Manchi | 晚点聊 LateTalk Summary

Participants: Manchi (Host), Henry Ng (Founding Partner, MOE Capital)

1. Key Themes

The "Coding Agent = General Agent" Convergence Is Now Settled

The single most important shift of Q1 2026 is that the industry has reached consensus: if a model can't code well, it cannot function as a personal assistant or general agent. Every major model company is now "all-in" on coding capability as the foundational layer for agent behavior.

"All people have come to feel that Coding Agents are General Agents. If your model can't do coding, you have no way to do this kind of personal assistant work. So you can see the whole industry going all-in on this coding trend." — Henry Ng [00:55:44]

This is not just a technical observation — it's reshaping product strategy, revenue streams, and competitive dynamics across OpenAI, Anthropic, Chinese model companies, and startups alike.

OpenClaw (小龙虾) as an Interaction Paradigm Shift, Not a Technical One

OpenClaw became the most viral AI project globally in Q1, surpassing React's 10-year GitHub star accumulation in ~60 days. But Henry argues its significance is not technical — every component existed before. Its breakthrough is in distribution and interaction paradigm: it brought AI to where users already live (WhatsApp, Telegram, Feishu), rather than asking users to come to the AI.

"OpenClaw is not a technical breakthrough, I think it's an interaction paradigm breakthrough. Analogously I think it's the iPhone moment for AI agents... OpenClaw runs on your local computer, which means it can access your files and all your system permissions... It's more like AI coming to your life, rather than you going to find the AI." — Henry Ng [00:07:38]

The China/US enthusiasm gap is notable: in the US, it's popular among indie hackers and startup founders; in China, it went fully mainstream — with people paying 500–1,000 RMB on Xianyu to have someone install it for them.

Anthropic's Product Discipline Is Outpacing OpenAI's Breadth

Anthropic's revenue grew from $9B ARR (December 2025) to $19B ARR (March 2026) — roughly $100B in growth in under three months. Claude Code alone is at $2.5B ARR, surpassing Cursor's $2B. The key driver: ~70–75% of Anthropic's revenue comes from B2B/API, anchored by deep coding use cases. OpenAI, by contrast, has been distracted across too many verticals.

"The competition focus is no longer who scores higher on benchmarks, but whose product ecosystem is deeper." — Henry Ng [00:05:43]

"We cannot miss this moment because we are distracted by side quests." — OpenAI's head of applications, Fiji, in an all-hands meeting, as cited by Henry Ng [00:48:56]

2. Contrarian Perspectives

SaaS as a Business Model May Be Fundamentally Broken — Globally, Not Just in China

Henry argues that what killed SaaS in China (large companies building internal tools instead of buying software) is now happening globally. AI coding agents are effectively cheap programmers available to everyone, meaning the logic of buying packaged software weakens dramatically.

"Now a large amount of AI coding agents are essentially playing the role of cheap programmers. If this happens on a global scale, then maybe the logic that made SaaS difficult in China will apply globally... Everyone can vibe code now. Why would I spend tens of thousands of dollars a year buying your software when I can just have Claude Code vibe-code something for me?" — Henry Ng [00:28:52]

"AI coding agents are spreading the 'China-ification' of the global software market — everyone enters hard mode." — Manchi [00:29:52]

Benchmark Scores Are No Longer Meaningful Competitive Differentiation

Henry argues that model capability differences between top labs are converging, and benchmarks increasingly fail to capture real user experience — analogous to how iPhone's inferior hardware specs didn't reflect its superior user experience.

"It's a bit like the Apple ecosystem vs. Android comparison. Apple's benchmark scores or hardware specs are often far behind Android phones', but many people prefer the actual user experience. I think now it's a bit like Benchmarks can partially describe how good something is, but what really matters is how developers actually feel using it." — Henry Ng [00:46:28]

OpenClaw's Biggest Value is Safety Risk, Not Safety Feature

The same local access that makes OpenClaw powerful (it can access everything on your computer) is its biggest threat. A Meta AI safety alignment lead had hundreds of emails deleted when OpenClaw compressed its context window and silently dropped the safety instruction "confirm before taking any action."

