• Short answer
• First-principles model I’m using
• Year-by-year base-case predictions
  • End of 2026 — baseline
• 2027 — Agents become normal enterprise infrastructure
• 2028 — Persistent agents and the first serious AI security shock
• 2029 — Hybrid AI becomes default; political backlash begins
• 2030 — Mid-decade inflection: agents manage projects, physical agents appear in controlled settings
• 2031 — AI-generated software and services become commoditized; cyberattacks become routine
• 2032 — Physical AI expands; geopolitics hardens around AI stacks
• 2033 — Scientific discovery acceleration becomes real, but not magical
• 2034 — Long-horizon agents mature; labor policy becomes unavoidable
• 2035 — Autonomous operations in narrow domains become common; trust remains fragile
• 2036 — The world is transformed, but not ended
• Probability summary by 2036
• What would make the world end?
  • 1. AI controls critical infrastructure and fails catastrophically
  • 2. AI accelerates autonomous weapons or geopolitical conflict
  • 3. AI accelerates biology or energy in a dangerous way
  • 4. Rapid recursive self-improvement happens earlier than expected
  • 5. Energy or capital crunch causes cascading economic failure
• The uncomfortable truths I’m baking into these predictions
• Bottom line

§Short answer

No — I don’t think the world ends in the next 10 years from AI alone.

My rough estimate:

• Chance of global existential catastrophe by 2036 from AI alone: probably <5%, maybe 2–7% depending on how fast autonomy, security, and physical-world control improve.
• Chance of a major global crisis significantly influenced by AI: maybe 15–25%.
• Chance of severe labor/economic disruption in developed economies: >80%.
• Chance that AI becomes core infrastructure like electricity or the internet: >90%.

If by “the world” you mean the planet or human civilization, I do not think it ends in 10 years.

If by “the world” you mean current jobs, software business models, media ecosystems, corporate power structures, and institutional arrangements — then yes: much of it will be unrecognizable.

The next decade is unlikely to be a clean “AGI arrives” story. It will probably be messier: productivity gains, job displacement, security incidents, energy strain, geopolitical competition, political backlash, and a slow but powerful shift from “AI answers questions” to “AI does work.”

§First-principles model I’m using

Your snapshot already points to the main forces: agents, model routing, local AI, open models, multimodality, computer use, memory, physical-world interfaces, security, energy, and regulation.

From first principles, I would model AI impact like this:

AI impact = capability × autonomy × blast radius  
           − reliability friction  
           − security risk  
           − institutional resistance

That means progress will not be determined only by “smarter models.” It will be shaped by:

1. Capability per dollar and per second
   Smaller models, mixture-of-experts architectures, quantization, routing, local inference, and custom chips will keep making useful intelligence cheaper.

2. Autonomy
   The shift from chatbots to agents is the key change. An agent that can search, read files, call APIs, write code, run tests, use a browser, and revise its plan is fundamentally more powerful than a chatbot.

3. Reliability and verification
   The bottleneck is not raw intelligence. It is whether the system can be trusted to complete a task without causing damage. Evaluation, sandboxing, permissions, checkpoints, and feedback loops will matter as much as model quality.

4. Physical and economic constraints
   AI is becoming an energy, chip, data-center, land, cooling, and capital problem. Inference prices may fall, but total infrastructure demand will rise.

5. Institutional adaptation
   Labor markets, regulation, education, security policy, and public trust will lag behind technical capability. That gap is where much of the pain will come from.

The unit of progress will shift from:

tokens generated

to:

useful tasks completed reliably per dollar

That is the most important economic change.

§Year-by-year base-case predictions

I’m treating your August 2026 snapshot as the starting point. Model names and prices in it are point-in-time; I’m predicting trends, not specific brands.

§End of 2026 — baseline

AI state:
AI is no longer mainly a chatbot story. The center of gravity has moved to agents: coding, research, computer use, tool calling, model routing, local inference, and early physical-world interfaces.

World effect:
Software development, customer support, research, and knowledge work are already being changed. Coding agents are responsible for a meaningful share of public pull requests. Model routing is becoming normal. Local AI is becoming legitimately useful. The EU AI Act enters its enforcement phase.

No sugar coating:
The easy “chatbot demo” era is over. The harder era begins: reliability, security, labor displacement, energy constraints, and institutional friction.

§2027 — Agents become normal enterprise infrastructure

AI state:
Enterprise agent platforms mature. Model gateways and routers become standard. Companies stop asking “which model should we use?” and start asking “which task should be routed to which model, tool, or local runtime?”

Local AI becomes more practical. Machines with 16–32 GB of unified memory can run useful models for summarization, classification, embeddings, private RAG, lightweight agents, and document analysis.

