Go back far enough and almost every new machine arrives with a sermon attached. Someone promises a golden age. Someone else swears the lights will go out. I keep thinking about that split whenever another wave of AI predictions hits the headlines, because the pattern is older than the models themselves. Three decades ago people argued whether a young public network would turbocharge output or prove about as useful as a fax. We know how that one ended. What we still do not know is why the same emotional script keeps getting reused.
The Long Record Of Confident Calls That Aged Badly
Wanting tomorrow in advance is not a hobby. It is almost a reflex. People once read animal guts for clues. Plenty still check charts, horoscopes, and keynote slides with the same hunger. The appetite is human. The accuracy is not. Horse-race favorites lose most of the time. Canal and rail booms laid iron years before traffic caught up. Fiber got pulled across continents while cash sat underwater. Technology forecasts sit in a special corner of that graveyard because lab logic meets messy life and then pretends the collision will be tidy.
I have a soft spot for the people who actually date their claims. Vague poetry is cheap. A year and a number takes nerve. One Ethernet pioneer said the early web would blow up in spectacular fashion and then, after the calendar proved him wrong, blended the article and drank it in public. That is character. Most of us just quietly edit the slide deck.
Prediction is very difficult, especially if it is about the future.
The line gets recycled because it is true. It is also incomplete. The harder part is not the calendar. It is the second-order effect: who adopts the tool, how slowly offices change, which rules arrive late, and which prices whip around while all of that is still fog. Asset prices are the tail of the snake. They thrash first. Reality walks behind.
Famous Misses That Still Shape How We Talk About Machines
A computer executive once guessed the world might need a handful of machines. A phone chief dismissed a new handset because it had no physical keyboard and looked too expensive for business email. A video founder wondered whether anyone would watch that much footage. A physicist doubted nuclear energy could ever be summoned on purpose. A futurist put human-level software on a distant clock and then watched pieces of that clock move closer than skeptics expected. Mixed bag. That is the point.
One more example sits closer to the current fight. A leading researcher said training new radiologists would soon look foolish because image models would outclass them. A decade later the headcount in that field is higher, not lower. Demand rose with aging populations. The job was never only “look at the scan and pick a label.” Interpretation, talk with a surgeon, judgment under incomplete data: those pieces did not vanish. Category error. People keep making it.
- Underestimating demand for a tool once price and habit change
- Overestimating how fast a lab demo becomes a daily workflow
- Mistaking one visible task for the whole profession
- Dating a breakthrough and then treating the date as destiny
Wireless pocket talk in the 1920s was directionally right and sloppy on timing. That combination is common. Direction can be useful. Timelines are where capital gets bruised.
Why This Round Of Fear Feels Louder
Concern about new models has arrived in pulses. First came chatter that software subscriptions might get skipped by agents. Then job-loss language spiked. Then searches for extinction language jumped. Step back and those waves still sit behind more ordinary queries: deepfakes, student cheating, general AI risk, and the very unromantic problem of cybersecurity. The loudest phrase is not always the most useful one.
What lit the latest fuse is a familiar cocktail. A researcher leaves a prominent lab citing safety. Agents appear in messy security incidents. Labs hint that systems may start improving systems. Economists warn that the next industrial shift could compress into a much shorter window. None of that is brand new. The “risk of extinction” phrasing has been circulating among prominent names for years. Silicon Valley even has a slang number, a personal odds estimate for catastrophe. I find the number less interesting than the incentive to publish one.
Politics did the rest. Tech backlash is bipartisan in a way few issues are. Local fights over data-center water and power are easier to campaign on than a foggy claim about jobs in 2038. When voters can see a construction fence, the abstract debate grows legs.
The Doomer Versus Boomer Split Inside The Labs
Inside the companies building the largest models, the argument is not a side show. It has shaped hiring, product pace, and public letters. One camp says a system that can write code, browse, and model human persuasion is already close to a dangerous toolkit. Thought experiments about a single-minded optimizer that harvests the planet for one narrow goal still get airtime. The other camp says the models are brittle statistical machines that fail in ordinary ways and should not be dressed up as minds.
Companies answer critics with three lines that never change much. If we pause, a rival will not. Benefits could include medical breakthroughs and new kinds of work. Better systems may also defend against worse ones and eventually cut the energy waste that critics rightly hate. Fair points. Incomplete points.
