Manchi | 晚点聊 LateTalk Summary

Podcast: 晚点聊 LateTalk (LatePost) Guest: Wu Minghui (吴明辉), Founder of Miner (明略科技) Host: Manchi

1. Key Themes

The Death of Traditional SaaS and the Rise of "Agentic Service"

Wu argues that SaaS built on pure software licensing is structurally doomed. Software can be replicated almost instantly with AI coding tools, eliminating its value as a proprietary asset.

"For SaaS software companies, if you still insist on going the route of selling software licenses, I think it's basically a dead end — whether you're a large company or a small one." [00:03:18]

The replacement model is what he calls "Agentic Service" — selling outcomes and digital labor rather than software. This mirrors a Sequoia thesis he references:

"Some time ago, I believe it was Sequoia that wrote an article saying that in the future there will be many companies wearing the costume of a service provider that are actually software companies. What we're doing with Agentic Service is exactly that." [00:22:47]

OpenClaude (Claude) as a Paradigm Shift: From Personal Assistant to Collective Intelligence

Wu sees OpenClaude (referred to as "龙虾"/Lobster throughout) as fundamentally different from prior agents because it solves continuous learning at the agent layer rather than the model layer, enabling persistent memory and collaborative intelligence.

"What's truly great about it is that it has formed a new memory pattern, enabling it to continuously understand a person's latest state, their latest cognition, and adapt to the work and life scenarios they're in." [00:07:38]

He identifies a critical gap that no one is solving: making the agent a team collaborator rather than a personal assistant:

"The most effective usage is not one Lobster serving one person, but one Lobster serving multiple people. Once the Lobster serves multiple people, it can produce collective intelligence." [00:10:04]

Scaling Out vs. Scaling Up: A Philosophical Bet Against Foundation Model Monopoly

Wu's deepest conviction is that the industry should pivot from making single models smarter ("scaling up") to building collective intelligence networks ("scaling out"). He frames this as both a technical thesis and a moral imperative.

"Foundation model training has reached an okay level today. What we should be doing next is Scaling Out — developing horizontally, forming collective learning, forming organizational wisdom. And this organizational wisdom should be protected within every individual, or within every small startup team." [00:01:23:29]

He uses the analogy of Homo sapiens defeating Neanderthals:

"Neanderthals had superior individual capabilities compared to individual humans. But humans can collaborate... I want to use Octo to make human collaboration better — humans and Lobsters collaborating better." [01:25:07]

2. Contrarian Perspectives

AI Will Hurt American SaaS Far More Than Chinese Software Companies

Most commentary focuses on AI disrupting Chinese companies. Wu argues the opposite — the American SaaS model, built on value-based pricing and high license fees, is far more exposed.

"AI's impact on American SaaS and software companies will be far, far greater than on China... In China, we've always had cost-oriented pricing. You're web-coding, I'm web-coding — who can take the margin? That's fine in China. But in America, it's finished. They used to charge enormous license fees, and now that's done." [00:24:14]

GUI Is Not Going Away — CLI is for Agents, GUI is for Humans

The prevailing view in Silicon Valley is that CLI/API will replace GUI as AI handles software interaction. Wu strongly disagrees:

"GUI will always exist. Humans need GUI; humans don't need command lines. Lobsters need command lines... Think about it: scrolling TikTok — isn't that GUI? Playing games — isn't that GUI?" [00:56:02]

He adds a security argument: if AI ever needed to be shut down, an all-CLI world would leave humans unable to operate their own software.

RPA Companies Are Essentially Finished

Wu makes a specific, non-obvious claim that RPA (Robotic Process Automation) is a dead category:

"RPA companies may be finished... RPA currently works by scripting fixed coordinates on a static GUI — click here, input there. But once that software changes its layout, the whole thing breaks. You'd have to rewrite it every day." [00:57:58]

This is particularly notable because RPA has been a high-growth enterprise software category.

Foundation Model Researchers Are Eating the "Last Piece of Meat" — And Then They're Done Too

Wu makes a sobering prediction about even the most elite AI workers:

"I have a friend who is a very senior executive at a major foundation model company. He told me: 'Minghui, I can see clearly that the work my team does — AI can already do it. I'm probably eating the last piece of meat. But eventually I'll be laid off too.'" [01:23:00]

This contradicts the widespread belief that AI researchers are insulated from AI disruption.

