Terminal-Bench 2.0
20% → 77.3%
Agent capability in early 2026, up from ~20% in February 2025.
RESEARCH BRIEF / RB-03 · V1.0
RB-03 · How to Read AI News
When Everything Is Changing
When AI Moves Faster Than
Our Ability to Understand It
From AI Progress to Human Uncertainty
Our Ability to Understand It
From AI Progress to Human Uncertainty
AI is getting better faster than most of us can make sense of what that means.
The hard part is no longer finding information — it is figuring out which information deserves to change our minds.
The Question
Which AI news actually deserves to change your mind?
Executive Summary
Three things, true at the same time.
AI is improving very quickly. AI adoption is spreading very quickly. And we still don’t know exactly what the economic and social consequences will be. That is not a contradiction — it is probably the most important fact about AI in 2026.
01
AI really is moving fast
Generative AI reached an estimated 53% population adoption within three years — faster than the PC or the internet. Organizational adoption reached 88%, with 70% of organizations using generative AI in at least one business function.
Stanford AI Index 2026
02
But capability is not economic impact
The productivity evidence is much messier: a large field experiment found time savings around email but no broad task change, while a METR experiment measured experienced developers taking 19% longer. The answer depends on the task, the worker, the workflow, and the organization.
NBER · METR · ILO
03
The anxiety is not imaginary
Globally, 52% say AI products make them nervous, even as 59% see more benefits than drawbacks. In the US, 71% expect AI to reduce jobs over the next 20 years.
Stanford AI Index 2026 · Pew
Main Thesis
AI is moving faster than we can absorb its effects.
AI capability is advancing faster than the institutions, organizations and individuals responsible for absorbing its effects.
An analytical position supported by the evidence below, not a single measured finding
01 — The Problem
The problem is no longer lack of information.
The underlying object itself keeps changing: a model released six months ago may already be obsolete, a benchmark that looked impressive last year may no longer tell us very much, and a job that looked safe yesterday may suddenly become easier to automate tomorrow.
So the old question — “Is this AI news true?” — is no longer enough.
I think we need to ask a harder question:
“If this is true, what exactly does it change?”
That distinction matters. A model achieving a higher benchmark score is a fact. Saying “therefore most programmers will lose their jobs” is not a fact — it is a chain of assumptions.
02 — Capability Is Accelerating
AI capability really is accelerating.
There is a tendency to swing between two extremes: “AI is basically autocomplete” and “AI is about to replace almost everyone.” Both descriptions are becoming increasingly difficult to defend.
FIG. 01 — Capability and Adoption Are Both Moving Fast
Sources: Stanford AI Index 2026
Adoption
53% population adoption in 3 years
Faster than the PC or the internet. 88% of surveyed organizations used AI in 2025; 70% use generative AI in at least one business function.
And this isn’t only about benchmarks. Some benchmarks designed to remain difficult for years are now becoming saturated within months. Generative AI has spread into actual organizations.
So What
The right response to AI anxiety is not “don’t worry, AI isn’t actually that good.” It is getting good — and that is exactly why the question has become more serious.
03 — Capability ≠ Replacement
Capability is not the same as replacement.
A benchmark can tell us what a model can do. It cannot tell us what an organization will do with it.
Suppose an AI system can complete a task that previously required one hour of human work. There are several possible outcomes.
A
Reduce labor
The company uses the AI to reduce headcount.
B
Produce more
The company keeps the same people and produces more.
C
Create demand
The company lowers the price and creates more demand.
D
Rework the job
The task becomes part of another person’s job.
E
Fail to integrate
The AI is technically capable, but the organization cannot integrate it reliably.
F
New business model
The task becomes cheaper, so an entirely new business model becomes viable.
All six are possible. That is why simply measuring AI exposure cannot tell us what happens to employment. The ILO’s refined global analysis estimates that roughly one in four workers globally are in occupations with some degree of generative-AI exposure — but because human input remains necessary for most affected occupations, job transformation is more likely than outright redundancy.
So What
This is much less exciting than “AI will replace 40% of jobs.” But it is much closer to what the evidence currently allows us to say.
