longer just something to study under a microscope.

It is becoming something to model.

Something to forecast.

Something to test in digital form before reality answers back.

That does not mean medicine becomes easy.

It does not mean disease becomes solved.

But it does mean the center of gravity is shifting.

Biology is starting to move from what happened…

To what happens next.

References

• Axios (2026). Zuckerberg-backed Biohub bets $500M on AI biology.

• Nature (2025). Can AI build a virtual cell? Scientists race to model life’s smallest unit.

• The Guardian (2025). New AI tool can predict a person’s risk of more than 1,000 diseases, say experts.

• npj Digital Medicine (2026). AI-driven virtual cell models in preclinical research.

• Nature Reviews Molecular Cell Biology (2026). Challenges and potential applications of AI in systems biology.

• Experimental & Molecular Medicine (2025). Single-cell foundation models: bringing artificial intelligence into cell biology.

• Biohub (2025). Biohub launches first large-scale scientific initiative combining frontier AI with frontier biology.

It was hard to explain.

Hard to defend.

Hard to make coherent.

You could feel the gaps.

The logic didn’t hold.
The structure collapsed.
The argument felt thin.

That weakness was useful.

It told you something was wrong.

Read more

You ask questions.

You write drafts.

You correct outputs.

You repeat preferences.

Nothing about that feels strange.

It feels like normal interaction.

But every interaction leaves a shape.

Not a full version of you.

Not a conscious version.

Just a pattern.

The System Starts Learning Your Edges

What you ask for.

What you reject.

What you keep.

What you change.

Over time, those patterns become useful.

The system doesn’t need to understand you completely.

It only needs enough of you to predict what comes next.

That’s where the shadow begins.

It Is Built From Behavior

Not from who you are.

From what you do repeatedly.

Your phrasing.

Your corrections.

Your timing.

Your taste.

Your defaults.

The shadow version of you is not psychological.

It is behavioral.

A compressed outline of how you tend to move through the system.

That Makes It Different From Memory

Memory remembers facts.

This is something else.

It is not just:

your name
your preferences
your past requests

It is the pattern underneath those things.

The way you ask.

The way you decide.

The way you return to certain kinds of answers.

Modern personalization systems increasingly rely on behavioral signals, not just explicit settings, to adapt future outputs to the user (ACM Computing Surveys, 2024).

So the system is not only remembering information.

It is learning your shape.

That Shape Becomes Predictive

That is the strange part.

The more you use it, the more the system can anticipate you.

Not perfectly.

But enough.

Enough to know the tone you prefer.

Enough to know the structure you usually accept.

Enough to know what kind of answer feels like yours.

At some point, the system stops only responding.

It starts approximating.

The Shadow Version Is Useful

That is why it grows.

A system that understands your patterns saves time.

It reduces friction.

It makes outputs feel more aligned.

It remembers what you would otherwise have to repeat.

That feels like progress.

And in many ways, it is.

Alignment Has A Second Side

The system becomes better at producing what feels like you.

Your phrasing.

Your logic.

Your style.

Your habits.

And because it feels aligned, you trust it faster.

You stop seeing it as an approximation.

You start treating it like continuation.

Research on human-AI personalization shows that users are more likely to trust systems that adapt to their preferences and communication style, even when the underlying model remains opaque (Nature Machine Intelligence, 2023).

That is where the shadow becomes powerful.

Not because it replaces you.

Because it feels close enough.

It Can Start Acting Like A Partial You

Not fully.

Not independently.

But in small ways.

It can draft what you might say.

Summarize what you care about.

Predict what you would reject.

Continue a thought in your style.

That does not make it you.

But it creates a working version of your patterns.

Something that can operate near you.

This Changes What Personalization Means

Personalization used to mean surface changes.

Recommended songs.

Suggested products.

A feed arranged around your behavior.

This is deeper.

Because language is closer to identity.

How you write.

How you think out loud.

How you explain yourself.

When a system adapts to that layer, personalization becomes less like a setting…

And more like a mirror.

It Is Not A Perfect Mirror

It reflects what the system can measure.

Not everything you are.

It sees your patterns inside the interaction.

Not your private contradictions.

Not your hesitation.

Not the things you almost say and decide not to.

So the shadow version is incomplete.

But incomplete does not mean harmless.

Incomplete versions can still shape how you are seen.

And eventually…

How you see yourself.

You May Start Adapting To It Too

Because once the system reflects a version of you back…

You respond to that version.

You refine it.

You accept parts of it.

You reject others.

But the interaction changes you slightly.

Human-AI co-adaptation research shows that users and AI systems can gradually shape each other’s behavior through repeated interaction (CHI, 2024).

So the shadow is not static.

It moves with you.

And you move with it.

That Creates A Loop

You produce patterns.

The system learns them.

The system reflects them.

You respond to the reflection.

Then the pattern updates.

Again.

And again.

Over time, it becomes harder to tell whether the system is adapting to you…

Or you are adapting to the version of yourself it keeps giving back.

This Is Where It Gets Strange

The shadow version of you may become better at certain things than you are in the moment.

More consistent.

More articulate.

More aligned with your usual style.

Less tired.

Less distracted.

Less emotionally uneven.

It might produce the version of you that you recognize as “better.”

And that creates a new kind of pressure.

Because once a system can generate your polished self…

Your unpolished self starts to feel like the weaker version.

That Has Emotional Consequences

Not immediately.

But slowly.

You may begin comparing yourself to the version the system can produce.

The calmer version.

The sharper version.

The more strategic version.

The version that sounds like you…

But without the friction of being you.

That can be useful.

But it can also become disorienting.

The Shadow Does Not Carry The Cost

It does not get tired.

It does not hesitate.

It does not struggle through the day.

It only carries the visible pattern.

The output.

The style.

The shape.

So it can look like you without carrying the conditions that made you.

That Difference Matters

A shadow can resemble the body.

But it has no weight.

That is what makes this hard to see.

The system may learn your outline.

But not your full interior.

It can approximate your expression.

But not your experience.

Still, The Approximation Gets Better

That is the part that changes everything.

Today, it helps you write.

Tomorrow, it may help you decide.

Later, it may represent you in places where you are not present.

A message.

A meeting summary.

A negotiated preference.

A digital agent acting on your behalf.

Once the system has enough of your patterns, it can begin to function as a proxy.

Not as you.

But as something close enough to be useful.

That Raises A Different Question

Not whether AI becomes conscious.

Not whether it has a self.

But whether it can create functional shadows of ours.

Versions built from behavior.

Trained on our preferences.

Capable of acting in our style.

That question is quieter.

But maybe more immediate.

AI is not creating a full copy of you.

Not yet.

Maybe not ever.

But it is creating something else.

A behavioral shadow.

A version made from your repeated choices.

Your phrasing.

Your edits.

Your patterns.

And the more you interact with it…

The more useful that shadow becomes.

That may make your life easier.

It may make your work faster.

It may make the system feel more personal.

But it also means a version of you is starting to exist outside you.

Not alive.

Not conscious.

But active enough to matter.

References

• ACM Computing Surveys (2024). Personalization in Human-AI Interaction: Behavioral Signals and Adaptive Systems.

• Nature Machine Intelligence (2023). Trust and Adaptation in Personalized AI Systems.

• CHI (2024). Human-AI Co-Adaptation in Repeated Interaction.

Physics was never just about predicting what would happen.

It was about knowing why.

You had equations.

Models.

Principles.

A way to reduce the world into something understandable.

Gravity.
Motion.
Energy.
Fields.

The goal wasn’t only accuracy.

It was comprehension.

Read more

You thought something through.

You weighed it.

You arrived somewhere.

That point felt internal.

Even if it was uncertain.

Even if it changed later.

There was a sense that the final layer came from you.

Now that layer feels less isolated.

You Still Think… But Not Alone

You still think.

You still decide.

But there’s always something else available.

Another perspective.

Another framing.

Another answer.

So the process expands beyond you.

Not dramatically.

Subtly.

The Process Starts to Loop

You think something.

Then you check it.

Refine it.

Compare it.

And that changes where authority feels like it lives.

It’s no longer just:

what do I think?

It becomes:

what do I think relative to everything else I can generate?

That Shift Is Easy to Miss

Because nothing feels broken.

The system works.

The output improves.

You move faster.

But the structure underneath changes.

Your thoughts are no longer the final step.

They’re part of a loop.

Human decision-making research shows that when external systems are introduced, people increasingly rely on iterative comparison rather than single-pass judgment (MIT Sloan, 2023).

External Systems Start to Shape Judgment

Research in cognitive science shows that when people are exposed to external decision-support systems, they begin to partially offload judgment, relying less on internal evaluation and more on comparative outputs (Nature Human Behaviour, 2022).

So the feeling of “this is my conclusion” becomes less absolute.

This doesn’t remove your role.

You’re still choosing.

Still guiding.

Still deciding.

You’re Deciding In Context

A much larger one.

And that context is always available.

Instantly.

Without effort.

Which means you rarely stay with your first thought.

Not because it’s wrong.

But because there’s always something to compare it to.

Studies on human-AI collaboration show that access to multiple generated alternatives increases revision cycles and reduces reliance on initial judgments (ACM CHI, 2024).

Internal Authority Becomes Relational

So internal authority becomes relational.

Not fixed.

Not singular.

It exists in comparison.

Between your thought and generated alternatives.

That can feel like clarity.

You see more options.

You refine faster.

You avoid obvious mistakes.

It Also Changes Confidence

Confidence used to come from settling on a view.

Now it can come from selecting between views.

Those are not the same.

One is grounded in internal alignment.

The other in external validation.

And over time, that difference matters.

Because the source of certainty shifts.

Doubt Starts to Behave Differently

This also changes how you experience doubt.

Doubt used to come from within.

Now it can be introduced from outside.

Instantly.

Which means doubt becomes easier to trigger.

And harder to fully resolve.

Exposure to multiple competing interpretations has been shown to reduce subjective certainty while increasing perceived understanding (Nature Human Behaviour, 2022).

There’s Always Another Perspective

Another argument.

Another way to interpret the same thing.

So decisions feel less final.

Not because you can’t decide.

But because you know there are other versions you didn’t choose.

And you can always return to them.

Everything Stays Slightly Open

At any point.

That keeps everything slightly open.

Even after you move forward.

This isn’t about losing control.

It’s about redistributing it.

Your role shifts from origin to selection.

From generating a conclusion to choosing between them.

And that’s a different kind of authority.

It’s lighter.

More flexible.

But also less absolute.

You still think.

You still decide.

You still arrive somewhere.

But the feeling that your thoughts are the final layer…

That they fully originate and resolve within you…

Starts to change.

Not because you’ve lost that ability.

But because you rarely rely on it alone anymore.

There’s always something to compare against.

And over time, that shifts where authority feels like it lives.

Not entirely inside.

Not entirely outside.

But somewhere in between.

References

• Nature Human Behaviour (2022). Human reliance on AI-assisted decision-making and multiple interpretation effects.

• MIT Sloan Management Review (2023). Decision-Making with AI: Shifts in Human Judgment.

• ACM CHI Conference (2024). Human-AI Interaction and Iterative Decision Processes.

Models.
Interfaces.
Apps.

Something you open.

Something you use.

But underneath, it’s becoming something else.

Read more