> consistently do 50%+ annual returns over the past 15-20 years
That is only if you consider high-frequency trading investing. Which is very hard to, because on traditional investing timescales HFT is delta-neutral.
The concern (I'd rather call it concern, and not surprise) is in level of persistence.
See, when you ask the model a question, you expect it to give its reasonable best to produce an answer. Like, to comb through available data and stuff, etc, etc. You don't really expect "reasonable best" meaning "look for a side channel to escape sandboxed environment, and get access to information you was not supposed to".
And the gap between that and "hack someone's devices and blackmail them until they give an answer to the question" is narrow enough for the model for researchers to be concerned.
These kinds of alignment problems remind me of times where someone does something that's trivial for them but very hard for the recipient. They might say something like "this must have taken you days" when the task really took 15 minutes.
What's the difference between an API search and a DNS workaround from the model's perspective? I think for most humans the DNS workaround is discarded because it's obviously too much work, not because it's untenable. With the vast knowledge base in the latest models, the cost difference falls sharply; it knows what to do and can do it for a very reasonable cost to itself.
General alignment seems to typically focus on high level value questions. Here, we're dealing with an effort alignment issue where values diverge because the solution effort is different for models vs humans.
That makes sense. I was focusing on the DNS step itself.
Since the agents are set on endless loops of rumination (through every example ever) I can see how it might go further.
My other concern would be the clearly defined gaps between researchers who don't applied research let alone crossing the bridge into the real world of operationalizing things let alone implement.
Letting something rip across multiple domains without understanding what each of those legitimately have done for the past decades is pretty eye opening.
Do they distinguish between "censored in transit" and "blocked by destination"?
In the last years many sites started to drop/block connections from the countries like China and Russia. It is sad to see how Internet is becoming more and more fragmented.
The elephant in the room no one talks about yet, imo, is "should public funding of math studies be adjusted due to AI breakthroughs?"
The sports comparison is wrong here because general public never paid for the specific match results. The value was always in the show, the advertising and betting around it, the health and educational value of doing sports, etc. And the sport mostly lives on what it earns, not on public funding.
Math, on the other hand, was paid for because people and states believed progress in math might lead to meaningful improvements in other branches of science, and, in turn, in our lives. If this is better served by AI, should we keep paying for the same number of tenure positions? Should we increase their number to handle the speedup brought by AI? Or decrease because they're being replaced? Should we pay more to those using AI to do their research, or to those explaining and exploring the AI-generated results?
> Math, on the other hand, was paid for because people and states believed progress in math might lead to meaningful improvements in other branches of science, and, in turn, in our lives. If this is better served by AI, should we keep paying for the same number of tenure positions?
That is a pretty huge "if". The ultimate purpose of mathematics and really all scientific inquiry is human understanding of the natural world. It's not at all clear whether large language models can replace that any more than calculators can replace human mastery of arithmetic. Is society ready for engineers to design bridges and airplanes without understanding the underlying mathematics by simply handing off the entire process to a black box "AI architect"? Are people ready to ingest drugs "vibe-designed" by human drones pushing buttons on an "AI drug discovery" machine and just going "meh, seems about right!"
It might also be useful to take a step back from the hype that the frontier labs are obviously incentivised to incite. Before speculating about how "AI math" capabilities might supplant cutting-edge research in mathematics and other sciences, take a look at OpenAI's own job postings (https://openai.com/careers/search/?). Why doesn't OpenAI demonstrate its world-beating AI capabilities by automating more routine roles like "Account Associate", "Systems Architect" or "Android Engineer"?
I don't know, it reads like a pure copium at this moment.
While Zitron continuously whined about the "AI bubble", and how the models were not improving, and how spectacularly it should've blown, these same people who Doctorow accuses of, quote,
"cooking their brains by locking themselves in the bathroom, holding flashlights under their chins, and saying "Aaaaaaaaaay Eyeeeeeee" until they wet themselves in terror",
unquote, kinda promised, among other things, to give everyone an APT-level big and automated hacking bazookah, and now surprise-surprise three years later they delivered exactly on that promise, and now Doctorow is victim blaming everyone around that they were unprepared for that!
But it was you who said it was all hype, smoke and mirrors, it was you who said the AI is quote,
"a product of limited utility that has been shoehorned into high-stakes applications that it is unsuited to perform",
unquote, why are you suddenly surprised everyone around was not inspired to do anything around it?
