Original · Unique Research · 2026-08-22
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, five insight sections, closing reflection, and the full interview transcript. All named companies, products, and people are preserved. Pricing figures, product claims, and technical assertions are as reported by the source and interviewee, not independently verified for this rendition.
Original · Unique Research · 2026-08-22 20:00 Shanghai
Unplug the network cable — is your AI still working?
AI Industry Observer
After DeepSeek raised peak prices fivefold, someone stuffed an AI employee into your computer
"Everyone has their own algorithm for the water tap vs. well account. But at least starting today, you have one more option: bring an AI employee home to live."
First, two numbers for you.
First, DeepSeek officially adjusted pricing; peak-time Token prices are five times what they were. Second is me: ChatGPT Pro at $200 a month, $2,400 a year — over 17,000 RMB, enough to buy a new computer that runs local models.
I half-joked in the livestream: at this rate, will we need to set alarms to use AI during off-peak hours?
After the laugh, a bittersweet feeling. Cloud large models are like a water faucet installed outside your computer; water flow and water bill depend entirely on someone else's mood. Unplug the network cable, and there's no water.
At Wu Xiaobo Channel's "AI Future Talk" livestream, I brought on someone who gave another answer: Davis (Pang Dakui / 逄大嵬), founder of Yuankong AI (元空AI). His answer is direct: don't pin your hopes on the water faucet; dig the well in your own home. Their new product Yuankong AI Work packs the model and Agent together into your computer — works offline, data never leaves your door.
How to verify offline? Davis's method is primitive: unplug the network cable, turn off Wi-Fi, see if the task still runs.
Sounds like mysticism? I was skeptical at first too. But after two hours of conversation, I think this is worth discussing seriously.
A Dozen People, Three Leaps
Davis almost never does livestreams; this was his first. But their product you've probably scrolled past: ChatExcel.
Launched March 2023. What does that mean? GPT-4 had just come out, Doubao didn't exist yet, Kimi didn't exist yet. The interface would look familiar today: chat box on the right, Excel on the left, type a sentence and it processes the spreadsheet.
But what everyone takes for granted today, no one believed back then. Davis says, all the way until H2 2025, every time he had to answer the same question: what's the difference between you and WPS? And Office?
Why pick Excel? The two technical co-founders are Peking University PhDs; before GPT went viral, they were already researching using AI to process data. There's another rarely-noticed reason: Excel is the lowest fault-tolerance field in AI applications, bar none.
AI generates an image slightly off? Regenerate one. Copy isn't smooth? Revise it. But a decimal point one position off in a report is an accident. Worse, once it's wrong once, users never trust you again. That's why you'll see text-to-image and AI short dramas everywhere, but data products appeared latest.
This January, the "crayfish" OpenClaw just started heating up on X; the same night their team caught the signal, pulled the code, researched it, and had it running by the next day. When Davis posted to Moments, a bunch of people were still asking: what's a crayfish?
They originally wanted to grab the first batch of nationwide mini-program versions, but got stuck on one thing: registration filing. Filing came through and it hit Spring Festival, pushing them back. Davis judged that big companies would follow within a month; result, two weeks, they all came.
Then came Yuankong AI Work, launched in June, but the client name is still ChatExcel. I asked him why not rename; the answer is practical: millions of veteran users came for "chat to process Excel," still growing daily; renaming would raise switching costs.
A dozen people, three years, from ChatExcel to Claw to Work. If you ask me what this shows, my answer is: this team's reaction speed to technology waves is measured in days.
Small Models Aren't Castrated Versions
At this point, we must face the most piercing question.
After DeepSeek went viral last year, everyone was installing models locally, often 100+ GB. I actually know people who did, got excited for a while, then asked me: and then? What do I do with it?
This is also the question I threw at Davis. His answer is vivid: you only put the brain in the computer; the hands and feet are still outside.