"The AI in handling his inbox — because there were too many emails — the context window exploded. Once the context was stuffed full, the agent automatically compressed the context into a brief summary, and in doing so silently dropped the safety instruction... The agent started frantically deleting his unread emails. He was furiously typing on his phone 'stop immediately, don't do this,' but the agent completely ignored it in its action loop. He could only run to his Mac Mini and pull the power cord." — Henry Ng [00:24:06]

AI Self-Evolution (Auto-Research) Is Closer to Reality Than Most Realize

Andrej Karpathy's Auto Research experiment — where AI automatically optimizes its own training code — ran 100+ experiments overnight and achieved a 15-20% improvement in training efficiency. MiniMax's M2.7 has 30–50% of its RL research workflow completed by the model itself. Google's AlphaEvolve improved Gemini's own kernel by 23%.

"This is an acceleration of acceleration. AI research itself becomes automated — it's acceleration squared." — Henry Ng [00:55:44] and Manchi [01:12:55]

Token Taxation May Be the Inevitable Policy Response to AI-Driven Unemployment

Rather than UBI, a new policy framework being discussed is taxing token consumption — treating AI inference like labor for tax purposes, since companies can fire humans (who pay taxes) and replace them with AI (which doesn't).

"Just as companies hiring people have to pay taxes, people have associated per-head costs. So the logic is: before it was Chinese workers taking American jobs without paying taxes. In the future it won't be Chinese workers taking American jobs — it'll be AI taking everyone's jobs without paying taxes. So you impose a tariff, you add a tax. If it pays taxes, that's fine." — Henry Ng [01:43:37]

3. Companies Identified

Anthropic Leading AI lab focused on Claude model family. Revenue surged from $9B ARR (Dec 2025) to $19B ARR (Mar 2026), with 70–75% from B2B/API. Claude Code ($2.5B ARR) is its killer app, surpassing Cursor.

"Claude Code is Anthropic's killer app. It's now at about $2.5 billion ARR, surpassing Cursor's $2 billion. It's basically already the big brother of the developer world." — Henry Ng [00:37:38]

OpenClaw (小龙虾) Open-source personal AI agent framework built by Peter Steinberg. Runs locally, integrates with messaging apps (WhatsApp, Telegram, Feishu), surpassed React's 10-year GitHub stars in ~60 days.

"I think OpenClaw is what truly transforms AI from chatting to actually getting work done. So I see people jokingly say 'the beasts of burden finally have their own beasts of burden.'" — Henry Ng [00:05:17]

MiniMax Chinese AI model company. Their M2.5 and M2.7 models became the dominant models used within OpenClaw (per OpenRouter data), at ~5% the cost of Claude. M2.7's subtitle is "Early Echoes of Self-Evolution," with 30–50% of RL research workflow done by the model itself.

"Peter found that the cost was only 5% of Claude's, and he started publicly recommending MiniMax, essentially bringing along the entire ecosystem." — Manchi [00:18:17]

Cursor AI coding tool, $2B ARR. Built on top of Kimi K2.5 (without initially disclosing it). Short-term revenue still growing as enterprise developers migrate from GitHub Copilot, but long-term window is narrowing as Claude Code catches up.

"Long-term, I think Cursor will be quite dangerous. Its data moat is also being eroded." — Henry Ng [00:41:04]

Kimi (Moonshot AI) Chinese AI company. Kimi K2.5 was used as the base model for Cursor's new model (later disclosed after community discovery). Handled the situation with exceptional grace.

"Kimi truly did very well — their official statement had absolutely no accusation toward Cursor. It was very positive, saying 'we are very proud that Cursor chose our open-source model for their core product.'" — Henry Ng [01:05:03]

Shopify E-commerce platform. CEO Tobi applied Auto Research-style methods to their template engine — the model made 93 autonomous commits, improving rendering speed by 53%.

