Here is what happened.

This car did not begin as a serious project.

At first, it was only an April Fools' joke BMW posted on social media: an image that understood exactly what car fans would get excited about, a piece of brand mischief that looked likely to end the moment it was posted.

BMW really turned an M3 Touring into a race car capable of running the Nürburgring 24 Hours.

Not a static show car.

Not a wish in the comment section.

Not an absurd April Fools' post that disappeared after the day ended.

It actually went on track. In March 2026, it already made its debut in the second round of the Nürburgring Langstrecken-Serie, winning the SPX class and finishing 13th overall. It is also set to run the main Nürburgring 24 Hours race on May 16-17.

When I saw this, my first reaction was not to look up the specs.

One phrase jumped straight into my head: evil big BMW.

In racing circles, people sometimes use "evil big BMW" to tease BMW. The joke works because it is not purely an insult, and not purely praise either. It feels like something unreasonable: you know this thing may be big, heavy, fierce, and filled with a strange kind of German engineering obsession, yet you still cannot help staring at it.

The M3 Touring 24H has exactly that flavor.

Evil. Too evil.

Before going further, let us straighten out the name.

Many people may casually call it the M3 Touring GT3. That name is easy to understand, because it really does look like a GT3-ized M3 Touring: wide body, rear wing, racing stance, Nürburgring narrative, the whole package.

But the official name is BMW M3 Touring 24H.

It runs in the SPX class at the Nürburgring 24 Hours, not the official GT3 class. It borrows from the technical system of the BMW M4 GT3 EVO; the engine, chassis, and racing development logic are all related to that set of things. But it is not a customer racing car built to replace the M4 GT3 EVO.

That makes me like it even more.

If BMW had simply built another standard GT3 car, it would still be cool, but not nearly as interesting. A standard race car has the logic of a standard race car: fast, stable, maintainable, sellable to customer teams, and capable of winning races. That is a serious industrial product.

The M3 Touring 24H is different.

It feels like a half-joking sentence from a car fan comment section was heard by an engineer, and then actually carried into a wind tunnel, workshop, testing program, and race track.

That is honestly absurd.

Think about it. A wagon is already a strange species. It does not occupy family rationality as naturally as an SUV, and it does not present itself as purely as a two-door sports car. The charm of a wagon is that it always carries a little everyday refusal to give in.

I can carry things.

I can pick people up.

I can go on long trips.

But I can also go mad.

That is why the M3 Touring is so attractive. It does not average performance and practicality into a compromise. It stuffs two desires that should conflict directly into one body. It can go to the supermarket during the day, to a mountain road at night, to a track day on the weekend, and still have a trunk full of completely unromantic daily clutter.

Sometimes I think people who like high-performance wagons all have a bit of this problem.

They say they have grown up, that they have started considering practicality, that they can no longer buy such inconvenient cars.

Then they turn around, see an M3 Touring, and their eyes become very honest.

Because it does not give you the feeling of compromise.

It gives you an almost shameless sense of completeness.

Honestly, BMW took far too long with this. The E46 M3 Touring Concept back in 2000 had already proved that M3 and Touring were not incompatible. That car was like a gun hanging on the wall, and it stayed there for more than twenty years.

Of course car fans kept thinking about it.

What if the E46 M3 Touring had gone into production? What if the E90 had a Touring? What if the F80 had one? These questions had no answers, which made them even easier to grow in the mind.

Not until 2022 did the G81 M3 Touring finally appear.

What feels best about it is that BMW did not turn it into a gentle version of the M3. It still follows the hardware logic of the M3 Competition: S58 twin-turbo inline-six, M xDrive, wagon body, M3 temperament.

It is not a grocery car with an acceleration pack.

It is a real M3 wearing a body that can carry more things.

So the later M3 Touring 24H works not because the April Fools' joke itself was clever, but because the line before it had been laid for too long.

The E30 M3 gave M3 its racing bloodline.

The E46 M3 Touring Concept left a blank space.

The G81 M3 Touring filled that blank.

The M4 GT3 EVO provided racing technology.

Then BMW lit the fuse on April Fools' Day 2025.

If it had only lit the fuse, that would not be special. Brand social media teams light fuses every day. They post an image, the comment section is lively for two days, then everyone disperses and the next round of content begins.

What is especially cheeky this time is that they did not stop at the image.

Even cheekier, the livery itself plays with the joke.

It is not just a random racing livery. You can see comments from the original social media post on the bodywork. In other words, the people who egged BMW on under the April Fools' post, wished aloud, and said "then actually build it" were not merely screenshotted by brand marketing for a report.

Their comments were put on the car.

That is exactly the "evil big BMW" flavor.

You thought you were just casually fanning the flames in the comment section. BMW turned around and carried the fire to the Nordschleife.

In March 2026, this M3 Touring 24H already made its debut in the second round of the Nürburgring Langstrecken-Serie, winning SPX and finishing 13th overall. As I write this on April 27, 2026, it has not yet run the main Nürburgring 24 Hours race on May 16-17.

So the story has not truly ended.

But it has already stepped out of the screen.

I think car fans are especially vulnerable to this.

Not because car fans are easy to fool, but because everyone knows too well the disappointment of concepts that remain only as images. You see a handsome concept car, a wild design study, a visual package that understands you perfectly, and inside you know there probably will not be a next step.

It is only responsible for making you repost.

It is not responsible for truly existing.

The M3 Touring 24H is different. It really has to face the Nordschleife. It has to pass technical checks, run NLS, and deal with endurance, tires, traffic, weather, and the risks of the night. The Nürburgring is cruel. You can be very good at social media, but once you enter the track, all posture is judged again by asphalt.

So I think the most precious part of it is that small piece of reality.

Looking across the landscape, it is also not the most rational answer.

The Audi RS 4 Avant is steadier, more like a fast blade you can keep around for a long time. You do not need to adapt to it too much. It knows what it is doing: all-wheel drive, wagon, fast, useful. Its emotions are less outward, but it is very mature.

