wongarsu

If by "distill" they mean "used it for fine-tuning" then they might have used it in the final stages of fine-tuning of Kimi K3. I image they might have already been using Opus, and when Fable became available it was easy to switch over to it

It would have been a tiny part of the overall training, given the timeline

I frequently have cases where `ls **/*.png` fails because the argument list is too long. Just happened yesterday with an rm I wanted to run.

Find works, but the syntax is arcane. Not bad, but unlike any other common cli tool, which makes it more difficult to remember if you haven't needed it in a while

It also burns quite well. If are in a location that burns trash in waste-to-energy plants you can get a decent amount of that energy back by throwing it away

While running the model at 9000 tokens/s is the more flashy demo, I imagine running 1000 concurrent requests at 150 tokens/s each is the much more achievable goal

Planners forgetting that pedestrians exist is sadly a common occurrence

People associate a lack of lighting with crime

I vaguely remember that studies showed that lack of streetlights does not create crime, but the crime that happens anyways prefers badly lit areas. But from a pedestrian's perspective that's not much relief, you'd still avoid those areas. Which is not what a city planner has in mind for a park

I got a decent response with the same prompt

But yes, this will happen. 1.7B is already a tiny model, running that at 1-bit is going to lead to some behavioral issues and will need tuning of temperature/topP/repeatPenalty/etc, and probably a software fallback that detects pathological cases like this and simply retries. Just like we all had to do in the times of GPT 2 and GPT 3

There have been plenty of instances of paper-based land registries being wiped out. It's just usually a very local problem because paper lends itself more to decentralized storage

From the NY Times, 1995:

Spying on allies for economic advantage is a crucial new assignment for the C.I.A. now that American foreign policy is focused on commercial interests abroad. President Clinton made economic intelligence a high priority of his Administration, specifically information to protect and defend American competitiveness, technology and financial security in a world where an economic crisis can spread across global markets in minutes

At the time "only" to have the upper hand in trade negotiations, and presumably when "having the President press foreign countries to award business to American firms", as they call it. But that was 30 years ago, and trust that these these things aren't directly forwarded to US private industry is in short supply these days

https://www.nytimes.com/1995/10/15/world/emerging-role-for-t...

https://archive.is/LmEA1 (same site without paywall)

The market will quickly adjust to the point where spending $50 on tokens will on average give you a vulnerability valued at $50. Maybe $100 to account for your edge in having a better prompt and the risk of prison time you take in selling the exploit

In fact that may well be true today. OP didn't try to sell it to discover the true price the market is willing to pay. And we all know that "somebody paid $$$ for something similar in the past" is no guarantee that somebody else is willing to pay any significant sum for your thing today. If it was, startups would be a lot easier

In Europe (at least in Germany) we already have a lack of truck parking spots along many routes. And that's with parking spots that are just a bit of Asphalt and maybe a portapotty, or big asphalt parking lots attached to a rest stop

If we had enough chargers that any trucker could pull into a truck stop at their prescribed rest period and find a free charger, electrifying trucks would indeed be a no-brainer. But in the world we currently live in we don't have enough chargers to make that reliable. We barely have enough non-electified parking spots, and in your world they would all have to be electrified as well. It's a mix of a chicken-and-egg problem on the commercial side and a lack of will on the government side

But there is never a lack of will from the government to try all kinds of pilot programs that look cool but are mostly caused by the need to somehow sidestep the lack of sufficient charging infrastructure

In most realistic settings it's even easier. For example a Project Daedalus probe only accelerates for four years before running out of fuel. So you could decelerate in just four years. Maybe a bit more, since you only have the smaller second stage engines. But essentially you are accelerating for four years, coasting for 392 years and decelerating for another four years. Accelerating for the whole time and turning around in the middle would be faster, but we don't have the fuel for that

The issue is that in the original architecture without breaking you burn 50k tonnes of fuel to get 1k tonnes of payload up to 12% lightspeed. If you want to break all the way back to zero, you need to 50k tonnes of fuel to break. But that means you need to accelerate another 50k tonnes of fuel up to speed.

Which means you need 50 times for fuel to get from 0.11c to 0.12c, and you need to accelerate that fuel to 0.11c, so you need more than 50 times the fuel for the step from 0.10c to 0.11c, and an even larger factor more to accelerate from 0.09c to 0.10c, etc. So you don't just require another 50*50k = 2M tonnes of fuel, but an exponentially larger amount. The tyranny of the rocket equation

It's also just a very cherry-picked framing

Mozilla exists because one company owned the front door to the web, and another company abused their market position to push their free (as in beer) browser. Mozilla is the phoenix rising from the ashes of that first company.

