Tu Jinhao: Starting from 2026

👦🏻 Author: Richard Tu

🧑‍🎨 Layout: NCon

This Monday, Crossing released a new podcast episode featuring Richard Tu, where they discussed his predictions for Agent and AI development in 2026 and beyond.

This article is a blog post Richard wrote after recording the podcast (originally published at: https://www.richardstu.com/blog/2026-and-beyond#some-hard-questions)

It covers the core predictions discussed in the podcast, plus additional reflections he didn't have time to expand on — from proactive Agents and Memory, to model-as-product, Agent capability boundaries, and shifts in alignment and human-AI interaction, and onward to longer-term, harder questions.

Below is the translated original text:

🚥🚥🚥

I recorded a podcast with Koji for Crossing in mid-January, where we talked about my predictions for Agent and AI development in 2026 and beyond.

This blog is basically a written version of that podcast episode, covering the topics I discussed at the time, plus some additions I didn't get to mention.

So if you've already listened to the episode, you can treat this as a summary and extension; if you haven't, I recommend checking it out on Xiaoyuzhou or Apple Podcast~

We also recorded a video version of the podcast:

This is my projection for the coming years — what I expect to see revolves mainly around Agents. I've done some logical deduction about what will continue to happen this year and beyond, along with some forward-looking thoughts about the more distant future.

I know I've made several similar predictions before, but given how fast AI is evolving, I think it's still valuable to revisit them. The landscape already looks dramatically different from even a year ago.

2026

For 2026, I believe the central topic will still be Agents, but they will become more personalized and genuinely useful**.

🟢 Proactive Agent Looking at the overall evolution: initially we only had chatbots, just for conversation. Then people wanted them to access external data, so we got basic tool use — searching the web, looking up real-time data.

As models became more powerful, they could process more tools through reasoning, becoming more robust. Now we've given them reasoning capabilities, and this is what we call Agents.

Expectations keep rising. People want Agents to do more personalized things, to be more useful, to know them better.

But the limitations are clear: current Agents can only execute tasks humans explicitly request. They lack self-initiating ability.

We often say Agents will take action for us and save time, but the reality is that current Agents — whether text-based, GUI-based, or a combination — are slow. After you issue a command like "buy me some groceries," you have to wait a while.

This runs counter to the original goal of having them rescue us from tedious daily work.

To improve this, Agents must truly have the ability to execute tasks independently; in other words, they must understand our usage patterns and leverage this understanding to complete tasks quietly in the background, without waiting for humans to initiate and intervene.

They need to have what we want ready in advance. This makes me think of them as a more advanced form of "Autocomplete."

The only difference is the scale of the task. Ordinary autocomplete, like Cursor's Tab mode, handles lines of code across files; task-based Agents, like Manus, handle complete tasks.

To do this, they need to learn your usage patterns. For example: an Agent knows you always have it summarize emails on Monday mornings, so it automatically does this in the future; or it knows you're running low on household supplies and just handles restocking for you.

But we also need to make sure they're not annoying, which means timing is crucial — they can't be too intrusive, nor too hidden to be useless.

This means UI and UX need to change. We can no longer rely solely on a single input box interface.

The Gmail AI Inbox released last week is a good example: it doesn't fundamentally change how you use Gmail, but the AI features it adds do improve efficiency. AI-powered tools don't necessarily need a prominent input box; they should be bound to the context of the task itself.

If this develops well, it will greatly improve people's efficiency in using Agents, and people will begin to trust them more.

🟢 Memory The second key point is memory. Expectations for models are getting higher, and models need to know users better to feel truly useful. This directly connects to the "proactive Agent" I just mentioned.

Currently, even with several general solutions for memory. From a product perspective, there are mainly three categories:

• Models use tools to store information in memory spaces (e.g., ChatGPT, Gemini, Claude, Moonshot AI, Qwen, etc.).

• Models use conversation search tools to find specific topics from past chats (Claude, ChatGPT).

• Systems summarize user interactions daily, extract new information, and store it in detailed summary memory (Claude).

These all work reasonably well, and I've seen promising performance in products like Claude. But for more general-purpose Agents, they can — and should — do better. Memory isn't limited to basic information about us; it also connects to general preferences in our lives: shopping style, coding style, travel habits, and so on.

These all affect whether an Agent can complete tasks within your expectations. But mentioning or reiterating preferences over and over is annoying. So how products "form" these memories also needs innovation.

I've been thinking about one approach: let your Agent explore your apps and websites first.

Having an Agent learn your preferences is better than describing them yourself, and you absolutely don't want to repeat yourself; so you let the Agent log into these services, it browses and checks your previous orders, learns and summarizes your preferences — what do you usually order for groceries? Which airline do you always prefer?

Then the Agent compiles these into specialized documents. Each time it enters a specific app or website, relevant instructions automatically load, ensuring the model already knows everything it needs to know.

