AI Shift Field Guide

What six basics explain most of what you saw?

• Model vs app. The model is the brain (GPT, Claude, Gemini). The app is the car built around it (ChatGPT, Claude.ai, Cursor). The same brain can sit in many cars. Your product will be a car too.
• It predicts text. A model reads a huge amount of text and learns what usually comes next. That is why it writes fluently, and also why it can be confidently wrong ("hallucination"). Always test what it builds.
• It only knows what is on the desk. The "context window" is its desk. Whatever you paste or attach is on the desk. Anything else is forgotten between chats. In project #2 keep a short spec file (what the app does, rules, decisions) and hand it over every time.
• Tokens are the meter. Usage is billed by small chunks of text, like a phone plan priced per message. Long chats and big files cost more and run slower. It is also the price of every AI feature you put in a product.
• Chat vs agent. A chat answers. An agent acts: reads your files, writes code, runs it, sees the error, fixes it, repeats for hours. That jump is what lets one person build a SaaS today.
• Why the first project got messy. Big vague requests give big vague results. Small steps, a written plan, saved versions (git) and a quick test after each step keep it clean.

What happens in the next few years? (speaker's view, not facts)

• Costs fall, abilities rise, change accelerates.
• AI moves from words into science: new maths, drug candidates.
• Computing power stays short of demand.
• Office jobs come under pressure; people who learn AI do well.
• Risks named: debt-funded spending, cyber attacks, over-use, inequality tension.

What is this "AI" everyone talks about?

A computer helper that has read an enormous number of books, websites and conversations. You talk or type to it like you would to a very well-read friend. It can explain, write a letter, translate, plan a trip, or help with a recipe.

How did it change in the last few years?

• 2022: it could chat and write. Like a clever pen pal.
• 2024: it learned to think before answering. Like a student who checks the sum.
• 2025: it can take a whole job and finish it alone for hours. Like a hired assistant.

What does my son do with it?

He uses it to build small online services, a bit like building a little shop that runs on its own. Before, you needed a team of programmers. Now one person with a good idea and this helper can do it. Later he may sell those little shops.

Is it dangerous?

• Like electricity or cars: very useful, and it needs rules and care.
• Some serious people worry that very powerful AI could one day get out of control. Others think that is far off. Nobody knows for sure, and companies are now putting effort into safety.
• The everyday risks are more ordinary: scammers (fake voices and messages), mistakes it states confidently, and spending too much time talking to a machine instead of people.

Will it take people's jobs?

Some office jobs will change a lot. People who learn to use it tend to do better than those who ignore it. Jobs that need hands and presence, like nursing, building and repairs, are in demand too, for example to build the data centers that run it.

How can I use it safely?

• Good uses: translate a letter, explain a confusing form, suggest recipes, plan a trip, write a message to the family.
• Check important answers (health, money, legal) with a real doctor, bank or lawyer. It can be wrong while sounding sure.
• Never share passwords, bank details or ID numbers.
• If a call or message from "family" asks for money urgently, hang up and call them back on their usual number. Voices can be faked.

1. Which pool of money does AI go after?

Old software sold tools. AI can do the work, so it competes for the money companies pay people to do jobs. 10 divided by 0.65 is about 15x; 20 divided by 0.65 is about 31x.

2. How much will the world spend on tech in 2026?

Gartner, April 2026 forecast, via SaaStr. Software +15.1%, data centers +55.8%, total IT +13.5% year on year. Gartner's software figure covers more than the speaker's $650B slide number, so the two do not match.

3. How much is big tech spending on AI infrastructure?

2026 capex guidance as of Apr 29, 2026, via Yahoo Finance. Bars show the midpoint of a range; total about $725B ($ billions). The speaker's warning: this is now partly funded by debt.

4. How fast is "fast"? (growth multiples)

• $200M to $700M in a year: one high-value knowledge-work company (speaker, unnamed).
• $110M to $3.4B in a month: average of seven deals, the "two-step" pricing (speaker's data).
• 10% a day: 1.10 multiplied by itself 30 times is about 17.4. My arithmetic, to show what "10%+ day-over-day" means.
• $0 to $100M in 7 days: a classifier product, about 100x cheaper than a general model. His claim, called an outlier among outliers. His slide says "0 to 70%", so numbers may differ from the audio.

5. Are AI leaps getting closer together?

Nov 2022 to Dec 2024 is 25 months. Dec 2024 to Nov 2025 is 11 months. That is why he says app companies must reinvent about every four months.

0:28Tech comes in stacking waves

• Chips → systems → networks → internet → mobile → AI apps.
• Biggest wave yet: AI goes after software spend and the much bigger services spend.
• Fastest wave: because everyone already has a phone and internet. Distribution exists.
• Different shape: earlier waves changed how information is moved. This one changes how it is processed. The ground under builders keeps shifting.

2:32Three turning points in five years

• Nov 2022 ChatGPT: pre-training ("system 1" fast thinking).
• Late 2024 o1: reasoning ("system 2" slow thinking).
• Fall 2025 Claude Code + Opus 4.5: long-horizon agents. His claim: this is where AGI arrived.
• Their January 2026 blog said "2026: This is AGI". Laughed at then, consensus now (his words).

