I often have the experience of going to a museum, seeing a painting I like, and then not being able to find any image of it online - let alone a highres scan. In my casual searches I've found museums generally don't share high resolution scans of their collections. Try to find high resolution images of famous, but not super famous paintings. It's often virtually impossible.
I remember in college my art history professor had a personal digital collection of high quality scans that he'd procure somehow that were impossible to find online. The data seems to be highly guarded, I'm guessing b/c it can be used to make merchandise
So my impression is that the Rodin Museum is not really the exception here
Usually government data is CC0 or public domain while OSM is not (maybe Germany is different?), so I thought they'd be allergic to OSM's licensing. OSM actually has a lot of problems ingesting data b/c a lot of map data can't be relicensed under their license.
Ironically.. you can't contribute map data that requires attribution to OSM... even though they require attribution..
It's a great project.. but I think they really messed up the licensing in a misguided effort to make the project viral. Maybe trying to emulate Wikipedia. But text and map data are just fundamentally very different. It's really unfortunate. I hope an alternative emerged eventually
> It's a great project.. but I think they really messed up the licensing in a misguided effort to make the project viral.
OSM's sharealike provision has never been about making the project "viral". Free map data is enough to do that by itself.
OSM's sharealike provision is, and always has been, unambiguously aimed at stopping bigcos killing the project by scooping up the map data then contributing nothing back - the scenario where Schmoogle Maps takes all OSM's footpath data, combines it with their own motor-centric, sensor-fuelled dataset to create the one map to rule them all, and thereby attracts the flow of new contributors that OSM needs to survive.
The subtext seems to be a hostility towards people making money from what you've shared.. which I find sort of baffling - but is very on-brand for the general anti-capitalist sentiments in the memesphere. People building on your work feels like something to be celebrated, but I guess it's a difference of opinion. I'm just generally skeptical that this would be "killing the project" at all.. it feels like it'd be the other way around? If OSM data were for instance used by Google and Apple Maps.. it feels like it'd just have made it more foundational and would get people even more excited to contribute. It'd be the world's central repository of map data everyone builds off of.. like the Debian of maps. Instead it's a parochial project for map nerds.
Can you imagine if GIMP has a similar policy to watermark all it's output with a "Made my GIMP!" in an effort to "attracts the flow of new contributors" and to make it hard for people to build a product with it?
Oh interesting! So they manage to not watermark their maps. OSM seems to heavily imply you should watermark, but as I understand it, it's not strictly required
Sadly not always the case. The Irish government went with CCBY 4.0 as the primary license which causes some compatibility issues with OSM due to specific clauses. :/
I would say biggest issue with licensing comes from clear attribution being problematic at the scale of OSM not the virality or relicensing problems. That and OSM aiming for being squeaky clean, far away from any ambiguity in terms of third party data licensing.
It is simply impractical to attribute all the sources anywhere in UI or printed copies of map. And whether an attribution hidden somewhere deep in wiki is considered acceptable is grey zone.
The fact that OSM themselves want attributions is another reasons why they have high standards for what's considered an attribution which they can't achieve for third party data sources.
In case of software License.txt and Help/About is considered standard practices. But that doesn't necessarily translate to other mediums of copyrighted work. Something like books or research papers have their own generally accepted practices of how attributions are handled. For maps digital and physical text in the corner is often used practice. You can see it even for something like a building plan posted next to construction site, listing additional map sources used for preparing the drawing. So it's not exactly unique invention by OSM. Back to comparison with software, software can't exactly be printed out so the concerns about attribution are different. Also software licenses typically require listing including a copy of license text not just attribution , which is simply impossible outside separate file or dedicated UI.
In practice the OSM aiming for better than good and thousands of unverified contributors being uncontrollable lands it somewhere in the middle. But if they aimed for barely acceptable all the contributors would definitely push the bar bellow legally acceptable.
On the topic of government data, CC-BY is also common which is somewhat problematic. But more often they have no idea under what license they are releasing their data. They come up with complicated schemes of metadata, which never gets properly filled or parsed, thus resulting in conflicting information about license being used. And if you ask them to clarify they will just say, "yes yes it's open data you can reuse it", with the government employee having no understanding about differences between various licenses and that not all open data is equal.