The Mercedes-AMG C 63 S E Performance Estate is more like a technical declaration of the regulatory era. Four cylinders plus an electric motor, very strong on paper, and highly engineered. But the C 63 name used to carry too much V8, sound, and roughness. It is strong, but some people will still get stuck on that one feeling inside.

The Porsche Taycan GTS Sport Turismo is another answer entirely. Electric response, low center of gravity, a beautiful long roof, thoroughly modern. It is of course fast, and fast in a very clean way. But for people still attached to engines, gear shifts, and mechanical connection, it lacks a certain bodily memory.

The M3 Touring stands in the middle.

It is not as restrained as the RS 4, not as technically aggressive as the C 63, and not as future-facing as the Taycan.

But it has something hard to replace.

Its identity is connected.

You see it and know it is an M3. You see the wagon tailgate and know it is not an ordinary M3. You see the large rear wing and racing body of the M3 Touring 24H, and your mind automatically connects M3, Touring, the Nürburgring, endurance racing, and car-fan wishes.

That is enough.

As performance cars develop, parameters alone are becoming less able to move people. Horsepower can be huge. Acceleration can be very fast. Electric cars can beat many fuel-powered performance cars into the ground. Lap time still matters, of course, but lap time increasingly feels like a private language among serious players.

What ordinary car fans truly remember is often a clear story.

An M3 Touring delayed by more than twenty years finally goes into production, and then is pushed onto an extreme stage that originally did not belong to wagons.

That story is very clear.

It is even a little childish.

But the most valuable things in car culture are often exactly this kind of childishness. Can we fit this in? Can we stuff that engine in? Can we make a wagon run the Nürburgring? Can you stop only showing me pictures and actually build it?

Very often, adults say they want rationality, but deep down this is what they most want to see.

You actually did it.

Those five words hit harder than a pile of specifications.

Of course, I am not saying projects like this can be copied endlessly.

If every year BMW turns another April Fools' joke into reality, and every time says they heard the fans, and every time builds a social-media highlight, the flavor will fade quickly. What makes the M3 Touring 24H work is not the rhetoric. It is the engineering investment, and the fact that it really has to face the track.

One less piece, and it collapses.

So I prefer to see it as a rare highlight.

It may not become a regular racing product, and it should not. The M4 GT3 EVO is the proper weapon in BMW's customer racing line. The M3 Touring 24H is more like a sudden bright interlude in the story of the M brand.

But this interlude matters.

It reminds me that car brands can still occasionally do things that are not so spreadsheet-shaped. Not every project has to immediately turn into sales volume. Not every answer has to be explained to a financial model. As long as the engineering is real, enthusiasm has somewhere to land.

I like this car probably because of that.

It is not rational enough.

It may even be a little unnecessary.

But in performance car culture, many of the things people cannot stop thinking about were never basic needs in the first place.

A large rear wing is not a basic need.

A wide body is not a basic need.

Sending a wagon to run the Nürburgring 24 Hours is certainly not a basic need.

But people like cars because they are often struck by exactly these non-essential things.

They make a machine more than a machine, a brand more than a brand, and make someone scrolling images in front of a screen suddenly feel that their childish love has been taken seriously once.

That is already rare.

I am putting this vertical image at the end purely out of personal preference.

If you have read this far, you probably did not click in by mistake.

You likely understand the feeling too.

A car may be impossible to buy. A Nürburgring 24 Hours race may only be watched through a screen. A pile of technical details may be forgotten in two days.

But a certain image, a certain stance, and the moment when an absurd idea actually lands can stay.

For me, the M3 Touring 24H is exactly that.

Not the most rational BMW.

But very much like the BMW in my heart that has not yet been worn smooth.

On April 25, I played with gpt-image-2 again.

This time I did not continue making resume posters, nor did I seriously test any complex infographic. I only did a very private little thing: I generated a set of posters for several F1 drivers I like.

And once again, it hit something inside me.

A few days ago, I had already written once about gpt-image-2. In that piece, I talked about why it is no longer just good at drawing images, so I will not repeat model release details, parameters, or capability boundaries today. Those things are certainly important, but for me, the truly interesting change is happening somewhere smaller.

It is the moment when you suddenly realize it can not only help you make a visual draft for work.

It can also turn your private taste into an image you genuinely want to keep.

F1 is a subject that suits posters very well.

The drivers themselves are figures, and the teams bring their own colors, symbols, and visual language. Red Bull, McLaren, Ferrari, Mercedes. These names do not need much explanation. People who like racing see them and automatically recall a whole set of things: liveries, helmets, pit garages, start lights, radio messages, podiums, and the tension of staring at a live broadcast on a weekend night.

This is different from simply "generating a good-looking racing image."

An ordinary racing image can probably get away with being fast, flashy, and full of speed. But a driver poster cannot. It has to handle two things at once: public symbols and private emotion.

Public symbols are the parts everyone recognizes. Team colors, sponsor blocks, the sense of speed in motorsport, and the central position of the portrait. If these are lost, the image no longer feels like F1.

Private emotion is the part only you care about. Why you like this driver, why you are willing to turn his image into a standalone poster, why a certain color or posture feels right to you. These things are hard to write directly into a prompt. You can describe them, but description is never complete.

So when the generated image appears, if it happens to line up, you very clearly pause for a moment.

That pause is not simply because it "looks so much like him."

It is more like this: in your mind, there has always been a vague display cabinet containing drivers, teams, race memories, and your own aesthetic preferences. In the past, these things could only exist separately, in wallpapers, screenshots, or the emotions left after a race weekend. Now gpt-image-2 suddenly gathers them into one poster.

That feels really good.

Honestly, I increasingly feel that the most easily underestimated value of AI image generation is not how much design cost it saves, nor whether an image goes from 70 points to 90 points.

It is more like helping ordinary people complete a kind of expression that used to be hard to complete.

You like F1. You like certain drivers. You want a visual collection of your own. This need is not big or grand, and it may not be worth hiring a designer to make a whole set. But it matters to you.

