IHLayman

Wait. You are taking the RF response, sending it up to the cloud to interpret, and then sending the result back? Hard pass from me, sorry. Cool tech, but I don’t want my internal thoughts anywhere near the cloud. Also, this would mean I couldn’t use something like this on the subway when the phone signal goes out intermittently, right?

... and this is why the training canary[0], or anything like it, apparently to protect us from RSI, will never work. Please allow me a small digression from the main topic.

This team was smart enough to realize that this is the kind of technology that could be used for RSI, because they call it out in their post near the bottom. I imagine they know why a lot of other researchers care about it. And they evidently don’t care to use the canary being used on other research published like it, like the recent post by OpenAI about two agents using the CPU cache to communicate [1, near the bottom for the cache experiment and the canary at the very bottom]. In fact, that post by OpenAI was the first time I learned to look for such a canary, and wondered whether such a scheme would work if it was so non-obvious to think to add such a thing or how to agree on such a canary or when to add it.

This is probably one of the times when it would appear that such a canary is warranted, and it’s not like these researchers are not in the know, right? Let me know if this is the first time you are also hearing of such a mechanism. It feels evident to me that it must be useless to attempt to implement. How much of us talking publicly about the possible jailbreaks from a container will bring exactly those scenarios to fruition? It’s almost like we’re doing the jailbreak research for the LLMs.

[0]: https://alignment.org/canary [1]: https://openai.com/index/securing-our-frontier-model-trainin...

Does playing Wingspan count?

Cyber operators I agree would tell you that physical airgaps are necessary, and I know they are being used for running agents. But ML engineers training agents? That is far more complicated. At that point sensibly you are talking about physically airgapping an entire datacenter, and when moving around petabytes of training data, model weights, training results, researcher code? The process moves too slow for most companies to tolerate. This is how we got into this mess in the first place. I think the industry woke up to the idea that the security threat model of training ML models has fundamentally been rewritten, and there is some serious non-trivial scrambling to come up with a better workflow.

100% on point. We're talking about Artifactory here, and even though people love to hate on it, I don't think anyone is claiming that using Artifactory to isolate these agents while giving them a way to run pip/cargo/whatever to install packages they need to write and test scripts.

Yes, the AI labs are making security mistakes. But, choosing a reasonable package repository (regardless of individual opinions on Artifactory, it is widely used and usually considered a safe choice) that their LLMs found 8 zero-days in? That was not one of them. Even if OpenAI were totally on top of their game, these paperclip-maximizers are going to find holes in the sandboxes they are given. And fully airgapped systems are frankly unrealistic for most systems. This whole narrative of blaming the frontier labs for shoddy security, while being quite an attractive target narratively (they made the mess, they should clean it up), misses the fact that from a cybersecurity perspective these LLMs, no matter what they are, are inherently dangerous and require far more resources to secure, not to mention expertise that I am unconvinced actually exists.

Wait wait wait. I don’t accept this premise: “But anger is different from anxiety because anger needs to be directed somewhere.” Citation needed; where did this come from? I think it is very possible to be angry without a source or obvious target of that anger; IANAP but my 10-minute googles could not find any authority in the field that would claim this statement.

From the perspective of the article, it seems that not having a target distinction between anxiety and anger would make the two feelings indistinguishable and thus the argument would fall apart. BUT, I do think there is a difference between the rage that comes from anger and the unease that comes from anxiety, and that may be triggered by the uncanny valley feelings from AI or the loss of job security from AI or the feelings of replacement that may come from AI or the environmental concerns that come from AI or the capitalistic underpinnings of the current implementation of AI, the list goes on. Some people may take anxiety from these things, and others may take anger. I don’t see one emotion as being less valid than the other.

Also, I don’t see any use in saying someone shouldn’t be angry or anxious in a particular situation. You can’t control your emotions like that. You can control how you _react_ to your emotions, but not the emotions themselves. I do think that emotional regulation is important in work (and also in life), but that doesn’t mean that you don’t feel what you feel.

“ What will happen to democracy in a world where 100% of the population are 27/7 consumers?”

…we’ll add three hours to our day?

Bu seriously, I support what you are saying. This is why the entire consumer system needs to change, because in a world with no jobs it is by definition unsustainable.

> "Smart Local Control" home devices work as expected until the electronics fail

Recently one of my Zigbee-controlled thermostats started pumping cold air constantly. To fix it, all I had to do was open and examine the board; one of the varistors got some battery acid on it when I had an alkaline battery burst in the unit. Because it was a no-name with an actual PCB, I was able to solder a new varistor in place, and it works good as new.

So I would say that "Smart Local Control" isn't the problem, but rather the ability to repair the thing. Also, the thermostat was $45 when I purchased it 5 years ago, so it was a good investment IMO. I think that's why everyone is upset about the Nest gen 1 and 2 sunsetting; there should be no reason that these devices should be breaking now (no failing electronics) but they die anyway because the company is too cheap to keep an extra endpoint running.