I understand that there is precedent and there is a logic to it... but to me it rubs me the wrong way. I think text and code are not very analogous situations. If I think in terms of software licenses..:
- I sort of respect the copyleft ethos. You put out a thing with the understand everything it touches is also going to be openly available. I think attribution isn't really part of the central idea there.
- I also respect the MIT/BSD style thing where you're just putting stuff out there in the public domain and it's part of the corpus of human knowledge. You leave your mark so to speak.
The middle ground of "You can use this but you gotta promote our service and stamp our name on it" just feels icky. I don't want to contribute to that. I feel I'm helping OSM the organization and.. I don't know them .. are they good stewards of the data I'm giving them? Are they going to be good stewards in 10 years? Hopefully that kinda makes sense? I guess the same can be said for Wikipedia, but Wikipedia for better or worse is very siloed. You have to attribute stuff you copy from Wiki but ..
A: Realistically nobody is actually copying wiki articles other than lazy high school students. It's just not very reusable outside of Wikipedia
B: This is more of a plagiarism issue. It encourages people to disclose it's not their own words (this is not an issue with maps.. nobody thinks you surveyed your city to draw the map)
If Wiki became CC0 tomorrow nothing would really change. If you needed to fork Wiki and put an attribution in the footer, it'd be very innocuous..
I don't get why you consider MIT/BSD fine when both of them have "The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software." but when OSM does the same it somehow becomes "you gotta promote our service".
> For practical reasons we require that users of OpenStreetMap data attribute the project as a whole and we in turn provide attribution to third parties via the "Contributor" pages. Licences and terms of use that require attribution of the third party data source directly on derived works are incompatible.
"require attribution of the third party data source directly on derived works" is exactly what OSM requires. It's hypocrisy of a sorts
Is there evidence that LLMs generate better code in more popular languages? I get the sense the "experience" translates between languages and it can reason in any language just fine. I write Clojure code using a rather esoteric framework (Pathom3). There is probably very little similar code out there (it's definitely a tiny fraction of the training dataset) but it seems to do just fine
Not saying you're wrong, just curious if there are numbers backing this up.
I built a brand new language to test this[1]. Not only is the language different to basically any other language but it also tries to be adversarial against LLM understanding.
The best models can still make sense of it[2], though the tasks so far have been pretty basic. But I do think it gives some evidence that languages which aren’t well represented in an LLM’s training can still be reasoned about and written well by LLMs.
In my experience LLMs are way better when they "know" a language "instinctively". It's just that unless your language is very niche, the corpus is usually good enough. I tried using Claude to write my own personal language a while ago (I wrote a toy compiler decades ago) and it struggled a bit, because you could see in it's reasoning it had to "repeat" the syntax equivalence to itself while it read the code. It didn't just "know" it could use a given construct to do something; conversely Astra, when carefully instructed to do so, can plop down esoteric template code that works the first time, because it just "knows" it's the right stuff to write
I've never hit anything with Qwen. Very rarely it will silently switches from 3.8 to 3.7 (which makes it notably dumber) but you just click the drop down at the top to switch back. It's rare though. I have >week convos with no switching
The issue is that there is ultimately a limited market for having access to space. SpaceX's business model is premised that if you make access cheap enough that there will be a market for multiple launches a day. But to me demand seems very finite? You can make it 10x cheaper or 50x and it sort of doesn't matter.
There just isn't much to do in space. You have:
- Telecommunications.. Can scale only so much. Starlink is cool. Compared to Iridium it's way way cheaper. But demand is .. modest? It seems to only beats fiber on the ground in special situations
- Telescopes looking up. This is limited by science budgets that aren't going to change much.
- Telescopes looking down. Those people have infinite budgets and already have all the satellites they'd want. If you're not in NATO then can't depend on SpaceX anyway - so they're not capturing that market.
- Tourism. Vomiting for several hours in zero-G will be a thrill for some for a bit, but I don't see this being a huge industry. Maybe on par with basejumping or something.
- Mining.. seems impractical scifi.. premised on wanting to build huge structures in space - but again, for what goal ultimately?
It's sort of like if you made access to Antarctica 100x cheaper. It'd be great, but it's hard to imagine it's going to suddenly open up a ton of demand.
Indeed SpaceX is boiling multiple oceans at the same time and betting the timing will align.
