返回资讯中心
外部精选
软件工程
#开发者实时流

ChatGPT Dot Review: Your Chief of Staff Is Still an Intern

🧱 I’m Mete, a product designer, builder, and inventor with 10+ years of experience at companies like Netflix and Peloton. I write to help connect the macro and the micro forces behind AI - from ways of thinking to ways…

Hacker News 最新young_mete2026/10/15 · 2026/10/1530 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Mete Polat

Oct 15, 2026

1

1

Share

🧱 I’m Mete, a product designer, builder, and inventor with 10+ years of experience at companies like Netflix and Peloton. I write to help connect the macro and the micro forces behind AI - from ways of thinking to ways of working. If you’re looking to understand and use AI beyond the surface-level, you’re in the right place.

I’ve been test-driving Dots daily since its release last week. As someone who’s been using Hermes and test-driving lots of other agentic consumer products, I had a clear need and a place in my workflow for something like Dot. I loved experimenting with Hermes, but ultimately most of my work still happened inside Codex. Over time, my life got integrated deeper and deeper into Codex to the point where it was more useful, convenient, and present for me than Hermes, which required constant debugging and optimization.1

Yet, Hermes gave me a taste of what it’s like to have one thread as an agentic abstraction for your computer. It was so clearly the right model that it was only a matter of time before every frontier lab introduced its own version of it. It didn’t take long for the avalanche of products to hit the market - GrokBot, Gemini Spark, Instinct, Muse, and now Dot (and dozens of other smaller players). They’re all an embodiment of the same fundamental idea - a personal agent is the new abstraction layer on top of your technology. Having used a lot of these products, I have little doubt it’s the future - for tech broadly and for OpenAI specifically (more on that below). But does it live up to this promise in the present moment? Not quite. Or at least, not yet.

From a product perspective, Dot right now is basically a light wrapper on top of Codex rather than a rethinking of what ChatGPT can be as Dot (more on that below).

From a more general usefulness perspective, Dot is like a PhD-educated intern. They don’t know where anything is in the system. When they mess up, they over-communicate and provide overly technical excuses. They occasionally drop the ball and completely forget stuff. But you can clearly tell they’re brilliant and destined to run things around here. But for now, you’re babysitting.

The best framework that bridges the past and the future here is Jobs-To-Be-Done: customers don’t buy your tools - they hire them to get a job done. Each SaaS app was effectively a toolbox. Our job as designers was to understand the job at hand, surface the right tools in a given context, and make those tools easy to use so you can accomplish the job. With LLM agents, a SaaS app is now more akin to a coworker that goes out and does the job. And how do you talk to a coworker? You chat with them.

In other words, it’s an entirely new kind of interface. We’re handing the vast technological toolbox to another entity that can use those tools better than us. More than that, they will soon be able to better anticipate what we need before we know it and discover needs we didn’t know we had. It doesn’t mean that we’ll never touch the tools ourselves; we may still need to tinker in the shed alongside our agents. But now they’ll be able to forge us tools on the spot that are perfectly tailored to us and the task at hand.

The simplest metaphor to understand it is like having your own personal team of assistants with their own computers. Many early agentic products like OpenClaw and Hermes showed the potential of this idea, but Muse was the first mass-market player that gave everyone their own persistent computer in the cloud that their agent can use. OpenAI’s Dot followed suit - it has its own cloud computer as well as the ability to control your local machine.

The problem with the OpenClaws of the world is that, while the promise is vast, the work it takes to set up, maintain, and evolve is also vast. That makes it truly useful only to a small subset of power users who either enjoy the pain, value portability, or have business workflows that offer some return on the time investment.

The new generation of consumer-friendly assistants largely eliminates this complexity and provides a smooth paved road you can ride. You can now focus on the destination instead of tuning your setup, studying confusing maps, and feeling every bump in the road.

This leaves a lot of people with a question - what is the destination? In other words, what do I use these things for? This is where a large imagination gap opens. Very (very) roughly, there are 3 distinct groups of people here that differ in how much potential they perceive there to be:

People who are privileged enough to have their own personal assistants

Group #1 already knows how much their life can be augmented by having a dedicated assistant (or a team of them). Group #2 is just trained to see every gap in life as a hole to fill with software, so that imagination naturally extends to AI agents. Group #3 is 99% of humans who have zero idea what this technology has to offer them - they have their way of doing things, they (mostly) work, so an AI assistant seems like a fancy vitamin vs a painkiller. As Ben Thompson argues, they’re just not interested in personal productivity - they just want to be entertained (OpenAI’s demo use case choices also reflect that they don’t get or want to accept this part).

From my experience, Muse is built for group #3. I talk more about it in my “What Muse Got Right” breakdown and I might put out a dedicated review soon (let me know in the comments if you’re interested in that).

Dot is for group #2. If you’re already a Codex/ChatGPT power user, the needs are clear, pain relief is instant, and the potential is infinite. And so are the frustrations.

Parallelism

As the models got better, I’ve generally found that the way I work with AI has slowly shifted from very structured and mediated by the interface to increasingly fluid and unstructured. Dot has played into this transition really well for me.

