7  7. What AI Can Do Today

AI products change every month, but the big ideas behind them change slowly. This page gives you a quick tour of what AI can do in fall 2026, the vocabulary to talk about it, and one of the year’s biggest business stories: cheap, capable open-weight models from China.

7.1 What counts as “AI”?

Artificial intelligence (AI) has no single agreed definition. A practical one is: the science of making machines do tasks that we used to think required human intelligence.

That makes AI a moving target. Playing chess well was once a famous AI goal. In 1997, IBM’s chess computer Deep Blue beat the world champion, Garry Kasparov. Soon after, chess programs stopped feeling like “AI” and became “just software”. The same thing happened to spell checkers, spam filters, and turn-by-turn directions.

Most of today’s headline AI is generative AI: systems that create new content (text, images, audio, video, or code) from a request written in plain language. That request is called a prompt.

7.2 What is an AI model?

As of Oct 6, 2026 · Official docs: Claude models overview

An AI model is the trained “engine” inside an AI product. Physically, it is a huge file of numbers called weights (often billions of them) plus the software that runs them. The next chapter explains where those numbers come from.

People often mix up three different things:

Term What it is Examples (as of October 2026)
AI company (or “lab”) Builds and trains models OpenAI, Anthropic, Google, Alibaba
Model The trained engine, with a family name and a version number GPT-6.1 Sol, Claude Opus 5.5, Gemini 3.8 Flash
App (or product) What you actually use; usually lets you choose a model ChatGPT, the Claude app, the Gemini app, Claude Code

One app usually offers several models (bigger and smarter, or smaller and faster), and one model can power many apps: Microsoft’s Copilot, for example, offers models from both OpenAI and Anthropic (Microsoft, Sept 2026). New model versions appear every few weeks; app names last longer.

A large language model (LLM) is a model trained on enormous amounts of text that reads and writes text (many can also read images). Chat assistants and coding tools such as Claude Code are built on LLMs.

7.3 A quick tour of what AI can do (October 2026)

As of Oct 6, 2026 · Official docs: Claude Code overview · Get started with Claude in Chrome

Names change monthly

Treat every product and model name in this section as a snapshot from October 2026. Some will be renamed, replaced, or shut down during this semester. Learn the categories, and look up current names on each company’s own website.

7.3.1 Words, code, and data

  • Chat assistants answer questions, write text, analyze files, and search the web: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Grok (xAI), Microsoft Copilot, Meta AI, Perplexity, and the Chinese assistants DeepSeek, Qwen (Alibaba), and Kimi (Moonshot AI).
  • Reasoning (“thinking”) models work through a problem step by step before answering; you can often choose how hard they think. In May 2026, leading mathematicians confirmed that an OpenAI model had disproved a famous 1946 conjecture by Paul Erdős (arXiv).
  • Deep research agents search the web for several minutes and write a report with citations (ChatGPT deep research, Gemini Deep Research, Claude Research). They still make factual errors.
  • Coding agents read a whole software project, write and run code, and fix what fails: Claude Code (this book’s main tool), OpenAI Codex, GitHub Copilot, Cursor, Google Antigravity. See The AI Coding Tool Landscape.
  • Agents for office work turn a goal into finished documents, spreadsheets, or slides (ChatGPT Work, Claude’s Cowork features, and the Cowork agent in Microsoft Copilot).

7.3.2 Images, video, voice, and music

  • Images from a description: ChatGPT Images, Google’s Nano Banana, Midjourney.
  • Video clips, many now with sound: Google Veo, Kling, ByteDance’s Seedance, Runway. Products come and go: OpenAI shut down its well-known Sora video app in 2026 (NBC News).
  • Voice and music: real-time spoken conversation in the major chat apps, synthetic voices (ElevenLabs), and complete songs from a prompt (Suno, Udio).

7.3.3 Acting in the world

  • Computer-use and browser agents operate a browser or desktop for you: they read the screen, click, type, and fill in forms (Claude in Chrome, ChatGPT’s agent features, Gemini in Chrome, Perplexity’s Comet browser).
  • Robots: humanoid robots are moving into factories. Boston Dynamics began manufacturing its Atlas robot in 2026 for Hyundai and Google DeepMind (Boston Dynamics).
  • Self-driving cars: Waymo was giving about 500,000 paid driverless rides a week by March 2026 (TechCrunch).