You don’t need an LLM for this. Use `kubectl` to create a simple pod/service/deployment/ingress/etc, run `kubectl get -o yaml > foo.yaml` to bring it back to your machine in yaml format, then edit the `foo.yaml` file in your favorite editor, adding the things you need for your service, and removing the things you don’t, or things that are automatically generated.

As others have said, depending on an LLM for this is a disaster because you don’t engage your brain with the manifest, so you aren’t immediately or at least subconsciously aware of what is in that manifest, for good or for ill. This is how bad manifest configurations can drift into codebases and are persisted with cargo-cult coding.

[edit: edit]

Anaconda was a good idea until it would break apt on Ubuntu and make my job that much harder. That became the reason _not_ to use Anaconda in my book.

venv made these problems start to disappear, and now uv and Nix have closed the loop for me.

Oh sure! While we’re at it, let’s increase the existing dystopia by giving employers the ability to track our stress levels and let them “compensate” as they see fit…

Alphabet will definitely try to do that (within their business interest and all that), but I still choose to believe in the precept that “the net interprets censorship as damage and routes around it”, as old and outdated as that sounds.

A number of my privacy-minded friends choose a bi-modal approach: have two phones, one for work and one for personal. They don’t get the recent model (costing half as much), hold onto the old phone for as long as they can, use one phone for “required” apps (Okta, Slack, those websites that only work on Chrome…) and the personal phone for everything else.

As annoying as it is, i think that compartmentalized devices/accounts/apps are the only way forward.

I love the idea of this flake to run Ollama even on Windows, but just pointing people to your _everything_ flake is going to confuse people and make it look harder than it is to run Ollama on Nix.

If you are using a system-controlling Nix (nix-darwin, NixOS…), it’s as easy as `hardware.services.ollama.enable=true` with maybe adding `.acceleration=“cuda”` to force GPU usage or `host=“0.0.0.0”` to allow connections to Ollama that are not local to your system. In a home-manager situation it is even easier: just include `pkgs.ollama` in your `home.packages`, with an `.override{}` for the same options above. That should be it, really.

I will say that if you have a more complex NixOS setup that patches the kernel or can’t lean on cachix for some reason that using the ollama package takes a long time to compile. My setup at home runs on a 3950X Threadripper and when Ollama compiles it uses all the cores at 99% for about 16 minutes.

Is it because pi isn’t measured, but calculated? The wikipedia article (https://en.m.wikipedia.org/wiki/Physical_constant) makes a distinction between a mathematical constant and a physical constant, stating that the latter cannot be calculated but instead needed to be measured experimentally… Pi could be measured experimentally, but it has an exact definition and can be calculated outside of any experiment.

How this article discusses reproducibility in NixOS and declines to even mention the intensional model or efforts to implement it are surprising to me, since it appears they have done a lot of research into the matter.

If you don’t know, the intensional model is an alternative way to structure the NixOS store so that components are content-addressable (store hash is based on the targets) as opposed to being addressed based on the build instructions and dependencies. IIUC, the entire purpose of the intensional model is to make Nix stores shareable so that you could just depend on Cachix and such without the worry of a supply-chain attack. This approach was an entire chapter in the Nix thesis paper (chapter 6) and has been worked on recently (see https://github.com/NixOS/rfcs/pull/62 and https://github.com/NixOS/rfcs/pull/17 for current progress).

“Even are ads are private…”

I don’t think that’s possible. Modern advertisers require those sweet sweet metrics, which can never be truly anonymized, and in most cases attempts to anonymize are half-hearted at best.

One book not on that list that really threw me recently was The Anomaly [0]. The less you know about it going in the better.

[0] https://en.wikipedia.org/wiki/The_Anomaly_(novel).

Assume around $3/hr per H100 (pretty generous pricing for GCP), that is $2250/month-gpu, or for their fleet of 8000 comes to $18MM/month or around $216MM/year in just compute costs alone, not looking at SSD, bucket storage, or egress. At their initial investment of 465-320=$145MM that means they can’t have operated that cluster for longer than 6ish months without their funds running dry or the got massive discounts somewhere.

Something doesn’t add up here.

I disagree… the article talks about defining the meter using the pendulum and the second. Other planets would have their own definition of second, but not their own definition of pendulum. Since one meter is prescribed though a pendulum because of the oscillation formula and dependent on the frequency alone (in this case, 2 seconds), no matter what planet you are on, how strong gravity, is or how long a second is, the pendulum describes a relationship between seconds and meters such that if using this method a planet’s scientists would always define their units such that acceleration of gravity was eerily equal to pi^2.

Makes me think of possible lunar scientists unwittingly making their meter 5/6th shorter(edit: english is hard) and then marveling at the same coincidence…

I can talk to the opposite side about why I am very happy going all in with NixOS. Prior to that I was running dual-boot Windows (for games) and Ubuntu (for coding). Despite all of the progress made on gaming on Linux, I was still having problems with the Nvidia drivers in Ubuntu, sway refused to work, hyprland glitched out, and I was stuck on i3 (because I am a full convert to tiling windows managers). I used home-manager with Nix to manage my dotfiles and had been quite happy with that workflow.