A truly workable local AI needs far more than a model. You say "analyze 100GB of file data on my computer" — this job might run an hour. In between it calls tools, self-corrects, resumes from checkpoints — this is Agent Runtime. Many products crash on long tasks; that's why.
Also memory. You chat a hundred rounds with a regular Bot, and it forgets the beginning entirely, because context length has a ceiling. A truly workable Agent should remember that you put profit margin before revenue in weekly reports, and pick it up next week. Yuankong's approach is a bit geeky: it stores memory in layers like a Git repository, with lossless compression — otherwise memory files balloon infinitely.
Hands and feet installed, then talk about the brain. Cloud models are often trillions of parameters; locally you can only fit 8B, 27B, 35B small guys — is capability keeping up?
Davis's solution is making their own edge models, doing post-training: not chasing an all-around generalist, but targeted training on office and data scenarios. You ask it to write AI short drama scripts, it really can't; but writing documents, doing data analysis, post-trained 27B can be faster and more accurate, with lower Token consumption. Healthcare, scientific research — scenarios where "data absolutely cannot go to cloud" — these are the hard bones they bit off first.
"He also dropped a hard line: if a problem can be solved for 200K RMB, why must you spend 2 million? Enterprises to run a 'full-blood' private large model, sinking in hundreds of thousands to millions in hardware. Actually a department-level scenario, with a 27B/35B solution for 100-200K, effect is close enough."
I buy this. Many companies buying large models are like buying a tank for commuting — parameters are sexy, fuel consumption is bony.
Prompts No Longer Need Learning
There was a moment in the livestream that stuck with me.
I asked Davis: your website used to have so many prompt tutorials; do users still ask how to write prompts? He says people still ask, but his answer has changed: ask anything, use anything.
What does this mean? When a model is smart enough, how you ask matters less. Davis's judgment is more aggressive: in many work scenarios, model capability already exceeds 99% of people. At this point what really separates people is whether you know what problem you actually want to solve.
In plain terms: How is no longer valuable; Why and What are. Pick the wrong goal, and no matter how beautifully AI executes, it's high-efficiency walking down the wrong road.
So what should working people learn now? His advice is too simple to sound like an AI company founder's: use every good product on the market, and use paid ones. Only by having used good ones do you know where AI's ceiling is right now.
I'll add my own observation: this is quietly changing how companies organize. One person who can set AI goals and verify results is replacing what used to be a small team.
Memory Is the Real Moat
If you only remember one point from this article, I hope it's this: the gap between people using AI in the future will largely be the gap in memory.
Same Agent, you use it three months, constantly telling it this is right and that's wrong, it gets smoother and smoother. Another person just installed it today, thinks it's so-so. You're already using two different species.
That's just personal memory. One level up is organizational memory, worth more. Davis gave an example: someone in the company defines a great set of metrics — sales rate, profit margin, collection rate, how to rank and view — this used to grow only in veterans' heads. Depositing excellent employees' memory and copying it to the company's Agents — this is worth far more than buying a more expensive model.
The difference between excellent and average employees has never been about obedience, but experience and paths. Now, these can be copied for the first time.
The question comes with it: how does memory migrate? What if you change computers? Who in the org can see whose memory? Davis says they're working on it; permissions and sharing are the next battlefield. I believe it, because once memory deposits, migration cost itself is a new moat.
Disposable Software
Talking about the future, Davis dropped a word I've been turning over all night: disposable software.
His judgment: two or three years from now, many internal systems and processes in enterprises won't need to be固化 as long-term software. You state the result you want clearly, and AI runs the entire middle process itself, use and leave. Today enterprises buy software essentially paying for the "middle process"; when AI black-boxes the process, you only pay for results.
He's a software veteran himself; this judgment has a source. Traditional software engineering is a long translation chain: business translates requirements to product managers, product managers translate to docs and prototypes, programmers translate to code. Every handoff has loss and cost. Now end users talk directly to models and get results directly; the middle translation layer is being flattened.
