It is appealing to think that special-purpose AIs could do a better job in specific subject areas than general-purpose LLMs and use fewer resources at the same time. But it may be hard to pull off in practice.

Sure, I want a better, cheaper Python coding assistant. But I want it to understand natural language instructions, complex real-world processes and relationships, and a lot of computer science principles. I'm not sure you can get there without a general-purpose model. Maybe that's why we don't see a lot of cheap, effective, domain-specific models being offered.

Kind of a minor point, but it may reduce confusion in this part of the conversation: when discussing climate policy, "mitigation" has a well-established meaning of âoereducing GHG emissions" and "adaptation" means âoegetting better at living with climate change." The parent and grandparent have used "mitigation" to mean "adaptation", and someone further down thought they were using the standard meaning.

TFS and TFA omit the link to The Conversationâ(TM)s article that mentions "smearing" by the big tech companies. That in turn links here.

They avoid the catastrophe of repeating the last second of the day (which has time run backwards briefly) by instead extending the duration of each "second" by a few microseconds during the 24-hour period that includes the leap second. I assume thatâ(TM)s short enough that the network time service on individual computers just says "oh, my local clock is running a little fast" and fixes it, instead of seeing a bigger adjustment which registers as an error or messes with the logs.

Most Canadians live south of 45.7Â N.

this... tracks.

I've been using Claude Code at work for a low-stakes project so that my company sees me being all "AI forward" and stuff (They're really drinking the Kool-aid) and yeah I'll admit its been doing OK on the project.... and yeah I see where you're coming from, there is a certain prose about how it outlines/summarizes things that feels familiar here.

I would really like a true new browser that doesn't use WebKit, Chromium or Gecko (and don't get me started on Goana) but I question such a tiny team being able to do so securely and in a manner that is actually maintainable. I WANT it to be true but I have my concerns and the AI summary thing just kind of puts the icing on the cake.

Or how about we switch to standard time for the winter but use DST for the summer? That way we get light in the morning in winter (which you want), plus light in the evening in summer (which other people want).

(Sorry if this is a goofy answer. It just seems like we already have the solution that works best for people's natural clocks and when the sun is available.)

As I understand it, there are a couple of main claims here:

(1) OpenAI heard Buckmaster and AlpÃge were making major progress on the Navier-Stokes problem with AI methods (and maybe had some inkling of their path from prior published work), then rushed to scoop them on it, using about $23M of compute time in 88 hours. This seems to be true (and unethical/shady), but I donâ(TM)t know if it would be considered academic malpractice.

(2) Buckmaster speculates that OpenAIâ(TM)s model may have been trained on his unpublished work (completed in August and September) because he processed drafts of his paper with OpenAI tools. As far as I know, OpenAI hasnâ(TM)t publicly refuted this claim, but oddly enough, there is breaking news on it in an episode of the Odd Lots podcast, where OpenAI Greg Brockman said the latest training data for their model was from July. If true, that seems to foreclose this claim.

The fact that they're so proud of "her writing 100,00 lines" no, she prompted AI and it wrote 100,000 lines of code. And if she wants to add a new feature, chances are good AI will just tack on more to it without following SOLID principles, leading to massive bloat and duplication, and a codebase that nobody can maintain - that AI just basically has to rewrite or kludge shit in.

Congratulations, you've built a kludge tower....

We're now adding a ton of very bloated, completely opaque AI vibe coded projects to the dependency stack in a way that makes the they don't reinforce each other and just sort of like a "code stripeset" may well increase speed (of development) but at the cost of incredible fragility.

I just need to wait till XKCD says it in a far more clever way so I can point and say "YEAH THAT"

https://www.cs.wisc.edu/2025/0...

And DARPA has a whole program for it:

https://www.darpa.mil/research...

I don't buy it. They've made it to massive scale without anyone going bankrupt; clearly no one is paying 0.3% of every transaction ($100 billion/year) to keep the current system afloat. The internet itself is evidence that you can have billions of secure, complex transactions where each party just pays a minimal amount for access to the network and no third party needs to be paid for every transaction (roughly $0.06 for a couple hundred bytes in this case).

This is just banks trying to muscle in on some nice-looking transactions.

Chevrolet stopped providing (selling) updates to my 2016 Colorado's nav system, so I bought a Garmin standalone unit for maps and just use the colorado screen to connect speakers to my phone and stream.

If it ever started doing that ad thing (not likely since they aren't doing updates on my truck) I'd rip out the built in and put in a third party unit.

Honestly, part of me wants to go in and rip out the OnStar box so they don't decide to monetize my driving telemetry without my consent. (they charge too much for the onstar crap)

Yeah, but you have to regard this as a probabilistic effort: when you spend the $2,000, do you have a greater than 0.2% chance of solving the problem? If not, you're throwing your money away. And chances are, you donâ(TM)t know your odds when you start, so then things become really speculative.

This line is really chilling. The fact that it passes without comment shows that we've now gotten used to living in an authoritarian state instead of a democracy. In a country that actually supported free enterprise, a company wouldn't have to worry what the president thinks of their perfectly legal business decisions. Nor would they need to be on the president's good side to run their business as they see fit in the future. âoePolitical ramificationsâ of this sort have no place in a free, non-corrupt society.

These reports always assume a 6 oz. serving of coffee, which isn't how people have drunk it in the U.S. for a long time. A 12 oz. serving of brewed coffee contains 140-240 mg caffeine, so 400 mg/day is actually roughly 1.5 to 3 servings.

Also, as usual, this looks like an observational study, so they can't actually say whether people are healthy because they drink coffee or people drink coffee because they are healthy or some third factor drives both.

Well, there is one, but it's been a long time since I needed it. If you get your vehicle semi-stuck in a deeply rutted road (or off-road), with a manual transmission you can often "rock" it free: let out the clutch to start moving forward, push it in to roll back and up the other side, repeat to go forward again, increasing the travel each time until you get up and out of the rut. This can work well when you're stuck in a gravity well without enough friction to pull yourself out or even to build the momentum at the bottom to get yourself to the top. But it's not common to need this, and careful drivers can usually avoid it.