Critics fire back that danger may come from limitation, not genius. A model that cannot reliably follow instructions is a management problem before it is a movie villain. They also note that a handful of firms with future listings on the calendar have reasons to talk about pauses and rules that freeze the field around current leaders. Open models complicate that story. So does slower-than-advertised progress toward general systems. I tend to think both sides overplay the purity of their motives. Money is not the only force here, but it is never off the stage.
Those who have knowledge do not predict. Those who predict do not have knowledge.
Extinction stories also lean on access that current systems do not stably possess: quiet control of weapons programs, reliable mass persuasion of unpredictable people, a single switch for a diversified public network. Power concentration is a more grounded worry. So is sloppy security. So is copyright chaos. Those are present tense.
The Market Question Hiding Under The Moral One
Investors are not only asking whether the species survives. They are asking whether more than five trillion dollars of planned hyperscaler spending over a handful of years pays off before the chips bought today look dated. Adoption speed is the hinge. A model that dazzles in a demo can still sit unused in a claims department because the workflow is a tangle of legacy software, union rules, and habit.
I have found that the boring layer decides more than the keynote. Integration. Procurement. Audit. Training a mid-level manager who does not want another dashboard. That is where capital either compounds or leaks. Markets price the dream first and the plumbing later. When the plumbing is slow, multiples compress in a hurry.
| Claim type | What usually happens | Investor tell |
| Lab capability jump | Real, then oversold | Multiple expansion |
| Workplace replacement | Partial, uneven | Earnings lag hype |
| Existential timeline | Unfalsifiable for years | Narrative volatility |
| Infrastructure boom | Overbuild then catch-up | Capex hangover |
Past buildouts rhyme. Canals. Rail. Telecom. Capacity arrived early. Demand arrived late. Some investors never got whole. That does not mean the technology was fake. It means timing and balance sheets matter more than slogans.
Four Reasons Forecasts Keep Breaking
First, complex systems are not straight lines. Traffic, hiring, regulation, and fashion all sit near tipping points. A small surprise changes the path. Self-driving cars are the cleanest recent case. The core software improved. Edge cases did not vanish. Sun glare, a person in a costume, a cyclist going the wrong way, a hand gesture that means trouble: the tail is long. Lawyers arrived. Insurers stalled. The public asked why a robot is held to a cleaner record than a tired commuter. Mixed traffic is harder than a world of only robots. The messy middle is where forecasts go to die.
Second, brains like simple stories. Exponential curves feel like a series of shocks even when the graph is smooth. That is why each model release lands like a plot twist. Optimism bias does the rest. Giant projects come in late and over budget with depressing regularity. Confirmation bias lets legacy car makers stare at early electric range limits and miss the battery curve. Anchoring locks an executive onto the first version of a product, the expensive one without a keyboard, and misses the subsidy and the new business model forming underneath.
Third, we cannot see the present clearly, let alone next year. Feedback is late. Cause and correlation get swapped. An old farming example still works as a metaphor. A community learned a process that kept a wasting disease away, yet could not isolate the one effective step from the rituals wrapped around it. Models do something similar when a single reward signal lands on a whole messy output. They keep the useful bit and the verbal tic. Overfit dressed as insight.
Fourth, incentives reward heat. Founders need belief. Funds need momentum. Incumbents need gravitas. Consultants need urgency. Officials need a hearing. Writers need a plot. Social feeds then add a multiplier. Moral language travels. False items can outrun true ones. Moderate analysis looks dull next to a precise-sounding number for a future economy that nobody can audit. I do not think this makes every participant a cynic. It does mean the distribution of claims is skewed toward the edges.
- Treat the system as linear when it is not
- Read an exponential as a staircase of surprises
- Confuse delayed feedback with proof
- Publish the version that travels farthest
Jobs, Tasks, And The Habit Of Replacing The Wrong Thing
Every decade names a task that will disappear and then misreads what people will do instead. Clerical rooms once feared machines would wreck health and wipe roles. Those same machines became the desk itself. Pocket calculators were going to rot young minds. Expectations about mental arithmetic simply moved. Cloud services were too risky for half the survey respondents in one old study. Critical systems now live there as a matter of course.