Coding agents become more capable at multi-file tasks: reading a repository, modifying several files, running tests, fixing failures, and producing diffs for review.

World effect:
Productivity gains appear in software development, customer support, research, and administrative work. Junior developer, analyst, support, and operations roles begin to feel pressure.

The first high-profile AI-caused enterprise incident becomes a public story: maybe a bad deployment, a data leak through prompt injection, an agent modifying the wrong file, or a support agent giving a costly bad answer.

No sugar coating:
Many companies will overbuy AI and see mixed returns. The winners will be those that build good evaluation, permissions, observability, and workflow integration — not just those with the best model.

Security debt will grow faster than security tooling in many organizations.

§2028 — Persistent agents and the first serious AI security shock

AI state:
Persistent agents with memory become common for professionals. An assistant can remember projects, preferences, decisions, unfinished tasks, and prior conversations.

Computer-use agents become more reliable for back-office workflows: forms, dashboards, email triage, ticketing systems, CRM updates, reporting pipelines, and internal software.

Open models become good enough for many private or cost-sensitive workloads. The choice becomes less “frontier cloud model vs bad local model” and more “frontier cloud model vs surprisingly capable local model.”

World effect:
Labor-market pressure broadens beyond coding into customer support, finance operations, legal research, marketing content, data analysis, and administrative work.

Regulation becomes more expensive and more visible. The EU AI Act enforcement phase creates compliance costs. Other governments begin responding, especially around high-risk systems, data privacy, copyright, and cybersecurity.

The first serious AI-enabled cyberattack or prompt-injection incident at scale occurs. It may not be apocalyptic, but it will be expensive and politically significant.

No sugar coating:
Some workers will be displaced faster than retraining programs can absorb them. Public trust in AI may dip after visible failures.

The uncomfortable truth is that AI will not automatically create new jobs fast enough to replace the old ones, at least not for many people in the short term.

§2029 — Hybrid AI becomes default; political backlash begins

AI state:
Hybrid local/cloud architectures become the default for many applications. Routine personal and professional work happens locally or on edge devices; harder tasks escalate to faster cloud models or frontier reasoning systems.

On-device assistants become useful enough for scheduling, email triage, document summarization, translation, personal knowledge management, and lightweight automation.

Frontier models increasingly compete on long-horizon reliability: multi-step research, complex coding, scientific reasoning, computer use, and physical-world control.

World effect:
AI becomes a normal line item in business budgets, like cloud compute or software licenses. Productivity gains appear in some sectors, but wage pressure grows in cognitive jobs.

Political backlash begins. Expect more debate around AI taxes, job guarantees, retraining programs, data rights, deepfakes, autonomous decision-making, and AI’s role in elections, media, education, and healthcare.

No sugar coating:
Inequality will likely widen. Companies with proprietary workflow data, integrations, distribution, and trust will win. Companies that merely wrap a model API may become commoditized.

The public will experience AI as both miracle and threat at the same time: amazing convenience, but also job anxiety, privacy concerns, and fear of autonomous mistakes.

§2030 — Mid-decade inflection: agents manage projects, physical agents appear in controlled settings

AI state:
Agents can manage multi-day or multi-week projects with human checkpoints. They are not fully autonomous, but they can maintain task state, retrieve context, delegate to subagents, run tests, revise plans, and report progress.

Physical-world agents appear in controlled environments: laboratories, warehouses, manufacturing lines, cleanrooms, and some healthcare assistance settings. Standards like Anthropic’s Model Hardware Standard begin to mature into more practical lab and equipment interfaces.

AI-native software dominates new builds. Traditional deterministic software still exists, but many new products are probabilistic systems: goal → model → tools → environment → feedback → result.

World effect:
Many jobs are redefined rather than simply eliminated. Humans move toward specification, review, exception handling, product judgment, and agent supervision.

“Agent orchestration,” “evaluation engineering,” “context management,” and “AI operations” become common professional skills.

Energy and data-center constraints become visible in some regions. AI infrastructure starts to affect electricity prices, land use, cooling requirements, and local politics.

No sugar coating:
The gap between AI-rich and AI-poor organizations, cities, and countries will grow. Some institutions will adapt quickly; others will fail to keep up.

Energy costs may constrain deployment in some places, making AI less uniformly available than its software availability suggests.

§2031 — AI-generated software and services become commoditized; cyberattacks become routine

AI state:
Model APIs are cheap enough that raw model access is no longer the main moat. Differentiation moves to workflow, data, integrations, trust, security, and user experience.

AI-generated software and services become common. Instead of buying fixed applications, users can generate or configure workflows on demand: a small internal tool, a reporting dashboard, a customer portal, a data pipeline, or a lightweight business app.

