Mathnerd314

There are some state laws - California, Virginia and Colorado, a bunch of others. And the SECURE Data Act currently in Congress. https://statescoop.com/house-subcommittee-secure-data-act-pr...

So I'm looking at Figure 2 of the study and it shows that risk is negative (I.e. it's a benefit) below 7 drinks per week or 1 drink per day. And then they say "There was no protective net effect of alcohol observed at any level of alcohol consumption." And then they discuss the observed protective effects and just say "this body of evidence [that we used] has substantial limitations". They also say "Readers should therefore consider both the point estimates and their associated CIs when interpreting the risk thresholds presented in this study." but then their CI indicates only 2+ drinks per day can cause statistically significant issues.

Overall, a very boring study. It would be more interesting if they published the code and we could play with the numbers, seeing as all the data they used is public, but we can't even do that.

This may help https://www.workboat.com/white-house-budget-would-slash-noaa...

The administration wanted to eliminate everything ocean-related, seems like they are doing it program by program and this is yet another. Probably explains the lack of context, this is like article 20 about program closures (and maybe 200+ more to go)

It is probably like with smart TV's where the value of the telemetry data ends up subsidizing a significant fraction of the hardware. Car manufacturers seem to be doing a lot of experiments with what they can charge for in terms of ongoing subscriptions. I am sure if they could show ads without it being considered distracting they would.

Maybe we need Uber for airlines. Pilots don't lose their skills, passengers always have demand, the issue is that pricing is too predictable. You could see this with skiplagged, there really is room for fare pricing innovation. Start by capturing the private luxury market, work down to commodity.

The author sounds like he actually responds to feature requests, though. Typical behavior I'm seeing is that the maintainer just never checks the issue tracker, or has it disabled, but is more likely to read PR's.

website is not up to date, https://www.youtube.com/watch?v=lqVN1fGNpZw is not on there

"Inference consumes 60–90% of total AI lifecycle costs." So shovel is not the right analogy, more like GPU = coal burning engine. And yes, coal was a big railroad expense, more so than financing construction debt.

The author is named Alok, so I would expect alokscript to be a self-authored programming language. But I checked the GitHub profile and I don't see anything.

Well, it's a little more than that, like they agreed to issue him a debit card

Yeah, it's like just ignoring the actual conclusions. I mean, it's easier to hit and then it hits less hard, which is actually good because professional baseball players generally hit it out of the park when they get a good hit. So the conclusion is actually completely opposite what the title of the article is - it's not the same. It is a substantial improvement.

to quote: "in the Persian Gulf today, the Navy grasps the reality of the circumstances, recognizing that it simply can’t sail into the strait without risk getting blown to smithereens by Iran’s missiles. Today, its carriers are stationed well outside the Gulf and the ranges of Iranian missiles."

It is on Wikipedia under T-Bank, this seems the best source that announces the resolution: https://web.archive.org/web/20220905212700/https://www.tinko...

Well, if you come at it from the mindfulness angle, there are real studies showing that mindfulness works. https://pmc.ncbi.nlm.nih.gov/articles/PMC8083197/ and similarly, if you come at it from the religious angle, you can trace a lot of the aspects of mindfulness back to the Buddha's original teachings as recorded in canon. And if you ask if there is a fundamental point beyond those, I think the answer is that there is none recorded - the best description I have been able to get of Nirvana is that it is a state of perfect mindfulness.

There's a specific writing style for globalized English that AI's use. And then this post also had none of the stylistic flourishes that a real author might add. And then simple things like constructing a table of 68 libraries or whatever organized by relatively subjective categories. That is something that nobody is going to do by hand.

Well, I mean, you can certainly say economic value doesn't capture all of the value. But you can also say that there are metrics of value that do capture everything. Thermodynamic entropy, for example - its steady march to zero is statistically unstoppable. You can't measure a child's economic value without making a lot of assumptions, but you can measure a child's thermodynamic heat production with a few simple experiments. It might sound a little out there, but I've been looking at the maximum entropy production principle and some books on thermodynamics, and there really is a lot that is applicable to calculations about human systems. Viewing humans as dissipative structures designed to maximize entropy production really explains a lot about how the world works. Notably, some questions about our energy usage patterns. AI may not be useful economically yet, but it's excellent at dissipating heat.

