When GPT-4 (?)1 was released, they teased at the model’s ability to help make and edit slides. As someone who spends a painful amount of time on slides (record so far is 2 months for a 45min talk), I was excited and very much looking forward to trying it out. Here we are in 2026, and LLMs have yet helped me with making slides in any significant capacity. This probably has something to do with my presentation style, and I thought it’d be interesting to write a non-technical blog to document how I make slides and my current experience with using AI tools for such tasks.

Why making slides takes me so much time and is difficult to automate

I make slides with keynote, which is known for being unfriendly to equations (feature, not bug). A few common time-consuming tasks I run into:

  1. Visualization, usually involving concepts such as distributions, vectors, matrix factorization, etc. I use animations whenever it helps to show connections between concepts; in the example below, the green and blue vectors on the left will move to the right figure to show connection between the density space and the feature space.

  2. Color coding: if I put equations on the slides, I will often use different colors to highlight the important terms. Text descriptions of a mathematical term and its visual representations should use the same color.

  3. I use a lot of free-form textboxes to provide definitions and remarks, as they are generally more effective and take less cognitive capacity for the audience to process than an inline bullet point.

  4. Maintaining consistency: if I refer to a concept mentioned in an earlier slide, I often copy-paste the corresponding visualization to the later slide as well, and make sure they look exactly the same. But if I need to update the earlier visualization, then I need to update the later one as well. Similar problems arise when I have too many animations (e.g., old contents disappear and new stuff come in) that I have to split a slide into two, then the overlapping parts need constant maintenance to ensure consistency.

  5. Other tweaks: baseline and size adjustments for inline equations, moving objects into foreground/background to make sure occlusion works properly, etc.

With the level of detail I attend to, I don’t think current LLMs are able to significantly improve my efficiency. In fact, I asked ChatGPT-5.5 to recreate a slide from screenshot, and the following is what I get (left: medium; right: extra-high)—not quite there yet, but better than I expected. Interesting that ultra-high did exactly what I did to create the blob for D, i.e., ellipsoid + white box. Plus, neither gets italics fully right (variables italic, brackets non-italic). Also note that this is the outcome when the final appearance of the slide is given in screenshot, and I imagine that the task will be more challenging when I need to specify the slide (for which I may not have a concrete picture in mind) via prompts.

After all, I am not even sure if better LLMs applied to the current software are the right solution; some of these issues should probably be addressed within the keynote software:2 for example, equations copied from LaTeXiT are always higher than the surrounding text that I constantly need to adjust their baseline (and then adjust again for the subsequent text), which seems to be easily fixable for keynote. Perhaps we can hope that AI can accelerate personalized software development, so that it is easier to develop presentation tools suitable for highly visualized and animated technical talks.

Is it worth it? Some might question whether this level of obsession is necessary. Delving into the philosophy of scientific communication is perhaps for another day, and for now I will just say: (1) audience easily get hang up on completely irrelevant minor details, (2) people often (vastly) over-estimate how much can fit into the short-term memory of the audience, and I am willing to spend 30 min tweaking the slides to ensure a smooth 5 sec on the stage. But most of the time I am just fighting with something super stupid, like dragging a shape so that its boundary aligns with another line, which is what I hope LLMs can help with. So far, my conclusion is that this seems very well within reach, but the current tools are not quite there yet.


Footnote Remarks

  1. I can’t remember exactly; could very well be GPT-3.5. ↩

  2. When I was an undergrad, the professor who taught Signals and Systems had fancy slides, which were made by a software tool he wrote himself. ↩