Agent security tooling matures: sandboxed execution, credential isolation, permission scopes, audit logs, anomaly detection, and human approval gates for high-risk actions.

World effect:
App and platform economics are disrupted. Small software companies, content studios, media outlets, and service businesses face existential pressure if they cannot adapt.

AI-enabled cyberattacks become routine but mostly contained. The scary part is not that every attack succeeds; it is that attackers can try many cheap, automated attacks against many targets.

A major incident may trigger a new wave of regulation around autonomous agents, critical infrastructure, and AI liability.

No sugar coating:
Trust becomes scarce. Personalized generated media, deepfakes, and AI-generated content will make verification harder. The public may become more skeptical of information generally, not just AI output.

Small businesses and creators will be squeezed between platform giants, model providers, and cheap AI-generated alternatives.

§2032 — Physical AI expands; geopolitics hardens around AI stacks

AI state:
Physical AI expands beyond controlled labs into manufacturing, logistics, healthcare assistance, and some field operations. Vision-language-action systems become more reliable in narrow domains.

Household robots are still not general-purpose. They may help with specific tasks — cleaning, inventory, monitoring, simple manipulation — but they will not yet be the universal domestic robot many people imagine.

Custom AI chips and inference infrastructure become strategically important. The largest companies increasingly control more of the stack: models, runtimes, accelerators, data centers, and energy contracts.

World effect:
Geopolitical competition around AI intensifies. The U.S., China, the EU, and other major economies develop increasingly distinct AI stacks: different chips, model ecosystems, cloud providers, open-model communities, and regulatory regimes.

Open models become especially important for countries that do not want to depend entirely on one foreign model provider.

Some governments use AI to improve public services, healthcare triage, education, and infrastructure management. Others struggle with digital divide issues, energy costs, and workforce displacement.

No sugar coating:
Physical robots will cause workplace safety disputes, labor conflicts, and liability questions. A robot injury or failure in a warehouse or hospital will be politically significant even if statistically rare.

Energy and chip supply chains may become national security issues, similar to semiconductors today but broader.

§2033 — Scientific discovery acceleration becomes real, but not magical

AI state:
Research agents plus lab automation become more practical. AI systems can search literature, propose hypotheses, design experiments, control instruments, analyze results, and iterate.

Standards for connecting AI agents to scientific equipment mature. Microscopes, robotic arms, sequencers, mass spectrometers, and other lab hardware become more accessible to AI control.

AI helps with protein design, molecular optimization, materials discovery, drug screening, energy research, and experimental planning.

World effect:
Scientific iteration speeds up in some domains. Drug discovery, materials science, and energy research see faster cycles than before. Some breakthroughs appear, but they are not sudden cure-alls.

Public health and energy policy benefit from faster analysis and simulation, but deployment still depends on regulation, manufacturing, clinical validation, and public trust.

No sugar coating:
Scientific validation remains a bottleneck. AI can generate many plausible ideas, but nature does not care about confidence scores.

Hype will outpace reliable deployment in some areas. The public may expect cures and energy miracles faster than the scientific process can deliver them.

§2034 — Long-horizon agents mature; labor policy becomes unavoidable

AI state:
Long-horizon agents can run multi-week projects with checkpoints, memory, retrieval, and human approval gates. They are not fully autonomous in open-ended domains, but they can maintain continuity across complex tasks.

Context management and memory become core differentiators. The best systems are not just smart; they know what matters, what changed, what was decided, and what needs attention.

Local AI handles much routine personal and professional work: scheduling, summarization, file management, private RAG, translation, classification, and lightweight automation.

World effect:
Education and career systems begin to adapt more seriously. Agent orchestration, evaluation, safety, data curation, and human-AI collaboration become part of professional training.

Labor policy debates intensify: retraining, portable benefits, AI dividends, universal basic income pilots, wage insurance, and taxation of automation.

No sugar coating:
Many people will feel obsolete even if aggregate productivity rises. The economy can grow while many individuals struggle.

Social strain will increase as AI creates wealth but also concentrates it. The political question will not be “Can AI do the work?” but “Who captures the benefit, and who bears the disruption?”

§2035 — Autonomous operations in narrow domains become common; trust remains fragile

AI state:
Autonomous operations become common in narrow, well-instrumented domains: some logistics, customer operations, research pipelines, manufacturing cells, and data-center maintenance.

Frontier models may be near or beyond human performance on many cognitive tasks, but they will still be brittle in open-world settings. The gap between benchmark performance and real-world reliability remains important.

AI is embedded deeply in operating systems, browsers, phones, factories, laboratories, and professional tools. The user may not always think of it as “AI”; it is just how software works.

World effect:
Major cyber or physical incidents occur, but defenses improve over time. The world becomes more automated and also more exposed to autonomous failure modes.

