In 2026, the most core concept in the AI field at present is this.
God Translation Bureau is a compilation team under 36Kr, focusing on fields such as technology, business, workplace, and life, and mainly introducing new technologies, new ideas, and new trends from abroad.
Editor's note: AI is shifting from the "dialogue box" to the "self - evolution cycle". When 99% of knowledge work is proven to be just redundant scaffolding, the ability to clearly define and verify "intentions" will become the dividing line between top elites and mediocrities. This article is from a compilation.
After about a week of thinking and participating in the RSA Conference during this period, I believe that several core AI concepts will change the status quo in a profound way that outweighs everything else.
Autonomous component optimization
Transformation to "intention - based engineering"
From ambiguity to transparency
Realizing that the vast majority of work is just "scaffolding"
Dissemination of professional knowledge to public knowledge
1. Autonomous component optimization
This concept is closely related to concepts such as the transformation from the "status quo to the ideal state", algorithmization, and general verifiability.
What makes this concept truly come to life and accessible is Karpathy's Autoresearch project.
His project focuses on AI research itself, that is, "automatically researching the'research process' in AI research". This means automatically handling those tedious and time - consuming chores, such as adjusting model parameters, fine - tuning fragile environments, and various option combinations.
In his release version, you just need to write down some ideas in the `PROGRAM.md` file, and the system will automatically handle all those messy chores. You can just go to sleep, and the system will use machine learning optimization techniques to produce better results than what you have at hand.
Expanding Autoresearch
Now, there has emerged "Autoresearch for X", which means it is becoming a paradigm and a movement, essentially turning into a general tool.
It has sparked a lot of people's thinking:
Can I apply a similar method to my current work?
This is truly extraordinary.
Combining Autoresearch with my research
I have always been concerned about the overall concept of "general verifiability", or "general hill - climbing algorithm (optimization)". This again draws on what Karpathy said a long time ago in "Software 2.0" and a recent tweet of his, in which he mentioned that the future of software lies in everything being verifiable.
Therefore, what I do in the PAI algorithm is to try to break down all content into "ideal - state criteria". These criteria essentially build an ideal blueprint for the results I want.
Based on this, the algorithm can continuously optimize (climb the hill) towards this goal.
Everything can be evaluated (Evals)
Related to this is the concept of "everything can be evaluated". This is very similar to my general verifiability or general hill - climbing optimization. The core idea is that everything we do becomes measurable, and more importantly - improvable.
What makes "everything can be evaluated" possible is transparency.
General optimization cycle
This will become the standard operating mode for every company, organization, government, and individual. This cycle is as follows:
You draw all the tasks you want to complete in a goal - oriented structure (mission, goals, workflows, SOP). These workflows are executed by AI agents. All content will be widely recorded - including output content, dialogue processes, final results, and quality conditions. Whenever errors, failures, or quality problems are captured in the logs, they will be collected at the entity's "problem collection point".
This collection point is the source of nutrients for the self - optimization algorithm. Agents extract problems from here, create execution tasks similar to Autoresearch to troubleshoot faults, try solutions, verify through evaluation, and optimize. Once a repair solution is found, they will update the SOP to ensure that the problem does not occur again. Then, the cycle repeats.
Everything has such a life cycle. Define the goal. Let the agent execute. Record everything. Collect failures. Improve autonomously. Update the SOP. Repeat continuously - and each time faster than the previous one.
2. Transformation to "intention - based engineering"
The real power of AI lies in bridging the gap from the "status quo" to the "ideal state". Define the status quo, define the goal, and then let AI narrow the gap between the status quo and the goal. The concept is simple, but there is a prerequisite before everything works: you must be able to clearly express what you really want. It turns out that this is extremely difficult. If you cannot describe what is "good", then no amount of tools can help you.
This is a huge problem for companies. If you ask a CEO what an ideal security solution looks like, you will only get a vague description. If you ask a team leader what "completion" of a project means, you will get a piece of text, and three people will have three different interpretations. This "expression gap" exists not only between experts and AI but also between leaders and their organizations. Most companies cannot clearly describe what they are doing, let alone break it down into measurable or optimizable components.
What I am building inside the algorithm is precisely this ability - a method of reverse - engineering any request into discrete, testable "ideal - state criteria". Each criterion is eight to twelve words long, and the judgment standard is a black - and - white "pass/fail". Once you have these, you can optimize (climb the hill), evaluate, and achieve automated improvement. But the starting point of all this is the ability to accurately say what you want. This is the new engineering skill - not writing code, nor writing prompts, but expressing intentions clearly enough to make them verifiable.
3. From ambiguity to transparency
Companies have never really seen what is going on inside themselves. What is the actual cost of the process? How long does it actually take? What is the quality of the output? Who is doing the core work, and who is doing the "scaffolding" work outside the core work?