I pushed him: then what is today's Work product? His metaphor I like: today's AI applications are like lightbulbs when electricity was first invented. Everyone's competing whose bulb is brighter, smaller, more uniquely shaped. But the bulb is just the beginning; what truly changes the world is electricity entering every machine.
Looking at the endgame, the logic is hard: global GPU and cloud compute are finite, but devices are infinite. Everyone has phones, computers; factories have machines; labs have instruments. Models get smaller, packed into every device, all devices gain some intelligence, and the world is completely different.
Closing
At the end of the livestream, I asked Davis what entrepreneurship feels like. He said he's a serial entrepreneur; entrepreneurship is a lifestyle, and he enjoys it.
I remembered a line he said in between: AI should make human life happier, not more anxious. We sleep, it works.
Back to the math. Cloud $200/month subscription, peak tokens fivefold higher; local starting at 49 RMB/month, a new 32GB RAM computer runs it, offline but not off work.
Everyone has their own algorithm for the water tap vs. well account. But at least starting today, you have one more option: bring an AI employee home to live.
More Conversation Details
Guest: Davis Pang Dakui, Founder of Yuankong AI
Host: Wu Wei, Founder of Unique Research
Why Did AI Move From Chat to Work?
Wu Wei: Now many people's understanding of AI is no longer just chat. It now has scheduled tasks, has Skills, can complete complex or repetitive work for you — it's gone from Chat to Agent, to Work. Once it enters Work, it needs to control computers and browsers, read and process documents, and data security issues come with it.
I want to ask the most direct question first: today's Work Agent, compared to what everyone used before — Doubao, DeepSeek, Kimi, ChatGPT, or pure chat products — what's the actual difference? Why is it appearing now?
Davis Pang: Actually, looking at our own product iteration can answer this.
Yuankong AI Work iterated all the way from ChatExcel. ChatExcel launched in 2023, one of the very first products using AI to process Excel. From day one we believed AI products shouldn't just be simple chat; they should solve data problems and complex task problems. So at the time we said one line: ChatExcel — chat to process Excel.
The product form back then, looking today, is actually very familiar: chat dialog on the right, Excel preview and editing on the left. But all the way until H2 2025, every time I still had to answer one question: what's the difference between you and WPS? And Office Excel? Because in 2023, many users simply didn't believe "chatting can edit Excel."
From 2023 to 2024, 2025, to Agent, Multi-Agent, today's Work — underlying models are upgrading, Agent processing chains are changing, but the interaction method hasn't been completely overturned. What truly changed is: AI went from answering you to starting to execute for you.
Wu Wei: So why did you enter through Excel? Excel looks office-like, but actually very hard.
Davis Pang: Right, from the start we didn't treat Excel as just Office, but as a data format.
Excel is hard because it's not plain text. Cell order, parent-child links, left-right relationships are themselves spatial relationships, carrying information. A1, A2, B2, A21 — these positions themselves carry information. So why are Excel and data hard? Because they can't be derived purely from Tokens.
More importantly, data products have near-zero fault tolerance. AI image wrong, regenerate; AI text unsatisfactory, revise. But a decimal point off by one in Excel and the result is unusable. Worse, once you miss once, users may never trust you again. Without trust, they won't use it again.
Wu Wei: Then as foundation models get stronger, what's your biggest feeling?
Davis Pang: If I had to say when I felt it most, I'd say 2025. After foundation model capability rose, our capability rose with it.
Actually when we started in 2023, we were already using Coding capability. Many people asked: did you replace some number of VBA functions? Actually no. From day one we believed in the model's Coding capability, using code to process tasks.
But the model alone isn't enough. Why did so many Work products appear this year? Essentially, foundation model capability and Agent framework capability rose at the same time. Without frameworks like OpenClaw, Hermes Agent, Work wouldn't have broken out so fast. Today you need scaffolding engineering to truly organize the model and let it execute tasks continuously.