3:59Good, bad, ugly

• Good: apps that finish jobs; expert help for rural doctors; education (Alpha School as a preview); blue-collar jobs from data-center building.
• Bad: big tech now funds spending with debt; worse cyber risk ahead; "AI psychosis" in heavy users; most white-collar jobs under assault.
• Ugly: doom fear-mongering (he claims data-center scare posts trace to a state-sponsored effort by China; his claim, unverified); wealth gaps raising social tension, possibly worse.

6:34Inside the AI labs

• Occam's razor: labs are bad at PR, so when they sound odd, they mean it literally.
• Target moved from AGI to ASI.
• RSI: models build themselves; "for much of the last year".
• Faster improvement raises P(doom); labs now take alignment seriously.
• Models move beyond language into science.
• Coming: efficient architectures, custom chips, continual learning.
• Business: fierce competition for tokens, "deploy cos" (consultants sent into big firms), compute shortage, cooler talent wars, more M&A by the labs.

9:39The diffusion gap is your opening

• Application builders, mostly startups, live in this gap.
• Best bet: high-value knowledge work (coding, cybersecurity, healthcare, finance, accounting). Each domain gets a giant company.
• New systems of record are forming, like the Salesforce/Workday wave of the cloud era.
• Frontier tech (drug candidates in days) and domain-specific "neolabs".

9:52Startups: menu, growth, operations, valuations

• Own your intelligence: move workloads from big models to tuned open-source models, cheaper for the same quality.
• Reinvent every ~4 months as models change.
• "Strongest link" teams: protect the best people and let them run.
• Org charts look like a network of agents.
• Valuations are wild: about $110M to $3.4B within a month across seven deals.

15:04Close

Pre-training vs reasoning vs agents

• Pre-training: the AI reads a huge chunk of the internet and learns patterns. Like school years.
• Reasoning: time to think before answering. Like showing your work instead of blurting.
• Agents (long-horizon): works on a task for hours using tools and checking itself. From a quiz answer to a finished project.
• Why it matters: each layer made earlier "AI features" look small. Products built for the old layer got swallowed.

AGI vs ASI

• AGI: can do most knowledge work about as well as a good human. The speaker says it arrived in late 2025. Others disagree; there is no agreed definition.
• ASI: clearly smarter than the best humans at nearly everything. The labs now aim here. No date given.
• Analogy: AGI is a great employee. ASI is a whole company of geniuses that never sleeps.
• Why it matters: do not build around a weakness that may vanish next quarter.

Occam's razor in AI

• General meaning: the simplest explanation that fits is usually right.
• In the talk: labs sound dramatic and contradictory ("AI is dangerous" and "we build it faster"). Simplest explanation: they mean both, and are poor at PR.
• Take it literally: when a lab says it is racing to superintelligence, believe it and plan for that.
• For your products: choose the simple reading (AI gets better and cheaper) and ask which of your ideas survive it.

RSI, takeoff, P(doom), alignment

• RSI (Recursive Self-Improvement): AI that helps build better AI, which builds better AI. A flywheel. The speaker says it has partly been happening for a year.
• Takeoff: improvement too fast for humans to follow.
• P(doom): your estimate of the chance AI takes control of, or wipes out, humanity. Experts give wildly different numbers.
• Alignment: making AI reliably want what we want. The speaker says labs realized it was under-funded.

Language to STEM

• First wave: words (chat, writing, summaries).
• Now: science, technology, engineering, maths. New maths theorems, drug molecules in days.
• Why it is big: words are a small slice of the economy. Science and engineering hold huge value.
• For builders: next niches are technical professions where an expert's hour is expensive.

System of record (SoR) in AI software

• Simple: the official place a business keeps the truth about one area. Salesforce = customers. Workday = employees. ServiceNow = IT tickets.
• Analogy: the filing cabinet everyone must use. Hard to replace because the history and habits live in it.
• In AI: as AI does the work, the place that holds the work, data and decisions becomes the new record.
• Defensible: an AI skin on someone else's record is easy to replace. Being the record is sticky.

Other words from the talk

• Token: unit AI usage is billed in (roughly a word chunk).
• TAM: total addressable market, the money a product could go after.
• Diffusion gap: abilities exist, few use them yet.
• Own your intelligence: run a tuned open-source model for cost and control.
• Hyperscalers: the giant cloud companies.
• Continual learning: models that keep learning after release.

Five moves to make an idea more unique

• 1. Sell outcomes, not seats. "Your invoices are reconciled" beats "a tool for reconciling invoices".
• 2. Pick a boring, expensive niche. Accounting, compliance, health admin, security.
• 3. Become the record. Hold the data and decisions so leaving hurts.
• 4. Build a data loop. Every use improves the product in a way a clone cannot start with.
• 5. Own distribution. Audience, partnerships, an umbrella brand with trust.

Reality check on selling (the author's read, not from the talk)

• The talk is about venture-scale companies with extreme valuations. Small product sales are priced on revenue, retention and risk.
• A buyer will ask what stops the product dying at the next model release.
• Labs and incumbents buy data, customers and distribution, not UI.
• Favor recurring revenue, customer data, a niche, and low running cost.