As a concept, the pendant comes up short, as it wouldn't know where your head is pointed, let alone tracking gaze.

Glasses I can see, though in both cases the "glass hole" phenomenon seems a risk, folks not liking being constantly recorded... Further, so far I've always had to not only make sure the viewfinder is just so, but also crop to let the software know exactly what I'm talking about, centering on the subject is frequently not enough.

I've found proxmox quite nice. Almost everyone else doing linux virtualization leans on libvirt, and proxmox decided to simplify and prioritize less awkward abstraction and so I find it more straightforward of an architecture than the usual "something over libvirt over qemu" approach.

Note the prediction is given for all office jobs, not just tech workers. That would be over half of the work force.

If things get to the point that is displaces senior developers, then it can displace the vast majority of those workers too.

However, regardless of how AI works or doesn't, I do think either way the software industry is in for a reckoning. The industry broadly just isn't innovating in most areas and many of the areas where there is innovation doesn't have a lot of interest. Much of what people are doing is just remixing common stuff to cater to pointless bespoke sentiment (which is reasonably within the grasp of LLMs usually). Rather than AI, look to the ever more pressing business mandate of "offering must be cloud based and user must be locked in". Businesses know they can't sell people on upgrades anymore, they have to keep the users paying for the software they already have. Some shops realized keeping the lights on for a cloud service doesn't require a robust development team and already cut down to skeleton crews.

Exponential compute resource does not mean exponentially more capable.

In fact, capability is more logarithmic, just obscene investments in resource produce marginally better results. The biggest gain is how long the models can keep up generating tokens in a coherent looking way, but the quality of the product represented by those tokens remains flawed in now familiar ways. Most painfully the flaw are non obvious.

Gave claude opus 4.6 a cakewalk of a task today, maybe 100 lines of code to generate. The output was a failure, "passing" test cases by adding clauses to swallow errors instead of fixing the mistakes, like invoking methods and resources that didn't exist. The output was salvageable insofar that it did get some tedious plumbing roughly right, apart from swallowing all errors and I had to fix that and rip out the upper layer of stuff and do it myself. But all this was, in the scheme of stuff I do, a super easy task. Usually I wouldn't trust it even this far, but it was so mind numbingly easy I thought it might just do it and save me tedium.

Tech execs in the midst of a hype cycle are always just insufferable.

I have seen codegen accelerate stuff, but it's still just dumb as all get out. It can successfully generate longer streams of code, but subjectively the lines per mistake seem similar, and I have my best result keeping it on a short leash or maybe 6 or 7 lines of code. I have to correct that code at least half the time and I don't have the attention span to review and repair a batch of a few hundred lines of codegen when it isn't quite right.

Exception being if I'm doing something very easy yet tedious, then the models usually can go further... Though even then I get surprised at some of the mistakes it makes in that easier context. This is also the realm where prompt driven code can actually work for me, for most of my work it can code complete generally way better than traditional approaches, but the result to prompting almost never works,

In a mirror network, the TLS cert is a weaker assurance than the signed payload from the presumably trusted source. A mirror host is more exposure and more risk that the actual content gets modified despite having 'a' valid certificate.

So the TLS assurance is redundant and less specific and less rigorous than the content signature validation from the originator. Further, unencrypted protocols are friendlier to things like proxying, which can dramatically reduce load on the internet hosts if proxies take on a good part of the burden from certain institutional clients that may use an HTTP proxy to improve their performance.

So yes, if it's worth downloading, it's worth validating as best as possible, and for a lot of content the best as possible is something like GPG validation not TLS validation.

This is a pretty nonsense comment, didn't make sense..

But to reiterate, the ars article just outright invented quotations for the guy claiming them to be true. That should be the kiss of death for any "journalism" endeavor

Exactly, after your death you keep posting and oh look, you suddenly will endorse product X....

Salivating at selling ghosts to do testimonials for you...

