It is a characteristic of neural nets that they do not have insight into their own functioning.
It is arguably a characteristic of any intelligent system, that at least some part of it must be opaque to itself, but the previous sentence is more defensible than a generalized claim.
If you don't understand what that means, tell me from your own metacognitive insight what parts of your brain are being used to read this. Not because of learned knowledge about what parts of the brain do what, through your own insight in your own functioning. You can't, because you don't have any.
This isn't just that human rationalize a lot. This is below that. This is that even if you notice yourself rationalizing, which is something you can train yourself to do, you have no access to the underlying computations/processes of the rationalization itself, or the process of noticing you are rationalizing.
There is arguably still a sense that we experience in which we humans could reasonably say "No, I'm pretty sure I used addition-with-carry to answer you", so that is perhaps not the easiest example to think about the experience of. But there will always be some question of "how did you do that" to which you can give no answer because the answer is in the firing of the neural net itself and you, who is in one way or another the product of that firing, do not have access to that. How did you quickly catch that ball that someone unexpectedly threw at you? You just did, as far your neural net is concerned.
(Also, while I've expressed this in terms of your conscious experience, this doesn't have anything to do with "consciousness". Neural nets in general do not get this feedback and do not and can not have arbitrary metacognition about their own functioning. This is an artifact of my writing text to address conscious beings.)
Yep. An effect cannot in itself reason about its cause. Some effects can suggest causes though. For example you tend to have memory of an algorithm you just executed, when the usage was at least somewhat conscious or intended, which can lead you to be able to guess which algorithm you used (and potentially even correctly)
The economic tension of these "run by default" AIs is quite hilarious once you see it.
On the one hand, the relevant KPIs of whoever is driving this product needs to be able to show AI usage is increasing, because AI usage is obviously the Platonic embodiment of goodness [1].
On the other hand, these things are expensive, so while it's mandatory that Google searches stuff these things in our faces, they are also horribly underprovisioned. If my only exposure to AI was the various search engine popups or the other free AIs, or even the bullet-point AIs that I'm nominally paying for but not really, like in Office, I would also have a pretty negative view of AI. I use DuckDuckGo more than Google but whatever model they may nominally be using to power their search result summarizer, it is de facto at least two years behind the state of the art in a very fast moving industry. It frequently gets things exactly backwards and is clearly leaning on its internal model a lot more than the links it has supposedly read, and clearly has a thinking budget of "indistinguishable from zero", and I don't know what kind of summarized web page content is being fed to it but it must be getting brutally dismembered in whatever summary is being fed to the AI.
The debate about how useful AI you pay for may rage on, but at least at this point in 2026, I'd say the AI you can get for free is every bad thing anyone says it is.
[1]: I believe there is a lot of useful things current AI can do, but there is no level of quality AI can ever reach in which AI usage for the sake of using AI will ever be a terminal good. Honestly any manager, whether they be a line manager or a CEO of a multinational company, that has ever pushed that in any capacity, should be fired for demonstrating gross incompetence for that position. It's "second or third week of Econ 101" or so that you learn about why it's never a good idea to just open the checkbook and spend an unbounded amount of money on something, and nothing you'll learn further down the line will ever contradict that.
That is not a fair summarization of their point because that is not the grammar. There's commas, slashes, asterisks, combinations, and then if you want randomization you need to put it in the command itself because cron can't do it. (Some crons can, but it's not a general capability of cron.) Writing a non-trivial cron spec is not easy.
I am familiar with the syntax, so I am biased ("*/3" and "12,14,20" makes sense if you are familiar with Unix tools), but it is still more intuitive to me than the systemd unit file syntax and usage. I know that I just have to edit /etc/cron or throw any executable file into /etc/cron.d/monthly and it will work on my system, but I cannot write a systemd timer file from scratch without looking it, and to do that I first have to find the directory where the other examples are located. /etc/systemd doesn't appear to be it.
