Where This All Goes Next — and How to Keep Learning

This is the last post.

Thirty-two posts ago, we started with "what AI actually is" — four nested circles, one honest definition, and the promise of a ladder from "vaguely worried about AI" to "quietly confident." If you've read them all in order, you've now walked through how machines learn, how neural networks work, how transformers reshape sequences into meaning, how LLMs are trained and fail, and how the whole thing fits into your work and the world. You know more about this stuff than almost anyone you talk to about it.

Which leaves one question: what now?

This post is the honest answer. I'm going to tell you what the field is actively working on — what problems serious researchers think are next — and then I'm going to give you a small, deliberate guide to how to stay curious without drowning in AI news. The landscape is going to keep moving. The parts of your mental model that will age well are the ones we've built carefully. The parts that will age fastest are the specific model names and capability demos. Hold the first loosely and the second very loosely.

No calculus. No code. One finish line.

---

Where the frontier is actually pointing

Let me be specific about what labs and serious research groups are spending time on right now. Not the headlines — the papers. Five clusters.

1 · reliability at long horizons. This is the wall most of Module 6 was about. Agents that can take 20 correct actions in a row. Models that don't drift across 50,000-token contexts. Systems that recover gracefully when a tool call fails. Most of the next wave of usable AI products will be shaped by how fast this curve improves, and nobody knows exactly how it will. The field is trying a lot of approaches — better planning, self-correction, explicit reasoning traces, tool use, verification loops — and some combination of them will stick. If you wanted to pick one technical story to follow over the next two years, pick this one.

2 · real reasoning, not just pattern matching. The reasoning-model generation (o1, o3, Claude extended thinking, Gemini deep thinking) is a first step toward models that do meaningful multi-step reasoning inside a single response. They're genuinely better at hard math, competition programming, and tricky logic than the previous generation, and they do it by burning more tokens on internal reasoning before answering. Whether this scales all the way to "can reason about novel open problems" is an open question. The early results are promising enough that every major lab is investing heavily.

3 · longer-lived memory. Current LLMs have no memory across sessions. The whole "what happened last conversation" is handled by stuffing history onto the scroll, which doesn't scale. Real progress on persistent memory — so the model can learn from your previous interactions without retraining — would change what AI products feel like. Several labs have released early versions; none of them is great yet. This is where the next big user-facing shift probably comes from.

4 · embodied and real-world AI. Robotics, self-driving, physical-world manipulation. For years these were separate fields that barely interacted with language AI. Recently, they've started merging: robots driven by language models, vision-language-action models trained on video and robot data, multimodal systems that can plan in physical space. Progress here is slower and more expensive than progress in language, but if it accelerates, the implications are larger. The next decade will probably include at least one breakout "AI does physical thing we thought was hard" moment, and it will come from this cluster.

5 · interpretability and oversight. How do we understand what's happening inside a neural network? Right now, we mostly don't. We can see the weights and the activations but not "what is this circuit computing." Interpretability research is trying to change that, with some real early wins (circuits that implement specific behaviors have been identified in small models, some patterns generalize). Progress here matters for everything else, because it's the precondition for being able to trust what the models are doing at large scale.

Those are the five main directions. None of them is guaranteed to pan out. All of them have serious teams working on them. If you wanted to follow the actual state of the art, you'd track these five — not via Twitter, but via the papers labs publish, which are usually better written than their marketing.

---

What the next few years probably look like

A cautious prediction, flagged as opinion.

In the near term — the next two or three years — I expect the things you notice as a user to be:

• Agentic features that actually work, on narrow tasks, most of the time. Code editors that can implement whole features. Research assistants that can do multi-step searches and summarise. Customer support bots that can actually close tickets. The frontier will keep moving in this direction and the products will catch up.
• Much longer context windows becoming standard, cheaper, and actually usable in the middle. The lost-in-the-middle problem will get slowly better. Models will genuinely handle book-length inputs.
• Better integration into everyday software. AI features inside word processors, email, spreadsheets, CAD, legal tools, medical records. Most of them will be clumsy at first and useful within a year or two.
• A slow but real arrival of voice as a default interface. Voice input and voice output are getting good enough to replace typing for many casual uses. This will matter more in the global south and on mobile than in any single flashy launch.
• The cost of frontier inference continuing to drop by roughly 10x a year. What's expensive today will be normal in 18 months.

And some things I'd bet against in the same timeframe:

• AGI in any meaningful sense. The systems that will exist in three years will be much better than today's, and still clearly not human-level-general. People who tell you otherwise are usually running a hype cycle, a fear cycle, or both.
• A sudden event where AI "takes off" and changes everything overnight. The trajectory has been steady and fast, and will probably continue being steady and fast. No singular moment is likely, even if one feels imminent at any given time.
• Mass unemployment caused directly by AI. Jobs will shift, as we discussed in the last post. The "half of jobs gone in five years" claims have not survived contact with reality in any previous technology wave, and I don't expect this one to be different. Slow, uneven, real changes are more likely than fast catastrophic ones.

