The gist
Right now we're in the position of someone living in 1995. The internet already exists. Email works. Some people have already registered their first domain name. But the iPhone hasn't been invented yet. Nobody has heard of Facebook, Uber or Netflix streaming. Those are still 10 to 15 years away.
It's the same with AI in 2026 (this lesson was written in May 2026 and updated in October 2026). The basic models work. Agents write code. But AI's "iPhone moment" is still ahead. Maybe in 2027. Maybe in 2029. Maybe it won't look anything like what we imagine.
This lesson is a map of the known and unknown territory of the next 3 to 5 years: 7 trends the author is most confident about, 4 disputed trends where smart people disagree, and a list of things that, in the author's view, almost certainly won't happen before 2030.
The first half of the lesson (trends, disputes, what not to expect, skills) is useful for everyone. The sections on strategies, the window for founders and the personal roadmap are written for people who want to build their own product or business with AI. If that's not you, feel free to skim them.
⚠️ Disclaimer (read this first)
This is a forecast, not a prophecy. The AI industry moves faster than any forecasting method. If someone had said in 2020 "a couple of years from now there'll be ChatGPT," hardly anyone would have believed them.
Read this lesson like a weather report:
- The next 2 to 3 years (2027-2028) are noticeably easier to forecast
- On a 5-year horizon (2029-2030) confidence is lower; these are possibilities more than predictions
- Beyond 2030 it's reading tea leaves
Action: read this lesson now. Put a reminder in your calendar to reread it one year from now. Compare it with reality: what came true, what didn't, what was underestimated. That's how you calibrate your own instincts.
🎯 7 trends the author is most confident about
This is the author's view, not an agreed forecast by analysts. The reference points behind it: the Stanford AI Index, Epoch AI and METR (links at the end of the lesson). The disagreements are mostly about speed, not direction.
In the year-by-year blocks below, the first line describes what already exists today. The lines for 2027 through 2030 are the author's forecast, not fact.
Trend 1: The cost of intelligence → close to zero
Today: Sonnet 5.5 costs $2 per 1 million input tokens (as of October 2026) Next: the price per unit of intelligence keeps falling; guessing exact numbers is pointless
Current prices and versions: What's current.
What this means:
Using AI may become as ordinary as using electricity. In the early 1900s, electricity at home was an expensive novelty. A few decades later it was simply there, and nobody thinks about what a light bulb costs to run.
By 2028-2030 you may well stop thinking about what it costs to "ask AI." Model pricing is likely to shift toward:
- Unlimited subscriptions for a flat fee
- Or free altogether for basic tasks (like Gmail after 2004)
- Premium pricing only for frontier reasoning (research, scientific discovery)
What it means for you: don't build a business model around "saving tokens." A couple of years from now that probably won't be a selling point.
Trend 2: Multimodal becomes the default
"Multimodal" means AI that works with more than text: images, voice, video and more, all at once.
2026: text, images and voice; some models already analyze videos an hour long or more 2027: analyzing long videos and meeting recordings becomes a routine assistant feature 2028: + 3D world models (spatial reasoning) 2030: real-time multimodal through smart glasses / AR
What this means:
Some of this already works today, though not everywhere and not always reliably. The author's forecast is that by 2028 it will be routine for an AI assistant to:
- Watch the recording of your meeting and give you a summary + action items (a basic version exists today; remember the lesson on meetings)
- Understand a 3D model of your home and suggest where to put the furniture
- Look through your phone camera and explain what's wrong with your car's engine (sharing your camera with the assistant already exists in Gemini Live)
By 2030: real-time multimodal interaction through smart glasses or AR (augmented reality). You look at a building, and AI tells you who the architect was, the year it was built and what the rent is.
What it means for you: by 2028, content probably won't be just text. Learn to make voice-first and video-first products. Text will stay, but it'll be one channel among many.
