Skip to main content

Log in to comment and vote

Comments6

  • Gray Crater

    implement BOTH $ and \( and )\

    Default should be $ anyway. That’s what most renderers render.

    Or go for one, and convert at your side.

    • Mark Florkowski

      Team

      We already take the approach of “go for one, and convert at your side.” we do not use $ … $ because it itnerferes too much with normal responses such as price ranges, and math rendering is a niche feature compared to the general use case of showing a price range.

      You can try adding some custom instructions in settings to remind the models to use the \( … \) formatting which often helps. We don’t do this by default because again it can affect other responses and the majority of users don’t end up using LaTeX anyway.

      • Gray Crater

        Math is less important than price ranges… We’re not going to fix this because most of our users add dollars, and math is for some minor majority of Geeks.

        I have to ask myself really hard why I’m using this.

        EDIT: go for one, convert at your side means:

        We detect the $…$ case the llm is output-in, and doing a seach-replace or in some way handling it so that the LLM out is rendered correctly.

        You are clearly NOT doing that.

        I’ve changed the custom instructions before I sent a bug feedback.

        I expected a comment about the actual bug.

        Besides, the whole argument with price ranges is bogus. You can be smart about it and fix.

        But claiming that you don’t care because it’s niche? Doing math work is niche for T3.chat… It is T3.chat that is niche, not math.

        • Mark Florkowski

          Team

          The issue you reported is real, and closing the report wasn’t meant to imply that math support isn’t important or that we don’t care about professional use cases.

          The reason we don’t currently convert $...$ by default is that $ appears very frequently in ordinary text, especially prices, currencies, shell commands, and code. A naive conversion causes incorrect rendering in many non-math conversations. We haven’t found an approach we’re comfortable enabling globally without introducing regressions. If we find an approach that does improve compatibility without regressing common cases, we’d be happy to revisit it.

          That doesn’t mean we’re opposed to improving it. It just means we don’t currently have an implementation that we think improves the overall experience enough to justify the tradeoffs.

          We did spend a fair amount of engineering effort on LaTeX rendering already because we do think it’s valuable. We simply prioritized supporting the syntax that doesn’t conflict with normal text by default.

          Your feedback about the bug report itself is fair. We could have been clearer that this was a prioritization decision and an implementation tradeoff, rather than implying mathematical workflows weren’t important

          • Gray Crater

            Thank you!

            This is a fair answer that I would have expected from the beginning.

            Look, I want T3.chat to grow and be good. I’m also using it to support Theo for his (mostly) fair and smart journalism on YouTube.

            I found a bug, actually traced down the root cause and reported it here.

            Your initial reply came across (to make crude analogy):

            Suppose I give a feedback to Facebook, complaining I can’t send a a complicated type message to my friend. And what I get from Facebook’s support is: Oh, why are you not watching cat pics? That’s what most of our customers do, we focus on cat pics. We won’t fix anything else.

            I understand, as a dev myself, that the request is not trivial to implement. I see regular expressions might not be enough for all edge cases. I understand that you may not have infinite resources, but please frame it like that.

            So please, kindly take care that your communication doesn’t make your customers feel like they’re doing something niche, unimportant, and not worthy of being taken seriously.

            Somehow, it is highly ironic for an AI chat platform to call math rendering “niche”, especially when technical users and developers are thought to be part of their core audience. Many devs, data scientists, students rely heavily on LaTex formatting.

            I hope this explains my initial reaction and frustration.

            It was the communication part.

            • Mark Florkowski

              Team

              Thanks, that makes sense.

              I think the disconnect was mostly around what I meant by "niche." I meant it strictly in terms of how frequently the feature is used relative to the rest of the product, not as a judgment about whether the use case or the users relying on it are important. While we have a lot of technical users and certainly care about serving them well, T3 Chat's audience is much broader than developers and technical folks. At this point, most of our users aren't particularly technical, and that necessarily factors into how we prioritize features and tradeoffs.

              Either way, I understand now how the original response came across, and I'll be more careful in my phrasing for the future.