Where is AI going?

The honest answer

The honest answer to "where is AI going?" is "nobody knows for certain, but here is what is on the table." If you encounter someone giving you a confident timeline for a major AI milestone, treat them with the same scepticism you would give a confident weather forecaster predicting the climate in 2050. The general direction is more knowable than the specifics.

The rest of this lesson is what is on the table. Four directions where consensus is strong, then four where it is not.

Four things most experts agree on

1. Multi-modal models will become standard

A multi-modal model is one that can handle text, images, audio, and video all in the same system. By 2026 the leading models can do most of this. The trajectory over the next few years is for this to become both more seamless and more capable. The same assistant that reads your email will look at the chart in the attachment, listen to the voice memo, watch the video clip, and reason across all of them. This is broadly agreed.

2. AI agents will move from chat to action

A chatbot answers questions. An agent takes actions in the world on your behalf — booking your flight, drafting and sending the email, navigating a website, calling an API. The early forms of this exist already (Operator, Claude Computer Use, AutoGPT-style tools). The strong consensus is that agentic AI will be one of the defining shifts of 2026 and 2027. The interesting disagreement is over how reliable agents will become, not whether they are coming.

3. Reasoning will keep improving

"Reasoning" models — which take more time to think before answering, and visibly show their working — emerged in late 2024 (OpenAI's o1) and have become a major focus across labs. Their strength is on problems that need step-by-step thought: maths, coding, planning, complex analysis. The improvement curve here has been steep. Most experts expect this to continue.

4. Costs will keep falling

The cost of running a given level of AI capability has fallen roughly tenfold each year over the last three years. This is a combination of better algorithms, better hardware, and increasing competition. The 2026 free-tier ChatGPT is more capable than the 2023 paid one, by a wide margin. The cost trend is broadly agreed to continue, though the exact rate of continued decline is uncertain.

Four things experts disagree about

1. How close is AGI?

AGI — Artificial General Intelligence — usually means an AI system that can do most economically valuable work as well as or better than a human expert. Definitions vary. Predictions vary even more.

One camp — most prominently Dario Amodei of Anthropic, Sam Altman of OpenAI, and Demis Hassabis of Google DeepMind — believes systems matching or exceeding human expert capability in most domains will arrive between 2026 and 2030. Another camp — including Yann LeCun of Meta, Gary Marcus, and many academic researchers — believes current architectures will hit a wall, and that AGI requires fundamental new ideas not yet in our possession.

The honest read is: we don't know. The capability curves of the last three years are extraordinary. Whether they continue at that pace or flatten out is one of the most-discussed open questions in technology.

2. How safely can we deploy AI agents

Agents that take real actions are more useful and more dangerous than agents that just chat. The disagreement is about whether the safety techniques (oversight, sandboxing, verification) will keep up with the capability. Optimists think yes. Pessimists think we are deploying agents into the world faster than we can make them reliable. Both camps point to evidence in their favour.

3. How big will the economic disruption be

Will AI cause widespread unemployment? Will it create more jobs than it destroys (as past technologies have)? Will the gains accrue to workers, to capital, or to a small number of companies? Forecasters who agree on what the technology can do disagree sharply on what the economy will look like in 2030. Lesson 5 looks at this in detail for individual jobs.

4. Should AI development be slowed down

Some researchers and policymakers argue that the pace of AI development is faster than humanity can absorb, and that some kind of slowing (regulation, coordination among labs, moratoria on specific capabilities) is wise. Others argue that slowing is both impossible (competition would continue elsewhere) and undesirable (much of the benefit accrues to whoever moves fastest). This is the most politically loaded of the four, and the disagreement runs across the AI community itself.

Three concepts worth understanding

Scaling laws

The empirical observation, first formalised around 2020, that AI capability improves predictably as you increase three things: model size (number of parameters), training data, and compute. For a long time, this was the closest thing AI had to a physical law. The 2025 question is whether scaling laws are starting to bend — whether each successive doubling of compute is producing smaller capability gains than the last. Researchers disagree on whether this is happening, and what to do about it.

Emergent capabilities

The phenomenon that, as models get larger, they sometimes display new abilities that smaller models simply do not have — translating between languages they weren't specifically trained for, solving multi-step problems, writing code. The phrase implies these abilities "emerge" rather than being explicitly built in. Whether this is a real phenomenon or an artefact of measurement is — like everything in this lesson — actively debated.

Compute as the bottleneck

The amount of computing power required to train frontier models has been doubling every six months for the last decade. This is faster than computer hardware can be built. The result is that a small number of organisations — those willing and able to commit tens of billions of pounds to compute — set the frontier of AI capability. Whether this changes (because the bottleneck shifts to data, or to algorithmic ideas, or to new hardware) is one of the most important strategic questions in technology.

What seems most likely (in our honest read)

Three things we would bet on, with appropriate humility.

1. AI tools will keep getting better at the things they are already good at. Writing, summarising, coding, image and video generation, scientific assistance, customer-service automation. The next two years will not be revolutionary in these areas — they will be steady, useful, and substantial. The cumulative effect over five years will be very large.

2. Agents will be the defining shift of 2026-2027. Moving from "AI answers questions" to "AI takes multi-step actions on your computer or in real systems" is a bigger change than the change from "Google search" to "ChatGPT". It will create both a lot of value and a lot of new failure modes.

3. The conversation about regulation, safety, and societal adaptation will keep getting louder. The EU AI Act, the US executive orders, the UK AI Safety Institute, and similar efforts elsewhere are early. Expect more of this. (See our other free course, The EU AI Act for Non-Lawyers, for the regulatory frame.)

What we are less confident about: whether the systems of 2030 will feel like "ChatGPT but much better" or like something genuinely new that we cannot picture from where we sit today. Both are possible. The history of technology is full of both kinds of decade.

Looking ahead

Lesson 5 is the most personal. We have covered what AI is, what it can do, what tools exist, and where it is going. The final question — and often the one people care about most — is: what does this mean for me, my work, my children, my life? We answer it as honestly as we can.