Trend 3: Agents → reliable enough for real work
2026: agents work but need supervision (human-in-the-loop) 2027: agents complete multi-day projects on their own 2028: teams of agents handle entire functions (marketing, support, accounting) 2030: companies built as "1 founder + 20 agents" are normal
What this means:
Today you wouldn't trust AI to run a marketing campaign completely unsupervised. By 2028 you may well do so, the same way you trust a bookkeeper to keep the books today without checking on them every day.
By 2030 a typical small business founder may:
- Not hire the first 10 employees
- Build "founder + a stack of agents" from day one
- Operate like a 20-person agency does today
What it means for you: if you're a solo founder, the ceiling on how big your business can get is likely to rise without hiring. Use that.
Trend 4: Local AI catches up with the frontier APIs
"Local AI" means models that run on your own computer instead of in a company's cloud.
2026: local models cover a noticeable share of everyday tasks 2027-2028: the gap with cloud models keeps shrinking 2030: practical parity for most tasks
What this means:
The author's forecast is that the big API providers (OpenAI, Anthropic, Google) will keep their lead mainly in:
- Frontier reasoning (scientific research, breakthrough discovery)
- The hardest math and coding at the edge of human ability
- Agent workflows that need very large context (hundreds of thousands of tokens at once)
Everyday tasks by 2028 → local by default (the author's forecast):
- Llama, Qwen, DeepSeek (or their successors) handle most everyday tasks
- Privacy + control over your own data + cost = a combination that's hard to beat for small businesses
- A powerful Mac or a PC with a good graphics card = a home AI server
What it means for you: if you're building your own product, by 2028 it should work with both cloud and local models. Don't build on a single vendor.
Trend 5: Regulation arrives worldwide
2026-2027: the EU AI Act takes effect in stages (deadlines for high-risk systems moved to December 2027) 2027+: a federal AI law in the US is possible 2028: regulation in China and other countries gets stricter 2029+: international agreements on AI are possible 2030: a compliance officer is a standard role at small businesses
About the EU AI Act: the Digital Omnibus package of amendments took effect on July 27, 2026 and moved the deadlines for high-risk systems listed in Annex III to December 2, 2027; some of the transparency requirements have applied since August 2026. More, if you want it, in the deep-dive library lesson AI regulation and compliance. This isn't legal advice.
What this means:
GDPR (the EU's data privacy law) arrived in 2018, and suddenly every online service had cookie banners, privacy policies and data retention policies. Within 2 to 3 years it was simply the norm, and nobody argued about it.
Something similar is likely with AI compliance by 2028-2030:
- An audit trail for every AI decision that affects people
- An "explainability" requirement for critical decisions (loans, hiring, healthcare)
- Mandatory red-teaming (deliberately trying to break the system) before launch
- Penalties for harm caused by AI
What it means for you: if you're building your own product, keep audit logs (a record of what the AI did) from day one. Not as an option, as the foundation. People who build compliance-first by 2027 will pull ahead of the ones scrambling to retrofit it.
Trend 6: Voice-first computing takes over
2026: voice modes available in the major assistants (Claude, ChatGPT, Gemini) 2027: voice = the default way to use personal AI on phones 2028: smart glasses (Meta already sells them) go mainstream 2030: typing to AI feels like dial-up internet
What this means:
Voice messages have long been a normal part of messaging apps. By 2028, voice may no longer be an add-on but the default interface for AI:
- At home: you talk to a smart speaker
- On the go: AirPods + voice
- In glasses: you speak under your breath and AI hears you
What it means for you: if you're building your own product, plan a voice mode for 2027-2028. Not as a gimmick, as one of the main ways to use it. That goes for business software too.
