Yes, but a lot of harnesses change previous context. E.g. the system prompt injects the current time/date, working directory, files in the working directory, etc. Compaction also changes the whole previous context. I _think_ changing the list of tools also invalidates cache, so invoking a subagent with different tools would invalidate the cache.

My vague impression is that it's in a similar vein to functional programming languages. It generally disallows doing things that lead to bugs (cache misses in this case), and presumably allows you to do those things in a way that makes it much clearer that this is likely to cause cache misses. I would guess that in this paradigm, you don't mutate your existing session, you derive a new session by mutating the prior context into a new context.

In that world, a) we are already at or close to having enough memory in local devices to do inference locally, and b) that memory isn't inference-specific and can be utilized for other things. Most devices come with enough memory to do some level of inference, and some come with plenty (eg a gaming desktop probably has 32GB+ of RAM in it).

You aren't going to run Kimi on it, but I think the reality for a lot of consumer inference is that it doesn't need to be. It's going to be a lot of things that are soft, and easily answered by a search API, so the LLM really just needs to be able to skim and summarize. Going a step further, we may even see some kind of hybrid approach where a local OpenRouter kind of thing decides whether the task is soft enough to do locally with models that fit in RAM or if it needs to be farmed out to a PaaS provider.

> I don't really understand why would the team that already got a full plate decide to also invent a whole new Linux system while they're creating their hardware device.

Honestly, I just wouldn't solve that. Nix makes the product way harder to sell, and home-building a solution is either a) an entire product all by itself, or b) shitty, in the "it only works on very specific happy paths" sense.

I also frankly just don't think it's a feature worth as much to consumers as it costs to make. At worst, it's a minor inconvenience to reflash a Pi card. If I'm really lazy, I just disable systemd services for whatever was on it and layer the new stuff on top. It's like 5 commands to get it back to "close enough to fresh".

In capitalism, the wealthy provide the capex and reap the profit. In China, the government provides a lot of the capex and claims a lot of the profit (either directly or indirectly). They can then re-invest that, use it to set salaries free from market forces, etc.

The US could do more of that without getting into the contentious issue of raising personal taxes. Ie the government could control the flow of GPUs into the country, asserting that it had the right of first purchase, distribute them to a tightly controlled private company (think power or water company, with margins set by law), and probably make a profit on the cheaper hardware because they’d be a massive purchasing block. We won’t, though, because Google, Amazon, Oracle, etc are currently posed to make a _ton_ of money off doing basically the exact same thing.

This is blatantly untrue in the face of even a cursory thought. If the deli didn’t exist, that worker would have done any number of other jobs. Subsistence farming, doing deli stuff out of his house, maybe even community funding a coop grocery store.

That was how things worked at many points in human history. Instead of “private citizen makes $thing and collects rent” it was some variety of “municipality funds $thing for the community”. That just doesn’t extrapolate to a global economy well, for better or worse.

My non-tech friends and family would probably be served perfectly fine by local models today, if they had a working web search tool. Their queries are often “soft” and don’t have an exact answer. My mom and aunt used it to pick a hairstyle, my mom used it to get an image of what a room would look like with particular drapes in it, etc. Stuff I think mid-sized local models like Gemma or smaller Qwens could do without issue. They just don’t have a device that will run them.

Businesses won’t move. They need a huge context so they can stuff a bunch of Confluence pages in it and 300 tools and it needs to read an entire codebase and yada yada. The hardware depreciation and electricity will probably make it a net zero or even cost more than paying for API access.

On the practical side, "learn Nix" is a _massive_ onboarding task. Without Nix, I'd probably pick one up assuming I'll find something to do with it. With Nix, I'd wait until I have a project I know is worth figuring out Nix.

If this were my project, I'd probably go with the absolute most simple answer: multiple SD card readers. Install the base OS on one card, allow hot-swapping the other card, do some mount point stuff with the other card (like maybe it auto-mounts to /usr/local, and have packages install into /usr/local). Or maybe some kind of overlayfs with the other card. SD cards are cheap, and I'd rather glue an SD card holder to the back of a Flipper than learn Nix.

The way your argument reads to me is “software as a category has such little utility that profit margins can only be derived from corner cutting”.

The reality of the landscape is that most companies don’t get hacked as the result of an incredible and novel Spectre-esque attack, it’s something bland and entirely preventable.

Eg https://nvd.nist.gov/vuln/detail/CVE-2025-31324

SAP got a CVE because they just flat out didn’t implement auth on an endpoint in an app architecture that will execute files just for being in a certain directory, and also didn’t prevent writing files to executable paths (or maybe that’s how the feature works, not a SAP person). For every 0 day with a novel root, there are like a thousand that are some kind of humdrum “didn’t enforce auth/SQL sanitation/XSS/other well known exploit with comprehensive solutions”.

I do think there are good reasons to withhold some classes of exploit. If a hacker writes a 14 page proof on how to beat some encryption we had no idea was vulnerable, that’s one thing. Getting owned for making an insecure architecture and then not even putting auth over it is a whole other issue.

E.g.

https://github.com/kstenerud/yoloai/blob/main/internal/fileu... <- that recursively creates directories, but will only change permissions on the innermost dir (user may be unable to cd into intermediary directories)

https://github.com/kstenerud/yoloai/blob/main/internal/mcpsr... <- all the json.Marshal calls in this file just suppress errors, so if anything un-marshallable ends up in there the app will return empty strings with no errors logged

https://github.com/kstenerud/yoloai/blob/main/runtime/regist... <- `Register` embeds a copy of the code from `IsAvailable` because of the locking; that could be replaced with a private `isAvailable` that has no locking that both use (after doing their own locking)

https://github.com/kstenerud/yoloai/blob/main/runtime/exec.g... <- these functions are identical except for the strings.Trim, one should just call the other and then trim the output

Just out of curiosity, I enabled some other linters and it looks bad. Excluding test files, there are 110 functions with a cyclomatic complexity over 10 and 7 that are _over 50_. The worst is at 86, which is mind-boggling.