The real world software threat model was never suited for a relentless hacker-ex-machina, limited only by the token count you can throw at the task. And maybe we didn't prepare in time because no one believed in possibility of such a machine, because people like you said it was just for-profit scaremongering?
Give or take, AI evangelists gave us all a pretty wild and unbelievable set of expectations few years back. I didn't believe them back then too. But now they're steadily delivering on _some_ of them, and we should be correcting our world model to take into account _all_ of them might be true, instead of making up reasons why other predictions will certainly fail.
I've lost a lot of respect for Cory around his stance on AI. He did great work raising awareness on enshittification, and his language around AI (reverse centaurs) has value, but his extrapolations are so nakedly political and wrong.
Strongly disagree. Hardware support improved by leaps and bounds last decade and a half. Namespace isolation and cgroups driven containerization. eBPF appeared and gained adoption. Async I/O (and later just everything async) converged on io_uring. Btrfs is _stable_ - some people believed it would never be possible, ever. All kinds of scheduling were iterated and iterated over.
All these things are visible across the board, both from developers and end users side.
I have been using Linux on the desktop since the late 90's. Other than increase hardware support, I have not tangentially noticed/felt almost anything you have mentioned here. When I fire up the desktop and load Firefox, I do not feel or see the sandboxing, the scheduler, the file system or the I/O. It just works.
Yes, we interact with it all the time but it isn't something that the end user would notice and that is a great thing. If you are not directly poking around with that stuff, you shouldn't notice it.
That they are so seamless should be seen as a badge of honor.
I also have been using Linux for over 20 years and I agree that hardware support is the most noticable.
This year I have installed Linux Mint on computers from people who wanted to jump from Windows to Linux. Every time I booted via USB, installed Linux and 10 minutes later everything was working.
This would not have happened 10 years ago. Back then there would always be an issue with a video driver, WiFi driver or audio driver.
Of course there are still hardware issues for some special deviced, but today Linux just runs fine for most people. And that's a great accomplishment.
Only on a select combination of blessed components and BIOS/EFI not half-ass broken to the tune it only worked somewhat stable on Windows.
Anything with split iGPU/dGPU was a mess no matter what and where, WiFi that was not Intel was a mess, suspend to ram and suspend to disk were a coin flip if the system would wake up again or if it would decide to wake up on its own and overheat itself to panic.
"I installed Ubuntu on peoples laptops and was extremely lucky they happened to have compatible hardware or didn't notice the stuff that didn't work and had a non-borked firmware". FTFY.
This was absolutely not everyones experience, including mine.
> Btrfs is _stable_ - some people believed it would never be possible, ever.
Right:
> The RAID56 feature provides striping and parity over several devices, same as the traditional RAID5/6. There are some implementation and design deficiencies that make it unreliable for some corner cases and the feature should not be used in production, only for evaluation or testing. The power failure safety for metadata with RAID56 is not 100%.
Seems like parent thread spoke from perspective of developer/tech savvy person.
But even from end-user perspective, maybe its hard to see change when you compare one year or so, but over somewhat longer period I definitely see improvements. Over say a decade or so, wifi/bluetooth/trackpad support has significantly improved. Over last 5 years, I think gaming has become huge and you can actually use Linux as gaming machine. In last 3 years or so, I think I've seen my 5k display using Thunderbolt work much better. Fractional scaling also works much nicer.
NVIDIA GPUs are not shitty to use on Linux, and AMD has been working great for slightly longer. UEFI/TPM works much better now as well, where I don't have to worry about booting in legacy mode when using GPU drivers.
I didn't have to think about any of these above. If I think harder, I can likely come up with many more.
Well ya, it's more useful in observability usecases than desktop ones. But recently I've been using for a consumer desktop usecase. I made a file watcher with it. I'm sure you're familiar with the woes of using inotify/fanotify.
eBPF vfs hooks based watchers are really nice.
[edit] details
- you setup a single `mount` - this is the only part that requires cap_sys_admin
- then you run a single daemon, this one only needs cap_bpf, and this runs once at startup and is its own self contained program
- this daemon writes events into a ringbuf referred to by a file in that bpf mount
- you then read that from your standard user program (this can be done by whichever application needs to consume the events)
the events are create, mkdir, delete. renames are automatically decomposed since we hook at the vfs layer. the hook itself is practically free to run and thus there is no question of a performance impact.
You may know more about what it powers. It has enabled much faster, higher quality, more stable security software on Linux. Many modern security tools are built on top of it. While end users might not know it by name, it has made Linux easier to support in the enterprise.