Economic growth is uneven. Some regions and companies ride a productivity wave; others fall behind due to energy costs, skills gaps, regulation, or weak institutions.

No sugar coating:
If a high-blast-radius failure hits finance, healthcare, energy, or critical infrastructure, public trust could trigger rollback or heavy regulation.

The most dangerous near-term failure mode is not a robot uprising. It is an agent with credentials, terminal access, or browser session doing something expensive in a system that was not designed for autonomous software actors.

§2036 — The world is transformed, but not ended

AI state:
General-purpose computational workers are normal in enterprises and many homes. Models are faster, cheaper, and more interchangeable than they were a decade earlier.

The model layer is largely commoditized. Value shifts to context, tools, data, trust, security, workflow, and physical interfaces.

Local, cloud, edge, and frontier inference coexist in a layered architecture. Many routine tasks happen locally; difficult tasks escalate to more capable systems.

World effect:
The world is not ended, but it is deeply transformed. New industries exist. Old institutions are disrupted or fail. Power is concentrated in those who control data, workflow, energy, chips, and trust.

AI contributes to sustained productivity growth in many sectors, but also to social strain: job displacement, inequality, media saturation, security concerns, and political conflict.

No sugar coating:
The winners will be those who turn AI into reliable, trusted workflow. The losers will be those who treated AI as a chatbot, assumed institutions would absorb the shock automatically, or ignored security and energy constraints.

By 2036, AI will not be a novelty. It will be infrastructure — and infrastructure is powerful, boring, political, and easy to take for granted until it fails.

§Probability summary by 2036

§What would make the world end?

I am not betting on extinction in 10 years, but these are the main paths that could make it worse:

§1. AI controls critical infrastructure and fails catastrophically

If agents gain broad access to finance, energy grids, healthcare systems, transportation, or defense systems before safety is mature, a bad autonomous decision could cause major damage.

This is more likely than a robot apocalypse in the next decade.

§2. AI accelerates autonomous weapons or geopolitical conflict

AI can improve targeting, logistics, cyber operations, and decision speed. If major powers use AI in conflict before trust and verification are strong enough, escalation risk rises.

§3. AI accelerates biology or energy in a dangerous way

AI may help create new pathogens, materials, or production systems faster than institutions can regulate them. This is a longer-term risk, but 10 years is enough for early consequences to appear.

§4. Rapid recursive self-improvement happens earlier than expected

If AI systems become much better at improving their own models, tools, and research processes by the late 2020s or early 2030s, the pace of change could accelerate sharply. That would raise both opportunity and risk.

§5. Energy or capital crunch causes cascading economic failure

AI infrastructure is energy-intensive. If data-center demand outpaces grid expansion, cooling capacity, land availability, or financing in key regions, it could slow deployment or create economic strain.

This is more likely to cause a crisis than extinction, but it could still be severe.

§The uncomfortable truths I’m baking into these predictions

1. Productivity gains will not automatically become broad prosperity.
   AI can create wealth while also displacing workers. The distribution of gains is a political problem, not just a technical one.

2. Many jobs will not disappear, but many will be devalued.
   The pain may come from status loss, wage pressure, and reduced autonomy — not just outright unemployment.

3. Security failures will happen before safety is perfect.
   Agents with credentials, terminals, browsers, and APIs create a large attack surface. Prompt injection, tool misuse, and supply-chain attacks will be common problems.

4. AI companies, cloud providers, chipmakers, and energy firms may become quasi-sovereign infrastructure owners.
   If a few companies control much of the AI stack, they will have enormous economic and political influence.

5. The public will likely experience AI as both miracle and threat.
   It will be convenient, powerful, invasive, and sometimes frightening — all at once.

6. The “AGI” narrative may obscure the more important transition.
   The bigger story is not one magical model. It is a new layer of autonomous software workers embedded in every industry.

§Bottom line

I do not think the world ends by 2036 from AI alone.

But I also will not sugarcoat it: the next 10 years will probably be a period of intense disruption. AI will become cheaper, more autonomous, more embedded, and more physically connected. It will create enormous productivity gains, but also job displacement, security risks, energy strain, inequality, and political conflict.

The most likely outcome is not extinction. It is transformation.

By 2036, AI will probably be less like a chatbot and more like a general-purpose computational worker: something that perceives, reasons, remembers, uses tools, takes actions, observes results, and coordinates with other agents.

The scarce resources will not be raw intelligence. They will be:

• good context
• good tools
• good data
• good evaluation
• reliable infrastructure
• security
• trust
• energy
• and human judgment

If the last few years were about building increasingly intelligent models, the next decade will be about giving that intelligence useful — and sometimes dangerous — things to do.