In the past, small needs like this were often suppressed.

Because the cost was too high. You had to find references, assemble images, adjust colors, cut out subjects, and design the layout. Once you really began, half your patience might be worn down before your enthusiasm faded.

But now it is different.

You can quickly throw that feeling to gpt-image-2 and let it give you something visible first. It may not be perfect, and it definitely still needs human judgment: whether the person is accurate, whether the team elements are appropriate, whether the text has problems, whether the scene is overly dramatic.

But it has already pushed past the hardest step.

The distance between "I want a poster like this" and "where can this poster be improved" has suddenly become shorter.

That is the strongest feeling I had this time.

When I wrote about the resume poster, what excited me more was that it had started to organize information. That was a workflow judgment. It turned a blank canvas into a draft that could be discussed, where you could point out what to change, what to delete, and what to strengthen.

This set of F1 driver posters excited me in a more personal way.

It made me realize that generative models do not only help us finish tasks faster. They also make some very small, very personal wishes, wishes that previously did not justify starting a complex process, suddenly executable.

Make a set of posters for drivers you like.

Make a series of collectible cards for a game character.

Turn a travel memory into a movie poster.

Generate a visual keepsake for a song, a car, a relationship, or a city you love.

From a business perspective, these things may not be large. From the perspective of personal life, they carry weight.

Because people naturally make room for what they love. Some buy models, some put up posters, some keep tickets, some save race screenshots. We keep these things not because they are rare, but because they fix a part of ourselves in place.

This is where gpt-image-2 amazed me again.

It did not simply make F1 look cooler.

It let someone who likes F1 suddenly own a set of images that could be placed into his own world.

That sounds small.

But I think it matters.

Because the most interesting moments in AI are often not when it appears all-powerful, but when it happens to catch a very specific thought of yours. You did not prepare a requirements document, you did not prepare to start a project, and you did not prepare to do anything very serious.

You just suddenly wanted to try.

And then it really gave you a surprise.

Just like this set of F1 driver posters on April 25.

I had only meant to play around.

In the end, it took me seriously once again.

Here is what happened.

On 2026-04-21, OpenAI released Introducing ChatGPT Images 2.0.

At first, I did not take it too seriously.

It was not that I was not looking forward to it. It is just that over the past two years, AI image generation has produced so many astonishing moments that everyone has become a little numb. You give it a sentence, and it gives you a beautiful image. The lighting is good, the atmosphere is good, and the details are quite full. And then?

And then, often, there is no "then."

Especially when it comes to real work, problems appear. You want it to make a usable poster, and it starts inventing text. You want it to organize a pile of information, and it turns the layout into a colorful PPT cover. You want it to balance a person, resume information, technical direction, and visual style, and it often tries very hard, then exposes itself very clearly.

So in the first two days after release, I browsed what other people had made with ChatGPT Images 2.0. It was indeed impressive, but I was still only watching from across the screen.

Until today, when I tried it myself.

I gave it my own information and asked it to make a personal resume poster.

When the image appeared, I was genuinely stunned for a few seconds.

To be honest, my expectations for AI image generation used to be quite fixed.

It can make atmosphere. It can make concepts. It can make those emotionally satisfying images that look great at first glance: cyber cities, future laboratories, rainy neon nights, lonely backs, people in trench coats standing in front of giant screens.

Those images are of course good-looking.

But between that and a real design deliverable lies a very wide river.

The truly difficult part is not whether it can draw something that looks right, but whether it can place information properly. Where should the name go? How should the technical direction be presented? How should project experience be layered? How do awards and skills avoid fighting each other? Does the whole image look like a unified visual system, rather than a pile of assets forced onto a canvas?

A resume poster is exactly a stress test for this kind of scenario.

It cannot only look good.

It also has to explain things clearly.

It cannot only feel technological.

It also has to look like a real design draft that could be shown to someone.

That is the hard part. Because these were exactly the areas where AI image generation used to fail most easily. Text went wrong. Layouts became chaotic. Information hierarchy collapsed. The portrait looked as if it had been borrowed from another model. The background turned into a mess of blue light, circuit traces, HUDs, and hexagonal grids.

I am telling you, that flavor is too familiar.

Everything looks very AI, but nowhere does it feel like a human is actually doing design.

But this image felt different.

What surprised me most was not how similar the portrait was.

Really, it was not.

The portrait matters, of course. But when I saw this image, my first reaction was instead: it is actually trying to handle information seriously.

It did not only stare at the face, nor did it just paste in the familiar technological trio of blue light, circuits, HUDs, and hexagonal grids. It placed the name, HPC, project experience, awards, and technical direction into a relatively unified layout. The whole image had priority, whitespace, and a visual center of gravity.

It is not perfect.

Small text definitely still needs manual proofreading, and some details cannot be used directly as a final version. For example, whether certain descriptions are accurate, and whether certain layout choices fit the seriousness of a real resume, still need a human to close the loop.

But the point is that it is no longer the kind of image you can only post to Moments with a comment like "so cool."

It feels like a draft that can be discussed.

That is a big change.

Because the most painful part of many creative tasks is not the final 20 percent of refinement, but the blankness at the beginning. You know you want something, but you cannot quite say what it should look like. You open the design software, the canvas is bright white, the mouse drifts around for a long time, and eventually you go looking for references. One hour disappears.

At this point, you should understand why I was excited.

What interests me about gpt-image-2 this time is not that it completes the design for you, but that it pushes the first visible thing in front of you.

You can say this text needs changing.

You can say this area is too crowded.

You can say project experience should stand out more, awards should not overpower the name, and the sense of HPC technology could be more restrained.

But at last you are no longer speaking into the air.

You are speaking to an image that already exists.

Frankly, this is where AI tools truly begin to become useful. Not by generating a perfect final product with one click, but by pulling your idea out of fog and turning it into something that can be seen, criticized, and modified further.

Sometimes I feel that discussions about AI creation go too easily to extremes.

One side says designers are finished.