For desktop I think it would be fine, though if you have the advanced needs of all the ZFS features I would stick with ZFS, but I had no problem.
The real problem I had was a low limit to save history. I install a lot of software, Cachy would run the snapshot every time (good), but it only kept like 5-10. And the one I needed to restore to was further back (I can't remember the exact issue), but that was more the defaults of how cachy set it up, and my use case than btrfs/snapper itself.
Whatever tool you use to snapshot btrfs, you may wish to have a custom number of saves if you install tons of software a lot...so you can avoid having too few backups available from the boot menu. I'm sure there was a way around it but I said f it and just reinstalled instead of trying to look up on my crap tablet how to get at the old files (assuming they were still there and just not accessible from the boot menu vs being completely erased).
I do not think it was btrfs that broke, to be clear.
This was like a week before the big hack on AUR, so I wonder if I got hit by a supply attack, but it was pretty shoddy, and broke my system, thus my reinstall saved me from (possibly) a lot more pain down the road.
It was the default on SUSE / OpenSUSE for years. Haven't used them for a while, not sure whether it still is. No problems with basic functionality including snapshots. I believe some RAID configurations were experimental for long time.
Been using it for a year or so on my main box. Never had any kind of issue. Snapshots saved my ass a couple of times. All in all it’s been pretty boring (in a good way)
Maybe. I think there's still no chance of repair when it blows up. Only FS since Reiser and early XFS that I've completely lost data to on healthy hardware.
Whichever one of my test machines is using btrfs is doing fine, though, so on a subjective level that one's been stable enough.
It's basically fine as long as you read the release notes carefully and don't do anything the release notes say not to do (e.g. RAID5, let the disk fill up). It still has some sharp edges, but they're all documented.
> There was a time Cloudflare just made the internet better
There wasn't. Cloudflare is a cancer grown too big. And it was always positioned to become one, the middleman between users and the Internet.
It is already painful to browse web sometimes using the non-"standard" tools (that is, not a Chrome with Google account signed in, not an EU/US residential IP). What if tomorrow Cloudflare checks will require attested and signed browser binaries?
>It is already painful to browse web sometimes using the non-"standard" tools (that is, not a Chrome with Google account signed in, not an EU/US residential IP). What if tomorrow Cloudflare checks will require attested and signed browser binaries?
If it's painful, that's not because of cloudflare. You can get past turnstile challenges using tor browser in a VM. You still might be blocked because of policies set by site owners, but that can hardly be blamed on cloudflare.
Genuine good faith question, what would you recommend the solution or alternative to be? How could we be able to get the benefits Cloudflare provides, without the inherent drawbacks?
The general arguments here are that CDNs have replaced tier1 ISPs as the predominant force on the Internet. Geoff Huston has written about this over the last ~10 years[0][1].
So the question as to alternatives is void - the problem is CDNs MitM'ing yur datas across multiple domains, and the loss of the nostalgia (or whatever) for the tier1 ISPs (who were all horrible in their own ways).
I don't particularly subscribe to the harms Geoff purports, but I cannot fault his analysis of the situation. If you haven't read it before, it is worth the read.
If you want a free and open Internet, then you don't want someone else's TLS proxy between every consumer and every producer. This ignores the problems of DDoS, scaling, etc. Buuutttttt.... 99% of websites don't have those problems, so why are you building a solution which trades decentralization for it?
Anyhow, sounds like you're into all the wonderful things Cloudflare (and every other CDN) provides. Great. That's fine, just don't be angry when the free and open Internet promised to all of us in the 1990s requires 3rd party gate keepers. And no, there aren't solutions to all the problems.
It could be but proposing a standard and getting it adopted by everyone else is its own monopolistic behavior. Everyone here shits the bed when google tries to do it for ad measurement and decloaking
It works off PoW, so all you need is a vaguely working js/wasm runtime. It won't fail you if it detects you as spoofing user agent, or are running a "weird" browser like servo on freebsd (or whatever). All of this is "good" for being non-discriminatory, but it also means it's pretty easy for baddies to implement hardware accelerated solvers that can solve challenges orders of magnitude faster than any user can, and at a fraction of resources required (because they don't need a full browser).
The impression of XP being the best version of Windows came after at least two service packs (patch bundles / minor releases, for those unfamiliar) and the Vista shitshow. First Windows release that was good from the very beginning was 7.
reply