广密 | 张小珺Jùn｜商业访谈录 Summary

Podcast: 张小珺Jùn｜商业访谈录 Participants: 广密 (Guangmi, Silicon Valley-based AI investor/researcher), 张小珺 (Jun Zhang, host)

1. Key Themes

Coding Is the Accelerant of AGI — Not Just a Vertical Use Case

Guangmi argues that coding has crossed from a niche developer tool into the central engine of AGI progress, comparable in strategic importance to GPU availability. The fundamental insight is that language describes the world, but code describes solutions — making code the most generalizable substrate for automating knowledge work.

"Do you believe that code can express the vast majority of tasks in the digital world? Because natural language is a description of the world, code is a description of the solution — language is the world, code is the solution. If this assumption holds, wouldn't a coding agent, done well, be able to automate most of the tasks of white-collar knowledge workers?" 00:11:30

"Coding is not just an industry, not just an application scenario, not just a product — it has actually become one of the most important accelerators in the entire AGI roadmap. Not having the leading coding model is like not having the leading GPU." 00:12:29

The Three-Act Structure of AGI — We Are in Act Two

Guangmi presents a clear framework: Act One was chatbot/ChatGPT (conversation, limited commercial value); Act Two is coding agents that can actually complete high-value tasks and accelerate AI research itself; Act Three will be automated AI researchers solving fundamental science.

"The first act we saw was chatbot — ChatGPT. Models could have dialogue and search the web. But the commercial value was still limited. The second act, today, has arrived at coding agents — agents that directly complete tasks, essentially able to do real work for us. Furthermore, they can accelerate the entire AGI process. If coding agents are realized, possibly 90% of AGI has already been achieved." 00:14:24

"The third act might be automated AI researchers. This is also what OpenAI most wants to do — everyone might have a powerful research assistant to solve all kinds of problems, perhaps then solving fundamental science problems: brain science, neuroscience, materials science." 00:14:51

Strategy and Culture Beat Technology — Anthropic as a Case Study

The most important lesson from this cycle is that organizational focus and strategic clarity — specifically what you choose NOT to do — is the primary determinant of who wins. Anthropic's dominance in coding was not accidental; it was the result of deliberate top-down strategy, data obsession, and stable culture.

"Anthropic is more like a strategic victory, or a victory of product focus, or culture. I think these invisible factors are quite important. Otherwise it would have been impossible for them to fight their way up from a trailing position." 00:30:51

"Data being fundamental is written into their DNA, carved into their bones. There's a rumor that their chief scientist personally led the negotiation for training data — that's quite rare." 00:18:12

"OpenAI's ChatGPT was too successful, making them focus on 2C while neglecting coding." 00:01:26

2. Contrarian Perspectives

DAU and C-End Traffic May Be Largely Irrelevant in the AI Era

Most tech investors and operators still benchmark AI companies using internet-era metrics like DAU and user growth. Guangmi argues these metrics are becoming structurally misleading — a few hundred thousand "tower-top" users generating massive token usage may be worth more than tens of millions of casual subscribers.

"Anthropic has already officially announced their ARR has surpassed OpenAI's. But more importantly, their top one or two million users contribute more revenue than OpenAI's fifty to sixty million paid subscribers. So you thought ChatGPT had already won on the consumer side, but it turns out the consumer side may not be as large as coding or agents." 00:09:21

"Today I feel the more interesting metric is no longer just chasing DAU or advertising scale, but pursuing token usage — especially from super-developers or high-value users. This feels more important." 00:09:49

Sora and Multimodal Were Strategic Distractions That Set OpenAI Back

Most observers celebrated Sora as a breakthrough. Guangmi's contrarian view is that it was a costly misallocation that delayed OpenAI's coding prioritization — and the fact that it was just shut down confirms this.

"Does Sora being shut down have any connection to this? I think it probably does — because GPU is too limited. Actually multimodal is quite GPU-hungry. And think about it: after Sora was released, so many people went to create content. What benefit does that have for OpenAI? It seems like none." 00:35:09

"OpenAI had a very serious strategic misjudgment on coding, including Google. Everyone made strategic misjudgments on coding. You shouldn't be using internet thinking — DAU and growth thinking — to look at these things anymore." 00:34:33

The Best AI Researchers Are Already Afraid They'll Be Unemployed in 1-2 Years

The conventional wisdom is that AI will first replace lower-skill workers, with researchers and scientists safe for decades. Guangmi's direct observation from frontier labs contradicts this entirely.

