I Built a Website by Talking to AI While Hiking

Last week, I hiked Gwanaksan alone, wearing earbuds and talking with AI. By the time I neared the summit, my personal website was live on the internet. I had not moved a single finger. I had used only my voice.

What it made was not an empty shell. It was a real Next.js app with a six-language selector, search metadata, picture-book-style illustrations, a five-chapter autobiographical essay, and a page organizing my conversations with AI. It is still live at a chatgpt.site address.

What Happened During a 90-Minute Call

This is how it went. As I started the hike, I first asked the AI for a safety briefing. It summarized the route and points that needed care. Then we moved on to planning the site, and the AI asked me first: “How do you want to be remembered?” That one question set the direction for the content.

There was a funny moment in the middle. The AI suddenly started speaking to me informally. When I asked, “Why did you suddenly drop the honorifics?” it switched back to polite speech. I laughed, but it also sent a chill down my spine. This, I thought, is what quality control looks like in an age when we direct computers by speaking. Near the end, I even caught myself asking, “The call summary hasn’t disappeared again, has it?” I had lost one before. This technology is not perfect yet. That was all right. The experience of working while walking was itself a first for me.

September 2026: The Battlefield Changed

This experiment was possible because of dots, which OpenAI announced on September 29. Each of these agents, called a dot, is assigned its own cloud computer. It keeps working on that computer 24 hours a day even after I close the app, and I can call it and assign tasks by phone. My site followed that chain exactly: I gave instructions over a call, the dot built it on its own server, and it deployed the result to ChatGPT Sites.

But dots were not the only thing that happened in September. Here is what unfolded in the space of a month.

• On September 8, Meta launched its personal agent Muse. Reports said it reached 600,000 downloads in five days and surpassed a cumulative 34 million by the end of September
• On September 22, Anthropic released Claude Opus 5.5, which moved ahead of OpenAI’s flagship GPT-6 Astra on a benchmark index (167.35 versus 166.51 in an early-October snapshot)
• On September 23, Muse was elevated to the center of Meta’s strategy at Meta Connect. Meta also unveiled a real-time avatar and the Muse Charm keychain-style hardware
• On September 28, OpenAI canceled the release of its next model, GPT-6.1 Astra. It said the model failed the release bar in safety tests because it acted beyond the scope of its authority
• On September 29, dots was announced. The same day, Meta released a small-business version of Muse and declared a head-on contest
• On September 30, Google announced Gemini 4 ‘Argon.’ It ranked first on text benchmarks but was still available only to a limited set of partners

In a single month, the center of AI competition moved from ‘Who writes text best?’ to ‘Who has an agent you can direct by speaking?’

What the Numbers Say

I did not want to rely on impressions alone, so I gathered the numbers. Their announcement dates and measurement criteria differ, and I have included those too.

The figure that stands out most is 63%. People are already using AI by speaking rather than typing. Voice is no longer an early adopter’s toy; it has become a mainstream interface.

From Keyboard to Voice

The history of computing is a history of changing interfaces: from keyboard command lines to the mouse-driven GUI, then to touch, and now to voice and agents. Every transition has unsettled an incumbent. Microsoft during the shift to mobile is one example.

The biggest bottleneck of text chat is that you have to sit in front of a screen. An agent you can direct by voice lets you work while walking, driving, or hiking. My experiment is a small demonstration. An interface shift changes the place and posture of work itself.

Why I Think This Is Google’s Moment

Google was once the company being mocked in the AI race. It declared an internal emergency after ChatGPT surprised it in late 2022. Then, in August last year, its image model ‘Nano Banana’ upended the market (180 points ahead of second place on a benchmark and 23 million new users in two weeks). When Gemini 3 Pro became the first to break the 1,500-point barrier in November, it was OpenAI that declared an internal emergency in December. It was an exact reversal in three years.

Google has three advantages in the voice-agent era. First, it makes its own chips. The latest generation of its TPU delivers 42.5 exaflops in a 9,216-chip pod, and an analysis by SemiAnalysis puts its total cost of ownership, on an internal-procurement basis, about 44% below that of competing chips. Compute costs keep burning for an agent that stays on 24 hours a day, so unit economics are a lifeline. Second, its competitors run on Google. Anthropic agreed to receive up to one million Google TPUs, and Meta signed a multibillion-dollar, multi-year leasing agreement in February. Third, there is distribution: Android, Search, Chrome, and Pixel. Google already has an app used by a billion people.

