mikepurvis

I was going to say the same about Jean; I think that one at least is hard to really draw conclusions from without additional data like languages spoken at home.

Anecdotally, the boy->girl names are almost all very popular names where I've met multiple each of those across different ages, whereas the other two categories feel like they have a lot more down-list entries.

You actually can sort of do this just inside a modern docker file, with COPY --link

That version of copy overlays the copied content without any knowledge of its underlay, so if the underlay changes, the hash of the layer created by COPY remains identical.

Obviously there are lots of situations this doesn't work for, but say you have a pool of Ubuntu packages (the system layer) and then you separately copy a CMake installspace or built uv project from a build stage, that project layer can be identical in hash if it's copied onto an updated system layer.

In practice, it's hard to get this to work because it's very easy in docker to accidentally trigger a rebuild for some random reason, but all the pieces are there.

It's first-principles thinking. You have to have some sense of how you would do it in order to reasonably engage with a machine's idea of how to do it. Otherwise you're just nodding along, like yeah sure bro uhhhh lgtm.

I think the number one thing that bites a lot of container workflows is when you have a separate "base image" build that depends on external, mutable package sources using tools that don't support locking, such as apt.

This gets worse for each additional such layer, for example if you start from ubuntu:whatever and have a system deps layer, then an application deps layer, then static assets, then your app's code. Unless you think about it upfront when designing the pipeline, it can be challenging to trace all that state backwards and figure out how to safely re-build thing-from-x-weeks ago but with a single unsafe package bumped forward.

Every shop with competent devops people present should be set up for this, but it usually doesn't come for free.

It's normally an opt-in thing I think, since it's not an unconditional win: you have to figure out which code paths are worth it, and then where to trade off binary bloat vs the runtime cost of dynamic dispatch. In rust for example:

https://doc.rust-lang.org/beta/core/arch/#dynamic-cpu-featur...

That sounds like a scary investment to make on behalf of what's effectively a single customer. Without some kind of contractual commitment that locks in volume and price, they would really have you over a barrel.

Should it also predate Docker by a decade? That would be crazy.

Isn't that the world that Docker gives us, basically?

On some level, yes, though when I choose a saas product, what I'm really choosing is not a basket of features, but rather someone else to be worrying about scaling, monitoring, security patches, uptime, etc.

Like in theory your agents can build all that too, but all the ancillary bits benefit a lot from being spread across a pool of users with different time zones, release cycles, whatever.

It can also just turn into a waste if you build all the automation to turn around and only use it once, recalling https://xkcd.com/1205/

Now that said, the economics have changed in a world where many trivial automations can be thrown together by agents themselves in a fraction of the time it would have taken a human.

I bet we'll see more KVM-for-X type products appearing, basically multiplexers for various storage devices and peripherals, as ways of giving agents control over them.

I think this tool is meant to be something you impose on your colleagues who treat all code like the second bucket.

That's horrifying to hear. I also went on a multi day dive trip at gbr and the coral sea in 2007 and it was spectacular; I can't imagine going back and seeing it degraded.

It's definitely not nothing, and the numbers bear that out: https://www.bbc.com/news/business-59008000

Whether this is one of those "we accept some number of school shootings cause gun freedoms are just so great" issues or not is I guess its own matter for debate. But the fact that this happens, and then as dudes we get to discuss it at arms length on the internet and opine about whether "99% of trips completed without incident" is reason to dismiss it all out of hand may help explain why an individual woman getting home from the bar might choose to walk or take waymo vs getting in a cab and risk joining the unlucky 1%.

I guess? I was on a date with someone and we took an uber back to her place after a concert, and she mentioned off-hand afterward that had I not been there she would have just walked... even though it would have been over an hour on foot, it just wasn't worth the risk after multiple bad/creepy experiences taking carshares home late at night.

Maybe it's a bit like cops: they're not all bad, but the system is kind of set up so that bad ones thrive and good ones burn out. It's going to be a particularly kind of person who is willing to spend their late Friday night driving drunk people home from the club, and honestly if you're a vulnerable person, maybe the guy most willing to do that job is the one you least want as your particular driver.

I'm not sure if I buy this, it seems like it's a bit of a turtles-all-the-way-down discussion. Like a vulnerable person ordering a taxi also has a phone, which is the ultimate surveillance endpoint, and opens them up to potential abusive behaviour from employees of their phone maker, their telco, and every app they have on there, including the rideshare ones which by necessity have precise location and other permissions.

At the end of the day, a man driving them home and knowing their home address while right there is a much more immediate and specific point of threat than "I'm being surveilled while I travel in a Google-controlled robotaxi".

Also, like it or hate it, but Google actually has a pretty decent track record with protecting their customers from their own employees. Gmail and gsuite content stays private, and there are enough protections in place to satisfy insurance companies, medical institutions, banks, and even governments. If a Waymo kidnapped or killed someone under the control of an overseas remote operator, it would be national news and congressional hearings; a person getting assaulted by a cab driver is nowhere near that level.