-- SpaceX is building a starship factory, not just starship. The market won't deliver enough demand, they need to create their own
-- V3 Starlink constellation targets 100k sats. With 50 sats per flight (100t to orbit) you need 2000 flights to build constellation and then 400 flights per year to replace sats after 5y lifespan.
-- There was that idea of ballistic human travel but this is unlikely to materialize
-- There also was solarsat/powersat where solar power is beamed from space and if you massaged your math the right way the numbers were ~promising
-- AI saves the day. We "need" 1mln AI sats. That solves the demand and justifies the spending. for now
> Telecommunications.. Can scale only so much. Starlink is cool. Compared to Iridium it's way way cheaper. But demand is .. modest? It seems to only beats fiber on the ground in special situations
It is not designed to compete with finer.
Go somewhere truly remote and see how it’s changing the world very quickly. I’ve seen it first hand in Arctic Canada, Australia and North Africa. Before starlink the best they had was dialup. Now they video chat family and doctors, do school online and so much more. Easily a few billion people will use it.
You also forgot about AI data centres in orbit, on which SpaceX have pegged virtually all their valuation in.
the demand for bandwidth is infinite… this used to mean copper -> fibre, but it’s that squared for orbital bandwidth, as they lower and close up the sat formation, the antenna gets smaller and the data rate gets faster … I think Musk killing all terrestrial telcos is quite likely
Even my really underpowered PHEV has a 65kW electric motor. No "off the shelf generator" can generate 65kW, in fact it would be a huge unit that could.
We could of course say - well, it doesn't need to match peak usage, it just has to charge the battery faster than it depletes - you still probably need around 20kW for steady motorway driving, and that's a pretty hefty generator.
I don't think you need to have it match the depletion rate. Even at say 10kW you'd doubling the effective range. The point is people want to be able to use their car for that twice a year road trip, or drive to the inlaws for Thanksgiving. If you can attach something that extends the range for these occasion then that's probably enough.
Furthermore, the generator can run 24/7, but realistically you will at most drive ~6 hours a day. So for a leisurely road-trip this may be enough to avoid needing stops at charging stations
Thanks for writing this up! It still assumes a lot of prior knowledge, but it's a good starting point to approaching these problems. How JavaFX works on different platforms really looks like dark magic from the outside. I wrote an app in Clojure using cljfx. I can jpackage it and run it Linux/Windows/macOS .. it's quite easy to get working. But once you step outside of the standard platforms it gets complicated really fast. Trying to get it to Android or a native build and suddenly you have to be an expert in several build systems and JVM internals.
Oh hmm, I'm honestly not super sure then. I know Gluon is the one that's seemingly has a business providing JavaFX support, while Oracle.. as far as I know doesn't offer a license or anything services in that direction? But I could be wrong. Gluon is mentioned in the article.. they have a very Qt-style confusing licensing situation where it's probably all free.. but you're never quite sure :)
The whole space is a lot more icky than one would like it to be.. Which is unfortunate b/c while it doesn't have a billion features, it's a nice composable reactive framework to use (at least from cljfx/clojure)
There was a time Oracle decided it wasn't worth the investment, note that it started as E3 scripting under Sun and it was only finalised after the acquisition.
So a few Java champions took over it at Gluon.
Then when Java 26 was released, Oracle reintroduced support for JavaFX,
Oh huh, I guess it's pretty recent news. I wonder if it'll lead to new features.
It's nice to see people are working on Native builds for instance,
and that Graal integration is now an official priority of sorts.
is there a good metric of model degredation over time?
Im a bit lazy and only use the free models different companies host and the biggest difference i see is that some models (Gemini, OpenAI) get progressively stupid in long chats. You end up having to start a new session every oncr in a while. Or they get really hung up on a theme and cant shift to a new topic.
By contrast, Ive been impressed with Qwen. I have some chats on research and code architecture that have stretched for weeks without any noteable change in quality (though occassionally it seems to "rush" to an answer)
Im just looking at all the listed benchmarks and im unsue which i should be looking at
I remember in college my art history professor had a personal digital collection of high quality scans that he'd procure somehow that were impossible to find online. The data seems to be highly guarded, I'm guessing b/c it can be used to make merchandise
So my impression is that the Rodin Museum is not really the exception here
reply