Trend 7: Programming as a profession transforms
2026: AI writes a significant share of the code on dev teams (with oversight) 2027: most code is written by AI with a senior reviewer 2028: AI architects + 1-2 human reviewers per company 2030: programming = "instructing AI," not "writing code"
What this means:
A developer in 2030 isn't a developer from 2020. By analogy:
- 1980: a programmer "writes assembly"
- 2000: a programmer "writes C/Java"
- 2020: a programmer "writes TypeScript/Python and uses frameworks"
- 2030: a programmer "orchestrates AI agents that write the code"
The top-level skills will shift:
- ❌ Memorize syntax → ✅ Design system architecture
- ❌ Write boilerplate → ✅ Define acceptance criteria
- ❌ Debug character by character → ✅ Debug patterns in agent behavior
- ❌ Code review for typos → ✅ Code review for emerging risks
What it means for you: if you're learning syntax, switch to system design, taste and product thinking. Those will keep their value.
🤔 4 disputed trends (medium confidence)
This is where smart people seriously disagree. Keep several scenarios in mind.
Spectrum 1: The AGI timeline
When will AI become a "general intelligence" comparable to a human at most tasks? (AGI stands for artificial general intelligence.)
| Camp | Forecast | Who says so |
|---|---|---|
| Optimistic | before 2030 | Dario Amodei, Anthropic's CEO (wrote in 2024 that such AI could arrive as early as 2026, though it could also take much longer); Sam Altman, OpenAI's CEO (wrote in January 2025 that the company knows how to build AGI) |
| Moderate | the 2030s and 2040s | A survey of 2,778 AI researchers (AI Impacts, 2023): a 50% chance by 2047 |
| Skeptical | not in the next few years, and not with today's language models | Yann LeCun (lecture at Brown University, April 2026), some academic researchers |
What to prepare for:
- If AGI arrives by 2029 → the industry will change beyond recognition. Professions will reshape themselves within 3 years. Don't build on a 5+ year horizon; the world will be different.
- If AGI arrives by 2035 → the changes are incremental, and your 5-year planning holds up
- If AGI comes in 2040+ → today's trends (agents, multimodal) continue at a steady pace
The author's bet: closer to moderate. Preparing for the 2030-2032 scenario, with a fallback plan in case it happens in 2028.
Spectrum 2: How many jobs get displaced
How many jobs will disappear by 2030?
| Scenario | Forecast | Argument |
|---|---|---|
| High | AI takes over a noticeable share of office work, and some jobs disappear | The Goldman Sachs report (2023) on how much work AI could take over; the "AI 2027" scenario |
| Moderate | Some tasks disappear, new AI-related roles appear | The annual Stanford AI Index reports |
| Low | Net positive (like computers in 1990-2010) | The historical pattern, human adaptation |
How to prepare: focus on skills AI can't replace YET:
- Sales and relationship management (trust = human)
- Strategy and vision (taste takes years to build)
- Creative direction (not creative execution, direction)
- Crisis management (judgment under uncertainty)
- Hands-on work (electrician, dentist, cook: AI won't fix your outlet in 2030)
Where the risk is higher (the author's view):
- Mid-level coding without senior expertise
- Routine analysis (BI dashboards, basic accounting)
- Tier-1 customer support
- Translation and transcription without a specialty
This isn't a verdict on these jobs. These are the areas where the way of working will have to change sooner than elsewhere.
Spectrum 3: Open-source AI vs. proprietary
| Scenario | What wins |
|---|---|
| Open wins | Llama / Qwen / DeepSeek catch up with the frontier, local dominates |
| Proprietary wins | Regulation pushes against open-weight models, a "safe API only" mandate |
| Hybrid (likely) | Coexistence: proprietary at the frontier, open models in the middle tier |
How to prepare: don't rely on one side.
- Have a local fallback (a Llama setup)
- Have API access (Anthropic + OpenAI + Google)
- Build an abstraction layer (an LLM router) that lets you switch providers within a day
It's worth building this principle in from the start: your system is ready to work with several models and isn't tied to a single provider (no vendor lock-in).