7.3.4 Science, health, and business

  • Science: AlphaFold predicts the 3-D shapes of proteins; its creators Demis Hassabis and John Jumper shared in the 2024 Nobel Prize in Chemistry (Nobel Prize).
  • Health: the US FDA has authorized over 1,600 AI-enabled medical devices, for example to support disease detection and diagnosis (FDA).
  • Finance and business: fraud detection, credit scoring, customer service, and increasingly agents that handle paperwork-heavy back-office work.

7.4 AI is built into everything

As of Oct 6, 2026 · Official docs: Claude for Microsoft 365

More and more often, AI is not a separate app but a feature inside products you already use:

  • Office software: Copilot inside Word, Excel, and PowerPoint; Claude add-ins for Excel, PowerPoint, and Word; Gemini inside Gmail and Google Docs.
  • Search and browsers: AI-written answers at the top of search results, and AI helpers built into web browsers.
  • Messaging and social media: Meta AI inside WhatsApp, Instagram, Facebook, and Messenger.
  • “Invisible” AI you have used for years: spam filters, fraud alerts from your bank, product and video recommendations, and photo search on your phone.

So you often use AI without noticing, and it pays to ask what data an AI feature can see (Limits, Risks, and Responsible Use). For businesses, AI is becoming a standard feature, the way “having a website” was around 2000, and AI agents are starting to compete with traditional software.

Much business software is sold “per seat”: a company pays a monthly fee for each employee who uses it. If AI agents can do part of that work, a company may need fewer seats, or may drop the software entirely. In early February 2026, after Anthropic released industry-specific plug-ins (legal, finance, sales, and others) for its Claude Cowork agent, investors sold software and data stocks heavily. Legal-information company Thomson Reuters fell 18% in one day, and a broad index of software and services stocks lost about $830 billion in market value in a week (ITPro, Feb 2026). Whether those fears turn out to be justified is still an open question.

7.5 Open-weight vs. closed models

A closed (or proprietary) model can only be used through its maker’s apps or paid API; the model itself never leaves the company’s servers. An open-weight model is one whose trained weights are published, so anyone can download it, run it on their own hardware, and fine-tune it (train it further on their own data).

Closed models Open-weight models
Examples (October 2026) OpenAI’s GPT-6 models, Anthropic’s Claude, Google’s Gemini, Meta’s Muse Spark DeepSeek, Qwen, Kimi, and GLM (China); Google’s Gemma, NVIDIA’s Nemotron, OpenAI’s gpt-oss (US); Mistral (France)
How you use them Through the maker’s app or API Download and run them yourself, or use a hosting company’s API
Cost Pay per use, or a subscription Free to download; you pay for the computers that run them (or for a hosted API, often cheap)
Your data Sent to the model maker Can stay on your own machines
Safety and updates Handled by the maker Your responsibility
Capability The very best models are currently closed Strong, and the gap is narrowing

Open-weight is not the same as open source. The Open Source Initiative’s definition of open-source AI also requires the training code and detailed information about the training data (OSI), which most open-weight releases don’t include. Licenses differ too: some (such as MIT and Apache 2.0) allow almost any use, while others add conditions, for example for very large companies. Read the license before you build a product on a model.

7.6 The 2026 headline: cheap Chinese open-weight models

Chinese labs such as DeepSeek, Alibaba (Qwen), Moonshot AI (Kimi), Zhipu (GLM), and MiniMax release strong models as open weights and charge low prices for using them. In 2026 this became one of the biggest stories in the AI business:

  • Popularity. By spring 2026, Chinese models made up 41% of downloads on Hugging Face (the main site for sharing open models), overtaking US models. On OpenRouter (a service for reaching many models through one account), the six most-used models in mid-2026 were all Chinese open models (TechCrunch).
  • Quality. Moonshot claims its Kimi K3 (July 2026) competes with Anthropic’s top model of the time (a vendor claim), and it arrived months sooner than analysts expected (Fortune).

Effects on US AI companies:

  1. Price pressure. US labs cut prices in 2026, which the trade press calls a price war with China’s labs (Capacity). OpenAI, for example, launched GPT-6 Sol and Luna at half the price of the models they replaced, crediting efficiency gains (TechCrunch).
  2. US products built on Chinese models. Cursor says its Composer 2.5 coding model is built on Moonshot’s Kimi K2.5 (Cursor), and Cognition’s SWE-2 is “post-trained from Kimi K3” (Cognition).
  3. Security and trust disputes. In February 2026 Anthropic said DeepSeek, Moonshot, and MiniMax had used about 24,000 fraudulent accounts to distill Claude: to collect millions of its answers and train their own models on them (Anthropic). Fast Chinese progress also feeds the debate over US export controls on AI.
  4. Different bets in the US. Meta switched its flagship from open Llama models to a closed model, Muse Spark (VentureBeat), although in August 2026 it also released a smaller open-weight model, Muse Glimmer (Meta). Google (Gemma) and NVIDIA (Nemotron) keep releasing open models.