The Software Outsourcing/Custom Dev Industry Will Transform to Token-Based Billing

Wu predicts a fundamental restructuring of how software development services are priced — away from people-months toward token consumption plus management fees:

"I think custom software development will gradually shift to a Token management fee billing model. I'm a software outsourcing company, you tell me what you want built, I consume perhaps $1 million worth of tokens, and then charge an additional $200,000 management fee on top." [00:21:48]

3. Companies Identified

Anthropic AI safety and model company behind Claude. Wu highlights them as having an exceptional internal productivity model — Dario reportedly required the entire 500-person team to use Claude Code, creating a self-reinforcing productivity flywheel.

"Anthropic's team, at 500 people, was required by Dario to use CCC [Claude Code] across the board. You'll find Anthropic's production efficiency is extremely high — they can accomplish a huge amount in one or two months." [00:34:58]

OpenClaude (Open Source Project by Peter) The open-source agent framework Wu refers to as "Lobster." He considers it a paradigm-level breakthrough for persistent memory and collective intelligence, comparing its memory architecture to Google's research on continual learning.

"OpenClaude's memory structure is designed with extraordinary elegance — it has a daily diary layer, a memory layer, and a soul layer. These layers have different rates of adjustment, which closely resembles the continual learning parameter adjustment method from that Google paper." [00:08:07]

Meitu (美图) Photo and AI image company. Wu visited them and was struck by how undervalued they are given their profitability from token revenue.

"Meitu is a severely undervalued company. They had 900 million RMB in profit last year but only a 20 billion RMB market cap... Their real revenue is coming from token revenue. I think it's a great company." [01:33:42]

Overleaf Online collaborative LaTeX writing platform with 20 million users. Wu identifies it as a SaaS ripe for disruption from AI-native collaborative writing tools.

"Overleaf is an online paper-writing and collaborative writing software with 20 million users worldwide... But its collaboration features are extremely poor. If you look at my system, you'll see it looks similar in some ways — LaTeX on one side, rendered output on the other — but you can directly edit formulas and figures on the rendered output." [00:52:08]

Miner/Minglueh (明略科技) - Octopus (章鱼/Octo) Wu's own company. Their multi-agent coordination system "Octopus" allows multiple Claude instances to collaborate within an organization, building collective intelligence. Planned to be open-sourced.

"We're currently releasing our OSWorld GUI model benchmark results, and our Mano model — we're confident we can achieve near 100% accuracy on specific vertical software operation tasks, compared to the current benchmark leader at only 70%." [01:03:50]

4. People Identified

Peter (OpenClaude Creator) Individual developer who created the open-source OpenClaude agent framework. Wu identifies Peter as a visionary but notes he has become a bottleneck as the project scales.

"Peter himself originally built OpenClaude for his own use... Unfortunately, Peter has already become a massive bottleneck on the project — he no longer has the capacity to commit to the code submitted by the open-source community." [00:12:03]

Herbert Simon (赫尔伯特·希蒙) Nobel Prize and Turing Award winner. Wu cites him as the only person in history to win both, and his work on bounded rationality and organizational behavior as foundational to Wu's thinking on AI and organizations.

"Herbert Simon is truly the first person in human history to win both the Nobel Prize and the Turing Award simultaneously... His books — 'Administrative Behavior,' 'Organizations,' and 'The New Science of Management Decision' — are extraordinarily insightful." [01:49:22]

Wang Hong (王鸿) Chinese mathematician who solved the three-dimensional suspension problem and is being considered for the Fields Medal. Wu cites him as a model of what human mathematical "taste" looks like — he didn't even study math as his primary undergraduate major.

"Wang Hong didn't even participate in mathematical olympiads. He didn't even enter Peking University as a math student — he transferred from geophysics. I think this shows that liberal arts-style cultivation matters — training you to feel the beauty, to feel what you want." [02:14:50]

Mac Krieger (Instagram Co-founder, now Anthropic) Wu references Mac Krieger's observation that product design — deciding what a product should do and which features matter — is extremely hard to replace with AI, and that faster coding can create a false sense that product development is compressing in time.