Instead of trying to find (or build) a perfect interface to manage all the threads, structure my prompts, design my context, etc., I embrace the single ongoing thread with voice notes and calls with all my thoughts mashed together. I can tell it about my idea for an article (so it can create and file a note), tell it to start my Japan visa process, while also rambling through some feedback for a landing page I’m working on. My Dot (named Satori) is able to keep track of it, extract the actionable parts, and spin up separate threads for all of it.

On the one hand, I feel like this emboldens some deep-seated ADHD tendencies I already have and lets me work on more things at once, faster. On the other hand, I was going to try anyways and it would have been more painful. In that sense, Dot has been helping me unburden my mind in a way I haven’t experienced before.

In my article “The Chatbar Isn’t Lazy Design”, I argued that proactivity is a natural next step for agentic AI:

This “coworker” mental model neatly and naturally extends into the future:

The best coworkers are proactive. And so the best agentic SaaS will be proactive - it will anticipate your needs, catch issues before you do, and go out and solve them, only looping you in when needed.

Dot has already started delivering on this. For example, it spotted a discount code I hadn’t used on an order. It emailed the company on my behalf and got a partial refund. All I did was say “Sure!”

This is just one of the examples, but there are plenty of smaller ones - it highlights calendar conflicts I may have missed and reminds me about some things that I’ve ignored for too long. Like a real assistant. I also generally appreciate that it’s not trying too hard. In other words, it’s not spamming me with recommendations just to get engagement (I’m sure this won’t last long and may get ruined the same way push notifications have been ruined for us).

In some ways, a personal AI assistant is the ultimate realization of an “Important” filter. Each app in our lives tries to surface the most important information to us in its own way - email inbox categories, priority notifications, etc. They fail mostly because (a) they fall prey to engagement incentives (looking at you, LinkedIn) and (b) they just don’t know us and what’s important to us. The AI assistant solves perfectly for (b) because it knows everything about us. As for (a), time will tell.

OpenAI’s launch video for Dots reflects their idea of how we’ll interface with their agents and it’s mostly through calls:

While most tech company demo use cases these days seem disconnected from what people actually care about, I think they’re onto something here. When they worked, I found the calls to be delightful and extremely useful.

Like many of us in the modern world, I have a hard time shutting my brain down. I’ll keep thinking through problems, to-dos and plans on my walk, while doing chores, etc. It’s not a bad thing in and of itself, but at some point you need to touch grass. I’ve found that calls help me unload my mind while doing something else so I can actually move on from it. So now I’ll often spend 10 minutes on the morning dog walk talking through my plans for the day, knowing my Dot will act on all of it by the time I’m back home.

That said, the call feature is still extremely rough around the edges:

It was mostly broken for the first few days. And because they made it so it fake rings for a good 10-20 seconds before trying to connect, it felt like being ghosted by an emotionally unavailable girlfriend.

It’s not good with silence. If you’re taking a second too long to articulate a thought, it’ll chime in with some useless affirmation. And then, after it tells you it’ll go do something, it just goes silent for minutes until you ask it what’s going on. It still doesn’t feel quite natural.

It couldn’t get my name right for the life of it. I’m not a stranger to this problem at all, but what’s really frustrating is that it would get it right occasionally. Which means it can absolutely say my name right! Each time I corrected it, it would apologize, assure me it remembered it this time, and then greet me with “hey meat” on the next call (my name is pronounced meh-teh). It’s a small thing, but this made it feel less personal.

Is it something I said?

Where It Struggled

The Interaction Model (With No Mental Model)

It seems pretty obvious that Dot has been bolted on top of ChatGPT and Codex using their existing primitives. What this means is that the Codex / ChatGPT app is still an absolute mess:

Would you like a toggle with your toggle?

While the intention is for you to use Dot to streamline your experience, in reality it just added another choice to the plate. Should I start a new chat (chat or work) or a new Codex thread (local or cloud) or a chat with Dot? And if it’s related to a project I created, will Dot know about it? Does Dot have its own memory or is it using the overall ChatGPT memory? All these things have answers, but the point is that it’s too many questions to ask.

It’s clear that OpenAI is trying to help me out by adding my Dot’s icon into general ChatGPT chats but it feels like they’re using a bucket to bail out my ocean of confusion.

This is a regular ChatGPT chat thread that surfaced my Dot’s figurine next to tool calls. Does it mean it’s my Dot working on it or…?

I suspect that OpenAI is just being conservative here and wants to integrate Dot gradually instead of blowing up the ChatGPT app and messing with people’s ingrained habits. Instead of rethinking from first principles what ChatGPT would look like as a personal assistant, they just threw it in as yet another tab of the app.

I’m sure they considered whether to go for a big redesign or even release Dot as a separate app. Contrast it to Muse, which was designed from the ground up instead of throwing it into, say, WhatsApp (like they tried with Meta AI). Instead of ripping off the Band-Aid, OpenAI chose what they probably think is short-term confusion so they can be lean,

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →