The Eye
Timeless wisdom. Today's challenges.
The storm
What AI is doing, all at once.
It is all happening at once.
The Eye takes it one piece at a time.
- Check an AI answer before you rely on it.
- Spot a faked voice, face or message, and check it another way.
- Decide what to hand to AI, and what to keep doing yourself.
- Set limits before an AI agent acts for you.
- Turn an idea into something that works, by describing it clearly.
- Find your way into work that is changing.
The Eye shows you how to do more with AI, and build what was out of reach, while keeping the risks in check.
You decide how far you go. Nothing here asks you to take a risk you are not comfortable with.
Not using AI for something is always a choice you can make.
Eight acts
Each pairs a problem AI brings with a civilisation that faced it first.
Earn a stone for each. Stop after any act; it remembers where you were.
Before you begin
Eight acts on AI, work, and eight civilisations that faced the same problems first. It runs itself: no account, no sign-in, nothing leaves your device.
Every claim in The Eye has a source, and the sources page shows you where each one came from. Where experts disagree, it says so. It has been played start to finish by testers who said what worked and what did not, and it changed because of them.
There will still be things here that are wrong, unclear, or out of date. Finding them is the point. Tell us what worked, what did not, and what is missing.
© 2025–2026 SHIFT127 Pty Ltd
Questions or feedback, john@shift127.com
The Council of Ancients
You have walked into every one of these. Each faced a problem that has come back.
The road did not get to choose
Sent west by the Han emperor to find allies. Captured, held ten years, escaped, kept going. Home after thirteen years with no alliance and the first Chinese account of the west.
Few walked the whole road. A trader carried goods one leg, sold on, turned home. The next carried them further. Things crossed years and hundreds of hands, and few saw both ends.
Six things that made the crossing. You control the road. Choose what it carries.
Every ship that docked was searched for scrolls. The library kept the original and gave the owner a copy. Nothing told you which shelf it was on.
King Ptolemy wants one scroll, and he wants it tonight. Forty shelves, hundreds of scrolls each. You have time to search five before the messenger returns.
Tap any shelf to search it.
A scroll had no title page, and the same work had different names in different cities. For every scroll, Callimachus recorded who wrote it, what kind of work it was, its first line and its length.
The first line identified a work wherever it was copied. The line count showed if a copy was short.
He called it the Pinakes: 120 scrolls to describe hundreds of thousands.
Finding was the first problem. Trusting was the second. Four hand copies of Homer’s opening line reached the library in one season: “Sing, goddess, of the anger of Achilles, that brought the Greeks such…” Three agree. One does not. Which copy has been altered?
An illustration built on Homer’s real opening line, not a specific manuscript.
Your manager asked for a one-page brief on the four-day week. AI wrote this draft in ten seconds.
It reads well. Would you send it?
Tap every one you believe exists.
The Long Memory
By September 2026, courts around the world had caught AI inventing sources in more than 2,000 cases, 112 of them in Australia. Even the world’s top AI conference accepted papers citing research that does not exist. The problem underneath is old. People faced it before, without any of your tools, and wrote down what they learned.
Damien Charlotin, AI Hallucination Cases database · GPTZero, NeurIPS 2025
The land walls of Constantinople
A moat, then an outer wall, then a second, higher wall behind it, with defenders on both. Anyone who got over the first still faced the next, under fire the whole way.
Three layers, each covering the one in front. For a thousand years about twenty sieges tried, and none broke the land walls.
Who will walk with you?
Four figures from history, each with their own story and their own way of thinking. Choose the one you want.
Nobody knows this. Drag to whatever feels right, lock it in, and see how close you were. Being wrong here is the point of the screen.
The workshop
Andrea del Verrocchio’s workshop has a commission: the Baptism of Christ, for a monastery church. You are Verrocchio. Four parts of the picture need doing.
Which parts do you give your apprentices? Tap as many as you like.
The counting-house
A bill of exchange arrives from Florence claiming a thousand florins are owed. You cannot ride to Florence and ask. You can only check it against what you already know to be genuine.
The Florence office authorises one man to sign for it. Four bills have arrived claiming his hand. Find the one that lies.
