A bigger potential issue is that 100% means the metric was saturated--and we don't know by how much. Sol 6.1 could have had 1.01x the capabilities it needed to hit 100%, or 10x.
There's usually someone offering a reward around $1000 for verifiably pre-2019 writing that scores over 75% AI on Pangram, you should pick up some easy cash! So far, the only writing to do so is a very bizarre-looking poem that repeats a line many times.
The Internet, in its current form, provides far more value to the world than copyright, in its current form. So it is, rather, incumbent on copyright fans to change it to fit the state of the world.
I can't tell you how the parent commentor does it, but I saw an excellent reproduction of some historical building done by Astra in headless blender, using photographs, blueprints, and contemporary descriptions. Frontier models are smart enough to incorporate just about any kind of evidence you can give them about what you want, so give them everything.
That reminds me of one of my idiosyncratic policy wishes: An enormous pigouvian tax on any advertising beyond a spreadsheet listing quantitative technical data about the product. Companies don't lose anything, since the relative spend and impact remains the same; consumers gain a lot of information and public funding, and lose daily aggravation.
Pigovian taxes were invented to compensate for externalities. Where's the externality here? You go to a website. The website owner says "watch this ad to access my content". It's a free agreement (or disagreement) between you and the website owner, nobody else is harmed.
Since advertising spending is largely zero-sum, additional spending on advertising may shift market share, but it does so without delivering new factual utility or consumer surplus. The fact that there's a pareto-superior equilibrium makes it a symmetric reciprocal externality: Robert H. Frank showed, in "Positional Externalities Cause Large and Preventable Welfare Losses" (2005) that positional arms races are shaped exactly like environmental pollution.
But, even if there's a better equilibrium, isn't loading a website crammed with obtrusive ads still a coasean bargain between two parties? Simon P. Anderson and Stephen Coate studied this question in "Market Provision of Broadcasting: A Welfare Analysis" (2005). They identified unpriced externalities like the common-pool attention tragedy, as has been raised in other comments here, and asymmetric information & hold-up, which increase transaction costs.
Divide advertising spend into two components: Informative specs, and persuasive/obtrusive spend. The margial external cost is the sum of profits foregone by rival businesses, and aggregate viewer nuisance. A pigouvian tax with an exemption threshold for factual spreadsheets and apunitive rate on expenditure above that (or visual intrusion metrics) both prices the cognitive nuisance imposed on readers, and collapses the positional arms race between rival firms; transforming the industry from a high-waste prisoner's dilemma to a low-cost, high-information pareto optimum.
Obviously, you still have to watch out for firms goodharting the spreadsheet metric, but that's both a partially solvable problem, and a much better situation than the current nash equilibrium.
The Mere Recognition effect exists in pretty much all humans, yes--but it's completely possible to train a reflex of "if I recognize the brand name, it's the same product as the store brand next to it for $150% of the price." My family also did this when I was a kid.
I gave a safety talk at an ML conference that also had a skeptical representative from HuggingFace. Like Delangue, he didn't seem to have seriously considered the threat model--myopically focused on the capabilities of current agents (a myopia which, while less excusable then than in 2015 or 2021; was at least more excusable than Delangue's in this article since HF hadn't yet been pantsed by an agent swarm).
It's not that long ago that sites were pantsed by bored college students, rather than agents. It didn't stop because sites got more secure really, it stopped because those students started getting real life consequences.
Gives you a real easy solution to stop the agents too.
> It will make adding features easier once the debt is paid off.
I'm not so sure: DJB-style secure microkernel/microservice architectures are hard to extend, and I have little hope that today's monolithic code bases are securable without retreating to a small, proven, and well-tested TCB--even by ASI.
The first such news is pretty old: In ARC's 2023 eval of GPT-4, it hired a TaskRabbit worker to bypass a CAPTCHA out of the small research budget allocated to the agent. Even the rentahuman.ai site which forms a market for meat puppets is kinda old now.
If you set up a VPN endpoint on a Hetzner box, you're trusting Hetzner not to MITM your DNS queries. They're likely more trustworthy than your local ISP (although that leaves you more identifiable to folks running the servers or snooping on server-side traffic, and Cloudflare will hate you).
That threat model tradeoff is good for many people.
Not if they are only using it to route the DNS traffic (though in that case a cheaper VPS might be better), which will work if the ISP is only poisoning DNS replies and not (which is often the case) blocking traffic to the target servers entirely.
The problem with this is you are adding a moving part to your DNS lookups. It might be worth adding two on separate VPS hosts and have your resolver fall back to the second if the first goes down.
Another, usually more minor, problem is that some services (like CDNs) use the source of your DNS query to modify the response so you get sent to a topologically near-by end-point. This means you might notice such services being slower as you are sent to a PoP near your VPN host not one closer to your actual location.
Not for DNS lookups. Hetzner forwards complaints from copyright holders, they don't proactively snoop traffic to make sure you don't access 1337x at all. It's legal to access the site.
Oh well OP was doing shady stuff ... then yeah they should continue with using VPN provider. VPN providers are shady AF, I don't think there is VPN provider I would trust more than my local internet provider.
Well that wouldn’t be running your own vpn. Hetzner is. They’d let you do this? Seedbox is against their tos hard to imagine they’d be a good choice to run a vpn for your linux isos.
Can't speak for Hetzner, but considering VPN use in Russia and China - I'd imagine most EU\US vps providers have someone's 3x-ui, remnawave or naked v2ray installed.
Plus - how are they going to detect your remnawave node sitting on 443 port? For them it's just some traffict.
And if you are worried about the amount (torrents) - you don't really need to torrent though them. You just need an access to a torrect tracker to get your magnet link and meta info.
Most default configs for 'ray have torrents locks for this very reason - you communicate with other people (seeds\peers) directly.
Nothing. They'll only act if they receive a complaint. If you're downloading or seeding 1337x torrents from your Hetzner IP address, they will receive a complaint and act on it.
reply