The other side says AI will never understand real aesthetics.

I do not particularly want to stand with either judgment.

Because what is really happening may not be that dramatic. It is more like one very annoying part of the workflow suddenly becoming thinner. It does not mean there will be no designers from now on. It means that the pain of forcing out the first draft from zero may be reduced a lot.

You still need judgment.

You still need aesthetics.

You still need to know what information should be emphasized, what details must be removed, and which parts look high-end but are actually cheap routines.

It is just that these abilities used to have to pass through a large blank canvas before they could take effect. Now the tool may first give you a target.

That is very different.

Back to gpt-image-2.

The official help center says ChatGPT Images 2.0 now covers all subscription tiers, and the developer documentation lists gpt-image-2 as the latest GPT Image model, with a snapshot available at gpt-image-2-2026-04-21. This information is certainly important, but simply repeating parameters and release cadence is not that interesting.

What is really interesting is that it is pushing image generation from pretty pictures toward complex visual tasks.

Posters, infographics, handouts, resumes, event materials, product images, course covers. These things were not impossible to make with AI before, but they often made you sigh halfway through. Because it was very good at generating, but not very good at organizing.

This time, for the first time, I felt quite strongly that it is starting to organize.

Of course, it is still early to say this.

One image cannot represent everything. Especially with a newly released model, it is easy to get excited, and only after calming down and testing a dozen more rounds can one draw a serious conclusion. I will probably keep trying other images later, especially harder scenarios such as multilingual text, complex infographics, and project showcase pages.

But today's resume poster is worth recording separately.

It made me realize one thing.

The most valuable direction for AI image generation may not only be helping ordinary people make a beautiful picture.

More importantly, it lets ordinary people more quickly obtain something they can begin to modify.

From blank page to first draft.

That step alone is already huge.

On the evening of Saturday, April 18, I met up with friends at "Chuanxi Huoyanshan Xichang Firepot BBQ (Xindu Green City Garden Branch)." Calling it a gathering makes it sound more formal than it was. We simply found a place to sit down, gathered around a brazier of charcoal, ordered a few things we wanted to eat, drank a little, and slowly opened up the recent bits of life.

The best part of this kind of meal is that it does not hurry you. Once the skewers are placed beside the brazier, you cannot eat them immediately. You can only wait as the charcoal slowly draws out the fat. When the mountain-raised pork, spicy tender beef, grilled pork belly, and grilled pork skin were laid out one skewer after another, the table already had a solid, smoky warmth to it.

I have always felt that firepot barbecue suits a gathering of friends better than many forms of barbecue that arrive already cooked. It naturally slows the pace down: someone watches the fire, someone turns the skewers, someone warns that a piece is about to burn, and someone takes the first bite. Before the meat is ready, the conversation has already connected; by the time it can be eaten, the table is warm as well.

What impressed me most during this meal was, unexpectedly, the seasoning plate. When eating barbecue, the most familiar seasoning is of course chili powder: fragrant, spicy, direct, and immediately present once you dip into it. But this time there was also soybean powder on the plate. I tried dipping grilled meat into it for the first time, and the result was surprisingly good.

Soybean powder does not seize the flavor the way chili powder does. It is more like placing a soft layer of aroma underneath the charcoal flavor. The grilled meat has its own fat, and the charcoal carries a little roasted bitterness. After dipping it in soybean powder, the bite becomes rounder, and the aroma grows thicker. It is not simply "heavier tasting"; it is the sudden realization that barbecue can be eaten another way: less aggressive, but very lasting.

Fenjiu was poured slowly into small cups, the measures neat and just right. A plate of crisp peanuts sat at the corner of the table. There was no need for deliberate small talk. We peeled a few casually, sipped in the warmth of the charcoal, and let the conversation drift. The night was quiet, the liquor light, and the talk wandered without an edge. An ordinary evening gained a few more gentle flavors.

Vegetables also fit this meal well when eaten slowly. Eggplant absorbs flavor, lotus root stays crisp, onion becomes sweet after grilling, and lettuce can wrap just-cooked meat. Under the slow heat of charcoal, the fresh squid gradually softened from its original firmness into something tender and elastic. Wrapped in the clean sweetness and bite of onion, the smoke entered just enough.

Two Drinks Worth Noting​

We had Fenjiu at the table this time, and also Asahi beer. Putting them into the same barbecue meal was actually quite interesting: one is a representative of light-aroma Chinese baijiu, while the other is a beer built around a dry, clean finish.

Fenjiu is a classic representative of light-aroma baijiu, closely tied to Xinghua Village in Fenyang, Shanxi. The feeling of light-aroma baijiu is usually not heavy or stacked, but cleaner and sharper. At a barbecue table, drinking it slowly from small cups means it does not overpower the skewers or seasonings. Especially in a meal carrying the smell of charcoal, a baijiu that is too sweet or too thick can easily steal the scene. Fenjiu's cleaner style can meet the fat of grilled meat quite well.

Asahi beer takes another route. Super Dry was born in 1987 and centers on a Karakuchi style: dry, crisp, and quick on the finish. Its fit with barbecue is also straightforward. Skewers have fat; chili powder and soybean powder enrich the flavors in the mouth. At that moment, a sip of cold beer with a clean finish can gently carry away the oil and spice from the previous bite, making you naturally want the next skewer.

So the drinks were not there to make the night extravagant. They added ease to the table. Fenjiu suits slow, small toasts. Asahi suits cleansing the palate after grilled meat. Together, they matched two rhythms in a gathering of friends: one for slow conversation, one for continuing to eat.

An Ordinary Night​

Looking back at these photos, what I remember is not only what we ate and drank. More specifically, it is the red glow in the charcoal, the roasted aroma slowly forming on the edges of the skewers, the unexpected delight of dipping into soybean powder for the first time, and the ease of sitting beside friends without having to search for a topic.

Very often, the things worth keeping in life do not need to be grand. A barbecue, a few skewers, a plate of soybean powder, a little drink, and several people willing to sit together and wait for the meat to cook are already enough to become a real footnote to an evening.