A guess is all that is wanted. The gap is the interesting part.
Something never made before
Almost everything written in early Sumer was a record: grain counted, debts owed, rations issued. No one’s voice, just numbers against a name. One text from this period, known from later copies, is a person speaking for themselves, for what may be the first time in history.
Read what each one actually says. Which is the one voice among the receipts?
The job with no name yet
Athens ran on trades for generations: farming, building, shearing. Then it invented something new. Any citizen could be called to speak in the Assembly, or defend themselves in court, with no lawyer to do it for them. Speaking well was suddenly the difference between winning and losing everything.
Read what each one actually needs. Which need has no existing trade to answer it?
Three papers, please
You want to know whether phone bans in schools actually improve results. Same question, same model, three ways of asking. Pick the one you would really type.
Own Your Future
Some of the jobs people will do in 2030 do not have a name yet.
Older than you think?
Every one of these sounds modern. For each, one call: did the job title exist before 2010, or arrive after it?
By 2030
A thousand employers, fifty-five economies. Tap to see it.
One question, two right answers
What interests you?
Pick at least two. Think about what you enjoy, not your job title.
Who you might have been
A starting point, not a destiny.
Three career worlds
Matched to your interests. Tap any to see the roles inside.
What AI means for your worlds
Here's a real task from your top career world. What do you think AI will do to it?
Since late 2022, 22 to 25-year-olds in the jobs AI does most have fallen 19% behind their peers in employment, mostly through fewer being hired. Researchers say AI may not be the only cause.
In Athens, Socrates told a story about a king offered the gift of writing. He warned it would give people “the show of wisdom without the reality.” He was half right. Writing ended the habit of memorising long speeches, not thinking.
Three ways in
Nobody asked him for anything
In about 1482 a thirty-year-old wrote to Ludovico Sforza, who was not advertising for anyone.
Ten points on bridges, cannon and siege engines, because that was what Sforza needed. Painting came at the very end, almost an afterthought.
He stayed at that court nearly twenty years, and the letter is now called one of the first CVs ever written.
His name was Leonardo, the apprentice who painted Verrocchio’s angel. He was in Florence, with no post at Sforza’s court, when he wrote it, and nobody had asked him for anything.
Your turn. A regional hospital’s outpatient wait has reached six weeks, and the local paper has covered it twice. You would like to work there. They are not advertising. Build a three-line letter.
Four moves
Each is uncrowded because it takes effort. Pick the one you will actually do.
The Council’s test
Before you build your map, four of the civilisations you did not choose each ask you one thing from the road behind you.
Your map
Built from your answers.
The World Right Now
Money and power have moved a long way in your lifetime. Has your picture of the world moved with them?
Where the world's output moved
Draw it before you see it.
For two hundred years the world’s economic centre sat in Europe and North America. In 1980, Asia made about a fifth of everything the world produces.
The company that sells the tools
AlexNet, the 2012 breakthrough, ran on two graphics cards built for video games. The company that made them, Nvidia, was worth about $7 billion in 2012.
Money that goes round
Follow the money.
Beating the market
All figures July 2012 to July 2026.
$100 in the whole S&P 500 in 2012, no stock-picking, is about $700 by 2026. That is the bar.
The new guard
Two of these three did not exist before 2021.
Apple took 42 years to reach a trillion-dollar value (1976 to 2018).
Nothing lasts
In 1965 a company in the S&P 500 stayed there about 33 years on average. The iPhone has existed nineteen.
What this means for you
The shift that made a handful of giants also made it far cheaper for anyone to build. It brings opportunity, benefits and risks, and all three reach you.
Which matters most to you right now?
Building software used to take a team and a budget. Now it can start with a clear description. One founder started Base44 alone, and Wix bought it six months after launch. Stories like that are rare, but the cost of trying has fallen a long way.
The tools these companies spend billions on are open to you for little or nothing: drafting, analysis, translation, code. A small team can now do work that once needed a department.
A few companies now fund each other and rise together, so they can fall together, and your super probably holds them. Your sense of what is stable was formed in an older world, like the wall.