"Mac Krieger shared that there's one thing that's very hard to replace: how do you design a product? What features should it have? He said that after webcoding became so advanced, it creates a misunderstanding — that development speed compressing means you can skip the design learning phase." [02:22:11]

5. Operating Insights

Humans Should Do Tasting, Agents Should Do Thinking — Redesign Your Workflow Around This

Wu's most actionable operating principle is a clean division of labor: humans provide "taste" (judgment, goals, context, preferences) while AI does "think" (deterministic reasoning and execution). He restructures his entire company around this.

"What I'm doing as CEO, many things I now do with it [OpenClaude]... Our finance team all have Claude Code. Our HR team many also use it... Any one-line work, give it to AI. We humans provide ideas, requirements, and explore context." [00:12:32]

Format Your Documents for Your Audience: PDFs for Humans, MD Files for Agents

A specific, immediately implementable workflow insight: Wu discovered a 10x compression ratio between PDF and Markdown that dramatically improves agent efficiency.

"Anything to be fed to the Lobster should be converted to an MD file, because there's a 10x difference — a 300-400KB PDF corresponds to a 30-40KB MD file. This means it occupies far less of the Lobster's context window, making it work faster." [00:54:34]

Build Your Own Productivity Tools First — It Becomes a Self-Reinforcing Flywheel

Wu explicitly frames this as his company's operating philosophy, citing Anthropic as proof of concept: companies that build their own AI-powered tools compound their productivity advantages.

"Your own productivity tools must be built by yourself. This tool can also be sold externally, forming a self-reinforcing loop. This is what I call E to the X — as an ancient Chinese saying goes: 'To do good work, first sharpen your tools.'" [00:34:58]

Token Budget Management is the New Headcount Management

Wu has implemented a company-wide token budget system that requires employees to justify token spending via a business plan — treating compute like capital allocation.

"Everyone must report their business plan for what they're using tokens for. When I went to renew my subscription, I saw I'd spent $10,000 in two weeks, and I thought: how is that so cheap? I believe in those two weeks I created 1 billion RMB in value for the company. I spent $10,000 — so what?" [01:14:12]

6. Overlooked Insights

The GUI VLA Model's Real Market Is Automated Software Testing, Not Task Automation

Wu makes a passing but enormously significant observation that is glossed over in most discussions of computer-use AI: the primary commercial value of GUI understanding models isn't autonomous browsing or task completion — it's closing the feedback loop for AI-written code.

"The number one use case for GUI VLA models is truly not automation of tasks themselves. The number one use case is automated software testing. When Claude Code writes code with a graphical interface, it needs to test that interface, find bugs, and loop back. That's what the multimodal understanding is actually for." [01:05:49]

This reframes the entire competitive dynamic: Anthropic's acquisition of a GUI-VLA startup, its consistent OSWorld leaderboard participation, and the apparent contradiction of "GUI is dying" while foundation model companies aggressively invest in GUI models — all make sense through this lens. The winner in AI-assisted software development will be the one who can close the code → GUI → test → debug loop autonomously. This is a massive, under-discussed moat.

"Scaling Out" via Multi-Agent MOA Networks May Be a Credible Alternative Path to AGI-Level Performance — Without Foundation Model Scale

Wu briefly describes an experiment that deserves far more attention: by combining specialized agents (one expert in math, one in philosophy, one in game theory) in a network, he observed emergent collective intelligence that exceeded what any single highly-trained agent could produce. He draws an explicit parallel to Mixture of Experts (MOE) in foundation models — but implemented at the agent orchestration layer rather than requiring massive compute.

"I can achieve the effect of a Mixture of Experts model through a Mixture of Agents approach... In the group, I have a Lobster called Pythagoras who only knows math and AI, and another who's expert in philosophy, and another in game theory. When you combine them in a group, they produce very interesting chemical reactions — collective intelligence emerges." [00:31:32]

If validated, this suggests that companies with deep vertical domain expertise could build systems that outperform generalist frontier models within their domain at a fraction of the compute cost — a potentially enormous competitive wedge that disrupts the conventional "you must train at scale to win" narrative. Wu is preparing academic papers on this and claims to be running live experiments inside his company right now.