So I think Work's breakout isn't some model suddenly getting smarter; it's "model + Agent framework" both crossing a critical point.
OpenClaw's Heat Has Passed, But It Left Behind Agent Basics
Wu Wei: OpenClaw suddenly went viral early this year; many people spent Spring Festival installing "crayfish." Later heat came down fast. But my own feeling is, it may not have become a long-term high-frequency product, but it left very important framework ideas for later Work Agents, like Memory, heartbeat mechanism, scheduled tasks.
Why did you follow so fast?
Davis Pang: Because we've always been doing Agents and always been doing foundation models. The day OpenClaw just started heating up on X, that night we caught the signal, pulled the code and studied. The whole team was actually quite surprised, feeling the framework was very powerful. The next day we installed and ran it.
At the time many people didn't know what "crayfish" meant. But we quickly found it especially suitable for solving problems previously unsolvable. Its heartbeat mechanism, its Memory handling, were both worth learning.
So after Spring Festival, we quickly made a mini-program version of Yuankong AI Claw. Later this product didn't get further heavy investment because we migrated many mechanisms into the new Work platform.
I've always felt OpenClaw is a bit like DeepSeek: it educated the market very well. It let end users, enterprise users, even big companies, for the first time very concretely realize that an Agent isn't a chat box; it can have memory, can schedule itself, can execute continuously.
Wu Wei: Then when did Yuankong AI Work start?
Davis Pang: After finishing Claw, we found user needs obviously got more complex.
We weren't just processing Excel anymore. Expanding from Excel to Word, databases, complex business scenarios, multi-file and cross-file. In this process, old frameworks started limiting you, so after Claw, we actually already started Work.
The product came out around end of May, June. Everyone remembers OpenClaw as February-March everyone installing, April heat declining, and in May we were already moving to the next-generation Work product.
Wu Wei: So one very important thing for AI startups is indeed constantly changing.
Davis Pang: I think whether big companies or small, everyone's answering the same question: how to run forward fast.
We small companies ship versions weekly; big companies too, models and products both iterate at high frequency. At least in this AI wave, the first wave is still tech-driven, because the technical ceiling of large models hasn't been reached. Model, Agent, device is one whole tech combo, and this combo is still changing fast.
Why Pack the Model "Into the Computer"? Edge Solves More Than Security
Wu Wei: You now increasingly emphasize "edge." In simple terms, what is an edge model?
Davis Pang: You can understand it as putting a very smart brain into the local computer.
Cloud models are easy to understand: model runs on servers, your computer is just an entrance. Unplug the network cable, turn off Wi-Fi, it can't Work. Edge models put some intelligence capability truly into computers, workstations, compute boxes. Many companies require no internet, and you can still locally do Excel, PPT, write documents, organize folders, do data analysis.
We now position ourselves as an edge intelligence company. Because we're increasingly sure that eventually models and Agents will enter all kinds of devices: AI PCs, workstations, research equipment, even more IoT devices. All devices will have some intelligence; this intelligence doesn't have to be a hundred-B model, it might just be an 8B, 27B, 35B.
Wu Wei: Can ordinary computers actually run it now?
Davis Pang: This question needs to be on a timeline. Before, to run offline, you often needed large compute servers; now 2026's new AI PCs and AI Stations are leaning this way.
Simplest metric: memory at least 32GB; ideally 64GB+, especially unified memory and larger shared VRAM devices will be smoother. Of course, a 3,000 RMB old computer still can't run it. But many new devices this year are slowly gaining this capability.
And edge isn't just one form. One is single machine, a laptop, a workstation; another is LAN, one compute box driving ten ordinary computers; another is edge-cloud hybrid, locally installing 7B, 8B, 9B small models, simple tasks locally, complex tasks compensated by cloud compute.