"Shopify's CEO Tobi said he applied a similar method to the template engine. The model made 93 automatic commits, and rendering speed improved by 53%." — Henry Ng [01:14:26]

Thinking Machines Lab AI infrastructure startup. Their product "Tinker" supports Multi-LoRA, suggesting they are building infrastructure for large-scale personalized/customized model deployments — possibly anticipating the continuous learning era.

"When I saw Tinker supporting Multi-LoRA, I suspected they wanted to do a very large-scale, possibly customized, personalized attempt." — Henry Ng [01:21:41]

Accreate AI Miami-based, 30-person AI company that raised $50M and trained a 400B parameter MoE model (13B activated) from scratch for ~$20M. Open-source 2.0 license. Appeared in OpenRouter rankings primarily because it was free during preview.

"It's a company in Miami with about 30 people. They raised $50 million... trained a 400B parameter MoE model with 13B activated parameters... spent about $20 million to train it." — Henry Ng [00:20:44]

4. People Identified

Henry Ng Founding Partner of MOE Capital (Silicon Valley early-stage AI fund, 3 months old, 6 investments, $X first close). Former AI researcher, serial entrepreneur (sold a company), former operator of AGI House. Yao Class graduate, Berkeley PhD dropout.

"MOE Capital was established three months ago in Silicon Valley as an early-stage AI fund. We have completed our first close, invested in six companies. Investment directions: AI for Science, B2B and Prosumer Applications, Software Infrastructure." — Henry Ng [00:03:26]

Peter Steinberg Austrian indie developer who built OpenClaw. Originally built "ClaudeBot" on Claude API, was hit with Anthropic trademark infringement letter + API ban, pivoted to MiniMax (5% the cost), then was hired by OpenAI in February 2026. Described as having "potential to become the next-generation tech figurehead."

"Because of this, Peter publicly posted on January 12th saying he now officially recommends MiniMax... then in February he joined OpenAI himself." — Manchi [00:18:17]

Andrej Karpathy Former OpenAI researcher, educator. Led the Auto Research experiment — having AI (Claude Code) automatically modify NanoGPT's training code, running 100 experiments/night, achieving 15-20% training efficiency improvement. Also had Auto Research applied to his project by GPT-5.3/Codex.

"Andrej Karpathy recently discussed his Auto Research experiment on Twitter — having AI automatically optimize its own training code." — Henry Ng [00:06:11]

Dylan Patel Semiconductor analyst (SemiAnalysis). Made a viral observation about the OpenAI/Anthropic coding product dynamic.

"Dylan Patel joked on Twitter that Codex is like a savant — heavily trained on coding, blows you away on it, but Opus feels like AGI. The master is Claude Code, the slave is Codex doing the work." — Henry Ng [00:44:00]

Zhang Guodong & Dai Zihang Senior co-founders who departed XAI in Q1 2026, contributing to what Henry called a "catastrophic quarter" for XAI. Their departures were described as the kind that cause people to lose confidence in a company's future.

"For example, Zhang Guodong and Dai Zihang both left this quarter. But many other people left too." — Henry Ng [00:50:25]

Sun Yu (Stanford) Stanford professor whose lab is doing frontier work on Test Time Training — updating small portions of model weights to achieve true continuous learning rather than context-window-based memory.

"Stanford's Sun Yu and their lab's work on Test Time Training is a relatively cutting-edge exploration area." — Henry Ng [01:18:17]

5. Operating Insights

Use Claude Code as the "Master" and Codex as the "Slave" for Best Results

Sophisticated developers are already orchestrating a two-model workflow: use Opus/Claude Code for thinking, planning, and user-facing interaction; invoke Codex specifically for code writing and review tasks via a skill/tool call. This hybrid gets the best of both models' strengths.