04 — The Job, Not the Title
The first real change may be the job, not the job title.
AI doesn’t necessarily eliminate a job. It can change what is inside the job.
FIG. 02 — How Exposure Is Changing
Sources: 2026 job-postings study · OpenAI (Jul 2026)
Reallocation
52%
of the observed change in exposure due to hiring reallocation between jobs
Redesign
39.5%
of the change due to within-job redesign — the same job changes shape
Cross-occupation
43.5%
of occupation-specific AI use involved tasks of another occupation
The first signal of AI disruption may not be “Company X fired 10,000 people.” It may be “Company X stopped hiring the same kind of junior people,” or “the same employee is now expected to handle tasks that previously belonged to another team.” A marketer can now do basic technical work; a salesperson can explore data; a small business owner can draft contracts; a researcher can write code.
This is not necessarily job destruction. It is something subtler: the boundaries between jobs are becoming less rigid.
So What
The first real change may be the job, not the job title — reallocation between jobs, and redesign within them.
05 — Involution
This is where 'involution' starts to change.
AI doesn’t only automate work. It also lowers the cost of producing work.
If writing becomes cheaper, everyone can write more. If coding becomes cheaper, everyone can build more. If research becomes cheaper, everyone can produce more reports. In Chinese, this is what 内卷 (involution) means here: more outputs competing for the same amount of attention.
The baseline moves. Yesterday, writing one decent article was enough. Tomorrow, everyone can produce ten. Yesterday, a prototype took a week. Tomorrow, everyone has one by Friday.
The scarce resource then moves somewhere else.
Judgment
Deciding which result to trust, which direction to invest in
Distribution
Being found — reach, users, attention
Taste
Deciding what is good, what is worth showing
Trust
Being believed when it matters
Direction
Deciding what is worth doing at all
This is an analytical interpretation, not a measured law. But it is consistent with the direction of current evidence: AI is lowering the cost of many cognitive tasks while simultaneously increasing the amount of work that people can attempt.
So What
The next stage of AI competition may feel less like traditional automation and more like acceleration — the scarce resource moves somewhere else.
06 — Productivity
Productivity is the part we still don't understand.
AI companies can demonstrate extraordinary model capabilities, and users can show impressive demos. But the economy ultimately asks a different question: did we actually produce more valuable things with fewer resources?
The answer is not yet obvious.
FIG. 03 — The Productivity Evidence
Sources: NBER · METR
| Study | What was measured | Result |
|---|---|---|
| NBER · Dillon et al. | 7,137 knowledge workers · 66 firms | ≈2 fewer hours/week on email · after-hours work reduced · no broad change in task quantity or composition |
| METR · 2025 RCT | 16 experienced open-source developers · 246 real tasks | AI-assisted tasks took 19% longer to complete |
| METR · 2026 update | Follow-up experiment | Selection effects — developers increasingly refuse to work without AI — prevent a clean estimate; authors believe AI-assisted developers were probably more productive by early 2026 |
The honest conclusion is not “AI doesn’t increase productivity.” Nor is it “AI has already transformed productivity.” The evidence says something more uncomfortable: the answer depends heavily on the task, the worker, the workflow, the organization, and the AI system being used.
So What
Even measuring the effect becomes part of the problem — the technology changes, then people’s behavior changes, then the experiment gets harder to run.
07 — The Productivity Paradox
The AI productivity paradox.
People often assume: better model → more productivity → fewer workers. But the real chain is closer to the following — and every arrow can introduce friction.
FIG. 04 — The AI Productivity Paradox
Every arrow can introduce friction
Better model
↓
New capabilities
↓
New workflows
↓
Organizational adaptation
↓
Worker adaptation
↓
New demand / expectations
↓
Eventual productivity effect
Microsoft’s 2026 Work Trend Index makes a similar organizational point: AI and agents may expand individual agency, but organizations have to redesign the way work is structured to capture that value. This may explain why the AI economy can look simultaneously incredible and disappointing.
So What
The models are getting better. The tools are getting better. People are using them. But organizations are still learning how to reorganize around them.