I just don't think "but Google/Waymo bad too" is a reason to poo poo someone feeling more comfortable sitting alone in a robotaxi than sitting alone in a cab or an uber.

That's a very different problem that probably has a generational timetable.

Letting women get home safely with a robot driver is a thing that can happen today.

At least for myself, business travel remains relatively simple: go to Perk or whatever the business-supplied travel portal is, put in the dates, book whatever flight and hotel are in-policy, when I get there report to the conference/customer/whatever, and buy whatever food is available when I get hungry.

It's travel for pleasure (especially with groups) where all the complexity comes in: there's a lot more stuff to schedule, more slop in what exactly happens when, more competing preferences, more need to have multiple things on deck so you can react to changing energy levels, etc.

Not too hard to see why a lot of people just give up and book an all-inclusive where every day is the same and everyone can just "do whatever", but with better tools, maybe rolling your own vacation could be made more practical and coherent for a normal busy person to put together.

I'm intrigued about the "Trip Planner" being a top-level app. Only tried it in a very surface level way, but the recent trips I've planned (Ottawa for 6 people, Cancun for ~20) have ended up as an ugly mishmash of spreadsheets + google maps saved locations list + random chatgpt sessions.

It feels like there's an unmet need there for managing something more flexible than a minute by minute itinerary, maybe like a list of must-do and want-do activities and dining options, with the system able to figure out what is open when and has other constraints like the weather, costs, what activities pair well together, which group members some things are more or less suited to, what the contact details are for things like a private tour, etc etc.

Oh okay, in that case mine were never sold. As in the sibling reply, my assumption is that the "retail" side of the mortgage (who you make your payments to) is fully detached from who ends up owning the mortgage backed securities that your debt is being financed by.

Is there a way that you can track this yourself? I've had three different mortgages [1] in my lifetime and as far as I know it was all completely opaque who/what was ultimately holding the underlying debt.

[1]: First was a decade ago with RMG (a Canadian monoline lender) and the other two with Scotia, the first of which was a fixed mortgage at purchase, and the second was a refinancing.

I would definitely encourage folks to seek out community theatre where you are. Not that I don't enjoy a full on broadway-style blowout every now and then, but it's also been shocking to me how much of the difference between a $40 seat and a $150 seat is pyrotechnics and spectacle that don't really make the experience for me.

The words and the performances, the chemistry and creativity of the actors, that's the real priceless stuff.

“Being an adult is staring at medium screens all day to come home and stare at a little screen while a big screen is on in the background”

https://x.com/andlikelaura/status/1392091221032460291

I kind of agree, and it's interesting how Nebula and Spotify (at least their video efforts) have been based on elevating and funding those ultra-budget YouTube creators. Like I watch Lindsay Ellis and Practical Engineering and SkillUp all on non-YouTube platforms now, despite YouTube being the recommendation algorithm that originally brought me there.

It's really the opposite model of the "major" streaming players who fund tentpole releases in the conventional Hollywood way, and then have a long tail of niche content (cooking reality shows and the like) that are clearly on a much tighter budget, but still basically modeled on being a cheaper version of a cable show rather than a much better funded version of a YouTube comedy troupe.

GPUs, wifi, power/thermal management, firmware. Especially in portable computing there is still a huge amount of hardware support surface area to contend with.

Game consoles worked a lot like this until around the PS3 era, where the disc/cartridge the game came on was shipping binaries for everything including all the hardware support.

This led to some hilarious implementation shenanigans for the Wii as a transitional console, where the Home button pause screen is not in fact a task switch to some underlying console OS but rather a piece of the SDK that is separately-delivered from each individual game.

ref. https://www.copetti.org/writings/consoles/wii/

Obviously they tried it multiple ways have brought receipts, but nonetheless it seems surprising that it wouldn't be of benefit to bring as much of that kind of metadata to the model as possible. You'd think depth and segmentation in particular would basically just be a straight shortcut without which the model spends its own time and effort re-deriving that stuff.

I'd also be interested in how post-processing fits in with this. Like if you've got weather effects, film grain, tone mapping, etc, I would have thought the model would do better working on the image before those processes.

You would think so, right? But as with others in the thread, I'd found directing the agent itself to prepare the handoff does deliver much better continuity.

I assume Anthropic & friends have noticed this as well and will change how they handle long running sessions, so the gap will likely close over time, but this is definitely where things stand today.

This is what I've been working toward as well. It's interesting how having the agent do its own reasoning about what it thinks is the most relevant knowledge to carry forward into the next pieces of work is vastly more effective (and even fast sometimes) than whatever the mystery-meat "compaction" process is.

Mentioned by the author in a recent HN thread, I'm also experimenting with automating this through a tiny issue tracker called epiq [1] that basically lets the agent sessions themselves file tickets with the follow-on tasks and relevant handoff right in them, and then a dispatcher automatically launches those tickets into new agent sessions.

[1]: https://ljtn.github.io/epiq/