Spectrum 4: AI safety incidents
Will a major AI-related catastrophe happen before 2030?
| Type | Probability (the author's estimate) | Impact |
|---|---|---|
| Major catastrophe (bioweapon, financial crash, election manipulation) | Low | Civilization-level |
| Smaller, recurring harm (scams, deepfakes, harm to individuals) | High | Everyday problems, livable |
| Negligible (everything goes fine) | Low | - |
How to prepare:
- Security-first architecture for your own products
- Audit logs for all critical AI actions
- Look into insurance against AI-related liability: some insurers already offer such products
- Don't launch a product that could harm a lot of people if the AI makes a mistake
❌ What WON'T happen before 2030 (the author's view)
This is for calibrating your hype filter. When you see a headline like "AI will soon replace all humans," check it against this list.
- ❌ A Terminator-style AI takeover. AI won't "decide to get rid of people." That's Hollywood, not a real risk model.
- ❌ Every profession replaced. Even the aggressive forecasts talk about some professions, not all of them.
- ❌ One AI monopoly controls everything. Antitrust law works, and Anthropic + OpenAI + Google + China + the open-source ecosystem = a balance.
- ❌ AI becomes conscious in 2027. The definition of consciousness is unclear, AGI isn't the same as consciousness, and the doubts remain.
- ❌ A crypto-AI hybrid (Web3 + AI) redefines the internet. It was overhyped in 2024-2025, and crypto's real use cases are narrow.
- ❌ Universal basic income rolled out worldwide by 2030. Politics moves slower than technology.
- ❌ Quantum computing breaks all encryption. In the author's view, a quantum computer able to break today's encryption won't exist by 2030.
🎯 Strategic positioning for AI builders
Four workable strategies for 2027-2030. Pick one (or a combination).
Strategy A: AI augmentation (the most workable)
The principle: build tools that make professionals stronger, not tools that replace them.
| Works | Doesn't work |
|---|---|
| AI legal research for an attorney | An "AI lawyer for consumers" |
| An AI diagnostic assistant for a physician | An "AI doctor" in the app store |
| AI research for an analyst | "AI replaces the banker" |
Why it works: a professional trusts a tool they control. A professional does NOT trust a tool that claims to replace their expertise. On top of that, regulation is on the side of augmentation: AI assists the licensed professional, it doesn't do licensed work for them.
Who it fits: you understand a specific profession (attorney, physician, accountant, real estate agent) and you're building a tool for your peers.
Strategy B: Deep vertical software
The principle: pick ONE narrow industry. Know it inside out. Build AI that understands that domain.
| An example that works | An example that doesn't |
|---|---|
| AI for bookkeepers who serve independent restaurants (deep integration) | "AI for every accountant in the world" |
| AI for veterinary clinics in Latin America | "AI for healthcare" |
| AI for Spanish-speaking therapists in the US | An "AI therapy app" |
Why it works: general-purpose (horizontal) AI tools compete head-on with OpenAI and Anthropic, and a newcomer will almost always lose. A vertical product built on deep domain knowledge is a moat that general models can't copy.
Who it fits: you have domain expertise or access to people who do.
Strategy C: AI infrastructure (sell shovels)
The principle: build tools for people who build with AI. Don't dig for gold; sell shovels.
Examples:
- Helicone (observability for LLM apps; acquired by Mintlify in March 2026)
- LangSmith (tracing)
- LlamaIndex / LangChain (frameworks)
- Vector databases (Pinecone, Weaviate)
Risk: a smaller total market (only AI builders), it takes technical depth, and the competition is tough.
Reward: if you get it right, high margins, customers who stay, and a chance of being acquired.
Who it fits: you enjoy building developer tools and have DevOps or platform engineering experience.
Strategy D: Local AI specialist
The principle: become an expert in on-premise AI (running on the client's own hardware) for regulated industries.
Markets:
- EU healthcare (GDPR + the AI Act = local preferred)
- Defense and government contractors (control over their own data)
- Private banking for high-net-worth clients (privacy at a premium)
Pricing: premium. This kind of work is valued above ordinary development, and the specific amounts depend on the market and the project.
Risk: a longer sales cycle, it takes enterprise sales skills, and the regulatory landscape keeps changing.