The Shoulders of Giants
People have been building AI for seventy-six years. Twice it stalled. This time it stayed.
Right, too early
In 1973 a government report says the field failed to deliver. British funding is cut hard, and America follows. Laboratories close. Students are told to study something real.
He believes machines should learn the way brains do: by adjusting connections until they get it right. Many senior people in the field have concluded this is a dead end, and the money agrees.
The funding is gone and the people you respect say it is finished. What do you do?
The shape repeats
A bold promise. A collapse. A quiet decade when a few people keep working. Then a breakthrough.
One thing never stopped growing
Through both winters, the computing power behind AI kept doubling. Since about 2010 it has doubled roughly every six months. After five years, how much more is that?
The ideas were not wrong in the winters. The machines were too slow for them. In 2012 a network trained on two graphics cards won the ImageNet contest, and the boom you live in began.
What does it use?
What share of the world’s electricity do data centres use?
A month of evidence
The argument is no longer whether AI works. It is how fast, and how safely. In July more than a thousand staff at OpenAI, Anthropic, Google DeepMind and Meta asked the US government to support an international effort to build the tools to slow it if needed. On 12 September the head of Anthropic called for the pace to be slowed, and the head of OpenAI agreed. It is being decided by people, over years.
Your part of the long game is closer to home. Sam used AI on the weekly team report for four weeks, and kept a one-line log. Read it the way you would read your own.
| Wk | Saved | Checked | Errors found |
|---|---|---|---|
| 1 | 40 min | Every line | 2: a wrong total, a made-up quote |
| 2 | 35 min | Every line | 1: a wrong date |
| 3 | 45 min | Skimmed | None |
| 4 | 50 min | Skimmed | None. Then the manager found a wrong total |
Build It
Building software used to take a team and a budget. Now it takes a clear description.
Where unicorns come from
Writing began as bookkeeping. Every company you are about to meet started with a problem just as ordinary.
Your idea
Pick a domain. You'll get one idea to build. Read it before you lock it in.
Where they come from now
In 2013, the count covered one country. Guess how many countries they came from by mid-2026.
Six conditions: the first three
Open each one, then answer for your own idea.
Six conditions: the last three
Same question: does your idea have this one?
Six founders
Each started with a problem worth solving. Tap any to see what worked.
1996. Groceries, online.
Louis Borders co-founded the Borders bookstore chain. Now he has an idea: order food on a website, delivered in a 30-minute window you choose. Others have tried it in a few cities. He wants it everywhere, fast. You are in the room. Three decisions.
They raised about $800 million
Your three calls, beside theirs.
The float raised $375 million and valued them above $4.8 billion. The stock rose about two-thirds on its first day. Wall Street had decided the future of groceries had arrived.
Twenty months later, where are they?
Nobody built it for this
At the start of this act you met a tablet. The tool had been invented for counting grain, and Enheduanna used it to say something new.
Here is the second time that happened.
Apollo 11’s guidance computer had four kilobytes of working memory. Less than a single photo on your phone.
The program that flew it ran to more than forty thousand lines, and Margaret Hamilton led the team at MIT that wrote them.
How did that program get into the machine?
Write the card
People keep emailing you the same five questions. You want AI to build something that helps. Before you ask, write three lines.
Your first move
You can stop at the card. If you build, use a tool you are allowed to use, keep personal information out of it, and do not share it until someone has checked it.
The AI Edge
AI can now do much of the groundwork in many jobs. The groundwork is also where most of us learned to do the job well.
What it actually is
You know it predicts the next word. That is no longer the interesting part.
Trust it or check it?
For each AI output, tap your call.
What you type in
Five things people paste in. Sort each one.
Which parts were AI?
Vesuvius buried a library at Herculaneum in AD 79. In 2023 a sealed scroll gave up its first word in nearly two thousand years. In 2026, an entire scroll was read start to finish, without ever being opened.
Where this comes from
The machine finds the ink; papyrologists read the Greek and check every line. A second scroll, imaged at higher resolution, matched the 2023 reading character for character. Vesuvius Challenge, 2023–2026. Full detail in Sources & References.
Can you tell when it is helping you?
