# AI Keeps Amazing Me. I'm Not Sure We're Ready.
**Summary:** AI progress excites me, but fast adoption and incomplete evidence leave me unsure how prepared we are. A personal reflection on why I still want to move forward.
- Canonical: https://markhuang.ai/blog/ai-keeps-amazing-me
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-11-01
- Section: Behavior & Mindset
- Tags: ai-adoption, ai-safety, generative-ai, long-context, jev
- License: https://creativecommons.org/licenses/by-nc/4.0/
---
*The nearby road is clear and the distance is obscured. That is close to how AI progress feels to me.*
AI can help write code, generate video, and make music. I am impressed by all of it. Each capability makes me wonder what else I could build, or what I could try without spending months learning another tool first.
What unsettles me in 2026 is that we are making these systems part of ordinary life while still studying what they can do and what their consequences might be. That includes the companies developing them, researchers, governments, and the rest of us trying to decide how much to trust them. I still believe AI can make life better. I have less confidence that our preparation is keeping up.
The picture in my head is a car accelerating while the road ahead is only partly visible. Sometimes that feels exciting. Sometimes I find myself bracing for an impact I cannot describe. It is a feeling, though, and I need to be careful about treating it as evidence that we have already lost control.
## The changes I can place on a timeline
When I think about technology since the mid-1990s, I can put names to different periods: Windows 95, Nokia phones, the iPhone, and eventually AI tools helping with code. Looking back, those changes feel easier to separate. That does not establish that progress was steady, or that people understood each technology before using it.
```mermaid
flowchart TB
accTitle: Selected technology milestones from 1995 to 2021
accDescr: Windows 95 launched in 1995, Nokia announced the 3310 in 2000, Apple announced the iPhone in 2007, and GitHub Copilot entered technical preview in 2021. These selected events do not measure the rate of technological progress.
W["1995: Windows 95 launches"]
N["2000: Nokia announces the 3310"]
I["2007: Apple announces the iPhone"]
C["2021: GitHub Copilot
Technical preview"]
W --> N --> I --> C
```
*Selected reference points, not a measurement of progress between releases. Sources: [Microsoft](https://blogs.windows.com/windows-insider/2020/08/24/looking-back-the-25th-anniversary-of-windows-95/), [Nokia](https://www.globenewswire.com/js/news-release/2000/09/01/1845525/0/en/Nokia-introduces-mobile-chat-with-the-Nokia-3310.html), [Apple](https://www.apple.com/newsroom/2007/01/09Apple-Reinvents-the-Phone-with-iPhone/), and [GitHub](https://github.blog/news-insights/product-news/introducing-github-copilot-ai-pair-programmer/). Arrows indicate chronology; spacing does not represent elapsed time.*
The recent AI sequence feels different to me because the kinds of work involved keep widening. A conversational answer, a generated video, and a tool acting on files ask for different kinds of trust.
```mermaid
flowchart TB
accTitle: Selected AI milestones from 2022 to 2026
accDescr: ChatGPT became public in 2022. Anthropic introduced Claude in March 2023, and Google introduced Gemini in December 2023. Gemini 1.5 Pro previewed a context window of up to one million tokens in February 2024. Google announced Veo in May 2024. Claude Code entered limited research preview in February 2025, Cowork launched in research preview in January 2026, and TypeSafe introduced Jev in early access in September 2026.
A["2022: ChatGPT
Public research preview"]
B["March 2023: Claude introduced"]
C["December 2023: Gemini introduced"]
D["February 2024: Gemini 1.5 Pro
Up to 1M tokens in private preview"]
E["May 2024: Veo announced
Private preview for selected creators"]
F["February 2025: Claude Code
Limited research preview"]
G["January 2026: Cowork
Research preview"]
H["September 2026: Jev
Early access"]
A --> B --> C --> D --> E --> F --> G --> H
```
*Examples of expanding capabilities, with their original access stages. This is a selection rather than a complete release history, and the gaps are not to scale. Sources: [ChatGPT adoption reference](https://www.stlouisfed.org/on-the-economy/2025/nov/state-generative-ai-adoption-2025), [Claude](https://www.anthropic.com/news/introducing-claude), [Gemini](https://blog.google/innovation-and-ai/technology/ai/gemini-collection/), [Gemini 1.5 Pro](https://blog.google/innovation-and-ai/products/google-gemini-next-generation-model-february-2024/), [Veo](https://blog.google/innovation-and-ai/products/google-generative-ai-veo-imagen-3/), [Claude Code](https://www.anthropic.com/news/claude-3-7-sonnet), [Cowork](https://www.anthropic.com/news/introducing-anthropic-labs), and [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev).*
The context-window example helps explain why this feels consequential. In February 2024, Google offered selected developers a private preview of Gemini 1.5 Pro with up to one million tokens of context. That means more information available during a request. It does not mean permanent memory or reliable understanding of every detail. Still, it changes how much material someone might give a model at once.
A builder could include a much larger set of documents instead of choosing a few excerpts. That may help an answer, and it can also change how much private context crosses a boundary in one request. I find that combination hard to separate: the feature that makes a tool more helpful can make the disclosure easier to overlook.
There is evidence of rapid adoption beyond my impression of release announcements. In their November 2025 update, researchers Alexander Bick, Adam Blandin, and David Deming reported that 54.6% of U.S. adults aged 18 to 64 used generative AI in August 2025, up from a revised 44.6% a year earlier. That is a dated U.S. survey finding, not a measure of how fast all technology is evolving. It does help explain why the adjustment can feel immediate. [Their analysis describes the survey and revision](https://www.stlouisfed.org/on-the-economy/2025/nov/state-generative-ai-adoption-2025).
