kacey's recent activity
-
Comment on A couple of posts about AI and the Independent Games Festival awards in ~games
-
Comment on A couple of posts about AI and the Independent Games Festival awards in ~games
kacey Link ParentI certainly hope so. I think we're starting to see cracks in it. Maybe this year, at the New Year's gathering, someone else will bring up AI, and I can see what other peoples' reactions are. 'Cuz...People need to re-enter reality and realize people will almost always opt for the lowest effort option available.
I certainly hope so. I think we're starting to see cracks in it. Maybe this year, at the New Year's gathering, someone else will bring up AI, and I can see what other peoples' reactions are. 'Cuz last year it was rough. Hopefully people understand that more than just scammers and ex-tech bros can use these tools, though, and that genuinely good things can come from them as well.
-
Comment on A couple of posts about AI and the Independent Games Festival awards in ~games
kacey Link Parent(I'm not unkz, just dropping in) If I may lovingly place a single word into unkz' mouth, I think that their point is that local inference -- or even just inference sat in someone else's data...(I'm not unkz, just dropping in)
What if inference does, against all odds, become too expensive?
If I may lovingly place a single word into unkz' mouth, I think that their point is that local inference -- or even just inference sat in someone else's data centre, at the pre-hype rates -- is already feasible, so it'd take like a bacteria that eats silicon wafers to make it stop. Apple's apparently shipping an app to run local models, too, so I can't see how things won't only get more stuck in from here.
-
Comment on AlphaProof Code: a breakthrough in verifying complex software at depth in ~tech
kacey LinkTo oversimplify, this is a human workflow + LLM agent process that helped generate a formal specification, and then autonomously proved its equivalence to the semantics of a library in Signal (in...To oversimplify, this is a human workflow + LLM agent process that helped generate a formal specification, and then autonomously proved its equivalence to the semantics of a library in Signal (in particular, on a single instruction architecture, for a single thread).
It should hopefully go without saying that this doesn't solve all bugs forever, but my hope is that it indicates the feasibility of generating similar proofs around some core security principles to prove the safety of the rest of the world's software + hardware stacks. Mostly useful for adjusting priors and all that for now.
-
AlphaProof Code: a breakthrough in verifying complex software at depth
3 votes -
Comment on The web is more than enough in ~tech
kacey Link ParentMmhm, mostly agreed. I think my concern around a "type 1" bubble burst (so to speak) is that much of the pricing of Anthropic/OpenAI/Google imply that only they hold the secret sauce to both (1)...A software industry where there's more people making software for each other would seem to imply they're buying a lot of tokens? But maybe the token API pricing is a lot cheaper, or people switch to local inference.
It would be sort of like how Sun was really big during the dot-com boom, but afterwards, people figured out that they didn't actually need Sun computers to build websites. Only much later did cloud computing take off.
Mmhm, mostly agreed. I think my concern around a "type 1" bubble burst (so to speak) is that much of the pricing of Anthropic/OpenAI/Google imply that only they hold the secret sauce to both (1) developing irreplicable, frontier AI, and (2) the means to use them. If other companies develop meaningfully competitive models (e.g. the Chinese open weight models), or demonstrate that they can host at lower profit margins (e.g. take a gander at OpenRouter), then it sucks the value proposition directly out of their companies. Since they're -- in theory -- being valued on capturing all the value from selling those tokens + training those models, that'll "pop" the bubble by correcting the market from their current valuation to a dramatically lower valuation, more in line with Moonshot AI or Deepseek plus a small buffer for operating on American soil. That would be ... bad ... for the industries that have leveraged themselves assuming that the American AI labs continue growing.
Er, perhaps more concisely, even if everyone is buying tokens in a hypothetical future, they need to all be buying tokens from X AI in order for their valuations to make any sense. If they aren't, then pop goes the bubble. I'm not an economist, though, and I may absolutely have missed a beat somewhere.