A person should first become brave, and then become strong. Many things do not come with so many standard answers. Do not trap yourself too early in questions like "Do I really like this?" or "Is this the most perfect path?" Liking something and making money from something can be separate. If you do not hate it, can keep doing it, and can earn from it, then keep going for now.

The more I think about it, the more I feel that many people have been misled too deeply by the word "passion." It is as if something is not worth doing unless you genuinely love it from the bottom of your heart. That is not true. Looking at reality plainly, anything that can support you, help you stand firm, and let you accumulate ability already has value. Survive first. Become stronger first. Then talk about freer choices.

If you really want to make money, do not begin by asking other people, "How do I make money?" Behind that question, you should first ask something else: why would they be willing to tell me? What they are telling me, is it a way to make money, or a way to turn me into their income?

In recent years I have become increasingly wary of one kind of rhetoric: on the surface, it teaches you how to enter the game; underneath, it monetizes your anxiety. There is a signal that is not absolute, but very useful, when judging whether an industry is starting to decline: are more and more people inside that industry beginning to teach others how to make money?

When an industry is truly still rising, most people keep putting their energy into doing the work, because doing the work directly is more profitable. Only when some people find that front-line money is no longer so easy to earn do they start dismantling their experience and packaging it into courses, communities, consulting, and methodologies for a second round of monetization. To put it bluntly, you may well be the final wave they plan to earn from before exiting.

So a person needs to be brave, and also strong. Bravery means daring to judge for yourself, instead of being easily led by other people's success narratives. Strength means being able to distinguish a real opportunity from bait someone placed while arranging their own retreat.

Here is what happened.

On April 7, I spent some time reading LangChain's article, The Anatomy of an Agent Harness.

At first, I was drawn in by the word "harness."

It has been appearing more and more often in the Agent world recently, but it is also a slightly awkward word. If you translate it directly as "tack" or "restraint gear," it sounds strange. If you translate it as "framework," it feels too light, as if it is only a few layers of code wrapping. "Exoskeleton" may be a bit closer, but it is still not exact.

In any case, it points to this thing:

the whole system outside the model that lets an Agent actually do work.

My strongest feeling after reading it was that, in the past, we may have stared too much at the model itself.

Which model is smarter.

Which model reasons better.

Which model writes code more fiercely.

These things are of course important. You cannot take a very weak model and insist that engineering alone can make up for it. That is unrealistic.

But if you have really used Claude Code, Codex, Cursor, or built Agents yourself with LangGraph, the OpenAI Agents SDK, or CrewAI, you quickly run into an annoying problem.

A smart model does not equal a reliable system.

It can think of the next step, but that does not mean it knows when to stop.

It can call tools, but that does not mean it knows which tools must not be used carelessly.

It can write files, but that does not mean it knows how to roll back a bad change.

It can read context, but that does not mean it will not tie itself in knots after a dozen turns.

It is like putting a very smart person into a car with no steering wheel, no brakes, and no dashboard.

The brain is good.

The car is not.

What moved me most in the LangChain article was that it made this problem very concrete. It was not vaguely saying that Agents need engineering. It took the harness apart and showed you planning, filesystem, subagents, stateful middleware, context engineering, tool permission, and evaluation loop.

None of these words look exciting.

But every one of them is critical.

Take planning, for example.

Many people may think a plan is just asking the model to write a todo list. Sounds ordinary, right?

But think about what most often goes wrong when an Agent does a complex task. It is not that it completely cannot do the work. It is that it drifts while doing it. It starts out fixing a bug, then halfway through begins refactoring the whole project. It starts out writing an article, then halfway through adds a pile of background material barely related to the topic.

At that moment, a plan is not decoration.

A plan is a rope.

It is not there to restrict the model for the sake of restriction. It is there so that every so often the model can look back and see what it is actually doing.

The filesystem is the same.

We used to think the longer the context window, the better. The longer the window, the more the model remembers, and the stronger the Agent becomes.

That is only half true.

Long context is useful, but you cannot stuff everything into it. Code repositories, logs, drafts, materials, previous attempts, intermediate results. Once everything goes into the context, the model looks as if it knows everything, but in reality its attention becomes scattered.

The filesystem is another brain.

Write down what should be remembered. Read it back when it should be read. You do not need to stuff the whole world into the prompt. You only need the model to know where things are, and when to retrieve them.

This has also become one of my strongest feelings recently.

An Agent is not a model that is better at chatting.

An Agent is a model placed into an environment.

Once the environment becomes complex, the problem changes.

The early ReAct pattern had already exposed this issue. The model thinks one step, acts, observes, and then continues thinking. Back then it was still a very simple loop of Thought, Action, and Observation, almost like hand-building a small operating loop inside a prompt.

Then AutoGPT had its moment. For the first time, many people saw a model decomposing tasks, searching, writing files, and continuing to iterate by itself. It felt astonishing.

And very soon, people saw the other side.

It gets lost.

It falls into loops.

It works very hard in a large circle, and finally hands over something no one knows how to evaluate.

That moment already proved that a model's ability to think is not enough. You still have to give it state, permissions, tool boundaries, error recovery, and places where humans can intervene.

Back to the harness.

I think what LangChain's article really wanted to say was not, "We have invented another new concept."

It was more like a reminder that a large part of Agent capability lives outside the model.

There was an interesting example in the article. They said that with the model unchanged, changing the harness moved a coding Agent from Top 30 to Top 5 on Terminal Bench 2.0.

That result should not be mythologized. A benchmark is always only a slice.

But it is enough to make the point.

The same model can perform very differently when placed inside different systems.

This is actually similar to people.

A smart person without schedules, documents, collaboration tools, review mechanisms, or reminders about priority can also become chaotic. Conversely, someone whose ability is not exaggerated can still produce steadily if they have good processes, notes, tools, and feedback systems.

Agents are the same.

The model is the brain.

The harness is the work habit.