None of this is certain. Predictions about technology are famously bad, and my confidence in any specific item is low. What I'm more confident about is the shape: a fast but not infinite curve, with the biggest surprises coming from combinations of existing techniques rather than single breakthroughs.

---

How to keep learning without drowning

Now the most practical part of this post. If you want to stay current on AI without getting exhausted or gaslit by the discourse, here's what I'd actually recommend. Opinionated, short, and biased toward signal over noise.

1 · Read one paper a month. Pick a recent important paper — from DeepMind, Anthropic, OpenAI, Google, Meta, a good academic group — and read it. Not skim, read. If it's too technical, use an LLM to explain the parts you don't follow. The reason this matters is that papers are where the actual thinking happens. Headlines are downstream reactions. You'll be better informed than 99% of the people quoting the headlines.

2 · Follow two or three careful people, not a feed. Social media about AI is a swamp. The signal is in specific individuals who are thoughtful and calibrated. A short list of people whose writing I find consistently useful: Simon Willison, Gwern, Jack Clark's newsletter, Jason Wei. Find two or three people whose judgement you trust, and let them filter the news for you. Don't try to keep up with everything.

3 · Use the models, constantly, for real work. Most people who loudly form opinions about AI aren't using it much. People who are using it seriously tend to have sharper and more modest opinions. The fastest way to calibrate your own sense of where the technology is is to keep a running diary of what you tried to do, what worked, and what didn't. Over six months this will give you better intuition than any amount of reading.

4 · Keep a "surprising results" file. When something happens that surprises you — a model can suddenly do something you thought was hard, or can't do something you thought was easy — write it down. Your list of surprises is the closest thing to a private benchmark of your own understanding. When the list starts growing again, it's a signal to update your model of the field. When it shrinks, you're calibrated.

5 · Play with the edges, not just the middle. The middle — "ask ChatGPT a question, get a helpful answer" — is familiar. The edges are where the interesting stuff lives. Try multimodal inputs. Try agent tools. Try long contexts. Try reasoning models. Try open-source models you can run locally. The technology looks very different from different angles, and spending an hour a month at the edges will keep your mental model honest.

6 · Resist the urge to have a strong Take. The temptation, when you know something about AI, is to immediately form a confident position on every AI topic. Resist it. The people I trust most on AI have a lot of "I don't know yet" in their writing. The confidence of people on Twitter is inversely correlated with the depth of their thinking, almost without exception. Hold your opinions lightly and update them when evidence comes in.

---

A final note, about what this course was actually for

We've gone through 33 posts. We've covered about 75,000 words of material. If you read one post a day, it was about a month of morning coffee. That's a meaningful investment and I don't take it for granted.

I want to end by being honest about what I think this course was actually for.

The stated goal was to teach you how AI works without math or code. The real goal — the quiet one behind every post — was to give you a mental model you can actually hold, so that when AI news crosses your feed for the rest of your life, you have a place to put it. A skeleton of understanding, durable enough to survive the next wave of hype and correction, so you can read any AI claim and ask useful questions about it.

That mental model isn't about specific models or architectures. It's about a few durable ideas:

• AI is a stack of nested ideas — from the umbrella goal of "doing tasks that used to need a person" to the specific trick of "predict the next token from a scroll." Every claim about AI lives somewhere in that stack, and the work is noticing where.
• Most of modern AI is learning from examples by walking downhill on a loss landscape. Not reasoning. Not thinking. Fitting a shape to dots, scaled up.
• Everything becomes a point in a space where nearness means similarity, and most interesting operations become distance or generation in that space.
• The capabilities are all about scale and data, and the failure modes — hallucination, bias, sycophancy, reward hacking — are all about what happens when you optimize a proxy for what you wanted instead of what you wanted.
• Progress has been and will continue to be real, and neither the doom nor the dismissal are supported by the evidence we have.

If those five ideas are in your head, cleanly, you're equipped for whatever the next several years bring. You can read a new paper and place it. You can hear a new claim and ask the right skeptical questions. You can use a new model and recognize what it's good at and where it'll fail. That's the ladder I promised you thirty-three posts ago, and you're standing at the top of it now.

The field is going to keep moving. Some of what we learned will age. The ladder won't.

Thank you for reading. Keep thinking clearly. Good luck out there.

---

Course navigation

📚 AI Zero to Hero · Course Home — all 33 posts, six modules.

---

Cover photo via Unsplash. This post is the conclusion of the AI Zero to Hero series.