For you, this means more choice and lower prices, but also more homework: before sending company or customer data to any AI service, especially a low-cost one abroad, check where the data goes and what your school’s or employer’s rules allow.

7.7 Narrow AI vs. AGI

  • Narrow AI is built for one task: a chess engine, a spam filter, a fraud detector, a self-driving system. It can be superhuman inside its niche and useless outside it.
  • Artificial general intelligence (AGI) is a hypothetical AI that could learn and do essentially any intellectual task a human can, and carry skills over to problems it has never seen. There is no agreed definition or test for it.

Today’s LLMs sit in between. They are far more general than classic narrow AI: one model can write an essay, fix code, and read a chart. But their abilities are uneven, or “jagged”. The same kind of model that helped settle an open math problem can still miscount the letters in a word (How Large Language Models Work) or confidently invent a fact. Whether today’s methods will lead to AGI, and when, is hotly debated.

People also talk about superintelligence: AI far better than the best humans in nearly every field. The first chapter of the AI Safety part looks at the argument that building it could be catastrophic.

7.8 Key takeaways

  • Keep the company, the model (the engine), and the app apart.
  • Learn the categories of AI tools; the product names will keep changing.
  • Cheap open-weight models, many from China, are reshaping prices and strategies.
  • Today’s AI is broad but uneven; AGI remains hypothetical and debated.

Practice Questions

Easy

Question 1 · Short answer

Why do people say that “AI” is a moving target? Use the Deep Blue example in your answer.

A common definition of AI is “making machines do tasks that we used to think required human intelligence.” Once a machine can do such a task reliably, people stop thinking of it as AI. Playing championship chess was once a famous AI goal; after IBM’s Deep Blue beat world champion Garry Kasparov in 1997, chess programs came to be seen as ordinary software. So the boundary of “AI” keeps moving as technology improves.

Question 2 · Multiple choice

Which statement best describes an AI model?

  1. The app or website you type your questions into
  2. The trained “engine”: a huge set of numbers (weights) plus the software that runs them
  3. The company that builds and sells AI products
  4. A database of facts that the AI looks answers up in

b. The model is the trained engine inside a product. The app (a) is what you use to reach a model, and one app can offer several models. The company (c) builds models. A model is not a fact database (d): its knowledge is stored indirectly in its weights, which is one reason it can get facts wrong.

Question 3 · Matching

Match each type of AI system with what it does.

Type of AI What it does
1. Reasoning model A. Clicks, types, and fills in forms in a browser or desktop for you
2. Coding agent B. Searches the web for several minutes and writes a report with citations
3. Computer-use agent C. Works through a problem step by step before answering
4. Deep research agent D. Reads a software project, edits code, and runs commands and tests

1 → C, 2 → D, 3 → A, 4 → B.

Question 4 · True or false

“Open-weight” and “open source” mean exactly the same thing for AI models.

False. An open-weight model publishes its trained weights so anyone can download and run it. The Open Source Initiative’s definition of open-source AI also requires the training code and detailed information about the training data, which most open-weight releases do not provide. Their licenses also vary, and some add conditions on use.

Question 5 · Short answer

What is generative AI? Name three kinds of content it can create.

Generative AI is AI that creates new content from a request written in plain language (a prompt). It can create text, images, audio (voices and music), video, and computer code. Any three of these are correct.

Question 6 · Multiple choice

Which of the following is an example of narrow AI?

  1. A spam filter that sorts your email
  2. An AI that could learn any job a human can do, and carry skills to new problems
  3. An AI that is far better than the best humans in nearly every field
  4. Any AI system that runs in the cloud

a. A spam filter is built for one specific task. Option b describes AGI and option c describes superintelligence; both are hypothetical. Where a system runs (d) has nothing to do with whether it is narrow or general.

Medium

Question 7 · Scenario

A friend says, “I don’t use AI.” Name three places where your friend is probably using AI anyway, and explain what it means that AI is “embedded” in products.

Examples include: the spam filter in their email, fraud alerts from their bank, recommendations on shopping, video, or music sites, photo search on their phone, AI-written answers at the top of search results, and AI features in office software or messaging apps (for example Meta AI inside WhatsApp or Instagram).