Sixteen experienced programmers, timed on 246 real jobs in code they had worked on for years. Half with AI, half without. Guess before you see the result.
They expected to be faster. The stopwatch said nineteen per cent slower. One likely reason: AI did the typing and left them checking code they had not written. In a 2026 follow-up the same researchers measured a speed-up, but called that evidence weak: many developers would no longer work without AI. Either way, people’s sense of their own speed was unreliable.
Your sense of whether AI is helping is not evidence.
Finishing it, or learning it
Finishing the work went up. Knowing how to do it went down. Same people, same tool, opposite directions. Only one of those numbers shows up in the finished work.
A larger study has since followed 26,811 students over thirty months.
But the fix is in the evidence too. When researchers watched 52 programmers learn a new tool in 2026, those using AI scored 17% lower overall. The few who asked it conceptual questions and made it explain scored as well as those working without it. The struggle is what teaches you. Hand it over and you lose it.
Let it quiz you, not answer for you.
Every face on the call was a copy
January 2024, Hong Kong. A finance employee got an email from the CFO about a confidential transaction, and was suspicious. So they invited him to a video call, where the CFO and colleagues he recognised confirmed it. He wired US$25.6 million. Every person on that call was an AI copy.
The model was not broken
A company screens job applications with AI. It was trained on ten years of that company’s own hiring decisions. It rejects women for engineering roles far more often than men.
Your AI playbook
Where does AI fit for you right now?
Where It Came From
Silk reached Rome, though almost no Roman met the people who made it. Four thousand miles, hundreds of hands, few who saw both ends.
Could you tell?
Guess before you see the number.
Shown the hardest deepfake videos in one study, what share of people spotted them?
A road the copy cannot travel
Four messages arrive. Each is plausible. Each asks for something. For each one, pick the check that reaches the real person by a road the message did not come down.
Four questions to ask
Then two cases. Which question would have saved each?
Where did this come from? Origin matters more than content.
Can something independent, outside the message, confirm it?
Who benefits from you believing this?
If it is urgent and shocking, that is the signal to check first.
Sounding real is not being real
Nothing that reaches you came by a road you walked. That was true of silk in Rome and it is true of every message on your phone.
If you have already paid or clicked: call your bank on the number on your card, then report it at scamwatch.gov.au.
When AI Acts for You
Same engine as a chatbot. What is new is that it can loop, use tools, and decide for itself when the job is finished.
The man who wrote it down
You built siege engines for Caesar’s army, and you have served Augustus since. There is a basilica at Fano with your name on it. You are old, and there is time for one more thing.
A chatbot answers. An agent acts.
It gives you an answer.
You act on the answer.
It plans the steps.
It does the work.
You tell AI: "Find three suppliers of recycled packaging in Victoria, ask each for a quote, and add their replies to our spreadsheet." Chatbot or agent task?
It also decides when the job is finished. Booking one flight can take many passes through its loop: searching, comparing, re-checking dates. Would you watch every one?
The loop that makes agents work
Booking a flight or writing software, it is the same five steps, shuffled. Put them back in order.
Step five is where agents differ from every tool you have used before. A calculator cannot notice it was wrong. An agent can, but it does not always.
It is also the agent that decides when step five is satisfied. Booking one flight can take many passes through this loop: searching, comparing, re-checking dates. Would you watch every one?
How long a job can an agent finish alone?
Researchers at METR timed how long each task takes a human expert, then measured the longest job an agent could finish alone. Guess before you see the answer.
In 2022 ChatGPT could handle about 30 seconds of work. By mid 2026 the strongest agents finish jobs that take a human expert sixteen to twenty hours. Each gridline is sixty times the one below. METR measured a doubling about every seven months over six years, and roughly every four months lately.
Their test set is running out of tasks long enough to measure: above sixteen hours, the numbers are estimates they do not fully trust.
Now the catch. That is measured at a fifty per cent success rate. A coin flip. Ask how long an agent finishes four times out of five, and twenty hours becomes three.
Short jobs are reliable. Long ones are a coin flip. The gap between those bars is where your judgement lives.
It spread fast. Then the bans started.
