Hacker News Reader

Comment by Sherveen | Hacker News Reader

People keep asking for this w/ Codex, too, and I really regret that both labs seem inclined to listen.

If you ask a thinking/research type question in 'Chat' versus 'Work' mode in these products -- say, something complex about politics, or do a multiturn business strategy, or want to work thru a new concept, you get very different answers.

The harness, steering, etc. in the chat/reasoning products is so much better for this type of question (that doesn't require code as a primary substrate).

Even as someone who mainlines like, 7 coding agents at all times, I regret that productivity fever will mean the regression of think-first-act-later AI UX.

View on HN

I hope not Chat is still all you can eat and Work uses Codex tokens. I can see why they would want people to think they wanted this without thinking about how it currently works. If Chat starts eating tokens then there is not much reason to use Chat other than most people don't live in the command line like developers. Claude Chat / Cowork already eats tokens either way so not much of a change really. I have never used Work since if I want access to local files the CLI is a much better interface combined with an IDE, but can see the appeal to non-developers.

This is my main reason for not using Work. I was tempted to use a few "scheduled tasks" in ChatGPT, but once I found out they used Codex tokens I stopped creating new ones.

I already burn through Codex too fast.

I recommend using Work sparingly because sometimes it is the better choice, but I do ration it. With Claude I simply don't use the web/app chat at all (and I have no idea what Cowork even does ngl, never used it) I think there's a Claude Design? No idea why I would need it (mostly, why I would spend tokens on it lol)

Claude Design is good at making a bunch of UI pages in a single design look consistent, and it's better at making an initial starting UI than just asking Claude Code. Think of it like Anthropic's version of Figma. It honestly works pretty good and I use it often and feel that it saved me a lot of time. It's pretty token heavy though.

I built a terminal control plane and added an MCP server that runs on my Mac. It lets MCP clients create terminal sessions, send messages, read the contents, etc on my Mac.

The terminal sessions are running Luna on XHigh (on Pi) and GPT 6 Pro on ChatGPT.com is the main orchestrator.

Since ChatGPT and Codex have different limits, I can have 6 Pro orchestrate the Luna sessions. Using voice mode to talk back and forth about the software factory is really nice. Will be open-sourcing soon.

The OpenAI folks have been clear that they are merging them by the end of the year, obviously TBD what they do with quota.

I like the web chat UI for quick one off questions where I don’t want it to have the context of a local project I’m working in.

GPT Chat seems smarter for business questions or planning that does not require iteration or wrangling data sources.

But Work gives the model a computer instance it can use for programmatic data access, API munging, etc.

Success cases I’ve had: chat - read the reports in my Gmail and put a counter proposal together.

Work - cross reference all the options with commute time using traffic aware map data. Ground it on drop off / pick up times from each option’s stated rules and hours. Create a table with commute and idle time break downs.

The fact that it doesn't work well today doesn't invalidate the user need. Not everyone wants a terminal style interface. Traditional UX makes more sense especially if say you are accessing existing plugins and say kicking off a scheduling task - no way would I want to do that via a chat interface if I have the option.

Agree, and it allows ordinary users to have a taste of both worlds, instead of being like: "eww, what is this black box and why doesn't my mouse work"

I agree that they should be separate modes, but presenting it as a toggle in the app is a terrible user experience.

The model itself should figure out how much “power” it needs to run the query. I’m just asking a question. Whether you need to do a simple internet search or spin up a VM and run scripts is up to you.

I see Work as the mode where you can generate and edit files without manual intervention.

I tell chat to recommend me some books based on some of my favourites. Chat does it. I read some of them, liked some, disliked few. I give chat that information, it recommends me based on those, now I don't know which of the entries are relevant or not.

I don't want lock-in, so I try my best to tell it to generate copy-able MD syntax that I have to manually paste to an editor.

With Work mode, I tell the model to just maintain an append only markdown file, where it writes "still recommended", "not recommended any more", etc. beside titles. It does this after each turn.

This is just an example. Work - I see it as a mode that can maintain, update, and edit a file.

It's very convenient.

> The harness, steering, etc. in the chat/reasoning products is so much better for this type of question (that doesn't require code as a primary substrate).

I had this same thought recently. Is it possible to 'recreate' ChatGPT/Claude Chat easily in Codex or Claude Code? I assume it would roughly be just swapping the system prompt. Anthropic seems to publish theirs but I don't think OpenAI does

https://platform.claude.com/docs/en/release-notes/system-pro...

I sometimes get very different answers if I ask the chat product the same question multiple times.

Isn't that expected since LLMs are inherently non-deterministic?

LLMs are chaotic pure functions. Their input is usually randomized.

There's deterministic enough, and then there's computer science non-deterministic. If I ask for a Todo app, I'm going to get a Todo app, even if the buttons get moved around and the background color of it is brown instead of purple if I ask today vs 6 months ago. If the AI completes the phrase "the capital of France is..." with anything other than Paris, something has gone more wrong than usual.

LLMs are inherently deterministic. The way everyone deploys LLMs leads to non-deterministic results, but there's nothing† stopping providers from offering deterministic evaluation if they choose.

All the sources of randomness are under the control of the provider, even if today's deployment structures mean providers introduce extra randomness due to the concurrent nature of the evaluation. Serialise the computation, feed it from a pRNG, and you have a fully deterministic result. But providers don't want to offer a deterministic result, and especially not one as fragile, expensive, and inefficient as a full serialisation would be.

†: For variants of "nothing" that include cost and deployment challenges.

This is technically true, but when people talk about randomness, its not only about same input-> different output, like temperature>0 and the things you said.

Its also about very similar inputs -> different outputs. Even with everything you said, yes, same input would result consistently into same output, but sliightly different input and you might get completely different/semantic answer.

That's "chaotic" rather than "non-deterministic".

In one of the earlier iterations of this they at least made the UX distinguish Codex vs. ChatGPT which I really enjoyed.

I must admit I spent a little time in ChatGPT this weekend, thinking I was in Codex. After some unexpected behavior (disappointed) I realized my mistake.

The fact that people asked for it means that there is a demand for it. Maybe you aren't the target audience? One size can't fit all. People just have to adjust and go.

Yes the point is that they wouldn't have to adjust unless they are merged into one. Precisely because one size can't fit all, yet they insit on one size now.

Reply on news.ycombinator.com