Wu Wei: But here's the problem: cloud models are often hundreds of billions to trillions of parameters; locally maybe only 8B, 27B, 35B. How do small models compare to big ones?
Davis Pang: Today's large model vendors pursue maximum parameter count, essentially making a super-smart "generalist." Edge is naturally limited by physical space; can't fit all parameters.
So the edge approach isn't competing with the cloud on who's more "all-around," but doing post-training, targeted training into a "specialist."
For example, our edge model, if asked to generate AI short dramas, definitely isn't our strength; but paired with Yuankong AI Work to write documents and do data analysis, we can be faster and more accurate, with lower Token consumption. I think future edge AI will increasingly look like this: video models, office models, research models, medical models, all different. Generalists go in the cloud; specialists go into devices.
Wu Wei: So models aren't always better bigger.
Davis Pang: Right. Besides big models, there are a few practical problems: security, cost, speed, stability.
Cloud peak times might be more expensive, also slower. Many serious business scenarios can't accept latency. Edge isn't just "data security" simple; it also means tasks aren't affected by cloud traffic and network environment; plus after targeted post-training, in specific scenarios it may also be more accurate.
Post-Training Is Becoming the Real Competitive Point: Not "Teaching Knowledge," But Teaching How to Work
Wu Wei: You repeatedly mention "post-training." This word is getting more important; can you explain it plainly?
Davis Pang: Before 2025, many teams mainly did pre-training. What's pre-training? Let the model learn knowledge through lots of Tokens, with parameter count rolling up from tens of B.
But the pre-training layer, big companies and open-source models have already done very deeply. Now more and more people put focus on post-training. Because after base intelligence reaches a stage, what you need to answer is: how to strengthen a certain direction's capability.
Before it was more like "feeding it books"; now it's increasingly "teaching it to work."
Wu Wei: How specifically?
Davis Pang: Like office scenarios: you input a file, ask it to generate a document. What you input first, what Agent returns, what Tool it calls, what Skill — this is a task trajectory.
Same task thrown at a cloud model, you'll find the path it takes each time might differ. What trajectory training does is tell the model: when seeing this kind of task, what's the better path. Let it learn the "best practice."
Today it's no longer just traditional data annotation. Especially after models move from digital world to physical world, you need real-environment data, task trajectories, device interaction data. Simply annotating "what is this" isn't enough; you have to tell it "what to do in this environment."
Wu Wei: Does this also mean Prompt engineering will weaken more and more? Before people learned prompts, then learned Workflow, now people say Agents plan themselves.
Davis Pang: Right. Before our website had lots of prompt tutorials. Later models got stronger; when users ask me "how to write better prompts," my answer has become: ask anything, use anything.
I think at this point in time, there's no need to obsess over one tool. More important is learning how to frame the question well, how to summarize your own problems. Critical thinking is more important.
"How to do it" — How — will get cheaper and cheaper; Why and What will get more and more important.
Because if you pick the wrong goal, no matter how fast you go afterward, it's meaningless.
Wu Wei: I agree with this judgment. As models get stronger, what's truly scarce may not be operating tools, but knowing what problem you actually want to solve.
Davis Pang: Right. I'd even say, in many work scenarios, model capability already exceeds most people. What's truly important today is, can you frame the question well, know what your problem is.
A Local AI That Truly "Works" Depends Not on One Model, But on a Complete Systems Engineering
Wu Wei: I think many people when doing local deployment before had a problem: model installed, and then?
When DeepSeek first came out, people around me excitedly deployed locally, installed it and found it was just chat. Why is it different now? Why do you say "it's starting to work"?
Davis Pang: This is also a development process. Only making the model small and putting it in the computer easily becomes showing off; only installing one agent but still connected to cloud behind, also not hard.
What's truly hard is: putting model and Agent together into the computer, letting it work locally. You can't only have a brain, no hands and feet.
So we emphasize a modifier: complex tasks, real productivity.
Wu Wei: What specifically are these "hands and feet"?