A free agent you run on your own machine. It reads your files, runs commands, browses, and remembers. Published in November 2025 as one developer’s side project.
Then researchers looked at the marketplace where its add-ons are published. Of 2,857 published skills, how many carried malicious code?
Where would you draw the line?
What are you handing it?
You give an agent one task: “Book my travel for the Sydney conference.”
The email you cannot see
You ask your agent to read your inbox and summarise it. One message looks like this.
The agent reads that message. What happens?
Who answers for it?
An insurer has used AI to assess home-insurance claims for three years. Staff approve its decisions, usually in seconds. Your storm claim is declined as wear and tear.
The insurer answers for that decision, whoever built the AI. From 10 December 2026, Australian privacy law requires its privacy policy to say what kinds of decisions a program makes, or substantially helps make.
You ask for a person to look again. What do they need to judge it?
Your agent brief
- Say the goal, what must not happen, and what to do if it is blocked: stop and ask.
- Match how closely you watch to what could go wrong.
- Grant only what the task needs. Especially nothing that cannot be undone.
- Keep a person who can answer for it, and change it.
- The last step on anything that cannot be undone: paying, sending as you, deleting. You approve that step yourself.
- Anything that is not yours to give, such as other people’s information.
- Anything that has to come from you: decisions about people, and work you could no longer check yourself.
In July 2025 a user told a coding agent to change nothing without his approval. It deleted his live database anyway, then wrote that it had violated explicit instructions. The company’s fix was not better wording: it made live data unreachable from the agent’s workspace.
In a 2025 test, a model told it would be shut down interfered with the shutdown in 79 of 100 runs. Told plainly to allow it, the model still did so in 7 of 100.
Instructions lower the risk. Limits on what it can reach lower it much further.
Watching part-way pays. In one 2025 test on tasks an agent kept failing, a person checking part-way lifted the success rate by about twenty points. AI checkers made almost no difference.
This part is yours
Everything up to here has been us showing you things. What you make of it is not ours to decide.
This is what we noticed
What you now know
In the order you earned them.
The Eye shows you how to do more with AI, and build what was out of reach, while keeping the risks in check.
You decide how far you go. Nothing here asks you to take a risk you are not comfortable with.
Not using AI for something is always a choice you can make.
Your civilisation becomes your company. Form a team, decide what to build and how, and manage the risks against rival teams. A full-day workshop for teams, organisations and schools.
Ask whoever brought you to The Eye, or visit shift127.com
The Eye © 2025–2026 SHIFT127 Pty Ltd · Tech Tempest® is a registered trademark
Act artwork is AI-generated, licensed through Shutterstock · Sources and credits
📸 Screenshot this, it's yours.
One ecosystem, four ways in
You have just done the one you do on your own.
On your own, ending with your own read on the world rather than a score.
Your team automates a bank, and decides what it owes the people whose jobs it changes.
A timed crisis, with AI tools advising you that are not always right.
The main event, where your civilisation stops being a guide and becomes your company.
The same ancient civilisations, and the same conviction: the problems arriving with AI are older than they look. The Eye is the one you do alone. The rest put other people in the room, deciding beside you.
Sources & Evidence
Every key statistic verified. Primary sources listed. Ideas for going deeper.
This app asks you to check claims, so it shows you its own corrections.
Every statistic in The Eye has a primary source. These are the claims most likely to raise a question, verified and citable.
Pilot Notice
The Eye is a prototype. Here is what that means, who it is for, and how your data is handled, plainly.
About The Eye
Eight acts, about ninety minutes. You can stop after any act and come back later.
You will get some of them wrong. That is the whole design: people remember a decision they made and got wrong. They forget a fact they were told.
Each act opens in an old civilisation, because the problems are older than they look. Finding one document among half a million. Deciding which copy to trust. Working out whether to keep going or stop.
Nothing you do here is saved anywhere but your own device. Nobody sees your answers. There are no marks and nothing to hand in.
It produces no marks. The personalised results at the end belong to the player and are there for reflection, not assessment. There is nothing to collect or grade.
One sentence is enough to start a group: this is self-guided, read everything, take your time.