Yeah, noticed that too, he felt the need to clarify 'best'. That sounds super odd, why would specifically only the best be able to claim that? I would have assumed it more likely for lower level developers to claim that achievement. So if you have some people writing code but only the 'best' not writing code....

Of course, I suspect it's like some executive I recently dealt with who felt that 'developers' were beneath him until they left that stupid 'coding' behind and became just executives. So he would be very excited to declare a 'coder' that professed to not writing a line of code in months to be the best.

Well, for a company like Spotify, the downside isn't so scary because their software isn't exactly doing rocket science either. Their business is internet radio with on-demand capability. The technical piece is relatively basic enablement of that direction that isn't difficult and others can and have easily competed on technical design. See the cited three features, *super* easy sounding stuff.

But I *have* seen a software sales guy fail to understand the point you just made. He was excited because now when a customer asked for some software, he thought his non-technical sales team members could just put that request into a gemini prompt and then sell the customer the software the AI just generated. He seemed to completely fail to recognize that Gemini isn't some exclusive secret to his team and if someone without skill can do it on his team, then the customer could just do for themselves. For this vision to work even in theory, it has to be software for yourself, or enabling your actual business objective where the 'stickiness' comes from.

Yeah, they speak to stupid shareholders, and shareholders that might not be stupid, but are willing to bank on the stupidity of others, either way, currently money wise the money flows to the hype.

But very good point that this *should* be a double edged sword. Our software can be completely constructed low skilled using LLM, so what might prevent competition from just eating their lunch on the technical front? Of course, it is just spotify, and it's not exactly a technical marvel to begin with, basically riding marketing and content rights and the technical piece is pretty basic anyway.

Very consistent with my experience. Sure, it can accelerate certain tasks, but it will blunder along the way.

Even if it gets something mostly right but I see a mistake, it sounds like they say explain the mistake to it to let it try to fix it (which for me when I tried was very unreliable, and more work than just manually amending the code, since I also know that correcting it's mistake won't even pay dividends because it won't 'learn' from that interaction). I have similar experiences with people, it's sometimes a bit of work to explain to them, but at least there I know they get better because of that direction. There was a person I just cut off because he never learned anything and kept making the same mistakes over and over again, and LLM usage feels like that. LLM is better than that guy because at least it is immediate in the proposed code and I don't have to feel bad about failing to 'teach' by going in and just fixing the mistakes directly.

I note the weasel word 'best' in front of developers, so he pretty much decided that by definition the 'best' developers are the ones exclusively working through AI.

The two party stuff is just ridiculous to cover the gamut of what people want. Need more diversity of viewpoints and avoiding the potential for one party to go so off the rails and enjoy so much control without having to really work with anyone outside their circle.

Well, there was the Steve Bannon proposal about identifying "risky" voting locations and deploying ICE to 'manage' the situation, and the White House refused to deny that course of action. More vaguely they talked about GOP "taking control" of voting locations more directly from Trump, though no one knows exactly what he meant. Around here they are shutting down a number of college campus polling locations and other democratic hotspots to make it more difficult to vote. They both are trying to impair mail-in voting and reduce early voting.

Then there was what they pushed for in 2020 but Mike Pence wouldn't carry through on it of just discarding the election results.

So a lot of that may fall under voter suppression, but it is quite a bit to possibly hurt the elections.

Of course, even if they get a majority of the House, they still can't do *too* much to an executive branch that runs rogue unless the republicans in the senate work with them to do something about it. They can make the formal funding story more complicated, but an administration willing to ignore norms can do a lot...

Yes, the folks who murdered a US citizen instead saying that it was related to the immigrant that was allegedly in some building around, but neither the woman being pushed around or the man who was shot was in their way....

Getting all pissy about getting filmed and yelled at and abusing and killing citizens instead of even trying to do the job they were allegedly there for.

I certainly wouldn't trust them with any equipment that requires a modicum of responsibility, judgement, and capability.