I don't think things will go back to "normal" either, but I have no idea what the new tech industry's economics will look like. I wouldn't have predicted that Google could earn so much from ad revenue for so long.
Oh, agreed. I pulled the ripcord on the tech industry because all I could see was fire and brimstone; everything looks hazy but scary to me. I don't like unpredictability, and extrapolating from progress in the AI field, it didn't sound like the cushy, prestigious, salary-competitive software dev market would survive this. Not that I have a crystal ball either, just that every plausible path I could think of would be miserable or unprofitable (or both!).
-
Comment on The web is more than enough in ~tech
kacey LinkJust as an aside to this article, I'm increasingly getting the impression that people are attaching different semantics to the term "AI bubble" than I am? My read is that there's shaky financials...I put maybe twice as much effort into building a web-based office suite as I might normally put into writing a long newsletter post. I’m not telling you this because I am a very special boy. I’m telling you because if you want to, you can learn to do it, too.
[...]
This is one of the places where “are we in an AI bubble” gets fuzzy for me, because it’s totally clear all these giant AI companies are huffing AGI paint fumes, but at the same time, if we actually unlocked the economic transformation of, say, 100 times as many people making software for each other, that would blow an enormous crater into the current tech economy.Just as an aside to this article, I'm increasingly getting the impression that people are attaching different semantics to the term "AI bubble" than I am? My read is that there's shaky financials (e.g. massive circular financing and cutthroat markups) involved in combination with some unresolved questions about the future of the industry (e.g. preliminary share pricing seems to imply a winner-takes-all, frontier models-first, high profit margin market). So the "bubble" popping will still result in all of today's technology continuing to be deployed, development will just slow down a little until labs retarget cost efficiency over boundary pushing.
By contrast, some friends/folks have used the term to seemingly describe how all AI usage is cult-like and completely useless? So the bubble "popping" will be a return to ~2020 normalcy: AI image generation will disappear, all slop articles online disappear, the jobs come back, etc.
By the quote above, I get the impression that the author is rhetorically targeting this article at that second interpretation, since the economic argument about overvaluation wouldn't be affected by the statement: "if we actually unlocked the economic transformation of, say, 100 times as many people making software for each other, that would blow an enormous crater into the current tech economy." If anything, that would blow up the "shaky financials" bubble, especially if the tooling ends up owned by non-US AI companies, since it directly refutes their proposed business models.
Anyways. Maybe it's food for thought?
-
Comment on Qwen 3.8 27B on my laptop can interpret model weights, too in ~tech
kacey Link ParentAs ruspaceni noted in the other thread, mixture of experts models -- such as Qwen 3.6 35 a3b -- work fine on relatively low amounts of VRAM by being able to efficiently spill out into system RAM....As ruspaceni noted in the other thread, mixture of experts models -- such as Qwen 3.6 35 a3b -- work fine on relatively low amounts of VRAM by being able to efficiently spill out into system RAM. Before I blew $1300 Canadian loonies on an Arc Pro B70, I was running it on an RTX 2060 6gb card.
We've seen a few other, smaller shops train up MoE models that seem to be approaching the performance of some of Qwen's older, smaller models (e.g. Mellium2.1 12b seems to strike in the range of Qwen 3.5 9B, for instance), so it seems reasonable to assume that we'll see models that can run with less memory as time goes on. It's just that they fell out of vogue recently, presumably because the larger labs realized that they could court more favour with 100B+ models that compete directly with the present day frontier, rather than replicating performance from 6-12 months ago.
-
Comment on Qwen 3.8 27B on my laptop can interpret model weights, too in ~tech
kacey (edited )Link ParentI'm largely quoting the article, and added "average" in order to underline that a multi-million dollar datacentre was not used in this effort. The assertion of this post is that the startling...I'm largely quoting the article, and added "average" in order to underline that a multi-million dollar datacentre was not used in this effort. The assertion of this post is that the startling result from the other day was, in fact, relatively mundane. The LLMs aren't suddenly evolving into Skynet or anything; their capabilities are just slowly improving/are stationary (given that Astra is surely like a hundred+ orders of magnitude larger than Qwen 3.8 27b and provided less than a hundred orders of magnitude better results).