“Embedded” means the AI is a feature inside another product rather than a separate app you open on purpose. You often use it without noticing, which is why it is worth checking what data such a feature can see.

Question 8 · Short answer

A classmate says: “I always use the same chat app, so I’m always using the same AI.” Explain why this is not quite right.

The app and the model are different things. A chat app usually offers several models (bigger and smarter vs. smaller and faster), may pick one for you depending on your plan or settings, and switches to new model versions every few weeks or months. So the same app can give you different “AIs” over time, and different users of the same app may be using different models. (The reverse is also true: one model can power many different apps.)

Question 9 · Multiple choice

A hospital wants AI to summarize patient notes, but its rules say patient data must never leave its own data center. Which option fits best?

  1. A closed model used through a free consumer chat app
  2. An open-weight model running on the hospital’s own servers
  3. A browser agent that copies the notes into a web form
  4. A deep research agent that searches the web for similar notes

b. With an open-weight model the hospital can download the weights and run the model on its own machines, so the data stays in-house. In exchange, the hospital must handle the hardware, setup, updates, and safety itself. Option a sends the data to an outside company. Options c and d also send data outside and don’t solve the summarizing task safely.

Question 10 · Compare

Give two advantages of closed models and two advantages of open-weight models.

Closed models: the very best models are currently closed; nothing to install or run yourself; the maker handles updates and safety; easy to start through an app or API.

Open-weight models: free to download and often much cheaper to use; you can keep data on your own machines; you can fine-tune them on your own data; you are not locked in to one vendor’s service.

(Any two from each list are fine.)

Question 11 · Scenario

A tutorial from last year tells you to use a particular AI video app, but you can’t find it anywhere. Give two possible explanations, and describe how you would find a current alternative.

Possible explanations: the product was shut down (OpenAI shut down its Sora video app in 2026), it was renamed or merged into another product, or it is no longer offered in your country or on your plan.

To find an alternative, think in terms of the category (here, “AI video generation”), then search for current options and check the companies’ own websites and the dates on articles. Be careful with AI answers about this: a model’s knowledge may be older than the change.

Question 12 · Short answer

In February 2026, Anthropic accused three Chinese labs of a “distillation” campaign. What does distillation mean here, and why would the company that built the original model object?

Distillation means training your own model on the answers of a stronger model: you send it huge numbers of questions, collect its responses, and use them as training data. Anthropic said the labs used about 24,000 fraudulent accounts to do this with Claude. The original company objects because the copier gets much of the value of an expensive model cheaply, it breaks the terms of service, and (according to Anthropic) the copies may not keep the original model’s safety protections.

Question 13 · Multiple choice

Which statement about open-weight models in 2026 is supported by this page?

  1. US companies stopped releasing open-weight models entirely.
  2. Chinese open-weight models overtook US models in Hugging Face downloads.
  3. Open-weight models are always more capable than closed ones.
  4. Open-weight models can only be used inside China.

b. By spring 2026, Chinese models made up 41% of Hugging Face downloads, overtaking US models. Option a is wrong: Google, NVIDIA, OpenAI, and Meta have all released open-weight models (although Meta’s flagship, Muse Spark, is closed). Option c is wrong: the very best models are currently closed, although the gap is narrowing. Option d is wrong: anyone can download open weights and run them anywhere (subject to the license).

Hard

Question 14 · Long answer

Explain three ways that cheap, capable Chinese open-weight models are affecting US AI companies. Give an example for each.

A good answer covers three of these, with examples:

  1. Price pressure. Customers can get “good enough” results much more cheaply, so US labs cut prices (for example, OpenAI launched its GPT-6 Sol and Luna models at half the price of the models they replaced; the trade press calls 2026 a price war).
  2. US products built on Chinese models. Some US companies fine-tune Chinese open-weight models instead of training from scratch (Cursor’s Composer 2.5 is built on Kimi K2.5; Cognition’s SWE-2 is post-trained from Kimi K3).
  3. Security and trust disputes. Anthropic accused DeepSeek, Moonshot, and MiniMax of distilling Claude through fraudulent accounts, and the rapid progress feeds the policy debate over export controls.
  4. Different strategies. Some US companies respond by going closed (Meta’s Muse Spark), others by releasing their own open models (Google’s Gemma, NVIDIA’s Nemotron).

Question 15 · Scenario

A startup must choose between (1) a closed frontier model through its maker’s API and (2) a much cheaper open-weight model offered by a hosting company in another country. List at least four factors the founders should weigh, and say how they could decide.