08 — The Anxiety
The anxiety is not irrational.
If you feel behind, the numbers suggest you are not imagining the pressure.
FIG. 05 — Optimism and Nervousness, Side by Side
Sources: Stanford AI Index 2026 · Pew (Aug 2026)
Benefits
59%
say AI offers more benefits than drawbacks
Nervous
52%
say AI products make them nervous
Fewer jobs
71%
of US adults expect AI to reduce jobs in 20 years
Aged 18–29
73%
of young US adults say the same; only 5% expect more jobs
Both numbers can be true — and that is probably how many people actually feel. You can think “this technology is amazing” and at the same time “I don’t know if my job will still look the same in three years.”
That doesn’t prove those expectations are correct. But it tells us something important about the psychological environment people are operating in: students are choosing what to study, young workers are wondering whether entry-level jobs will still exist, and experienced workers are wondering whether ten years of experience will remain valuable.
So What
People are making career decisions under uncertainty — and everyone is watching the benchmark charts go up.
09 — Wonder vs. Worry
Optimism and fear are happening together.
Ipsos’ 2026 global AI Monitor describes the situation as a “Wonder vs. Worry” divide — excitement and nervousness have become nearly equal globally, and often the same people experience both emotions.
This is important, because the person who is excited about AI may also be the person most worried about AI.
A software engineer can love Claude Code and still worry about junior developers. A researcher can use AI every day and still worry about scientific reliability. A business owner can see huge productivity potential and still worry that competitors will move faster.
So What
That is not hypocrisy. It is what uncertainty feels like.
10 — The Feedback Loop
The information environment makes the anxiety worse.
AI news has an unusual feedback loop.
FIG. 06 — A Normal Technology Story vs. the AI News Loop
AI news increasingly feeds itself
A normal technology story
New technology
↓
Adoption
↓
Economic effect
AI news feedback loop
New model
↓
Benchmark jump
↓
Headline
↓
Social-media reaction
↓
Investment
↓
New product
↓
New benchmark
↓
New headline
The cycle feeds itself. And because attention is rewarded, the extreme interpretation often travels faster.
“AI can complete this task” becomes “AI can replace this profession.” “AI company raises billions” becomes “AI is the future of everything.” “AI benchmark reaches human level” becomes “AGI is here.” “AI deployment failed” becomes “AI is overhyped.”
The actual evidence is usually somewhere in between.
So What
The information environment makes the anxiety worse — every step of progress turns into a headline faster than it can be understood.
11 — A Better Way to Read AI News
A better way to read AI news.
I now think every major AI claim should pass through six questions. This is where my earlier “how to read the news” notes become useful as a research method, not a news-reading trick.
Step 1
What actually happened?
Separate the event from the interpretation.
Step 2
Who is telling me this?
Model companies emphasize capability, investors emphasize growth. Incentives matter.
Step 3
What was actually measured?
Benchmark? Experiment? Survey? Job postings? They are not interchangeable.
Step 4
What didn't the study measure?
Good research becomes more useful when we understand its boundaries.
Step 5
What would change my mind?
If the answer is 'nothing,' we are discussing a belief, not a hypothesis.
Step 6
What changed because of this?
Who has to behave differently now? This turns news into systems.
So What
A benchmark measures capability — it does not directly measure productivity. A survey measures beliefs — it does not measure future employment. Asking what would change your mind turns reading news into forming a hypothesis.
12 — Three Layers of Uncertainty
Three layers of uncertainty.
I think AI uncertainty can be separated into three different problems.
01
Capability uncertainty
What will AI be able to do next year? Progress is becoming easier to observe, but the frontier remains uneven — the 2026 AI Index describes a “jagged frontier”: models can achieve impressive results in some areas while failing surprisingly basic tasks in others. So “AI is smart” is not a sufficient description; we need to ask, smart at what?
02
Translation uncertainty
How does capability become productivity? A model can complete a task, but integrating it into a real organization requires people, software, processes, security, data, review, responsibility, and incentives. This is where many AI predictions become too confident.