At this point, I suddenly understood why products like Claude Code are so compelling.

It does not only give you a model that writes code better. It gives the model a very concrete work setting. It can read a repo, edit files, run tests, inspect errors, confirm dangerous operations with you, and keep moving through a task.

Behind those experiences is the harness.

The OpenAI Agents SDK is moving in the same direction. It brings concepts like agent, handoff, guardrail, session, and tracing into the SDK. Anthropic's MCP is, in a sense, trying to standardize the way models connect to external tools. LangGraph feels more like a controllable Agent runtime for developers, so that state, branching, human intervention, and durable execution do not all have to be handmade from scratch.

You can see it: everyone says they are talking about Agents.

But the real competition increasingly looks like a competition over whose shell is better.

This matters to ordinary developers too.

If you are only playing around, pick whatever product you like. If Claude Code feels good, use Claude Code. If Cursor suits you, use Cursor. If the OpenAI Agents SDK is convenient, use the OpenAI Agents SDK.

But if you really want to put an Agent into your own business, you cannot only ask whether the model is strong.

You also have to ask where its memory lives.

Who manages tool permissions.

How failure recovers.

Whether logs can be replayed.

Whether humans can step in at critical points.

Whether this workflow can be carried over when you change models later.

This question will only become more important.

Because after using Agents for a long time, the most valuable thing may not be any single answer, but the working style accumulated around it. It knows how you write code, how you organize materials, your team's process, which commands are dangerous, and which files should not be touched.

If all of that sinks into a closed-source product, you may feel very comfortable in the short term.

In the long term, you may feel uneasy.

This is not to say closed-source products are bad.

I use them too, and in many cases they are indeed good.

But we need to know what we are handing over.

We used to say that model vendors sell intelligence. Now it seems that, in the future, they may sell an entire work environment. The longer you work inside it, the more it understands you, and the harder it becomes to leave.

Big times, my friends.

So my mindset when looking at Agents is no longer quite the same.

When I saw a new Agent demo before, I would first ask what model it used.

Now I first ask what its harness is.

Does it have a planning system?

Does it have external memory?

Does it have tool permissions?

Does it have observability?

Does it have rollback and evaluation?

Is it a smart brain that can chat, or a small system that can truly do work reliably?

That is the difference.

After reading the article on April 7, an image kept staying in my mind.

A model stands at the center of the stage, and everyone is watching how smart it is.

But lighting, sound, teleprompter, backstage coordination, stage machinery, and emergency brakes are all hidden in the dark.

The audience sees the actor.

What really decides whether the performance can finish smoothly is the whole theater.

Agents are the same.

We will of course keep chasing stronger models.

But perhaps from now on, we also need to look carefully at the theater beneath the model's feet.

On Thursday, April 2, 2026, 19:30 - 21:00, I attended a talk by Professor Qiang Xu from the Department of Computer Science and Engineering at The Chinese University of Hong Kong. The title was sharp: Paper Is Dead, Hardcore Research Must Rise: Self-cultivation For Researchers In The AI Era.

"Paper is dead" is of course a deliberately provocative phrase. But what it really pierces is not whether papers have value, but a deeper question: when AI is systematically rewriting the research workflow, what should researchers use to prove their value? And what should count as the real deliverable of future research?

After listening to the full talk, my strongest feeling was that the AI era has not made research easier. Instead, it has made the question of what truly counts as research much clearer.

When Papers Are No Longer The Only Endpoint Of Research​

For a long time, papers were almost the most typical representation of research output. Whether a study "stood" often depended on whether it could be written into a paper and published at a sufficiently good conference or journal. Papers are of course important. But if the endpoint of research is completely equated with publication, that paradigm is now being rapidly challenged by the AI era.

Professor Xu repeatedly emphasized that research outputs with real vitality in the future should not only be papers. They should also include tools, APIs, benchmarks, ground truth, and other infrastructure that Agents can call. In other words, the value of research is no longer reflected only in citations. It is also reflected in whether the work truly solves real needs, whether it can be called, reused, extended, and eventually enter a long-running technical ecosystem.

This is, in essence, pulling research back from "writing it out" to "building it out."

Around this point, the talk also mentioned an evaluation framework worth remembering: ROSE. A good research work should solve a real need, be original, be significant, and retain a certain elegance. Compared with asking "where was this paper published," this standard is obviously harder, but also closer to research itself.

Because truly good research does not necessarily stop at a paper. It is more likely to continue becoming tools, systems, and methods, and to keep living inside later work.

The Stronger Agents Become, The More They Need A Regulatory Kernel​

The second main thread of the talk was system governance in the Agent era.

Today, when many people talk about Agents, they focus more on the expanding boundary of capability: Agents can understand tasks, decompose workflows, call tools, and in some scenarios even show signs of near-autonomous action. But the more this is true, the harder it is to avoid one question: if Agents become increasingly powerful, who ensures that their final behavior remains safe, reliable, and controllable?

Professor Xu's answer was a "regulatory kernel."

This concept is very interesting. It is not simply adding a few rules to a model. Rather, it tries to build a constraint structure between the large language model and the external world, somewhat like an operating system. The large model is responsible for open-ended semantic output, for proposing plans and creating possibilities. The regulatory kernel maps those open semantics into deterministic instructions that can be inspected, constrained, and executed, and then uses boundary checks and rule interception to ensure that the system does not cross lines or lose control.

In one sentence: the LLM thinks; the kernel guards.

Behind this lies a crucial cognitive shift. A mature Agent system cannot be built only by stacking "smarter" models. It must have a verifiable and governable underlying mechanism. System design in the AI era is not only a question of capability. It is also a question of boundaries.

AI Levels Techniques, Not Research Ability Itself​

The point I agreed with most in the entire talk was that it did not treat AI simply as an "efficiency tool." It went further and asked: when tools become stronger and stronger, what remains irreplaceable in a researcher?