Factors to weigh:

  • Quality on their task (the very best models are closed, but a cheaper model may be good enough).
  • Cost at their expected volume.
  • Data: what data will be sent, where it is stored, and whether customers, laws, or contracts restrict that (especially for personal data sent abroad).
  • License terms of the open-weight model (some add conditions for commercial use or for large companies).
  • Reliability and support of the provider, and the risk of the service changing or disappearing.
  • Security and trust concerns, and any company or government rules about particular vendors.
  • Flexibility: open weights let them switch hosts or run the model themselves later.

How to decide: run a small test of both options on real (non-sensitive) examples, compare quality and cost, and check the data and license questions before committing. Many companies use a cheap model for simple tasks and a frontier model for hard ones.

Question 16 · Long answer

Is a modern chat assistant “narrow AI” or “AGI”? Make an argument, using the idea of “jagged” abilities.

A strong answer says it is neither, at least by most definitions. It is much more general than classic narrow AI: one model can write, translate, summarize, code, and read images, and it can handle tasks nobody trained it for specifically. But it does not meet the usual definition of AGI, because its abilities are uneven (“jagged”): it may help with research-level math yet miscount the letters in a word or confidently invent a fact, and it cannot reliably learn and carry out any intellectual task a human can. Experts disagree on whether today’s methods will eventually lead to AGI, so a good answer acknowledges the debate.

Question 17 · True or false

“An AI model that can help settle an open math problem will also reliably handle simpler tasks, such as counting letters or getting basic facts right.”

False. AI abilities are “jagged”: a model can be excellent at some very hard tasks and still fail at tasks that look easy to people. How Large Language Models Work explains why counting letters is hard for language models, and Limits, Risks, and Responsible Use explains why they sometimes invent facts. You must check AI output even when the model is impressive elsewhere.

Question 18 · Scenario

Your manager wants to “use AI” for three jobs: (a) summarizing 50 competitors’ web pages into a report with sources, (b) building a small internal web app, and (c) entering the same data into a web form 200 times. Which category of AI tool fits each job, and what is one risk to watch for in each?

    1. A deep research agent. Risk: factual errors or misread sources, so open and check the citations.
    1. A coding agent (such as Claude Code). Risk: code that looks right but has bugs or security holes, so review and test it.
    1. A computer-use / browser agent. Risk: it takes real actions, so mistakes (wrong data, wrong button) have real effects; web pages can also contain hidden instructions that try to trick it (prompt injection; see Limits, Risks, and Responsible Use). Watch it closely and start with a few entries.

Challenge

Question 19 · Hands-on

Make your own “as of today” snapshot. For three AI companies, find (1) their current main chat app, (2) their newest flagship model, and (3) whether that model is closed or open-weight. Use only each company’s own website or official docs, and record the date and link for each fact. Compare your results with this page: what has changed?

Answers will vary with the date. A good answer:

  • uses official sources (for example a company’s models page, docs, or newsroom), not just an AI chat answer or a random blog;
  • records the date checked and a link for every fact;
  • correctly separates app names from model names;
  • notes at least one change since October 2026 (a renamed product, a new model version, a price change, or a shut-down service), or states that nothing changed and how that was verified.

Question 20 · Hands-on

Give the same prompt to two different assistants (or to two different models inside one app), for example: “In three sentences, explain the difference between an open-weight model and a closed model, and give one example of each.” Record which model answered each time, compare the answers, and check every factual claim against this page or an official source.

Answers will vary. A good write-up names the app and the model used for each answer (not just “ChatGPT” or “Claude”), points out real differences (accuracy, length, examples chosen, how current the examples are), and flags any claim that turned out to be wrong or out of date. Noticing that a model gives an outdated example is a good lead-in to knowledge cutoffs (How Large Language Models Work) and hallucinations (Limits, Risks, and Responsible Use).

Question 21 · Hands-on

Pick one open-weight model on Hugging Face (huggingface.co) and find its license on the model’s page. Is it a standard license such as MIT or Apache 2.0, or a custom one? Could a company use the model in a paid product, and are there any conditions?

Answers depend on the model. A good answer quotes the license name shown on the model page, opens the actual license text, and summarizes in plain English what it allows and requires (for example: commercial use allowed; attribution required; extra terms above a certain company size or number of users; restrictions on certain uses). It should also note that “open weights” does not automatically mean “anything goes”, which is why companies read the license before building on a model.

© 2025-2026 Y. Rosenthal. All rights reserved.