Perplexity Comet: A magical experience
August 16, 2025I just got access to Perplexity Comet - thanks to Accel being an investor - well-timed given that I just read Dharmesh's July 17 blog post about his own Comet experience, I was keen to try it out.
In short, it's an absolutely magical experience.
Like Dharmesh (we go way back :P), I spend ~80% of my work time on a browser, with the rest split between ChatGPT (my AI assistant of choice), Zoom (meetings), Notion (notes) and a few tail-end apps (Excel, Powerpoint, etc.). My role involves a lot of repetitive, low IQ but detail-sensitive tasks such as scheduling meetings with founders that are confirmed on email/ LinkedIn, updating meeting details (time, invitees, location, emails, etc.), logging founder interactions on CRM (when I remember to), etc. Perplexity Comet abstracts all of this just through prompts - magical! Two simple examples:
- I had a call scheduled with a founder but did not have his phone number. I emailed him, got his number, gave perplexity a prompt to read my email, take his number and update my meeting invite header with the same - and it did just that!!
- I asked Comet to scan my LinkedIn DMs for founders I've forgotten to respond to and list them with their company website. It produced a concise list of founders and their company websites which probably reduced 10-15 mins of my work of doing this manually
This has been my most frictionless AI experience as of yet, it has access to my data stores (Gmail, Calendar, LinkedIn) and it can actually follow directed actions from me taking tasks to completion! I think the biggest alpha generator from using Comet is that it saves cognitive load vs just time. These tasks are low-IQ but detail oriented, hence they consume units of cognition (is there an actual term for this?). This is the precise level of abstraction I need, the right balance of assistance and control.
However, simple searches for web artifacts, that I know exist on the web, feel painfully slow. Overall, latency is on the higher end and tasks seem to run serially vs in parallel - likely due to the non-determinism of consumer tasks makes them hard to parallelize, or maybe there lies an opportunity here.
I am very excited to continue using Comet and to start building (more to come soon)!