This is generally my only real complaint about systemd. I don't care if it is too monolitic, written in C or whatever, I just want a straightforward syntax for straightforward operations. I'd like it if systemd could recognize if a .target file is a shell script and just do "the right thing". Perhaps it would make sense for a timer file to recognize cron syntax as well. Or at least allow for a kind of extensibility so that I can have it supported.
If systemd had a little more respect for existing conventions, I am pretty sure it wouldn't be so controversial. After all, system administrators like it because they use it all the time, but a regular, full-timer user like me, who only deals with it when something is broken or have to use it as a means-to-an-end to set something up, then all friction is annoying and bad UX. (And no, using Nix is not the solution)
Yeah, it would be nice to have a folder like /etc/systemd-jobs/ where I could put them and where there are no files unrelated to job scheduling. There is /etc/systemd/user, but it does get a bit of pollution depending on the system.
Not sure if you're talking about cron or systemd, but cron definitely has that in /etc/cron.d where you can have arbitrary crontabs, or /etc/cron.{hourly|daily|weekly|monthly} where you can just place arbitrary scripts if you don't care exactly when they run, just the frequency.
Systemd lets you create templates that take an argument in from the scheduled service. It gets that from the value after the @. So you can write a unit file that schedules a task to run say every 3 days and in that unit file reference `jobs/%i`, then put your task in a file in jobs and say `systemctl start every-3-days@script1.sh` to run `script1.sh` on your schedule without needing to create a new unit file for each script. StepCA has a nice write up on their site about using these templates to schedule cert renewals for any arbitrary service
It's a mystery to me why everyone tries to use OnCalendar here, when "n amount of times within a certain timeframe" can be done much more easily with OnActiveSec, in this case that'd be OnActiveSec=6h.
$ systemctl cat public-inbox-watch@.timer
# /etc/systemd/system/public-inbox-watch@.timer
[Unit]
Description=Periodic fetch of public mailing list
[Timer]
# twice a day
OnCalendar=*-*-* 5,17:35
RandomizedDelaySec=1h
Persistent=true
[Install]
WantedBy=multi-user.target
1) Your production equipment doesn't have its TZ set to UTC? Enjoy dealing with the intermittent and irregular hassle of DST changes, I guess.
2) From the crontab(5) on my system:
The CRON_TZ variable specifies the time zone specific for the cron ta‐
ble. The user should enter a time according to the specified time zone
into the table. The time used for writing into a log file is taken from
the local time zone, where the daemon is running.
If you have a job you need scheduled in a different timezone, dump a new file in /etc/cron.d, alter its CRON_TZ variable and go to town, as it were.
That's simple but consider "run something 4x per day but randomize a delay by hour so all of the 200 servers doing that task won't run it all at once"
In cron, you basically have to either use your configuration management to generate those times, or have a random delay script running before the command
In systemd timers, it's just
OnCalendar=0/6:00:00
RandomizedOffsetSec=60m
and the offset generated will be stable for the job on a given machine (i.e. always same on this machine but different on others) so you will get nice uniform distribution of load.
If you add
Persist=true
the job will also be run once if there was one or more scheduled runs when the machine was down
> In cron, you basically have to either use your configuration management to generate those times, or have a random delay script running before the command
Nope. From crontab(5)
The RANDOM_DELAY variable allows delaying job startups by random amount
of minutes with upper limit specified by the variable. The random scal‐
ing factor is determined during the cron daemon startup so it remains
constant for the whole run time of the daemon.
That's from my cronie install, but it looks like this has been a feature of some crons for at least a decade. (Notice that the post date of [0] is in 2016.) Given that cronie is based on vixie-cron, and I think I was was using vixie-cron in 2002, I bet it's been a thing for at least twenty years.
[1] This question is sarcasm. SystemD is often like this... dead simple things look dead simple, but complex things are -if they're possible at all- at least as complex as they are everywhere else.
I found the systemd time spec syntax you referenced to be logical and well thought out.