As the story goes, Ubuntu had just released 24.04 and I pushed to upgrade. My linux partition was on btrfs. I had been bitten by the snapshots-locking-my-system problem before so I cleared out all my snapshots (and backed up my drive to local NAS). During the update Ubuntu did ANOTHER SNAPSHOT, it took up all the space on device (256GB, yeah I should have had more space) and crashed the system. At that point I was finally done with all of the shenanigans that Ubuntu pulls with apt (not updating everything unless you have their subscription, forcing some installs to go through snapd, separate driver management from apt…), and this last update crashing my system’s boot, even though I could have recovered with my backup, was the last straw.

I considered my experience with Ubuntu, and MSFT’s actions to integrate Copilot into the OS, and decided to wipe the entire 1TB drive and go all in with NixOS. I was able to move my home-manager config right over, get my configs all converted to flakes, and get hyprland set up all in less than an hour. Smoothest transition in a long time. I even was able to jump to the 6.8 kernel and things just started working again. I was finally able to play steam games well on my system again. I may even have a path to get my VR to work in Linux again.

Are there still problems? Sure. Hyprland is a compromise because sway still won’t cooperate with Nvidia GPUs. But, when I have made changes that get me to a non-bootable state, I can boot to a previous generation directly from GRUB, see what went wrong, fix it, and get back to a good state. Having that ability to recover from a bad boot situation without hacking into a safe mode or GRUB console or rip the drive out and put it into another system was incredibly refreshing. I am not going back anytime soon.

By the looks of it, the announcement webpage also breaks Tridactyl unnecessarily, so I doubt they are interested in supporting *vim... Joking aside, I would be shocked if someone doesn't quickly write a plugin for at least Neovim to support this.

One of the more interesting perspectives I gained about Lord Byron was when I learned of his support of the Luddites from reading the book Blood in the Machine[0], which I am surprised the posted article never mentioned once since it seemed to be a significant part of his fame, especially during his time in the House of Lords[1].

[0]: https://www.hachettebookgroup.com/titles/brian-merchant/bloo...

[1]: https://www.smithsonianmag.com/smart-news/byron-was-one-few-...

Fun fact: if seems that if you use Lockdown mode on Apple devices you can't open PDFs from a browser (no official documentation says it but there is anecdotal evidence). This would allow people with Lockdown mode to open Arxiv papers more easily.

IRL universities will probably develop a protocol to remove access to helper tools for this sort of thing, maybe metal detectors at the door and specially tailored computers if needed for quizzes. Online universities OTOH are probably going to feel the bite of this the hardest. The CKA exam I took a while back had draconian measures to try to prevent me from cheating, such as taking the test with a webcam on me the whole time, with tools in the browser to limit where I surfed, in a room that has NO PICTURES on the walls. The tech to provide answers is moving faster than the tech to secure a remote room from such outside answers. And if in-person proctoring is eventually required, I would wager that will be a death-knell for online universities, who will have a hard time obtaining that physical space on a regular basis while staying marginally profitable.

“There was a slight increase in risk of HZ in people receiving influenza vaccine in the first 1–15 days after vaccination.” That’s kind of overselling it, isn’t it? The OP study says that stroke chances are lowered for 6 months. I think many people would take that trade off.

… why have the operator sit in the cab? If truly teleoperated, then wouldn’t it be safer for the operator to be in a room far enough from the robot so that if there is some accident that crushes the cab or sends electricity through it, there is no damage to the operator?

You are looking for the price like I was, and many pages are trying to snare you onto “speak to a specialist” links. Let me spare you the trouble, as you are looking for this page: https://www.insight.com/en_US/shop/partner/magic-leap/store....

If you’d rather not click (and I totally understand) let me tell you that the Leap 2 costs between $3200 for the base model and up to $5000 for the enterprise model, which gives you a better viewpoint.

I will take a stab at this... it is about object localization via sound instead of sight, using very little power. For comparison, let's say you have a sound sensing setup on a Raspberry Pi, known for its low power draw, attached to a pair of microphones, looking to alert the location of an object by similar sound triangulations. That processing could take maybe 4-6W of power (continuous running of the Pi and attached mics, as a generous estimate), and could be quite effective.

However the technique in this paper is _ultra_ low power. First off, they model the design off of a barn owl, and using "neuromorphic memristors" (sounds technobabble to me but I didn't understand that part). But in the Results part of the paper they claim they can sense movement sampling every 1/10th of a second using only 250 microwatts of power, orders of magnitude more efficient than a naive approach, with only 22 floating-point calcluations per sample. Sounds quite impressive, but I wonder what the actual applications of this will be, even though I'd love to be able to track mice around my house in real time grrrr.

There’s also https://en.wikipedia.org/wiki/Italki. But I actually have shied away from it, mainly because I get nervous talking in front of a live person who will be judging my capability as I talk. I much prefer the idea of a non-feeling ML algorithm talking back at me, even if it makes GPT-3 like gibberish mistakes.