Reward: very few competitors, sticky enterprise contracts, multiyear deals.
Who it fits: you have a network in large companies or government and you're comfortable with long sales cycles.
🎓 Skills to invest in (2027 priority list)
| Skill | Priority | Why |
|---|---|---|
| AI system architecture | HIGH | Always needed. Putting agents together is the new "full stack" |
| Domain expertise (one vertical) | HIGH | AI without context is useless. Your domain is your moat |
| Sales / relationships | HIGH | AI replaces tools, not trust |
| Security / compliance | HIGH | Regulation = premium pricing |
| Creative direction (taste) | HIGH | Taste won't be automated by 2030 |
| Coding fundamentals | Medium | Useful as a foundation for orchestrating agents |
| Math / statistics | Medium | For senior AI/ML roles |
| Prompt engineering as a separate profession | Decreasing | Models keep getting better at understanding plain language. Knowing how to state a task clearly stays valuable |
Skills to let go of
- ❌ Memorizing syntax and API references (AI knows them better)
- ❌ Routine data entry (automated)
- ❌ Basic translation (AI does it for free and usually well; important texts still need a human check)
- ❌ Simple copywriting (AI does it at scale)
- ❌ Routine bookkeeping (AI takes over more and more of it)
- ❌ Tier-1 customer support (AI takes over more and more of it)
- ❌ Basic graphic design (Midjourney and Canva AI often cover it)
💰 The window for founders (the author's view)
This is the author's view of the market, not investment or business advice.
| Year | What to do |
|---|---|
| 2027 | There's still time for founders of AI vertical products and tools. The window is open. |
| 2028 | Most winners in large markets are already decided. Niches remain. |
| 2029 | Late-stage consolidation. Acquisitions speed up. |
| 2030 | The ecosystem matures. AI = normal infrastructure (like the cloud in 2020). |
The practical takeaway: the window for launching a brand-new AI product isn't open forever; the later you go, the more crowded the market. The window for niche and vertical-specific products stays open longer.
🦢 Black swan possibilities (low probability, high impact)
| Type | Scenario | Probability (the author's estimate) |
|---|---|---|
| Positive | A major scientific breakthrough through AI (a cure for cancer, fusion, a longevity drug) | Medium |
| Negative | A major AI-caused incident → severe regulation worldwide | Medium |
| Geopolitical | An AI sanctions war between the US and China splits the ecosystem | Medium or higher |
| Technical | A new approach overtakes transformers (Mamba? Liquid networks?) | Low to medium |
How to be ready:
- Don't build a business that breaks under any one of these shocks
- Always keep 6 to 12 months of runway
- Diversify: revenue from 2+ sources, a hybrid tech setup (cloud + local), markets in 2+ countries
🗺 Your personal preparation roadmap
Year 1: from today to 12 months out
- ✅ Build foundation skills (this course: you're already on it)
- ✅ Launch your first project with AI (if you haven't yet)
- ✅ Find your niche or vertical
- ✅ Find a professional circle around AI (X, formerly Twitter; LinkedIn; local meetups; Discord)
- ✅ Subscribe to the core sources (see below)
Year 2: 2027-2028
- Scale your project OR pivot if it isn't working
- Build deep domain expertise (focused on 1 vertical)
- Build relationships with paying customers (10-100)
- Keep up with the major monthly releases
- Set up your first steady source of income
Year 3: 2028-2029
- Either: scale the business you've established
- Or: pivot to a new opportunity as the landscape changes
- Hire (if you're scaling) or specialize further (if you're solo)
- Position yourself for the AI consolidation phase
Years 4-5: 2029-2030
- A mature business in an established niche
- Possibly: sell to an acquirer if the market consolidates
- Or: build a moat against the competition (data, brand, integrations)
- Plan your next chapter (the next bet, slowing down toward retirement, a passion project)
🌍 Where you build from matters
Your location and the languages you speak change the strategy. One example: a builder in Latin America (Ecuador, Colombia, Mexico).