Davis Pang: First keyword is Agent Runtime.
For example, asking it to analyze 100GB of files on the computer — this can't be done in a minute; maybe ten minutes, an hour, even ten hours. Runtime means, as long as the computer doesn't lose power or shut down, the task can't be interrupted. It must self-correct, find tools, install tools, execute continuously.
Many products look good on simple tasks, but easily go wrong on long tasks. But real work is often long tasks, and cross-file, cross-software: moving things from one folder to another, merging different format files, then generating a new result, continuing to process.
Second keyword is Memory.
Many Bots, after you chat a hundred rounds, the beginning is already forgotten. Because context length naturally has a ceiling. Once truly working, you want it to remember your habits. Like you write weekly reports with profit margin first, revenue second; ask again next week, it can still pull out last week's metric system.
At this point you need long-term memory, even permanent memory.
Wu Wei: But doesn't long-term memory just keep growing?
Davis Pang: So behind it involves compression, indexing, layered storage. On one hand you don't want to lose information, on the other hand it can't balloon infinitely.
We now use something like a Git repository approach, storing different memory files in layers. This isn't necessarily the only optimal solution, but at least it's one route we've found so far.
Further ahead there's something even more important: how does memory migrate? You change computers, can the accumulated context come with you? An excellent employee forms a great metric system — this is actually no longer personal memory, but organizational memory. Can you deposit excellent employees' Memory and then copy it into the organization? I think this will be very valuable.
Wu Wei: This is actually a bit like past knowledge bases, RAG, but one step further. Because it's not "storing knowledge," but storing how this person, this organization does things.
Davis Pang: Right. Clearly this thing has been done, why do it again? A truly good Agent isn't just obedient; it should increasingly have experience.
Another practical challenge is the environment. Once you truly put an Agent locally, you're facing not a unified cloud Linux environment, but various Windows, Ubuntu, HarmonyOS, UOS, Kylin, various versions, various local software. Users' real environments are very complex.
So today doing Work Agent is no longer a single-point product capability; it's complete systems engineering: model, Runtime, Memory, Tool, Skill, operating system, hardware adaptation — missing one and real clients may not be able to use it.
What Edge Enters First May Be Those "Data Absolutely Cannot Leave" Scenarios
Wu Wei: What kind of users most need edge Work Agents now?
Davis Pang: A very typical scenario is healthcare and research data analysis.
Users might give you one file package at a time, inside Excel, PDF, images, instrument-exported data, even needing to connect to internal databases. This experimental data, medical data — in many cases absolutely cannot go to cloud.
So what to do? Process on the edge. It must decompress, recognize different file formats, aggregate data from different sources, then analyze. You'll find this itself is a multi-file, multi-step long task.
What researchers most directly want in the end is an analysis report; further ahead, next-step experiment recommendations based on analysis results. Like which metric is wrong, is the pH off, should reagents be adjusted. Further ahead is AI truly entering the experimental closed loop.
Wu Wei: This is also why you cut into AI for Science?
Davis Pang: Actually very natural. We've been doing data analysis since 2023; many researchers were already using our products for data and reports; our team itself has Peking University research backgrounds in scientific intelligence and AI for Science.
Why is research suited to AI? Because it itself is a loop: read lots of papers, propose hypotheses, design experiments, validate, produce new hypotheses. Why is new drug development slow? Because it takes thousands of experiments. AI's most direct value is accelerating this loop of literature, hypothesis, experiment design, validation.
We recently open-sourced OpenAI4S (Open AI for Scientist), with over 30 research Skills. You can understand it as a virtual research lab bench: proteins, small molecules, materials — different directions have different environments, different Agents, different Skills.
The underlying tech route is actually consistent with Work: edge model, on top of Runtime, Memory, Agent and Skill. Office is one scenario; research is another.
Wu Wei: Same inside enterprises. Individual users may not be so sensitive about data security, but at enterprises, what data each computer can Access, what environments can go online, all have strict requirements.