This doesn't require any special model capabilities — it just needs the product or environment (the "Agent Harness") to be optimized to further expand the model's known information.

This approach applies to many other use cases as well. And unlike recent Skills or similar features, this doesn't require extra attention from the user or model. There won't be cases where the model ignores specific preferences, because they're loaded by default.

All of the above is product-level thinking, but we can also consider this from a more fundamental level. Sometimes models aren't aware of the importance of using user knowledge, so they simply skip it. (I should note: my ideas here could be wrong, because there aren't clear experiments proving these work exactly as described.)

• We can rely on SAEs (Sparse Autoencoders) from the field of mechanistic interpretability. Anthropic has used these techniques in some of their research. Typically, SAEs can identify feature points activated when a model generates specific tokens. If we can use this technology to detect a model's tendency to "seek external knowledge" (including user memory), then when this tendency is high, we can automatically inject relevant knowledge after that token. Once the model receives it, it can generate more useful responses.

• We can use fewer, more specialized experts in MoE (Mixture of Experts) models. For example, a model with only three or four experts, each for specific actions: one for thinking/reasoning, one for tool use, and one for responding. Perhaps another expert is needed to coordinate which expert to use at each step.

On the model-side memory mechanisms, there may be more innovations.

Either way, around proactive Agents and memory, we'll see many surprises. The key question now is how models can truly boost productivity — I think this is where they can create the greatest economic impact before truly having broad societal impact, and they should first have a huge and noticeable impact on individuals.

Trends

Additionally, I think there are some trends that will continue or begin to shift in the next year or two.

🟢 Model as a Product The first continuing trend is "model as a product." This is a long-standing pattern; I think it has two slightly different sides:

• Models have unique capabilities that can directly become a new product or feature (e.g., GPT-Image, Nano Banana, Sora 2, Genie 3, etc.).

• Models are powerful enough that people can build general-purpose products around them with some special framework/engineering work (e.g., early Manus based on Claude-3.7 Sonnet, Claude Code, etc.).

Among these products, I think Genie 3 (Google's latest world model), released to the public last week, has tremendous potential. You can create worlds, control how you "walk" through them — the entire process is customizable. This is much more fun than video models like Sora. And because it can generate interactive worlds, it (or its successors) has the potential to become one of the first reliable generative game engines.

I didn't play many games before, but if we can have reliable products based on robust world models, I might start playing, because creating my own experience sounds genuinely fun haha.

🟢 Agent Capacity The second trend is the upper limit of Agent capabilities. Models will become more robust — this is a clear trajectory. They will handle more long-tail tasks, even complex ones. They won't just accelerate software engineers (SWEs), but AI research itself; even automate parts of it. We're already seeing tremendous potential in scientific research and other fields. We also have relevant benchmarks tracking this, like METR Time Horizon, VendingBench, and so on.

METR Time Horizon V1.1

VendingBench-2 Scores (featuring Gemini-3 Pro, Claude Opus 4.5, and GPT-5.2) The curve is rising and will maintain steady growth.

🟢 Model Alignment The third point, and one of the most important, is model alignment. As models become stronger and people deploy them in more production environments, malicious consequences will become catastrophic. If a model can help scientists build fusion reactors, it can also help bad actors make nuclear weapons; if a model can help companies develop drugs, it can also create bioweapons; knowledge is fundamentally dual-use.

I've written about my thinking on this problem before, and there's much research in this area, but one promising approach I've found is the new Claude Constitution. OpenAI's Model Spec is similar, but more rule-based: what you should do, what you shouldn't do. The Constitution is more like teaching the model how to be good and do good — less like rules, more like parents teaching a child (I recall Dario describing it as "a letter from deceased parents, sealed until adulthood").

I think this is a promising direction, and I expect more companies to explore this approach.

🟢 Human-AI Interaction The final point is the changing nature of human-AI interaction. Currently, we interact with AI through apps, APIs, websites — all confined to phones and desktops. I think a truly good new entry point is AI glasses, because they can see what you see, hear what you hear. They can have their own niche — they don't need to replace phones or anything else. They can add something new: a different way to interact and coexist with AI; this isn't like Humane AI Pin or Rabbit r1, which tried to replace phones and failed.

Because AI glasses can perceive almost everything we can perceive, they will be an excellent add-on for the "proactive Agent" I mentioned earlier. They can recommend things or help complete tasks based on your real-world environment, and better memory systems become crucial here as well.

We're already seeing some products. For example, Pickle 1 looks quite promising — I've pre-ordered one and am waiting to see how it turns out. This also seems to be the direction Google is working toward, which Demis mentioned at Davos 2026. But it's still early stage.

We can focus on the earlier parts for now; glasses are more a matter of hardware, software, and ecosystems catching up.