I have already written about [the privacy trade I make while trying to automate more of my work](/blog/aggressively-trying-to-replace-myself-with-ai). Context makes AI useful, so sharing more becomes tempting. I do not think privacy stopped mattering. I worry about how quickly an acceptable disclosure becomes a habit, especially when the information concerns somebody else.
## The part we are still trying to understand
My first reaction is to ask why we released something before fully understanding it. But that question needs an honest objection: studying a technology after release does not, by itself, establish that the release was irresponsible. There will be uses and interactions that testing did not cover.
Public use can teach us something that a laboratory setup misses: which tasks people actually delegate, which explanations they trust, and where a small error becomes a repeated habit. I want that knowledge to exist. My hesitation is about how it is acquired. A problem discovered in a voluntary experiment can have a different cost from the same problem discovered in a service somebody depends on.
Nor would it be accurate to say that nobody prepared. Anthropic describes pre-release testing, external evaluations, and ongoing monitoring in its [transparency documentation](https://www.anthropic.com/transparency/voluntary-commitments). The existence of those processes does not establish that they are sufficient. It does give me something more concrete to examine than an assumption that everyone is guessing.
The [2026 International AI Safety Report](https://internationalaisafetyreport.org/publication/2026-report-executive-summary) describes an evaluation gap: performance in tests does not reliably predict real-world usefulness or risk. In a May 2026 report, independent evaluator METR also described how fast launch schedules can leave little time for thorough analysis. These are limitations in assessment, not proof that every released system is unsafe. They are still a reason to question how much assurance a successful test should give us. [METR explains the limits of its usual pre-release evaluations](https://evals.alignment.org/blog/2026-05-19-frontier-risk-report/).
When I ask whether we are prepared, I am also mixing several questions. A company can test a model, while a school still has to decide how its students should use it. An employer can buy an assistant before people agree on who checks its work. Governments face questions about what evidence to ask for and how to respond when harm crosses borders. Those are different kinds of preparation, and I cannot read one model report and infer that all of them are in place.
Unpredictability complicates that question because sometimes I want it. If I ask for a melody or an unfamiliar design approach, a surprising result can be useful. If I ask which methods an existing library provides, an invented method is a factual error. NIST describes confidently presented false content as [confabulation, often called hallucination](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf). The value of a surprise depends on the task and on what someone does with the answer.
Even perfect factual accuracy would leave other problems. A voice generator could reproduce a voice accurately and be used to impersonate its owner. That harm would come from the capability working, rather than from a hallucination. I cannot reduce the uncertainty around AI to whether its answers are correct.
TypeSafe's Jev interests me here. Introduced in early access in September 2026, it returns typed decisions and probabilities instead of free-form text. That is a different approach to putting AI inside software. Restricting possible outputs can make an interface easier to use, while leaving the question of whether a decision is right. TypeSafe's [confidence documentation](https://docs.typesafe.ai/confidence) explicitly discusses uncertainty and testing thresholds against a particular use case. I want to see where approaches like that help.
My unease grows when the consequences reach people who did not choose the tool. Someone can avoid a chatbot and still receive an automated decision or be targeted with generated content. A provider learning more about its model and a person absorbing the cost of that learning are having different experiences.
## We have learned after the damage before
I cannot fairly contrast this with a past in which we understood the dangers first. The email security history is uncomfortable for that argument.
After the ILOVEYOU outbreak in 2000, Microsoft announced an Outlook security update that would block certain executable attachments, warn when external programs accessed contacts or sent messages, and tighten default security settings. Microsoft acknowledged that the changes would limit functionality people used. [Its May 15 announcement describes the restrictions](https://news.microsoft.com/source/2000/05/15/microsoft-to-deliver-major-outlook-security-solution-to-help-protect-customers-against-computer-viruses/).
Antivirus protection already existed. These were additional defenses developed in response to damage, with their own cost in convenience and compatibility. That history gives me less reason to romanticize earlier technology. It also explains why having a response available is different from preventing harm in the first place.
The difficult question is how well that pattern carries over. Some AI mistakes can be corrected before anyone relies on them. Others can spread or become actions before the person affected has a chance to respond. Being able to fix a service later does not undo every consequence of using it today.
I have not arrived at one release rule that resolves those differences. I also cannot assume that withholding a useful capability has no cost. There are benefits people could miss, and real knowledge that comes from use. That is what makes the question harder than choosing between trust and irresponsibility.
## I still want to see what comes next
I remain interested in what AI can make possible. The excitement is part of why I feel responsible for questioning it. I am participating in this change, and my own enthusiasm can make an uncertain trade feel easier to accept.
I believe disagreement can improve what we build. I want competition where people can challenge a claim, test an idea, and share a result that did not work. Cooperation matters too. A useful correction should be able to travel beyond the team that discovered the problem.
That does not settle how quickly to release a model, how widely to distribute it, or who should make those decisions. It does leave me wanting room for questions after the excitement of an announcement, including questions from people who will bear consequences without getting much of the benefit.
I still want to try the next useful thing. When I picture that road, there is plenty ahead that makes me curious enough to continue. I just do not have a confident answer to the question that keeps coming back: what would being prepared look like when we are still discovering what we have built?