"The most brilliant AI researchers are all worried about not having jobs in one to two years. Possibly the next one to two years may be their only remaining window to work. Because after that, AI may automate the entire AI research workflow." 00:33:00

"Some of the recent breakthroughs in AI research were not brought by human engineers, but by Codex and Claude Code. I think this is an even more qualitative signal — AI today can significantly accelerate AI, and even bring breakthroughs." 00:07:32

XAI's Failure Is Elon's Impatience, Not His Team — The Founding Team Was World-Class

The narrative is that Elon assembled the wrong team. Guangmi argues the opposite: the founding team was exceptional, but Elon's sprint mentality is structurally incompatible with what model training requires.

"The core team has actually left — the founding team was all world-class, from autonomous driving, multimodal, all top-tier. I think short-term XAI is quite difficult... The source of the collapse is probably Elon was somewhat dissatisfied with the founding team. But actually the founding team was a world-class team — that was a bit impatient." 00:58:29

"Elon maybe has no patience — wants to see results in two weeks. So this culture causes teams to pursue short-term effects while sacrificing long-term quality." 01:00:27

3. Companies Identified

Anthropic Leading AI lab, creator of Claude models and Claude Code Cited as the clearest strategic winner of this cycle — top-down focus on coding, data-first culture, stable team, premised on high-value users rather than scale. ARR reportedly surpassed OpenAI's.

"Anthropic all-in'd on one thing — coding — and today it has become enormous. They also abandoned the consumer side, abandoned multimodal. Strategically it is very top-down... I feel their product sense is quite strong — putting model capability into product, into user experience, the efficiency is higher." 00:23:04

Cursor AI-powered IDE for developers Identified as the #3 AI product by ARR (~$250M), beneficiary of model capability overflow. However, flagged as strategically vulnerable if frontier model companies stop providing API access.

"Cursor's essence is still eating the red dividend of model companies' technology overflow. If model companies choose not to overflow anymore, that seems reasonable too — just handle it themselves." 00:38:28 "The best thing for Cursor is probably to sell to Microsoft for developer-focused work." 00:38:56

Manus (Mannus) Autonomous agent company, acquired by Meta Cited as having sold too cheaply, and as a "Harness pioneer" — the practical embodiment of the agent infrastructure thesis before it became fashionable.

"Manus sold too cheaply. Definitely sold too cheaply... Actually Meta, I feel, is the godfather of Harness — it's an independent team that built Harness very well. Today model companies are shouting the Harness concept like a religion, but Manus is more like the godfather of Harness." 00:56:44

Eleven Labs AI voice synthesis company Mentioned as one of the fastest-growing AI application companies, crossing $300M ARR.

"Eleven Labs and Suno have probably both exceeded $300 million ARR. Their growth is all very fast." 01:17:24

Lovable / Mindus AI application development platforms Both cited as exceeding $400M ARR with rapid growth trajectories.

"Mindus and Lovable should have both exceeded $400 million ARR. Growth is all very fast." 01:17:24

Open Evidence / Uverage AI healthcare companies in Silicon Valley Flagged as fast-growing, with Silicon Valley investing heavily in AI for medicine.

"In healthcare, Silicon Valley is investing quite a lot — for example Open Evidence or Uverage. These two are developing quite fast." 01:17:54

4. People Identified

Dario Amodei (CEO, Anthropic) Physicist-turned-AI CEO, co-founder of Anthropic Praised for technical grounding, top-down strategic decisiveness, AGI mission conviction, and an almost religious ability to align his team around a single goal.

"Dario and their chief scientist — two physics-background founders. They understand AI more from the angle of physical observation. They didn't aim to create a new Transformer architecture, but rather like physicists, they intuited the laws of AI — finding data efficiency, architectural efficiency, engineering efficiency." 00:28:08

Jared Kaplan (Chief Scientist, Anthropic) Co-founder and chief scientist of Anthropic, physicist Specifically noted for personally leading data negotiations — an exceptionally rare hands-on behavior for someone at his seniority, signaling how deeply data is prioritized at Anthropic.

"There's a rumor that their chief scientist Jared Kaplan personally led the negotiation for training data. This kind of thing is quite rare." 00:18:12

Boris (Founder, Claude Code) Creator of Claude Code at Anthropic Cited as a world-class coder who had a clear vision of delivering the best coding workflow, and chose terminal-native interface over an IDE when everyone else was building IDEs.