Professor Xu's judgment was clear: AI will indeed flatten many differences at the skill level, but it will not flatten research ability itself. What automation more easily replaces is repetitive labor, low-level "paper-churning," and procedural work built only on proficiency. What remains difficult to replace is insight into problems, judgment about directions, the ability to distinguish reliable information from falsehood, and the ability to move a vague problem toward a verifiable result.

This means that research training in the AI era has changed.

In the future, more important than "can you use tools" will be: can you propose a question worth studying? Can you communicate with AI at a high level? Can you judge which results are genuinely effective and which are merely hallucinations that look like answers? Can you push a piece of work from an idea into an outcome that can truly be reused, communicated, and deployed?

The tools are getting stronger, but the abilities that decide the ceiling are becoming more fundamental.

The Q&A Made The Researcher's "Self-cultivation" Clearer​

Several points from the Q&A after the talk are worth recording separately.

Someone asked whether AI would make the gap between people larger, especially whether those already good at organizing, judging, and leading would pull further ahead with AI support.

Professor Xu's answer was not pessimistic. He did not think research ability would become permanently locked in the hands of a few "naturally gifted" people. Large models may well amplify efficiency gaps. But research thinking itself can be trained. How to sense the value of a problem, how to judge whether information is reliable, and how to organize one's research process are not mysterious talents that cannot be learned. They are abilities that can be built through long-term training.

He also mentioned that when reading papers, the most important thing is not only to see what conclusion the authors reached, but to ask: how did they get onto this path? What a mature researcher truly needs to cultivate is not only a store of knowledge, but research taste and judgment. You do not only need to "understand a paper." You need to see how a problem was discovered, defined, advanced, and solved.

This is very important. AI can help you summarize papers faster, extract information, and generate ideas. But if you lack judgment, you can easily be led around by content that is fluent on the surface but empty underneath. The stronger the tools become, the scarcer judgment becomes.

In The AI Era, Education Should Not Mainly Train "People Who Can Solve Problems On Tests"​

The talk also included a discussion about education that left a deep impression on me.

With the help of AI, knowledge acquisition and skill learning will become increasingly democratized. Many things that once required long accumulation are now becoming easier to reach and easier to call upon. But this does not mean education will become lighter. On the contrary, it may become harder.

Because when knowledge itself becomes easier to obtain, what becomes scarce is no longer how much content you remember, but how you judge what is worth learning, what is worth doing, and what is worth believing. What truly matters in the future may not be only teaching students to master knowledge, but cultivating judgment, the ability to choose, critical thinking, and the ability to form a sense of problems around real interests.

To put it more directly, AI can become an executor, but only if you first are a person with a sense of direction. If someone has no problem consciousness, stronger tools will only let them spin in place faster.

Closing​

So what "Paper is dead" really means is not that papers are useless, but that the era in which papers alone serve as the endpoint of research is ending.

Real "hardcore research" in the AI era asks researchers, on one hand, to produce outputs closer to real-world needs, leaving deliverables that can be reused, called, and incorporated into systems. On the other hand, it also asks researchers to cultivate again those most basic and most crucial abilities: problem consciousness, judgment, research taste, and the ability to work with new tools.

The tool revolution has already happened. What will really decide how far a researcher can go next may not be whether they can use AI, but whether, in the AI era, they can still see which problems are truly worth solving.

Appendix: Original Meeting Notes​

The following preserves my note-style record from the talk for reference.

Summary​

This session centered on changes to research paradigms in the AI era. It discussed the reshaping of research deliverables, the regulatory kernel in the Agent era, and core questions around research value, AI tool use, and educational models.

1. Reshaping Research Paradigms And Deliverables​

The core value of research should not be limited to paper publication and citations. It should also be reflected in whether the work can be called by AI Agents and generate real social impact.

Future research deliverables should include infrastructure such as tools, APIs, benchmarks, or ground truth that Agents can call, ensuring the long-term vitality and ecological position of the work.

The "ROSE rule" was proposed as a new standard for evaluating research: solve a real need (R), possess originality (O), have significance (S), and demonstrate elegance (E).

2. Regulatory Kernel In The Agent Era: OS System​

To address uncertainty, uncontrollability, and potential risks in Agent systems, the talk proposed building a "regulatory kernel" (OS) as the Agent's "rein."

This kernel maps semantic output from large language models into deterministic instruction sets, combines them with rules for boundary checking and interception, and forms a neuro-symbolic system.

The goal is for the LLM to create possibilities and formulate plans, while the regulatory kernel guards the boundaries and ensures safe, reliable execution.

3. Key Issues And Views​

Research value and "water papers": the value of hardcore research lies in deep insight into problems and real solutions, not simple repetitive labor. Automated "water paper" systems lack cognitive gain, while high-quality research processes themselves benefit personal ability.

Use of AI tools: AI tools such as large models can effectively flatten skill-level gaps among researchers with different capabilities. But core "research thinking," such as how to propose valuable questions, how to communicate effectively with AI, and how to judge truth from falsehood, still requires training and cultivation.

Evolution of education: facing AI's impact on knowledge acquisition, education should shift toward cultivating students' core judgment, critical thinking, and cross-domain integration ability, rather than simply transferring knowledge. The role of the classroom should become guiding students to solve problems and engage in deep exchange.

To-do​

1. Open-source project release

Qiang Xu will soon release the regulatory kernel system (OS) developed by his team and invite the community to try it and provide feedback.

Q&A Notes​

Liu Weifeng: In the first year or two, my feeling was different from the last two years. AI seemed to flatten the research ability of mid-level and senior people, but recently it feels like the gap has widened. People who can use large AI models well are pulling ahead, and many of them have strong organizational and management talent, abilities they have had since childhood. These abilities let them make better use of AI. This gap seems hard to catch up with and may become fixed. How does Professor Xu view this?

Xu Qiang: I have a different opinion. The Industrial Revolution flattened differences in physical strength, and large models flatten differences in human skills, but I do not think they flatten research ability. The research way of thinking can be trained: how to sense the value of a problem, how to communicate effectively with AI, and how to judge whether information is true or false. If you consciously train research thinking, it will not be flattened by AI. So I deny that research ability in the AI era is strongly tied to leadership and organizational talent. Xu also mentioned in his book that the gap will not necessarily become larger; research has patterns and can be trained.