Future

I've written about the future many times before, but AI is developing so fast that the situation looks completely different from a year ago. So I still think it's worth sharing my expectations for the longer term — after reading recent articles, interviews, and doing my own thinking, I have some new ideas.

Before going deeper, I want to mention Dario's (note: referring to Anthropic CEO) new article The Adolescence of Technology. It's a serious piece depicting the risks we face and how we might address them. I respect how he handles these issues — carefully, specifically, rather than like a doomsayer. If you haven't read it, I recommend it. What I write here is more personal perspective, from someone who will live through this transformation. Again, these are all my personal thoughts and may be incorrect.

🟢 What I Want to See The world on the other side: a world where survival anxiety is no longer the default mode of human life. Scientific progress in medicine, climate, and longevity happens much faster than before. People can pursue what truly matters to them, not just money.

Dario calls this "Machines of Loving Grace." I think he's right about the possibilities. The real question is whether we can get through the transition period without everything falling apart.

I've envisioned this beautiful future countless times. Robots handle physical labor. AI (here I mean regular models, robots, and other forms) automates almost everything, their super-high productivity brings abundance; then abundance makes material scarcity less important. People are liberated from the constant pressure of "making a living" and can truly live. This sounds utopian, but I don't think it's impossible — just very hard to reach, requiring enormous effort.

🟢 Some Hard Questions

If AI creates more value than you do, what is your purpose?

This will soon become a lived reality for many people. As I said, models are getting stronger in every aspect, companies will gradually deploy and use them for work, and the math is clear: AI is faster, cheaper, better.

The rational move is layoffs. Many people will be fired, lose their jobs. If this happens at scale, the "AI for humanity" narrative collapses, because we want it to benefit everyone, not just a small group. If something makes you economically irrelevant without giving you anything in return, you can't benefit from it — it's harm.

I think to prevent this, companies and society need to reach some consensus: even if AI creates more value, we should still protect humans for the foreseeable future.

After a company extracts the value it needs to operate, it should return value to displaced workers — this should be more like a social contract. Value always has a source.

I know this is extremely difficult to implement, almost impossible. There's no enforcement mechanism, no clear policy, competitive pressures will hinder it — but that's precisely why the journey is hard. Technology arrives faster than our social systems can adapt; this is why I said in a previous blog post that we should be the ones adapting to the development of these advanced systems. We've almost never seen the two keep pace, and our existing frameworks weren't built for this.

🟢 Meaning Without Work Even if we solve the material aspects, even if displaced workers eventually receive income, there's still the question of meaning. People don't just want things, they want to be valued, to be needed. Work used to provide this, even when the work itself was boring.

I've thought about this a lot, and there's much discussion out there. In a world where AI handles most cognitive tasks, we need new structures to carry purpose. Creative work, community, exploration, caregiving — things that matter to us, even if they don't maximize GDP. But this won't happen automatically; we must consciously build it.

Maybe this sounds abstract, but it's actually very concrete. If you didn't have to work, what would you do? Not "what would you do" in vacation mode, but long-term, what gives your life structure and meaning? For me, I think it's exploring the unknown, experiencing different places, maybe creating some things. But many people haven't even had the chance to think about this, because survival comes first.

The transformation will force us to answer this question, and I think everyone's answer will be different — that's precisely the point.

The freedom to figure out what matters to you, rather than being dictated by economic need.

🟢 The Transition Itself Obviously, this transition won't be smooth sailing. As I said before, millions will lose their jobs, society may collapse in some parts. History tells us the Industrial Revolution caused enormous suffering before things got better. This may be similar, but faster and more widespread.

The question is whether we can make the transition as humane as possible. Not "acceptable sacrifices for progress" — a framework historically used to justify much harm. Rather: we acknowledge this will be hard, and try to take care of each other through it.

🟢 Why I'm Still Quite Optimistic I know the risks are enormous. There are many doomsday perspectives, and I understand where they're coming from. Powerful AI falling into bad hands, misaligned values, societal collapse, and so on.

But many researchers are working on alignment and interpretability. Some companies (like Anthropic and DeepMind) are actually taking safety and related issues seriously. The new Claude Constitution tries to teach models to be good, not just follow rules. People are having these conversations, not ignoring them. This matters.

I've been thinking about how to hold both of these ideas at once — the hopeful vision, and the knowledge that the path there will be rough.

Honestly, it comes down to one simple thing: I believe our world can be better, and I want to see that day come. Maybe even contribute to it. That's what I hold onto.

The coming years will be hard, maybe decades. But I always come back to this question: so what? What is there to fear?

There's much more that could be written, but I think this is enough for now — I'll save the rest for future posts.

Anyway, I hope the world keeps getting better in 2026 and beyond.

(Original link: https://www.richardstu.com/blog/2026-and-beyond#some-hard-questions)