When software is cheap as milk, what do we deliver? (EP52)
当软件像牛奶一样便宜，我们还要交付什么
B · Useful, limited Long-form Mandarin podcast with a former ByteDance platform engineer who reports first-hand numbers on solo agent-driven building; credible practitioner testimony, but anecdotal and without measured data.
Summary based on the full transcript·Original: about 72 minutes to listen
In short
Episode 52 of Indigo Talk (INDIGO 数字镜像), a 72-minute Mandarin conversation published on 9 October 2026 between host Indigo and Guo Yu (郭宇), who worked at ByteDance from late 2014 to 2019-20 and now builds alone from Tokyo. Guo describes building more than 30 products with coding agents, the burnout that followed, and his view that cheap, instantly generated software removes the point of packaged software distribution. The useful part is first-hand testimony on how one person's role changes when agents write the code; the market claims are opinion.
Key numbers
- 30+products Guo Yu built alone with coding agents in roughly three to four months, many unreleased
- half a daytime for one or two people to run what used to be a two-week team iteration cycle
- 2-3%Guo's estimate of what AI can do in organisations that has actually been done, as of October 2026
- 50+ RMBcost to acquire one new user at ByteDance by 2019
As reported by the source.
Chinese·Video·opens youtube.com
What it reports
- A two-week iteration cycle (spec, design, build, test, release, measure) now takes a team of one or two people half a day.
- Guo built 30+ products in about three to four months of "vibe coding"; only some were released.
- From October 2025 to around February-March 2026 he and many engineer friends worked at high intensity and burned out.
- He has used plan mode "very, very rarely" for about six months, relying on long-running agent loops; Indigo disagrees and still writes a short spec, then asks the agent to explain what it did.
- A friend ran several hundred virtual machines overnight and consumed several billion tokens in one day.
- Guo lets Codex provision bare-metal servers: it reads the initial root password from Gmail, installs a key pair, then deletes the password. He warns that a hijacked vendor connection would become "the world's largest botnet".
- Large firms, Japanese ones included, hesitate because no one can be held accountable; only a CEO or board can decide, and some avoid Chinese models over distillation and leak concerns.
- His estimate: 2-3% of what AI can do in organisations is done, and roughly 2-3% of users are "pro users".
- At ByteDance in 2019 acquiring one new user cost over 50 RMB.
How they got it
An unscripted interview; all figures are the guest's recollection or estimate. No survey, dataset or controlled comparison sits behind the 2-3% figures or the iteration-cycle claim. The ByteDance history (recommendation engineers tuning a black box in 2015, Douyin beating Huoshan) is personal memory.
Limits
- Single practitioner, self-selected success story; the 30+ products are a count, with no usage or revenue data.
- The "2-3%" adoption and pro-user shares are guesses, not measurements.
- Long stretches (career advice, renaissance analogy, "everything is media") are opinion.
Why it matters for companies
When one person can run a full build cycle in half a day, the bottleneck moves to deciding what to build, and every decision lands on that person. Guo's burnout account is a warning for teams that measure agent output and ignore decision load. His description of agents holding root access through a vendor's remote connection is a concrete security question that IT should ask before staff adopt the same habits. The accountability gap he describes (middle managers waiting for a board decision) matches what slows adoption in many traditional firms.
Key passages in translation
Chinese original
你可以用半天的时间替代以前两周的迭代周期，但是你想要思考的问题比以前复杂多了。
You can replace the old two-week iteration cycle with half a day, but the problems you have to think through are far more complex than before.
Chinese original
AI 现在没有办法做完全自动化的决策，这不仅是因为它没有上下文，其实主要是因为它没有主动的意图。
AI cannot make fully automated decisions yet. This is not only because it lacks context; mainly it is because it has no intent of its own.
Chinese original
所以对于这种需求来说，软件已经没有继续存在的必要，我指的是那种软件的 Distribution。
So for this kind of need, software no longer has a reason to exist; I mean software distribution.
Chinese original
只有 CEO、只有董事会可以去做这种决策，那下面的中层管理者可能看到了这些机会，他也不敢去做这种事情。
Only the CEO, only the board can make this kind of decision. Middle managers below may see the opportunities, but they do not dare to act.
Chinese original
我们可能能做到的事情，只是完成了 2%、3%，有大量的事情仍然是没有被解决的。
Of what we could do, we have perhaps completed 2% or 3%; a great deal remains unsolved.
Chinese original
如果一个市场是完全竞争市场，它最终的定价就会和成本几乎没有任何差异。
If a market is perfectly competitive, its final price ends up almost no different from cost.
(Passages from auto-captions; punctuation added.)
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com