I’m delighted to share that Qwen 3.8 27B running on my laptop can do this as well!
Understood that "average" might have tweaked some people; apologies for using such language.
-
Comment on Qwen 3.8 27B on my laptop can interpret model weights, too in ~tech
kacey (edited )LinkI was sniped by Simon Willison, and did not have a chance to try this myself XD it is a response to a post from the other day about how a frontier LLM could interpret weights in models directly....I was sniped by Simon Willison, and did not have a chance to try this myself XD it is a response to a post from the other day about how a frontier LLM could interpret weights in models directly. Simon Willison tossed Qwen 3.8 27b at the problem and it did OK (3/4 games identified by inspecting weights directly). IMO, the takeaway should be that (1) models have probably been able to do this for like a year, except no one asked, (2) it doesn't take a frontier, billion dollar model to do this -- the one which runs on your average laptop gets most of the way there.
-
Qwen 3.8 27B on my laptop can interpret model weights, too
23 votes -
Comment on A lukewarm defense of LLMs in ~tech
kacey Link ParentFor future reference, acronymfinder.com is still a perfectly cromulent way to find information without invoking a few hundred billion artificial neurons. The top results were "Sony Music...For future reference, acronymfinder.com is still a perfectly cromulent way to find information without invoking a few hundred billion artificial neurons. The top results were "Sony Music Entertainment", "Small and Medium Enterprises", and "Subject Matter Expert".
(edit) I've also heard the term in a couple other fields, so maybe it's moreso a regional thing? Although IIRC RoyalHenOil and I are about as physically distant as is reasonable on this continent.
-
Comment on Setting up own email FQDN (and maybe service?) in ~tech
kacey (edited )LinkNah. Although I'd note that typically people refer to this as "getting your own domain", even though technically the wording you used is more accurate, in case that helps searching for resources....if there's faulty logic in my mindset AND
Nah. Although I'd note that typically people refer to this as "getting your own domain", even though technically the wording you used is more accurate, in case that helps searching for resources.
what are the common irritants/nuances of getting down my FQDN and connecting it to an email hosting service and ensure everything works as expected and any emails I send don't wind up going to people's spam
Effectively nothing. Your email host will have a short series of instructions on how to set up the appropriate DNS entries which will inform other providers of your messages' providence. The only real concern is if you get a really weird TLD (e.g.
.tkwere known to be used for spam a while back), in which case, some sites might automatically assume it's spam/fake (so they'll filter mail going to them as well). But I haven't had that happen in years, and you're good on anything typical (e.g..com,.ca,.org, etc.)if I do wind up going the third route, I imagine it wont be possible to migrate all my emails with me when I have to do a migration cause a service is starting to go down the drain? Is there a way around that? I'd like to have access to all my emails in 1 places.
I don't quite understand your concern. Perhaps to illustrate, when you want to port out of an email host, you'd get the new one to connect over IMAP and pull all your existing emails out of the first (and/or CalDAV, WebDAV, and CardDav entires if your host supported those too). Email aliases -- which is the common term used to refer to your suggestion about generating addresses for different purposes -- are a per-provider sort of thing, and I'm not sure I understand how relevant they are to your question about switching to a different host.
Overall, the process of getting a domain set up is fairly straightforward:
- Buy a domain (if you'd like a
.cadomain, I recommend canspace; otherwise, tld-list might be useful to understand if your registrar of choice is ripping you off) - Sign up with an email host (I suggest fastmail, but if you want to go absolutely crazy, you can always host your own IMAP server and pay a pittance to use a third party SMTP server (to ensure that your emails aren't dropped due to low reputation). Highly recommend against that, though, unless you want to make 24/7 sysadmin'ing your hobby)
- Set up domain w/your email host (on fastmail, this is copy/pasting a few DNS entries into your registrar's default DNS host. You can use cloudflare instead, but tbh, if you're only setting up email, it's easier to just do this),
- Import all your existing emails using IMAP (it's a quick and simple process on Fastmail, and should be the same for other providers).