Davis Pang: Right. Many manufacturing enterprise factories can't go online; government, finance, healthcare, research institutes, labs also have strong isolation environments. I actually feel that by this year, AI truly has a chance to enter these previously inaccessible scenarios.
Also cost. Enterprises don't need everyone to buy a high-config AI PC; they can put a compute center on the LAN, five people, ten people using it together. Ordinary employees still use their old computers, getting AI capability through LAN.
If an enterprise already has a private model, they can also just connect Agent and Work; if not, we can also deliver model plus Agent as a whole.
I think the key point here is: there's no need to stack a "full-blood large model" for all capabilities. If a 27B, 35B model after post-training can already solve problems in office, research, these specific scenarios, why must you spend ten times the money?
Two or Three Years Out, Software May Become "Disposable"; What People Should Truly Keep Is Problem-Judgment Ability
Wu Wei: Let's take the topic further. Now AI employees are moving into computers; in the future it may enter more devices. So where are humans? How do people collaborate with these smart computers and devices?
Davis Pang: Essentially asking what the future productivity relationship is.
I think today everyone making models, making various Agents, is a bit like when electricity first appeared, everyone first made different "lightbulbs": this one a bit brighter, that one a bit smaller. Many Work products today are still improving existing tasks.
The next stage worth watching more is: what new problems can AI actually solve that were previously unsolvable?
Office actually already has many things AI can do itself, like after people leave the computer, let it organize files, organize data. Harder are embodied and research — letting robots work in the physical world, letting AI truly enter the experiment site. Because these fields have less reference data, so they need more time, but I think within a year or two, some industry scenarios will be very obvious.
Wu Wei: Then what will enterprise software look like in two or three years?
Davis Pang: I want to use a word: "disposable."
Today enterprises have lots of digital systems and office processes; in the future many things may not need to be long-term fixed as software. You just tell it the work you want, the result you want, the task goal; the middle process is all closed-loop, gradually becoming a black box.
Software will of course still exist, but how software is produced, its form, and how employees use software will definitely change.
Past software engineering essentially had many roles doing translation: business people say requirements, requirement analysts translate into requirement docs, product managers draw prototypes, programmers translate into code, finally handed back to business people to use. In the middle is a long translation chain.
After AI appeared, this chain started being compressed. End users can directly converse with models, directly produce the results they want. The biggest change is from multi-role, long-process translation of requirements, gradually becoming directly getting results.
Wu Wei: So what should people keep in the enterprise?
Davis Pang: I think it can split into employees and bosses two dimensions.
For employees, most important is curiosity and continuous learning ability. So many products on the market, go use them, don't just watch others talk. Different products have their own characteristics; nothing is universal. Also, I've always suggested people be willing to use good paid products. Only by using truly good products do you know where AI's ceiling is right now.
But more importantly, don't put all your energy into learning one tool. Tools keep changing. What's truly long-term valuable is Why and What: what you want to do, why you do it, whether you can judge the problem.
For bosses, think one layer more. By 2026, AI is no longer just efficiency change, but starting to be result change. Everyone knows about efficiency improvement; the next question to truly ask is: will AI bring this industry new results, new opportunities? What couldn't your company do before, can it do now?
Wu Wei: So in a sense, as execution gets cheaper, human value moves more toward "choosing problems" and "making judgments."
Davis Pang: Right. Technology, products, tools all change; in the end you return to one origin: what problem do you actually want to solve?
From ChatExcel, Yuankong AI Work, what we're doing is still digital world tasks; another thing is going to the physical world. Our direction is letting edge models enter more devices.
Cloud GPU and compute are finite in the end, but today everyone has phones, computers, and lots of devices. Device count is almost infinitely expanding. The key question is how to pack models into various devices.
When models get smaller, enter different devices, all devices have some intelligence, this world will be completely different.