dgacmu

Those not familiar with it might find the history of the Cod Wars interesting (and very relevant to this discussion): https://en.wikipedia.org/wiki/Cod_Wars

(literally, Iceland having several decades of slightly-military-involved conflict with the UK about fishing near Iceland. Iceland won.)

It's a pun on The Onion's repeating headline about gun deaths in the US ("only country where...") [1] . I thought it was pretty clear before I clicked it it was going to be talking about security vulnerabilities in C, but I live close to that world so I might be a biased audience.

[1] https://en.wikipedia.org/wiki/%27No_Way_to_Prevent_This,%27_...

That's only half the problem. OpenAI is contractually obligated, if you will, to believe that models will continue improving at an impressive rate for the foreseeable future (otherwise their valuation makes no sense).

If you believe that, then you should expect to get Sol-level performance out of a Luna-cost model within six months or a year. If you have a system with the weights baked in, that means you're going to end up serving that Sol-class model several times more expensively than it will take someone who comes along a few months later. (such as what recently happened with DeepSeek's update.)

And under that assumption of continuing advancement, baking things in doesn't make sense in general - it's a play you'd make if you think things are slowing down a lot. Which may be right but it's not OpenAI or anthropic's play.

I don't think he's mocking at all - the Pittsburgh scrapping culture is fascinating. (I've lived here for 20+ years now). I throw stuff on the curb all the time and like magic, it disappears before trash day. I've also picked up a bunch of stuff from neighbors - office chairs, desks, a really nice stash of PVC pipe I used to build a grey water system for my washing machine, etc. it's like a weird little no-internet freecycle. Everyone just knows it often works and goes along with it. There are people who focus on higher value items, people who grab metals, you name it

Most likely you're seeing the causality in the wrong direction - poverty creates a strong incentive to go for immediate benefits rather than absorb the risk of one that is delayed.

But that's why they're arguing that repricing the entire generation fleet is bad compared to having a different auction for new capacity. You don't need the same incentives to keep an existing, profitable plant online as you do to invest in a new plant.

Facebook runs a number of quite complex internal distributed systems - databases, caches, proxies, etc. all of these are amenable to various forms of formal verification, and verifying them is the kind of thing that helps prevent outages and data loss.

Your second two had high comments to upvotes, which tends to get articles downranked more quickly. It may simply be that the stuff you're posting is generating disagreement without corresponding upvotes.

Will do. I'm on vacation right now and losing my laptop for a few days, but seems worth trying. My recollection from the RNGs a decade ago (I'm dating myself) was that the AES approaches had higher latency but were quite decent, though slower than PCG. Curious how that's evolved.

Oh, that's interesting - good point, since it's filtered and not trained from scratch. My prior would be to assume it's just bs'ing as LLMs usually do but it seems worth exploring.

I prefer my 8yo's answer about quantum entanglement, asked just now: "I don't know. How would I know? It's not a thing!"

Even an 8yo has better metacognition, it seems. :-)

It's not critical but if you gave me something that behaved statistically like PCG (i.e., I didn't fret about whether it was going to cause me weird problems) but was twice as fast I'd be happy and would shift to it - it would speed up profiling and measuring and that would be nice. We still find ourselves often pre-generating a list into memory to keep the prng entirely off of the measurement path. It wouldn't be magic, but I don't need magic. I like nice things that make my life a little easier in a small corner of my research. :)

This is a very weird hill to die on.

I do a lot of testing and designing of things like hash tables and filters, and having a really fast, non-CS generator is incredibly useful for being able to clearly identify performance bottlenecks in designs. PCG has been spectacularly useful for that purpose for me.

The was Samsung, iirc: https://www.reddit.com/r/Android/comments/11nzrb0/samsung_sp...

In photos after you can toggle between the AI modified and original. The original isn't as blurry as the preview shown - there's still a lot of image stacking happening and it's often very decent in my experience.

Edited to add: I have a pixel 10 pro which has a better zoom lens, so I could be having a different experience than you...

There's been a lot of fun, practical work recently on optimal join algorithms! (Full disclosure, this advisor on this paper is one of Andy and my former students, so I'm slightly biased)

RPT: https://people.iiis.tsinghua.edu.cn/~huanchen/publications/r...

Like most academic work, this one builds on some work that's been done over the last few years on ways to make the Yannakakis algorithm actually practical.

It's disjointed?

The post that started this sub-thread asked:

> 1. How many total problems were given to the model, and what percent were left unsolved at what cost before giving up? 2. How many attempts did you give the model at solving these problems? 3. How expensive was the harness, e.g. did the model have access to a job cluster?