03
Distribution uncertainty
Who actually gets the benefit? Even if AI makes society richer, that does not automatically mean everyone becomes better off. The ILO’s 2026 evidence review highlights inequality, younger workers, job quality, and worker autonomy as areas requiring attention.
So What
The question isn’t only how much value AI will create — it is who captures it.
13 — The Junior Worker Problem
The junior worker problem.
The biggest risk may not be mass unemployment. It may be the disappearance of the first step.
FIG. 07 — The Career Ladder, and the Entry-Level Signal
Sources: Stanford AI Index 2026 · Economy
The traditional ladder
Junior work
↓
Repetition
↓
Experience
↓
Responsibility
↓
Expertise
AI can automate parts of the first layer.
The entry-level signal
−20%
US software-developer employment among workers aged 22–25, from 2024 — while large-scale job losses have not appeared across the overall labor market.
That creates an uncomfortable question: if AI performs the beginner tasks, where does the next generation get its experience?
This is not evidence that “programmers are disappearing.” It is evidence that the impact may be concentrated at the entry point of certain exposed occupations — which could become a much more important social issue than a simple headline about total unemployment.
So What
The biggest risk may be the disappearance of the first step — junior work being automated before anyone gets to learn it.
14 — The Competition Is Changing
The competition is changing.
The old question was: “Can you do the work?” The new question may become: “Can you decide what work is worth doing?”
If AI makes execution cheaper, execution itself becomes less scarce. The scarce part may move upstream.
And this is why I don’t think the correct response to AI is “learn AI tools.” That is too shallow. The more durable skill is probably: learn how to work when the tools keep changing.
So What
The durable skill is not mastering the current tool — it is learning how to work when the tools keep changing.
15 — What the Evidence Says
What the evidence says right now.
Not “AI will replace everyone.” And not “AI is overhyped.” The evidence points somewhere much less comfortable.
FIG. 08 — The Evidence at a Glance
A synthesis of the 2026 evidence reviewed in this brief
| Signal | What the evidence currently says |
|---|---|
| AI capability | Rapidly improving |
| AI adoption | Already mainstream |
| Productivity | Real, but highly uneven |
| Job displacement | Limited at aggregate level so far |
| Job transformation | Already visible |
| Entry-level pressure | A genuine emerging risk |
| Public anxiety | Clearly increasing |
| AI reliability | Still inconsistent |
| Organizational adaptation | Much slower than model progress |
| Long-term outcome | Still genuinely uncertain |
So What
The central argument of this brief is not that AI will destroy jobs, nor that AI will create unlimited prosperity. The evidence supports neither conclusion with sufficient confidence.
16 — What to Watch Next
What I would watch next.
If I want to understand where AI is actually going, I would watch six things — not model releases. At least, not only model releases.
01
Real productivity
Not benchmark scores — actual output per worker, actual time saved, actual revenue per employee, actual quality.
02
Hiring patterns
Entry-level hiring, junior roles, job descriptions, skills requested — the labor market may change before unemployment statistics do.
03
Organizational redesign
Are companies actually changing structures, or simply adding AI to existing workflows? The difference is enormous.
04
AI economics
Capex, compute utilization, inference costs, pricing, revenue, margins — eventually they matter more than another benchmark victory.
05
Reliability
How often does the system succeed? How expensive is verification? This is where agentic AI becomes more interesting than chatbot demos.
06
Human adaptation
What people learn, how universities change curricula, how companies train employees, how young people decide what is worth learning.
So What
How people adapt will determine the social impact of AI almost as much as the models themselves.
17 — Living With Uncertainty
How should we live with the uncertainty?
I don’t think the answer is to stop paying attention. And I don’t think the answer is to consume more AI news. Those two things sound contradictory, but they aren’t.
The problem is not insufficient information. It is poor information-to-judgment conversion. You can spend five hours every day reading AI news and still have no idea what is actually happening — because information without a model of causality is just noise.
What I am trying to build instead is a slower habit.
News
↓
Claim
↓
Evidence
↓
Context
↓
Incentives
↓
What changed?
↓
What remains uncertain?
