Block I · Foundations · Lesson 1

From 2023 to today: how much AI has changed (and what hasn't)

We look at how we went from a chat that forgot everything within minutes to agents that work alone for hours, use the computer and read entire documents. You'll see which advances are real, which are marketing and which limits are still the same, so you know what you can expect from AI today.

Go to the course

What you'll see in this lesson

  1. 2023: a brilliant chat, but locked in

    GPT-4 arrived in March 2023: it wrote very well, but it could only work with what you pasted into the chat.

  2. From twelve pages to several books of context

    The context window has gone from 8,000 tokens in GPT-4 to one million in the top models of 2026.

  3. The four leaps that are real

    Between 2024 and 2025 AI gained voice and sight, learned to think before speaking and to use the computer, and an open reasoner arrived.

  4. Agents that last for hours of work

    In 2023, GPT-4 completed half of the tasks of about 4 minutes; in early 2026, Claude Opus 4.6, those of about 12 hours.

  5. 50% isn't enough: look at the 80% horizon

    If you want the task to come out right almost every time, Claude Opus 4.6's horizon drops from about 12 hours to just over one.

  6. Using the computer: from 12% to 85%

    On OSWorld, a test with real tasks on a computer, AI went from 12% in 2024 to 85% in mid-2026.

  7. 2023 versus today, at a glance

    It doesn't just write better: now it sees, hears, reasons, uses tools and costs much less per token.

  8. Check it yourself: a 40-page contract

    A task that wouldn't fit in GPT-4 in 2023 can be done today in a minute by uploading the PDF to the chat.

  9. What that task costs: 2023 versus today

    Analyzing that contract through the API cost $1.68 with GPT-4 in 2023; today it costs about 12 cents with Claude Opus 5.5 (October 2026 prices).

  10. What's marketing (or close to it)

    Almost all the figures are true, but they're told in the most flattering way. Learn to read the fine print.

  11. How to read a record: three questions

    Before you believe a "beats humans" headline, ask it these three questions.

  12. What hasn't changed: it still makes things up

    2026 models make far fewer mistakes, but they still state false facts with the same confidence as true ones.

  13. Summary: what to expect from AI today

    Much more capable than in 2023, but not infallible: delegate more than before and review as much as ever.

Tools and concepts

Sign in with your Google account to access this lesson and the whole course.

Go to the courseSee the full syllabus

More lessons in this block