It's, like, an afternoon project at most, even if you're as flighty and distractible as myself. Definitely worth it in order to avoid getting locked into e.g. a gmail domain for the rest of your days.
(edit) also, sidebar, but you can't use underscores in domain names. It's like most unicode characters these days but not them, afaik.
- Buy a domain (if you'd like a
-
Comment on LLMs can natively read and write neural network weights in ~tech
kacey Link Parent"Awe", potentially? But the dictionary definition, not the colloquial one. I assume we used to feel it more when exposed to forces beyond our understanding and control, hence the dilution of the..."Awe", potentially? But the dictionary definition, not the colloquial one.
Awe is an emotion comparable to wonder[1] but less joyous. On Robert Plutchik's wheel of emotions[2] awe is modeled as a combination of surprise and fear. Keltner and Haidt, in a highly-cited paper, describe it as 'In the upper reaches of pleasure and on the boundary of fear'.
I assume we used to feel it more when exposed to forces beyond our understanding and control, hence the dilution of the term as it sought out a new definition that could apply to everyday life again.
-
Comment on LLMs can natively read and write neural network weights in ~tech
kacey LinkI haven't dug into the paper much, but from several examples, it looks like it effectively reverse engineered what the weights were representing for a variety of specific examples? Which makes...I haven't dug into the paper much, but from several examples, it looks like it effectively reverse engineered what the weights were representing for a variety of specific examples? Which makes sense that it should be able to do that ...
W.r.t. the Qwen math result, that's interesting! It did have 50 tries, though, and it sounds like it fixed one issue that gradient descent couldn't bump the model out of a local optima for. Might be interesting to see if this sort of approach comes into play for training, if it's relatively cheap?
Overall: I didn't have a guess on whether or not this was feasible, just a general sense that we don't have a firm grasp on how most things an LLM does are performed. Notably, single bits of behaviour (e.g. refusals) were known to be removable relatively easily via abliteration, so correcting individual behaviours doesn't seem novel at least. I guess this is a cool result, but it's hard to tell what it means (or where it's headed) without more papers or a better grounding in the field.
Maybe some experts are around who can comment?
-
Comment on Any tips for improving personal cybersecurity in 2026? (I'm concerned about rogue AI agents) in ~tech
kacey LinkNot really. I'd be more concerned with severe weather events, earthquakes, or getting hit by a driver taking a right on red. Nope. If you look up "software verification", you'll note that it's...How alarmed should I actually be about this, if at all?
Not really. I'd be more concerned with severe weather events, earthquakes, or getting hit by a driver taking a right on red.
Is it realistic to think that a "cyber pandemic/lockdown" might happen in the near future (some scenario wherein global internet becomes generally unsafe/inaccessible and we'll be restricted to small local networks)?
Nope. If you look up "software verification", you'll note that it's possible to mathematically prove the absence of exploitable vulnerabilities of the variety people talk about when these doomsday scenarios come up (i.e. access to something you shouldn't, deleting someone else's data, or preventing people from accessing stuff). It's just always been less expensive to buy insurance than to build secure systems, hence why this stuff is rarely used (except for situations involving life safety). But as soon as someone invokes the "sufficiently advanced AI" argument to justify fear about Skynet taking over, they should equally explain why anti-Skynet can't just lock everything down.
How are you updating your personal cybersecurity in 2026?
I'm not. If you're looking for a safe password manager, KeepassXC is fine; just find a way to sync it to your mobile device if you're into that (or throw your database onto a USB stick, and plug it into your phone as need be).
How concerned should I be about digital banking?
So one thing you should do is contact your bank and have them disable whatever voice-based identity verification they're using, since several of them are these days. That's already trivially hackable and will continue to get worse with time. You'll want to do the same for your mobile phone company (also ask them to enable porting protection, so it's harder to steal your number) and your email host.