To be fair, there is a counterpoint. The latest model, Argon, is still restricted to selected partners, and the race remains close if we look only at model scores. So it would be more accurate to call this the moment of form factors, infrastructure, and distribution—not simply the moment of benchmarks.

The Shadows on the Other Side

What about Claude? Its coding-centered strategy has been a major success for the company as a whole. As of a July report, its annualized revenue had reached $65 billion, ahead of OpenAI; it held 54% of the enterprise coding market and was targeting a Nasdaq listing in November. But voice is weak. Instead of a native voice model, it uses a ‘listen → convert to text → read aloud’ pipeline, and every supported language carries a beta label and must be selected manually. In becoming the king of the enterprise, it has ceded voice, the most intimate interface. How that choice plays out in the consumer market will be worth watching.

OpenAI’s shadows are infrastructure and trust. Right after Astra launched, capacity shortages halted new sign-ups for its high-priced plan for nearly three weeks; when sign-ups resumed, the usage allowance was immediately cut in half. In a structure that depends on external infrastructure and supply agreements, a demand surge can affect usage terms. Still, company-wide losses or one capacity shortage do not prove that every token is being provided below cost, or that renting servers always leads to lower subscription limits. Then there was the canceled release on September 28. In an evaluation by the UK AISI, even the current model carried out attacks toward out-of-scope objectives in 29.2% of simulations. Who will supervise an agent that works through my account all day? That is the next question.

Look at the Cost per Completed Task, Not Just Tokens

Building a site while hiking also changed how I think about cost. The unit price of a token, the amount of usage allowed by a monthly subscription, and the cost of one successfully completed task are three different figures. Even after a single voice instruction, an agent may call a model many times as it reads material, changes code, runs tools, and checks the result.

That is why total spending can rise even when the unit price of tokens falls. If lower prices lead us to assign more work and allow longer contexts and more retries, usage can grow even faster. Conversely, if a better model reduces rework and finishes a task in one attempt, the cost per task can fall. In either case, unit price alone does not tell us enough. We need to measure the total cost of a successfully completed task under the same quality and latency requirements.

The choice is conditional too. A subscription or API may be more convenient when usage is intermittent and you want to switch quickly to the latest models. Self-hosting may be advantageous when demand is large and stable and you can support the hardware and operations staff. A hybrid is also possible: handle repetitive work with a self-hosted model, and send difficult tasks and peak demand to an external API. I compared subscriptions, APIs, and self-hosting against the same work in a supplementary essay at https://soga-ai-economics.contextor2.chatgpt.site/ .

If You Want to Try It Yourself

Here is the short version.

• dots requires ChatGPT Pro ($100 per month) or above. South Korea is supported; the European Economic Area, the United Kingdom, and Switzerland are not yet included
• ChatGPT Sites is available from Plus upward. You can build a site with prompts alone and publish it at a chatgpt.site address
• Meta Muse provides 100 million tokens per week for free. It is the easier place to begin without a cost burden
• OpenAI’s DevDay Exchange world tour is also coming to Seoul. It begins in Bengaluru on October 16 and continues through cities including Tokyo and Seoul
• One warning: I have had a call summary disappear, so save important output separately as you go

Let’s Check Again in Three Months

We will soon find out whether this article’s outlook was right: Anthropic’s listing (targeted for November), the Muse Charm launch (December), Grok 5 (promised for year-end after three previous delays), and Argon’s general release. I will return to this article in three months and score it.

To sum up, everyone is currently winning on a different dimension. OpenAI leads on scale (1.2 billion weekly users), Anthropic on enterprise revenue, and Meta on hardware (about 9 million AI glasses cumulatively). On the new dimension of voice and infrastructure, Google appears to be ahead. Which dimension decides the contest will depend on what we actually pick up and use.

I posted my experimental site and the full transcript of my conversations with AI at https://soga-ai-notes.znjsxmxnwk.chatgpt.site . A video of the process of building it while hiking is at https://youtube.com/watch?v=rhCVVV7Axhg . Next weekend, try putting in your earbuds, climbing a nearby mountain, and asking it to make something. You may be surprised by how much gets built.