Coding Is AGI's Second Act: No Leading Coding Model Is Like No Leading GPU
The real dividing line for model companies isn't consumer DAU — it's whether you can organise a few hundred of the smartest people to go all in on the dirty work of data. Anthropic won on strategic focus; OpenAI lost by wanting everything.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The past quarter's progress exceeded all of 2025
Guangmi reduces the past quarter's most essential turning point to Anthropic's leap from Opus 4.5 to 4.6, which he considers an era-defining jump on the order of GPT3 to GPT4 — the model went from ‘something you can converse with and ask questions’ to ‘a real agent that can do high-value tasks’. Because task value rose, usage and value rose together. His felt sense is that the past quarter's improvement in model capability may have exceeded all of 2025's progress, and the acceleration is especially fast — there may be another GPT3-to-GPT4 leap before June or July this summer.
— Guang MiFrontier lab researchers have stopped writing code
Guangmi describes the real felt experience on the Silicon Valley front line: AI researchers and very strong programmers at frontier labs basically don't write code anymore. Last year maybe 70-80% of the code in a system was written by humans; this year it may be less than 1%, and the daily job is ‘AI writes, humans review’ — and even the humans' reviewing ability isn't sufficient. People he knows burn several hundred dollars of tokens a day, several thousand a week. The effect: going from an idea to working code used to take two or three weeks, now it takes one or two days. The more qualitative signal: many AI research breakthroughs aren't brought by human engineers but by Codex and Cloud Code — AI today can significantly accelerate AI.
— Guang MiOne or two million tip-of-the-pyramid users out-earn fifty or sixty million subscribers
Guangmi points out that Anthropic officially announced AR exceeding OpenAI, but more important is that the revenue from the one or two million users at its apex exceeds that of OpenAI's perhaps fifty or sixty million subscribers. This means the competitive metric has changed: no longer just chasing DAU or ad scale, but chasing Token Usage, especially from super developers or tip-of-the-pyramid users. He judges Coding's explosiveness to be steeper than ChatGPT's was back then; if the momentum holds, by year-end OpenAI and Anthropic's AR could reach $80 billion or $100 billion, and next year head toward $200 billion-plus.
— Guang MiModel companies that neglect Coding will drop out of the first tier
Guangmi's second strong view: if a leading model company neglects Coding, it will most likely drop out of the first tier — and there is no such thing as a Coding Model you keep only for yourself, because your own task data distribution isn't complete enough and you will certainly fall behind. More critical is the risk of supply cutoff — once you reach the first tier and pose a threat to Anthropic, Anthropic will most likely cut you off; OpenAI was cut off too, xAI was cut off too, probably most of Google was cut off, and Meta may be cut off some day too. So Coding is like GPU: no leading Coding model is like no leading GPU — using A100 versus others using GB is a huge difference.
— Guang MiThe hard part of doing Coding well is organisation and data, not technology
Guangmi believes doing Coding well comes down to two things: organisation and culture, and data. Coding isn't just a technical know-how problem; it's more a strategic and organisational problem. The difficulty is which AI lab can organise its few hundred smartest people to go all in on the single thing of Coding — because there's a lot of dirty, hard work in it, and the smartest person in every lab wants to make their own bet, do 0-to-1, wants to be Elia. What's different about Anthropic is that the founders realised from the start that data is the root of all problems; valuing data is in their bones, and there's a rumor that chief scientist Jared Kaplan personally leads the data negotiations.
— Guang MiAnthropic is a victory of strategy, not of technology
Guangmi reviews Anthropic: going all in on coding wasn't clear on day one either — on one hand there was no opportunity left on the consumer side, and on the other, releasing Sonnet 3.5 in the summer of 2024 gave positive feedback on coding. After that they went all in on just one thing, coding, giving up consumer and giving up multimodality; strategically it's very top down, whereas OpenAI is especially bottom up. Anthropic's founders are earnest and understand technology, leading everyone to do data and engineering, the team is stable, and everyone believes in AGI. They don't mythologise any single link — pretraining-hits-a-wall theory, RL-is-god theory, reasoning models — they didn't follow any of the trends; they're more like a collective, a team, an industrialised system.
— Guang MiOpenAI has a 50% chance of being AGI's ultimate winner
Guangmi isn't so pessimistic about OpenAI, believing there's a 50% chance the ultimate Winner of AGI may still be OpenAI. The logic: today's secret to victory may be the next era's poison — OpenAI's excessive success with ChatGPT made them focus on 2C and neglect Coding, and because they had to care about Inference cost, the model was never made very large. Anthropic's success is more gradual accumulation, doing the details well — a victory of execution, of focus or organisation — but it may not withstand OpenAI's next paradigm-level breakthrough. OpenAI's bottom-up culture opens up free exploration; today its Coding capability is very strong, and one or two people can accomplish earth-shaking things.
— Guang MiHuman knowledge and intelligence have become cheap
Guangmi on social impact: the most essential point is that human knowledge and intelligence have become cheap. In the past, acquiring knowledge by studying and reading could get you a job, but today that intelligence and knowledge are all in the models, hugely compressed into compute resources or tokens. He judges that perhaps 70-80% of people's value and meaning in society will undergo a subtle change. AI will also bring a lot of deflation — after using ChatGPT and Cloud, the need to hire consultants and buy other software is much smaller; one product satisfies many needs, and in the long run many SaaS will disappear too. This year is certainly the year humanity begins to face unemployment painfully; perhaps 30% of jobs will feel like they're just gone.
— Guang MiIn their own words · checked verbatim
Because natural language is the description of the world and Code is the description of the Solution — language is the world, code is the solution.
因为自然语言是对世界的描述 Code是对Solution的描述 就是语言及世界 代码及方案
Guang Mi1:01
If a leading model company doesn't value Coding, it will most likely drop out of the first tier.
如果领先的模型公司不重视Coding 它大概率会掉出第一梯队的
Guang Mi12:10
Not having the most leading Coding model is like not having the most leading GPU.
你没有最领先的Coding model 就像是没有最领先的GPU
Guang Mi13:11
Figures
| Anthropic products released in the past 50-plus working days | 70-plus products and features | 8:06 |
| Codex weekly active users | 3 million | 37:36 |
| Cursor AR | $2.5 billion | 1:16:58 |
| Perplexity AR | over $500 million | 1:16:58 |
| 11 Labs and Suno AR | both over $300 million | 1:16:58 |
| Minders and Lovable AR | both over $400 million | 1:16:58 |
Glossary
- Harness Engineering
- The engineering of building working environments and managing constraints for Agents, so that ordinary models can also do high-value tasks.
- Token Usage
- A measure of the total tokens a model consumes; Guangmi believes it reflects a model company's value better than DAU.
- Solta model
- Guangmi's term for the most leading frontier model; continuously doing it well is the key indicator for investing in model companies.
- feedback loop
- The cycle from task execution to receiving feedback; the shorter the loop, the faster AI improves, which is why Coding ran out front first.
How to listen
Founders and investors watching the competitive landscape of AI model companies; readers who want to understand why Coding is the AGI mainline and how to assess a model company's organisational capability.
After 1:17:58 the discussion of robotics, AI hardware and other directions gets broad — you can skip through quickly.