I'm considering at least having printed out records of bank balances, in case of a scenario wherein digital banks' records are affected.
As noted elsewhere, it's not harmful to do this, but it probably won't help. A quick search turned up a Reddit thread which noted that many banks use offline backups anyhow, so it's hard to imagine how an AI would wipe your account records. Without, like, wiping cities off the map or something.
How are you balancing cybersecurity measures with practicality / ease-of-use?
I barely think about cybersecurity. For the past ~30 years or so, it was a safe bet that anything hosted off of your local machine will be leaked or hacked eventually, so I just didn't put much out there. If your entire life is on the cloud that might not be as feasible.
-
Comment on Proton’s CEO cancels AI ‘re-interviews’ in ~tech
kacey LinkI wonder if senior leadership could be replaced by AI at this rate. Proton is one of those companies that could do nothing and win -- since "getting security right" is not rocket science, and...I wonder if senior leadership could be replaced by AI at this rate. Proton is one of those companies that could do nothing and win -- since "getting security right" is not rocket science, and their competitors are few and far in between -- so incentivizing risky customer acquisition behaviours by promising large incentive packages to C-suite execs who promise to find growth (via launching new features) is the exact opposite of what their existing, profitable, conservative, security-focused customer base would want.
Admittedly I don't use their services, so I could be misreading the target demographic.
-
Comment on How is your current TTRPG campaign going? in ~games.tabletop
kacey LinkI'll be running a "Film Reroll"-inspired one shot of Event Horizon in GURPS with some friends this Halloween ... I suppose I'll post back here once it's done. Broad strokes: a real time clock...I'll be running a "Film Reroll"-inspired one shot of Event Horizon in GURPS with some friends this Halloween ... I suppose I'll post back here once it's done.
Broad strokes:
- a real time clock counts down until the Event Horizon plummets into Neptune's core (a novel detail intended to impose a time limit, since the GM of the long running game this is interrupting wants to finish the campaign this year),
- we've used Blades in the Dark-style progress clocks to great effect before; I'll use it here to ramp through threats the ship throws at them (jump scares, system failures, hallucination threats, then daemons of the warp),
- since there's real-time pressure, I figured it might be fun to offer players an IRL time consuming challenge to pass failed rolls. Like, solve a rubik's cube, or finish a sudoku. Really hoping for moments like players panicking, trying to collectively solve a crossword to open the airlock before a hellbeast beats down the door.
-
Comment on How to make tubular bell wind chimes by Lee Hite in ~hobbies
kacey LinkThings have been a little crazy for me lately, and thinking about making chimes out of cheap EMT conduit sounded like a nice project. Granted, there's not much wind in an apartment, but I run into...Things have been a little crazy for me lately, and thinking about making chimes out of cheap EMT conduit sounded like a nice project. Granted, there's not much wind in an apartment, but I run into things often enough that I could hang it around a corner or something. Or in front of a desk fan.
-
How to make tubular bell wind chimes by Lee Hite
9 votes
Zoinks, that's, uh. Yeah kinda wild. MoE models mitigate how awful local hardware is w.r.t. memory bandwidth, but that's probably like a 1.375 bpw model. Or else one of those crazy sixteen slot Xeons.
But yeah, things will only get faster. There are hard upper limits due to memory bandwidth, as noted, but better draft models, improved quants (e.g. trellis, autoround, etc.), and -- someday -- maybe a change in architecture (e.g. diffusion models) should push local hardware even further.
Fingers crossed! That'll mean this is a bubble and I get to watch OpenAI/Anthropic implode in a fiery economic fireball (admittedly, this will singe everyone else too, but I really want to live to see them go down).
(edit) there's promising stuff coming for CAD modelling, and hopefully robotics will follow that. Fingers crossed, but that stuff is far more theoretical at this point than I'd be willing to bet on (although I'm hoping to get directly involved in it ...)