Most organizations operate based on "feelings" and spreadsheets. And AI makes everything visible. The actual work, cost, and quality - all these become measurable in a way that was simply impossible before. Once you see them, you can improve them. This applies to enterprises, governments, small teams of three people - any object you focus on.
And the first thing that transparency reveals is: how much work is actually not core work.
4. The vast majority of work is just "scaffolding"
AI is revealing a fact: 75% to 99% of knowledge - based work is actually "scaffolding" management overhead. In fields such as security testing, development, and consulting, most of the time is spent on maintaining tools, workflows, templates, and knowledge bases. Real in - depth thinking only accounts for a very small proportion, done by a very small number of people in very little time.
AI can completely crush this "scaffolding" part of the work. Agent Skills have proven that you can package all context, methodologies, and tools into a single skill, and AI's execution effect can not only match but even surpass most professionals. The core work is actually not difficult; the difficult part is maintaining the supporting scaffolding.
5. Dissemination of professional knowledge to public knowledge
There is an "expression gap" between the knowledge mastered by experts and the recorded knowledge. Most professional knowledge exists in people's minds. For example, Cliff, who is 62 years old, knows how everything works, but he has never recorded it. When Cliff retires, that knowledge will go with him.
What is happening now is that professional knowledge is spreading from the brain to skills, SOPs, context files, and open - source projects. Once this knowledge is captured, it will never be lost. It's like peeing in a pool (irreversible). Every released skill, every recorded process, every captured expert report - will permanently enter the collective knowledge base. It makes every AI instance smarter. Not one, but all, getting stronger at the same time.
There is already a large industry in this area dedicated to extracting expert knowledge into models, and most people are unaware of it. This is a one - way ratchet effect. It takes humans 20 to 30 years to develop in - depth professional knowledge in a single field. They will forget, retire, and leave their jobs. While AI can instantly absorb all the captured professional knowledge, never forget it, and can be infinitely replicated. The gap between humans and AI in the speed of professional knowledge accumulation is widening.
Impacts
Autonomous optimization changes the speed of everything
The progress speed in many fields is about to accelerate in an unimaginable way. When you can define what is "good" and measure and automatically iterate based on it, work that used to take months of manual adjustment can now be completed overnight. Autoresearch has proven this in machine - learning research. But this also applies to security solutions, consulting results, content pipelines, and recruitment processes. Anything with a definable "ideal state" will become autonomously optimizable.
Every entity - company, government, team, individual - will run the same cycle: define the goal, let the agent execute, record everything, collect failures, improve autonomously, and update the SOP. Entities that adopt this model first will progress at an extremely fast speed, so that those who lag behind will not be able to compete with them at all.
Intention becomes the bottleneck
💡 Those who can clearly express their intentions will have a huge advantage.
The new scarce skill is not writing code or writing prompts, but the ability to say what you really want. And this must be a high - quality intention. The quality of creativity is always the most important, but second is the ability to express creativity, define it as an actual goal, and make the whole company operate around it. Most leaders cannot do this, and most companies cannot either. Companies that solve this problem first will be able to aim all optimization tools at the real goal, while others are still talking about OKRs.
Everything will become transparent
We are about to witness the world shift from a fuzzy "feeling" to transparent and optimizable components. There will be fewer and fewer hiding places for swindlers and industry "gatekeepers".
This also makes competition in selling products or services more difficult. Because agents will first ask: "What are your metrics?" They don't care about marketing copy or customer reviews, but about actual, verifiable performance data. If you don't have this data, you will lose to competitors who do.
🔮 I call this "from magic to Excel".
Scaffolding is commoditized
The so - called "mystery" in certain fields and professions will be uncovered, revealing their nature as scaffolding - it's just that most people didn't understand it before. For example, how to set up a specific development environment and keep it running until code is written. The same is true for the legal, consulting, and other high - paying industries.
⚖️ This will never be a 100% completed process, but as it approaches 95% or even 99.99%, the minor differences are no longer important overall. However, this will provide a competitive advantage for those who have unique insights or knowledge that the model has.
Expert knowledge becomes public infrastructure
The knowledge that was only mastered by experts in the past will soon be accessible to everyone - most importantly, to AI. The advantage of people having 50 years of experience in a certain field will no longer exist. Because this content will be extracted and aggregated by themselves or colleagues around the world.
Summary and implications
The craziest thing about all this is that these will interact and amplify each other.
We can not only improve all these different components, but also the improvement speed itself will be increased.
Every company, every government, every organization will ultimately converge to the same cycle: define the goal, let the agent execute, record everything, collect failures, and then let the system self - optimize. Entities that achieve this first will generate a compound interest effect at an extremely fast speed, leaving everyone else far behind.
Among all these ideas, this is the most core implication.
I really can't put into words how crazy all this is going to get.
Translator: boxi.