What's different there:
- A big opportunity: in the author's observation, there's less competition here and AI adoption is growing
- A gap in Spanish-language content: AI with native-level Spanish = a competitive advantage
- Cost arbitrage: prices in the US market are higher, while local costs are lower
- Lighter regulation for now: easier to launch, but plan for it to tighten (the author expects countries in the region to borrow from the EU AI Act's approach)
Strategy:
- Deep vertical software (Strategy B) for small business markets in Latin America
- AI augmentation (Strategy A) for local professionals (attorneys, accountants, real estate agents)
- Cross-border: build the product in Ecuador, sell it in the US or Spain
- Founders who speak both English and Spanish have a unique edge: two markets
🎯 Final advice
What this means for you:
- Don't try to become an "AI superstar." Statistically, it's unlikely. Foundation models are a game for Anthropic, OpenAI and Google.
- Find where AI helps solve real problems for people in your field. That's your territory.
- Build for specific customers, not for "the AI market." "The AI market" doesn't exist; real people with real problems do.
- Learn to adapt. The landscape will most likely keep shifting every six months or so for years to come. How fast you learn matters more than how much you know right now.
- Stay human. AI won't replace smiles, hugs, real conversations or a sense of meaning. Those will stay premium.
- Picture a good future. AI like electricity: invisible infrastructure that lets you do what matters. That's a realistic default scenario.
🚀 What to do after this lesson
Right away (the next 7 days)
Subscribe to the core sources:
- Anthropic Research blog: https://www.anthropic.com/research
- Stanford AI Index annual report: https://aiindex.stanford.edu
- One independent analyst: Stratechery (Ben Thompson) or The Information's AI coverage (both are paid; Stratechery has some free articles)
Find an AI community:
- A local meetup or a LinkedIn group (search for "AI" plus your city or your field)
- Discord (a few servers in your area of interest: coding, AI builders and so on)
- X (formerly Twitter): follow 10-20 AI researchers (not influencers)
Within 30 days
- Do your first small project with AI, for example automate one task from your own job. Don't wait for perfect: get it working and show it to at least one person.
- Choose your strategic positioning (A / B / C / D from the lesson above), if you plan to build your own business
- Pick your domain: one vertical to focus on for the next 12 months
After that
- Keep going through the course. At the very end, the lesson Choose your path will help you decide what to study next
- Reread this lesson in 12 months and check how accurate the forecasts were
✅ Checklist: ready for 2027-2030
This is the lesson's 30-minute exercise: go down the list and jot down your own short answers.
Key takeaways
The AI industry in 2026 is the internet in 1995. The foundation is built; the "iPhone moment" is still ahead. The author is confident about 7 trends; 4 are disputed. The map isn't perfect, but going without one is less wise than going with it.
Don't try to become an AI superstar. Statistically, it's unlikely. The best strategy is to augment one specific profession or vertical. Sell shovels; don't mine for gold.
Skills that scale in 2027-2030: system architecture, domain expertise, sales and relationships, security and compliance, creative direction. Skills that are losing value: memorizing syntax, prompt engineering as a separate profession, basic copywriting, routine analysis.
Your action: do your first project within 30 days. Not perfect, but working. The best way to understand where AI is heading is to build with it.
Sources
- Anthropic Research: https://www.anthropic.com/research
- METR (Model Evaluation & Threat Research): https://metr.org
- Stanford AI Index: https://aiindex.stanford.edu
- Epoch AI (forecasts, compute trends): https://epochai.org
- Stratechery (Ben Thompson): https://stratechery.com
- The Information (AI coverage): https://www.theinformation.com
- AI 2027 (a scenario by Daniel Kokotajlo and co-authors, 2025): https://ai-2027.com
Next lesson
→ If you're continuing through the core course (the Earn with AI path): Making money with AI: 5 client cases, by the numbers
→ If your path is Use AI in my work, your next step is the course finale: The lesson you didn't expect
The mark stays in this browser only and is never sent anywhere. My progress