hey! I saw this and liked it a lot! It’s impressive how you pull in all the routes per tail - we considered doing it but were worried it would be too expensive. Definitely opens up cool options though.
> how stale does the tail assignment data get in practice, and do you have a way to detect when an enthusiast spreadsheet goes unmaintained?
These are updated almost every day so far, so they seem very up-to-date. Internally we track all changes/removals, so I'm not that worried about spreadsheets being abandoned yet. It's a good thought though.
> And what happens to your probability estimate when an airline swaps aircraft last minute, which seems to happen pretty often on regional routes?
Honestly our estimate right now is pretty crude. At the scale we're at right now it works, but I think you're right that we could make this more accurate by tracking equipment swaps & really drilling into the details of which aircraft get assigned to which routes.
Oh that's a cool idea! We wanted to do a variant of this, will add it to the list. The tricky part for us is getting a canonical list of all flights + body types on it.
I’d imagine you could seed it more easily by focusing on top 50 routes by passenger count in domestic USA. Then go from flight schedules for top airlines into tail numbers into body types etc.
Took me a second to parse that. It says 28 of 28 E175s have Starlink, but what I am hearing you say is that Alaska has more than 28 E175s.
Indeed, wikipedia says their fleet includes 47 E175s. Consider my hopes dashed :(. Oh well, I don't usually bother with wifi on flights that are only a few hours anyway, but free Starlink speed wifi would be fun!
Feature request: Put a disclaimer on the fleet page that the tracking is limited. Or pull enough data to say "28 airframes of 47 are starlink capable" which is what I think most people will be looking to know in the fleet info.
> Feature request: Put a disclaimer on the fleet page that the tracking is limited. Or pull enough data to say "28 airframes of 47 are starlink capable" which is what I think most people will be looking to know in the fleet info.
Oh, this one is very doable and makes sense! We track this internally anyway so it's just a matter of surfacing it on the fleet information.
I'm not actually sure myself, but I was really surprised to learn how profitable it is. SpaceX made $15b of revenue last year and $8b of profit. Starlink was 60-80% of that!
It turns out the demand for really good internet everywhere is huge.
There were article claiming "$8b profit" but relabeling EBITDA as profit. EBITDA only tells you that Starlink makes money on a satellite once it is already in space and connected to a user. It deletes the cost of building the satellite, launching the satellite, the user equipment manufacturing, and just about all other substantial expenses. Not to mention payments servicing all their debt and Starship development.
The fact a Starlink satellite only has a < 5 year lifetime and ~2 starlink sats burn up in the atmosphere every single day is entirely left out as well.
Dwarkesh definitely got big in Silicon Valley from his AI podcasts. He's one of the few people who can get famous researchers on and also have them say something genuinely new.
After that, he become well-known to the general public through his Sarah Paine podcasts (which are excellent).
I have! I agree it's very good at applying abstractions, if you know exactly what you want. What I notice is that Claude has almost no ability to surface those abstractions on its own.
When I started having it write React, Claude produced incredibly buggy spaghetti code. I had to spend 3 weeks learning the fundamentals of React (how to use hooks, providers, stores, etc.) before I knew how to prompt it to write better code. Now that I've done that, it's great. But it's meaningful that someone who doesn't know how to write well-abstracted React code can't get Claude to produce it on their own.
Same experience here! As an analogy, consider the model knows both about arabic or roman number representations. But in alternate universe, it has been trained so much on roman numbers ("Bad Code") that it won't give you the arabic ones ("Good Code") unless you prompt it directly, even when they are clearly superior.
I also believe that overall repository code quality is important for AI agents - the more "beautiful" it is, the more the agent can mimic the "beauty".
reply