I think it's an extremely relevant question to ask, because it helps us better understand the current state of AI being able to handle math, for exactly the reasons I outlined. I was arguing against the idea this is just a reactionary anti-AI kind of question to ask. It's not! You can be very impressed by what AI is capable of in math (I am) and still think those are really interesting things for OpenAI to disclose (I do).

OpenAI specifically called out a $2000 per problem average, which implies something that's probably not true ("if you throw $2k at us we'll solve an open problem for you"). It would be cool to know what the actual number is.

This seems very unlikely compared to the obvious answer that they added a crappy antenna to keep their BoM and assembly costs low.

The radios on these are weak and adding an Omni antenna is exceptionally unlikely to violate EIRP.

Could this design accidentally radiate something it's picking up from the board outside of 2.4ghz? Sure.

But practically it's quite unlikely to in this context.

This isn't really about delivering - it's more about helping to understand the shape of problems that AI can solve right now. If they took 1000 problems and threw the model at it and it solved these ten, is there something we learn about these ten problems and the kinds of things that current AI is good at? That's very different from picking ten problems _at random_ and solving all of them successfully, which would suggest a much less bumpy capability surface. It's interesting and it would be good science to release it.

The mining datacenters beat the AI datacenters to the punch on that, actually:

https://www.wsj.com/world/americas/bitcoin-mining-noise-driv...

The people to whom you are responding were talking about the lack of details in this article about the Guinness "trick", not the ketamine.

Without trying to answer your "I believe the setup circumvents the fee" implied question, which you should find an actual answer to...

See reddit's r/solarDIY. There are a lot of setups like this, either for full house power or partial. I have a relatively big server in my basemeent that I power this way, for example, with a battery system that acts as both UPS and a "charge the battery from solar and run off inverter when possible". Something like an off-grid EG4 is popular among the reddit crew as a way to manage this; the device tries to run locally but accepts a shore power feed from the grid so if you need it you draw grid power. No grid tie involved.

Indoor signal reception problems though

I'm close to DIYing it also. I have two of the la crosse clocks. One works well. The other consistently fails to update. It would be cheaper to buy a new one to see if it's a dud or that location, but I already gave up on it and modified it to have a programmable LED strip integrated into it to visually indicate to my son when it was ok to get up for the day (before he could read a clock).

So, ironically, I've ended up with a non-automatic atomic clock that instead contains a raspberry pi pico w that speaks ntp and has a programmable LED strip. That I have to manually set every DST transition, although the LED controller handles it just fine.

It's probably not. I tried to extract the commit hash from the version of bun packaged in claude code. the commit hashes from previous versions resolve, whereas the commit hash for the current 1.4.0 isn't available.

from strings:

   bun-v1.4.0
   f6d0fcd24abd48061873c2f1a6fb2a67eee487b8
    Upgraded.
   Welcome to Bun's latest canary build!

This build does have a different way of identifying itself than earlier builds so it's possible this string isn't correct.

(In contrast, earlier builds that I have locally have commit IDs that can be found in the public repo). I don't think it's particularly damning for them to vendor a canary release, mind you.

I used it for a bunch of remote monitor boxes to have a way of centrally managing ssh access to things that were often on- and off-line. It was simple and convenient and access was easily revocable.

> Recently, there's a lot of talk about AI allegedly finding security flaws in software. That is an unsubstantiated claim. As such, it would need to be verified by a non-machine, and arguably, the verification process would require the same amount of effort or more than would be required to find the issue to begin with.

This is simply not true. Security flaws are a great use-case for AI specifically because they're easy to verify. If you can drive a program to segfault based on inputs, you've got a good indicator it is, in fact, a security vulnerability (at minimum a DoS, but usually you find out later it was exploitable). You could even have the AI generate an exploit PoC. Shell? Valid hole. Done.

The bad use cases for AI are the ones where it's as or more expensive to verify correctness as it would have been to find the solution in advance.

Intriguing. I should benchmark my dust-gathering-stack of Titan V's, unless someone already has?

It's definitely cribbing from other papers.

https://scholar.google.com/scholar?q=W.T.%20Tutte%2C%20Perso....

Sloppy scholarship. On the other hand, it's simply a credit attribution of posing the problem, so it's not material in evaluating the results. I observe that the majority of references I can find that attribute this to Tutte are very indirect - i.e., citing sources that themselves claim Tutte was one of the people who formulated it - so it would take someone with a little more time on their hands (or perhaps an LLM) to track down the original...

I had meta's crawler hitting the pi searcher at something like 5qps for days on end, just ... querying for substrings of pi, ignoring robots.txt, etc. it wasn't enough to break anything but it triggered a lot of alerts.

I can imagine that sites with dynamic content and potentially unbounded query types or pathnames are in danger from particularly stupid crawlers.