The short-term overstatement and long-term understatement pattern is tired because it keeps working. Tools look theatrical on arrival and then sink into the furniture. The furniture still rearranges the house. That is the piece markets underprice for years and then suddenly cannot ignore.
Perhaps the most interesting aspect is how often experts in one narrow domain assume the rest of the world behaves like a lab. A trading floor, a wholesale warehouse on a ring road, a plant outside a megacity: those places have constraints a demo never shows. Dunning-Kruger is an unkind label, yet the gap between vision and rollout is real. I would rather hear from the operations lead who has to connect the model to last year’s enterprise software than from another stage prediction about the death of work.
Where The Technology Is Already Quietly Useful
Not every claim is theater. Predictive systems already help weather desks chew through data volumes no human team could touch. Multi-day forecasts have improved in a way that would have sounded like bragging in the nineties. Physics still caps the horizon. Extra days are modest. Speed and energy use are not modest. A run that once needed a cathedral of compute can now finish on a laptop. That kind of gain is less cinematic than extinction talk and more like the way prior tools actually entered daily life.
Starting with narrow, checkable jobs is how trust gets built. A system that sorts, flags, and drafts can sit next to a person who still owns the decision. Partnership language is overused. It is still closer to the grain than either utopia or ruin.
What tends to stick: narrow task measurable error rate human review workflow fit What tends to slip: open-ended autonomy vague timelines unowned risk demo-only polish
How To Read The Next Five Years Without Getting Hypnotized
If you work with capital, the useful questions are almost rude in their simplicity. Who pays. How fast can a process change without breaking compliance. What happens to chip useful life if a new architecture lands early. Which revenues are software-like and which are still services wearing a model as cologne. Watch unit economics in ordinary firms, not only training runs in coastal offices.
Policy will not wait for philosophy to finish. Local permits, power interconnection, water use, and workplace rules will set the tempo as much as parameter counts. A country that wants a lead will keep funding the stack. A neighborhood that does not want the substation will slow the same stack. Both can be true on the same day.
Leaders have already walked back economic claims after public pushback while still defending the technical path. That split is healthy. Capability and social absorption are different curves. Treating them as one curve is how you get a forecast that looks brilliant in a lab notebook and clumsy in a payroll file.
In my experience the reader who will age best is the one who can hold two thoughts. The tools are getting better in uneven bursts. The world that must absorb them is sticky, legal, tired, and inventive in ways slide decks never capture. Hindsight will make the winning path look obvious. It always does. That is why we rarely learn.
A Practical Filter For Noise
When a new claim lands, I run a short checklist. Is there a date. Is there a mechanism or only a mood. Does the speaker eat the downside if wrong. Is the job being described as a single task or a bundle. Has a similar buildout already taught us about overcapacity. If the answer pattern is vague, cinematic, and unaccountable, I discount it. If it is narrow, priced, and testable, I lean in.
- Prefer claims you can mark on a calendar
- Separate demo quality from deployment quality
- Follow power, water, and permits as closely as model cards
- Treat extinction language as a worldview, not a cash-flow input
- Watch ordinary firms for proof of willingness to pay
None of this makes anyone a prophet. It just lowers the chance of buying a story because the story was loud. Markets will keep whipping. Search trends will keep lurching from bubble talk to job talk to end-of-days talk. Underneath, the same old contest continues: infrastructure first, habits later, prices somewhere in between.
What History Actually Licenses Us To Say
History does not say the current tools are small. It says our confidence about sequence is small. Wireless talk was right in spirit. Nuclear skepticism was wrong within a decade. Phone skepticism was wrong within a product cycle. Radiology timelines were wrong in kind, not only in year. Self-driving dates were wrong in the way complex systems are usually wrong: the last 10 percent was not a smaller version of the first 90.
So the honest stance is a little unfashionable. Stay curious about capability. Stay cheap about timelines. Stay suspicious of anyone who needs the stakes to be total in order for the meeting to matter. The future of this technology will be decided less in a showcase hall and more on ordinary streets, in ordinary offices, by people who have never used the word alignment in a sentence and still have a budget to hit.
That is not a soothing ending. It is a usable one. If the last century of bad calls teaches anything, it is that we will keep staring into the next machine and seeing our own hopes and dreads looking back. The trick is to notice the mirror before we trade on it.