Could probably find more, but you get the drift. I'm sure it runs, but stylistically this is more along the lines of what I would expect an intern to do.

This is also sort of nit-picky, but like half the stuff in https://github.com/kstenerud/yoloai/blob/main/docs/dev/backe... isn't idiosyncratic, it's just the way those things work and a lot of them aren't even tricky. The one linked is particularly blatant; that's not limited to os.Stat that's literally just how permissions work. Denying permission on inodes is a property of the folder, not the file.

The device is large and makes the user look weird and non-present, which are net negatives.

The only benefit of the AR is showing the directional arrows, but they could get the same thing with much less weird looking non-prescription glasses with arrows sharpied on them. More realistically, anyone really using for mobility probably develops muscle memory for which direction to look to go where and then they don't even need that. At that point it's just a really expensive, really clunky camera.

The whole bug bounty thing is a mess, admittedly, but lacking a bug bounty program entirely feels like immediately losing the moral high ground on “you should have told us first”. There’s a lively debate about what bugs are worth, but it’s objectively not $0 for many classes because a botnet developer will buy them for some amount.

Personally, a big part of my view is formed by the educated assumption that security practices will never improve unless poor security becomes a liability. That’s unlikely to happen with “responsible disclosure” because it gets swept under a rug. Immediate public disclosure changes that risk calculus a lot. I think wed see a lot more downward pressure from vendors to their suppliers if $RandomSaaS had to worry about losing their pants because Oracle had a vuln published.

Full VR hasn't done well, but it does continue to make me wonder if there's a market for a stripped and slimmed device. I'd maybe be interested in a device that does optical controls if it fit in regular-sized glasses. I'd be super interested if it had a HUD system (even a super basic one that can only show a handful of symbols). Better still if it had some basic audio, but maintaining the "regular glasses" form factor is more important to me than the HUD or audio.

Someone with likely substantial qualifications put in time to find this. The company is in it for profit (at least partially). What’s fair for the company is fair for the individual. The company can either offer to pay for bugs under the terms they want, hire more security folks to find the bugs themselves, or just accept that researches get to do whatever they want with their findings.

I’d tell Mullvad, but there are companies I don’t respect enough to feel compelled to give them a heads up. Perhaps the author feels that way about Mullvad, it’s entirely within their right to use this to publicly shame Mullvad.

There’s a bunch of weird corner cases that come with that. What if the receiving account ID doesn’t exist, or the recipients bank flags the transaction, or the sender closes their account before the recipients bank has applied their end of the transaction.

I think some of these transfer systems are simpler because they’re global so both records are in the same database. Transactions can be technically atomic rather than organizationally atomic (ie atomicity is present in the technology, rather than approximated by processes).

It’s a breath of fresh air that practically every DSL takes either lists or maps so I can use very similar patterns to build their input rather than “every API wants me to method chain their custom types so every API needs its own special helpers library for common patterns”.

I never realized how much I hate classes until Clojure.

Have you tried typed Clojure? Curious about opinions on that vs having Malli do runtime validation. I tried it when I was a total noob and got overwhelmed. I feel like I have just enough context now to try again, and not sure if it’s nice or the overhead is so high it’s a boondoggle

It does not appear to be an effective deterrent. https://www.academia.sg/extra/death-penalty-research-appendi... This article has a criticism of the SG government report (Study 6 header) on the deterrent effect when they added the mandatory death penalty in the 90s. The big takeaway is that convictions didn’t drop notably (cannabis convictions dropped a single percentage point, opium convictions went up 2%. Average opium weight seized dropped a ton, but is still like 13 times the mandatory death penalty limit so hard to call it there).

I think this is under appreciated, in part because it’s incredibly hard to fix. There’s a void where the US cultural image of masculinity used to be. As an adult man, I couldn’t really describe what a prototypically masculine person looks like or does or thinks anymore.

There’s a loss of identity there that we haven’t really rebuilt.

I wouldn’t point entirely at the PhDs, though. There were some real issues that were called out (inability to communicate, over reliance on anger as an emotional outlet, etc), but the identity could have stayed largely intact.

The killing blows were from segments of men who doubled down on the most negative aspects of masculinity, and made the rest of the men flee from that image of masculinity to avoid any association. I would rather drop masculinity from my self-perception than be associated with Andrew Tate or Logan Paul or whoever else.

I don’t think I know any men that have “modern masculinity” (whatever that would mean) as part of their core identity. They’re either clinging onto a Chuck Norris kind of masculinity, or just don’t have a strong gender component of their self-identity. They’re not feminine, they just don’t do anything to be “manly”. Being “manly” matters to them about as much as whether they’re a Coke or Pepsi person.

I’ve not used it in production, but I’ve shipped a few side projects and stuff for family members in it. ClojureScripts React wrapper, Reagent, honestly makes more sense to me than React does. I used Hiccup to generate HTML, and your components are just functions within Hiccups DSL (which is really just lists) and it ends up looking incredibly clean. Static things look static, dynamic things are obviously so, and it felt much less magic than regular React.

The only things I found that felt bad were trying to use non-functional components I found on NPM. It’s not a deal breaker, but the code was ugly. Nothing I couldn’t fix with a wrapper, but some JS libraries are heinously ugly in cljs by default.