Then another company came along, abused their market position to push their free browser, demolished Firefox's market share but keep handing them cash to avoid the appearance of a monopoly

Yacy tried in the 2000s. I'm sure some magazines made headlines posing the question whether yacy is a google-killer

With variations on nuclear propulsion we could plausibly get to up to around 12% the speed of light. At least that's the number quoted for Project Daedalus [1], which is using nuclear fusion for the first stage and nuclear-powered ion engines for the second stage. With the cruder but more realistically achievable right now Project Orion design (riding the shockwaves of nuclear bombs) you could still get to ~3% the speed of light

But even at 0.12c, we are looking at 400 years to get there. And we'd be zooming by at 12% the speed of light. If we want to slow down a bit that'd add hundreds of billions to the cost.

It might be worth waiting another century to see if we can come up with a faster design in that time. Not like closer targets like Alpha Centauri, where the thing stopping us is mostly just the absurd cost

So more like enterprise software

Steel production is on the decline in most first world nations. In case of a large-scale war there simply wouldn't be a lot of surplus to send Britain's way

Recycling steel uses by and large the same industrial sites as making new steel. Most new steel has some amount of recycled steel mixed in, the trick is getting the ratios of various types of scrap right

In a defensive war, a steel navy sitting in spitting distance of your shores and an army of steel tanks will still do a lot to keep the enemy at bay. And you can swap the tank turrets for Gatling guns on some of your newly produced tanks to help against drones

You wouldn't be able to win the war in your scenario, but you could still do a lot to make sure the other side isn't winning either.

But that's on OpenAI for selecting that company name. EU trademark law hasn't changed, this was always going to be a problematic trademark if challenged

They still have the trademark on their logo

In the EU "well known marks" are protected even without registration, and block conflicting trademark applications

As such "everyone knows them" isn't a reason to allow a registration. It would just mean that blocking the trademark has no practical effect

There are plenty of super remote areas in Europe as well. Even more in the US, where population is more clustered to population centers

But the big selling points of Starlink are either as a backup connection (which the consumer plan actively enables: in months where you use less than 10GB you pay $10/month) or as a connection for ships and airplanes

The trivial fix is to execute not "git.exe" but "C:\Program Files\Git\bin\git.exe". With a configurable git binary

Of course that then leads you down the rabbit hole of whether the llm should exectute a known-good ripgrep, tree, etc. instead of depending on PATH resolution

That only activates when pressing shift five times. Which is admittedly likely when trying to press it three times in a panic

But having a stray sticky keys window open is not too conspicuous

There are cables with 90° plugs (cable to the side). Those are pretty robust, get the cable out of the way more reliably, and provide a much smaller lever for physical forces on the connector. They aren't the solution for everything, but they solve a lot of issues

For everything else there are magnetic USB-C connectors

USB A cables and ports have the inner plastic piece color coded. White for USB 1, black for USB 2, blue for USB 3 (5 Gbit/s), teal for 3.1 (10 Gbit/s). Red or yellow for always-on

Blue is the only one that's actually endorsed by USB-IF. For the most part USB-IF would prefer if everyone just used their little icons. Luckily manufacturers have a bit more sense and all follow the same color scheme (apart from the occasional black sheep)

The developers of Process Explorer have a lot more to do with the OS than the makers of Minecraft, even if the company name is the same on all three

At some level of performance reasoning becomes the most effective method to predict the next token

In the last ten years we have one (1) accidental satellite collision. In 2021 when a Chinese spy satellite in a 788km orbit collided with an old rocket stage from the 90s. There was also that Indian anti-satellite weapon hitting a microsat in 2019, but that was very much intentional (and much condemned because of all the fragments it caused)

In the decade before that (so 2006-2015) we had four accidental collisions and two anti-satellite tests (China and US)

Those are not frequent events. And if you attribute any statistical significance to them, it seems we are getting much better at avoiding collisions

And they made it really annoying to get for only one month. It's meant to be a subscription for locals

That coupled with the just much nicer experience of high-speed rail (which is not included in the Deutschlandticket), I'd still call the Interrail pass worthwhile if you are visiting. Not the only good option, but one of the good options