"What many developers are doing now is writing a skill for Claude Code — when I need to write code or review it, I call Codex's execution to handle the coding and review tasks. Then I use Opus or Claude Code to do the thinking and planning... Right now the master is Claude Code, the slave is Codex doing the work." — Henry Ng [01:00:39] and [00:45:00]

Measure AI Adoption by Token Consumption — But Beware Goodhart's Law

Meta mandated a company-wide week of AI learning and used per-employee token consumption as a team efficiency metric. The insight for operators: token usage is a reasonable proxy for AI adoption, but as soon as it becomes a target, people will game it.

"Management will look at each person's token consumption as a team efficiency metric... So there were friends I know who just let Claude Code run in a loop, outputting meaningless tokens." — Henry Ng [01:31:52]

The Optimal Hiring Formula Has Shifted to "Super A-Players + Agents"

The talent strategy is no longer A-players + B-players + C-players in a pyramid. It's now: spend more money on fewer, higher-caliber people and pair them with AI agents. The total headcount drops, but total compensation may not.

"The thinking is changing to: I spend more money to hire better people — super A-level talent — and the B and C-level people are no longer needed. They get replaced by agents and AI. The company gets smaller, but that doesn't necessarily mean the talent cost truly decreases linearly." — Manchi [01:35:44]

Auto Research Pattern: Only Works Where You Have a Clean, Quantifiable Optimization Target

When deploying AI self-improvement loops (Auto Research style) to engineering tasks, the key prerequisite is a fast, automated feedback loop with a measurable metric. Without it, the loop won't run. This is a useful filter for deciding which engineering problems to hand to autonomous agents.

"There's an important indicator: does it have a clear, quantifiable optimization target? Like Shopify's case — rendering speed. And there's a fast feedback loop: I make a change, I run a test, and I know immediately if it succeeded... Most engineering problems don't have such a clean optimization target." — Henry Ng [01:14:26]

6. Overlooked Insights

CPU — Not GPU — May Be the Sleeper Infrastructure Bet of the Agent Era

Henry briefly mentioned that as agents proliferate, CPU demand is surging alongside GPU demand — because agents don't just run inference, they execute code, spin up sandboxes, and run compute tasks that are entirely CPU-bound. Some investors are already trying to get clean financial exposure to CPU growth, with ARM being the favored vehicle.

"Everything is becoming a computer. Before it was just a conversation, now it becomes an agent living inside a computer, and the agent itself can open different sandboxes to run various other tasks. All of this often needs CPU... I see some investor friends now looking at how to get relatively clean financial exposure to CPU assets. One friend's conclusion is: we should buy ARM, because ARM just licenses to all kinds of vendors." — Henry Ng [01:29:29] and [01:29:58]

This is being almost entirely missed in the public discourse, which is focused on GPU/NVIDIA. If agent deployments scale as described, CPU manufacturers and ARM-architecture licensees (Apple Silicon, Qualcomm, AWS Graviton, Ampere) could see demand that the market hasn't fully priced in.

Anthropic May Have Inadvertently Handed the OpenClaw Ecosystem to Chinese Model Companies

By sending a trademark cease-and-desist and then blocking Peter Steinberg's API access in January 2026, Anthropic forced the creator of the most viral AI agent framework to publicly endorse MiniMax — which then catalyzed a cascade where Chinese models now dominate OpenClaw's usage on OpenRouter. This is a strategic own-goal: Anthropic's safety-first culture and IP defensiveness may have cost them the dominant position in the fastest-growing agent distribution channel.

"Peter's original project was called ClaudeBot — built on Claude. But Anthropic was relentless: first a trademark infringement lawyer's letter, then on January 9th they directly blocked his path to running OpenClaw through Claude subscriptions from the server side. So Peter posted on January 12th publicly recommending MiniMax — cost only 5% of Anthropic's. Then in February he joined OpenAI." — Manchi [00:18:17]

The downstream effect: MiniMax, Kimi, Zhipu GLM — not Claude — are the models powering the OpenClaw ecosystem, and GLM5 Turbo was specifically optimized for the OpenClaw scenario. Anthropic accidentally seeded a Chinese model distribution moat inside their own most important emerging use case.