"Claude Code's founder Boris is actually a very strong coder. He actually wants to deliver the best coding work approach, leading the entire development paradigm... Anthropic didn't build an IDE, but instead made a terminal-based form — today's Claude Code. This can probably better receive the dividend of model capability, because models grow exponentially, and your product needs to exponentially catch that." 00:26:49

5. Operating Insights

The "Token ROI" Metric as a Business Viability Test for AI Applications

Guangmi introduces a simple but powerful heuristic for evaluating any AI-native business: for every $100 of tokens consumed, can the output generate $110 in revenue? This reframes how to evaluate AI application companies and internal AI tooling investments — not by usage or engagement, but by economic closure.

"I recently heard a pretty good metric: if you consume $100 worth of tokens, can you earn back $110? You need to get this ROI positive. Actually many people haven't gotten this positive, or the loop hasn't closed. I think this may be an important metric." 01:20:41

Agent Infrastructure = Building the "Parallel World" for AI Workers

The "Harness Engineering" framework offers a direct operating insight: treat agents as first-class citizens with the same infrastructure humans need — accounts, credentials, working environments, management structures. Companies that build this layer well will dramatically improve agent performance even with average models.

"We should treat AI agents as first-class citizens, as one with us. Things humans should have, agents should have too. Human knowledge workers have working environments, computers, credit cards. In a parallel world, you also need to build a set of environments that agents, as first-class citizens, need." 01:01:57

"Agents need their management science and organization — that's the meaning of Harness. With Harness, even ordinary models can perform high-value tasks." 01:02:57

What You Choose NOT to Do Is the Most Important Strategic Decision

Anthropic's playbook — abandoning multimodal, refusing to build an IDE when Cursor was hot, ignoring Reasoning Model hype — offers a direct lesson for operators: the discipline to concentrate resources is more valuable than the intelligence to identify opportunities.

"The more important thing is what NOT to do. Anthropic completely abandoned multimodal — multimodal probably isn't on the main technological storyline. They also didn't follow the trend to make reasoning models. They abandoned 2C. They were relatively focused on coding." 00:18:41

6. Overlooked Insights

The Coding Moat Is Actually a Data Moat — And It Creates a Dependency Trap That Could Fracture the Entire AI Ecosystem

Guangmi makes a brief but explosive observation: if you rely on Anthropic's Claude for your coding workflows, and you ever become a competitive threat to them, they will cut off your access — and this has already happened to OpenAI, xAI, and reportedly most of Google. The implication is that the coding layer is becoming a strategic chokepoint similar to TSMC in semiconductors. Any company that doesn't own its own frontier coding model is building on sand. This also means the model that wins coding may achieve a near-monopolistic position on the most productive users in the world.

"There's no such thing as a coding model that only serves yourself — because your own task data distribution isn't comprehensive enough, you'll definitely fall behind. And if you as a leading model company rely too heavily on Anthropic, once you reach tier one and become a threat to them, Anthropic will almost certainly cut off your supply. OpenAI has been cut off, xAI has been cut off, probably most of Google has been cut off, and possibly someday Meta too. Coding could be the same as GPU." 00:12:29

The non-obvious investment implication: any AI application company or enterprise that is currently "Anthropic-native" for coding workflows is sitting on a single point of failure that could be weaponized against them if they ever become commercially significant. The race to build proprietary coding model capability — even if inferior to Anthropic today — may be existentially necessary for every serious player.

AI Research Breakthroughs Are Already Being Generated by AI, Not Humans — But Nobody Is Talking About the Compounding Effect

Guangmi mentions almost in passing that recent AI research breakthroughs at frontier labs are no longer coming from human researchers — they're coming from Codex and Claude Code. This is the recursive loop that defines the "intelligence explosion" in concrete terms, not theoretical ones. The rate of AI capability improvement is no longer bounded by human researcher throughput.

"The recent breakthroughs in AI research were not brought by human engineers, but by Codex and Claude Code. This is an even more qualitative signal — AI today can significantly accelerate AI and even bring breakthroughs... The model progress in the last one quarter may have exceeded all of 2025's progress combined." 00:07:32

"AI's rate of IQ improvement over the past quarter may be faster than humanity's IQ improvement over the past 200 years." 00:37:28

The overlooked implication: the standard investment timeline assumptions about AI development — even the most aggressive ones — may be built on the old assumption that humans are the bottleneck in AI research. If AI is now conducting its own research, the timeline to AGI compresses in a way that isn't yet priced into any public or private market valuation. Guangmi's year-end or early next year AGI declaration estimate may not be hyperbole.