There are too many papers to read, so you must select carefully. When reading papers, the main thing is to consider how the author arrived at this path, not just the research content itself. Improve your research taste and judgment, and then you can improve the efficiency and effectiveness of using AI for research.

What kind of model to use: use the best AI tools, and pay for the strongest models to improve yourself.

School education: with AI, knowledge and skills become more equalized, but judgment and the ability to choose still belong more strongly to experienced people, such as senior students and teachers.

Children's education: every era has its own needs. It should not be viewed through the lens of people like us who carry the strong smell of the old order. Knowledge and skills themselves are becoming more equalized, so education now needs to change in this respect. Children must have things they are interested in and love. If they love something, they will naturally have discoveries, needs, and ideas, and at that point they can use AI to work on them, making AI the worker for their own interests and passions. People with ideas and ability will not be replaced. But if someone has no interests and no ideas, then AI really may replace them to a large extent.

On March 20, two friends and I set out from school and went to one of their homes for a small gathering. Since all of us like racing, we first stopped by a simulator shop in the afternoon, then went to the market in the evening, bought groceries, and cooked dinner ourselves. Looking back now, the day did not have any grand plan, but from the steering wheel to the stove, and then to the slightly tipsy conversation at the table, every part of it feels worth recording.

That afternoon found its rhythm the moment we walked into the simulator shop. Once the triple screens, steering wheel, and pedals were set up, sitting down felt automatic. We drove from F1 to GT3 cars at the Nordschleife in Assetto Corsa, then switched to rally racing, and somehow even Euro Truck Simulator did not escape us. When people who all like cars stay together, happiness often needs very little setup. Circuits, mistakes, restarts, and teasing one another were already enough to fill the afternoon.

After the simulator session, we went to the market on the way and then returned to my friend's home. In the afternoon we were still fighting with circuits; by evening we were busy around the kitchen. That switch was oddly satisfying. I made garlic butter shrimp, steamed sea bass, stir-fried clams, and bitter melon with eggs. As several of us helped here and there and waited for the dishes to come out, the smell in the room slowly changed from the spring outside into something much more concrete and domestic.

When we finally sat down at the table, a gathering where we bought groceries and cooked for ourselves did feel different from eating out. The steamed sea bass had a calm freshness. The butter shrimp carried a rich aroma. A large bowl of clams immediately made the table feel lively. The bitter melon with eggs balanced the meal with a home-cooked flavor. The photos also show a few other simple side dishes, but what I remember most from that night is the ease that came after everyone had made the dishes one by one and could finally sit down.

By the time dinner started, two one-liter cans of draft beer were already on the table. We did not drink in any exaggerated way. It was just enough, from the first bite of food onward, to push the whole evening into a comfortable place: faces warming a little, people relaxing a little. Later, my friend brought out some of the model cars he collects. Some had hoods that could open; some even had removable body shells. They were so detailed that you could not help leaning closer and looking slowly.

Looking back on this gathering now, what remains is not only which games we played or which dishes we cooked, but how naturally everything connected that day. We talked about racing in the afternoon, cooked in the evening, and looked at model cars after dinner. Interest and ordinary life were not separated; they flowed into one full day. A gathering like this does not need to be grand, but it can make a day feel thoroughly lived.

When I read The True Story of Ah Q as a child, it was easy to see Ah Q as someone far away from me: a small figure from an old age, a little ridiculous and a little pitiful. But the older I get, the less distant he feels. Very often, people are not defeated by reality itself, but by their ability to patch up a story for themselves. They have clearly suffered a loss, taken a hit, or fallen behind, yet they quietly rewrite the plot in their own minds and cast themselves as the one who did not really lose.

That is the strongest and most dangerous part of the method of spiritual victory. It does not make a person truly stronger; it only makes them feel, for a moment, less embarrassed. It wraps failure in a nicer explanation, gives incompetence a respectable excuse, and packages retreat as the posture of "I never really cared anyway." On the surface, it looks like self-comfort. In reality, it is avoidance.

This has not disappeared at all today. You do badly on an exam, but instead of thinking through what you failed to understand, you say the grading system was flawed to begin with. You miss an opportunity, but instead of admitting you were not ready, you say you never wanted it anyway. You fail to do something well, but instead of reviewing where you went wrong, you first blame the environment, luck, or other people's lack of taste. The biggest benefit is that you feel better immediately. The biggest cost is that you stop right there.

Of course people need some self-comfort. Everyone has moments when they cannot hold it together. No one can be calm, clear-headed, and ready to review immediately after every failure. The problem is not comforting yourself. The problem is whether you turn that comfort into a long-term way of living. Taking a breath for a while is fine. What is truly frightening is that, after breathing in that story long enough, you begin to believe the words you invented.

The more I think about it, the more I feel that the method of spiritual victory ruins a person not by making them look ridiculous, but by taking away the desire to change. Once you can always explain away the past, it becomes very hard to force yourself to face the most painful question: where exactly am I lacking? That question hurts, but only after asking it can a person move forward. If you keep yourself upright only through stubborn words and self-anesthesia, it may look like you are protecting your dignity, but in fact you are slowly consuming the possibility of becoming stronger.

So what deserves the most vigilance in Ah Q is not that he is "funny", but that he resembles a common part of human nature. When people are hurt, they want to save face. When they fail, they first want to preserve their self-image. That is normal. But if a person is always willing to defend only that small inner dignity and never willing to face the loss in reality, then in the end that dignity is all they have left.

A truly strong person is not someone who never feels pain, never loses face, and never fails. A truly strong person is someone who does not rush to excuse themselves after failure. You can be sad. You can feel unwilling. You can even give yourself a little time to recover. But in the end, you still have to come back and honestly admit: this time, I did not do well. Only after that can improvement, growth, and reversal begin.