↓
What would change my mind?
That last question matters. Because in a period of rapid technological change, being wrong is unavoidable.
So What
The real skill is not being right — it is becoming less wrong over time.
18 — The Part I Find Most Uncomfortable
The part I find most uncomfortable.
There is something I don’t want this research to hide: I am also uncertain.
I use AI. I build with AI. I write with AI. I research with AI. And sometimes I look at what these systems can do now and think: this is moving much faster than I expected.
Then I look at the labor-market data and think: maybe the economic transformation is slower than the technology.
Then I look at the people around me trying to find work, learn new skills, or figure out what to do next — and the question becomes much more personal. Not “will AGI arrive?” But: “what should I learn now if I don’t know what the world will look like five years from now?”
I don’t have a confident answer. And I don’t think anyone honestly does.
So What
The goal is not to predict the future perfectly. It is to build a way of thinking that still works when the future refuses to cooperate with our predictions.
Final Thesis
Don't try to predict every headline.
Learn how to tell which ones should change your mind.
What Could Change My Mind
A single well-measured study that settles how capability converts into productivity and jobs — one that survives replication.
Conclusion
The future is not arriving as a single event.
It is arriving as thousands of small changes.
AI is moving fast. Our understanding doesn't have to move at the same speed — it just has to become better.
References
Evidence Sources
A synthesis of 2026 academic studies, institutional reports, and company disclosures, current through August 30, 2026. Company research is identified as such and is not treated as independent evidence.
01
Stanford AI Index 2026 Main report — capability, adoption, technical performance, responsible AI, benchmarks saturating in months core
hai.stanford.edu/ai-index/2026
02
Stanford AI Index 2026 — Economy Organizational adoption 88% / 70% · US software-developer employment among ages 22–25 down ~20% from 2024 core
hai.stanford.edu/ai-index/2026/economy
03
Stanford AI Index 2026 — Public Opinion 59% see more benefits than drawbacks · 52% say AI makes them nervous · 2026 AI incidents 362 vs 233 in 2024 core
hai.stanford.edu/ai-index/2026/public-opinion
04
ILO — Generative AI and Jobs: A Refined Global Index Gmyrek, Berg, Kamiński et al. · Working Paper 140 (2025) · ~1 in 4 workers exposed · transformation more likely than redundancy core
ilo.org · WP 140
05
ILO — The Impact of GenAI on Jobs, Productivity and Work Organization Merola, Ernst, Samaan et al. (2026) · empirical review: gains real but uneven · displacement limited so far core
ilo.org
06
NBER — Shifting Work Patterns with Generative AI Dillon, Jaffe, Immorlica & Stanton · WP 33795 · field experiment across 7,137 knowledge workers and 66 firms core
nber.org/papers/w33795
07
METR — Developer Productivity Experiment 2025 RCT: 16 developers, 246 tasks, 19% longer · 2026 update: selection effects prevent a clean average uplift hist
metr.org · Jul 2025
08
Wang et al. — Generative AI and the Reorganization of Labor Demand arXiv 2026 · US job postings · reallocation 52% / within-job redesign 39.5% · junior jobs adjust differently from senior core
arxiv.org/abs/2605.23159
09
OpenAI — How AI Is Expanding What People Do at Work Jul 2026 · analysis of 800,000+ work-related ChatGPT messages · 43.5% of occupation-specific use involved another occupation core
openai.com/index
10
Pew Research Center — Young Adults in the U.S. Are Increasingly Wary of AI Aug 2026 · 3,488 US adults · 71% expect fewer jobs in 20 years · 73% among ages 18–29 · only 5% expect more core
pewresearch.org
11
Ipsos — Global Attitudes on AI 2026 Wonder vs. Worry divide · 32 countries · excitement and nervousness nearly equal, often in the same people core
ipsos.com
12
Microsoft — 2026 Work Trend Index Company research · organizations must redesign work to capture value from AI and agents — treated as industry-side evidence core
microsoft.com/worklab
Open Questions
- Will job transformation stay ahead of job displacement?
- Does productivity follow capability — or lag it by years?