At least with Gemini, I found the trick is to add anything in any system instruction about a task list. Then the follow-up prompt will always be, do you want to add a task for that? Which is actually useful most of the time.

https://doi.org/10.1038/s41591-025-03622-w this is the paper they're basing the research on. So in primary care, the accuracy rates are in the 80s. So that's something like a 17% false positive rate. That's still like 5 to 1 odds of getting a correct result though. It's much better than nothing.

https://www.federalregister.gov/documents/2025/11/20/2025-20...

Well apparently Congress passed a law that said TSA could just demand money as long as they published a notice in the federal register.

If you are flipping through the reading to find a quote, then printed readings are hard to beat, unless you can search for a word with digital search. But speed reading RSVP presentation beats any kind of print reading by a mile, if you are aiming for comprehension. So, it is hard to say where the technology is going. Nobody has put in the work to really make reading on an iPad as smooth and fluid as print, in terms of rapid page flipping. But the potential is there. It is kind of laughable how the salesman will be saying, oh it has a fast processor, and then you open up a PDF and scroll a few pages fast and they start being blank instead of actually having text.

I tried a thread, I got that both LLMs and humans optimize for the same goal, working programs, and the key is verifiability. So it recommended Rust or Haskell combined with formal verification and contracts. So I think the conclusion of the post holds up - "the things that make an LLM-optimized language useful also happen to make them easier for humans!"

I get that this is essentially vibe coding a language, but it still seems lazy to me. He just asked the language model zero-shot to design a language unprompted. You could at least use the Rosetta code examples and ask it to identify design patterns for a new language.

I update Linux maybe once a year. Sure, there are security vulnerabilities. But I'm behind a firewall. And meanwhile, I don't have to spend any time dealing with update issues.

Logic minimization is kind of boring? I had to solve a problem once and the answer was still to use the espresso software from the 1980s. It is a pretty specialized problem and honestly I don't see how you would improve on it, besides integrating the digital circuit design research. But in terms of software, there is not really any reason to use a Boolean logic formula instead of just passing around the truth table directly.

I'm wondering if a firewall is a solution here. Don't mess around with the stupid device settings. Just block the Xbox store, and then presumably Minecraft uses different server IPs, so you can let those through.

Research from the University of Amsterdam’s IViR “Global Online Piracy Study” (survey of nearly 35,000 respondents across 13 countries) found that for each content type and country, 95% or more of pirates also consume content legally, and their median legal consumption is typically twice that of non‑pirating legal users.

Takedown letter: https://github.com/github/dmca/blob/master/2025/12/2025-12-1...

For me, the entire inbox is this DBTC folder. I have notifications set up on my smartwatch and I triage each email in real-time as it comes in. If it's urgent, I act on it. If it is important or I want to follow up, then I add it to my (separate) to-do list, with a Google tasks voice command. And otherwise I just ignore the notification and the email sits there in the inbox until I feel like dealing with it. I use the unread status and pick things off in occasional focus sessions. Some things never get "read", and that's because they don't matter. Zero bandit stuff because I know exactly what's in my inbox at any given time, at least up to what my analog brain can hold. It fits right into the old "I heard a noise. What is it?" routine humans used when we were hunter-gatherers.

Well, so what the actual ruling was was that use of the books was okay, but only if they were legally obtained. And so the authors could proceed with a lawsuit for illegally downloading the books. But then presumably compensation for torrenting the books was included as part of the out of court settlement. So the lesson is something like AI is fine, but torrenting books is still not acceptable, m'kay wink wink.

So, the takeaway I get from this paper is that if you have a language model and you set it up so it has an input and it generates an output that is towards some goal (e.g., "make this sentence sound smarter"), then it should converge, because it is following a potential function.

But I have used prompts like this a fair amount, and it is more like stochastic gradient descent - most of the time, once it is close to the target, the model will take a small incremental change, but when it is really close the model will sort of say "this is not improveable as it is" and it will take a large leap to a completely different configuration. And then this will do the incremental optimizations and so on. This could be an artifact of the sampling algorithm, but I think it is also an issue that the model has this potential function encoded, but the prompt and the structure of the model do not actually minimize this potential. So, a real lesson here is that there is actually a lot of work still left to do in terms of smarter sampling. Beam search like is used today is sort of the tip of the iceberg. If we could start doing optimization with the transformer model as a component, like optimizing pipelines of reasoning rather than always generating inputs and outputs sequentially, that is where you could start using this potential function directly and then you would see orders of magnitude smarter AI. There is stuff about prompt optimization, but it is still based on treating models as black boxes rather than the piles of math they are.