Cron syntax is simpler for the easy cases because cron tries to do less. It ignores years and seconds entirely, and doesn't try to adhere roughly to ISO8601 ordering and field separators, instead using space universally for field separation and euro-style least-to-most significant field ordering. I like ISO8601, so I get along with systemd's style better, despite it introducing slightly more cognitive load.
The only thing that threw me for a loop and seems like "special magic" was
> "Mon *-05~07/1" means "the last Monday in May."
But good luck doing that in one line in cron.
Some cron-style libraries seem to support L/W/# for last / nearest-weekday / nth of month, but I don't know if any system crons do. (cronie? dcron? I don't think so. fcron? bcron? I don't see it there either.) '#' is syntactic sugar for DOW + 7-day range, while L is covered by the above quoted syntax.
If your cron has that kind of syntax, then for a case like "weekday closest to 1st of month", "W" is more convenient than writing 3 systemd timer rules to cover the three cases (weekday day 1, monday day 2, friday last day of month), but that's a big if. Generally you'd have to write 3 rules in cron anyway.
> I found the systemd time spec syntax you referenced to be logical and well thought out.
I found this amusing when in combination with
> The only thing that threw me for a loop and seems like "special magic" was
but -regardless- a careful reader notes that I never said that the Timer scheduling syntax was illogical or poorly thought out. It's at least as complicated as crontab-style time syntax, which was my entire point.
Related: Not that it's ether part of the core scheduler syntax or necessarily as nice as having it in the core syntax, but my crontab(5) suggests that one can use things like date(1) to get more fine-grained control over the time of execution:
As noted above, skip values only operate within the time period they´re attached to.
For example, specifying "0/35" for the minute field of a crontab entry won´t cause
that entry to be executed every 35 minutes; instead, it will be executed twice every
hour, at 0 and 35 minutes past.
For more fine-grained control you can do something like this:
* * * * * if [ $(expr \( $(date +%s) / 60 \) % 58) = 0 ]; then echo this runs every 58 minutes; fi
0 * * * * if [ $(expr \( $(date +%s) / 3600 \) % 23) = 0 ]; then echo this runs every 23 hours on the hour; fi
Adjust as needed if your date(1) command does not accept "+%s" as the format string
specifier to output the current UNIX timestamp.
While I expect that you're not one of those people, I know that folks who are accustomed to working with extremely inflexible tools forget (or never learned) that these sorts of things are possible. I'm very aware that people sometimes cut off their own limbs with power tools, but that's not a good reason to ban their use.
Looking at the other examples on that page, I'm gonna say that it's only arguably easier to read for basic stuff... especially if you're familiar with the syntax. The complex stuff is -at best- just as difficult.
"He’ll do anything to move the goalposts and turn his failures into successes. There is no norm he won’t violate, no boundary he won’t cross."
Unfortunately, if you really start digging in to what is going on in the financial world, you will find he has violated no norms here. This is not a defense of Elon; this is a condemnation of the entire financial industry.
The whole thing scares me, honestly. It has never been a clean happy market where lots of honest people get together and are just honestly trying to make a better world for each other, there is no golden past where people were just nice or anything, but damn if computers don't let people build some structures that the robber barons of old could only have dreamt of. I'm really concerned that "index and chill" doesn't just have a "best by" date but that the best-by date could be in the past; I've heard of an awful lot of ways of exploiting it and other retirements schemes we have, this is just one. I find it implausible that these ideas exist but nobody is doing them.
You can find such accommodating customers if you listen to customers.
In most companies executives can bypass processes to make projects happen - when done well that allows long term investment to happen when the business case is too complex to reduce to an RoI - when done poorly you end up launching a lot of pet projects that have no market and never will. I feel uniquely positioned to have a good understanding of the maluses and benefits of authoritarian project creation from all three angles and the best solution I've seen is to let it happen but bring down the hammer if things get too absurd.
I started to find that the AI bit was the most useful part of Google Search. But the actual search results were terrible and now I use Kagi. I like being able to add a question mark and control what becomes AI and what doesn't. I use normal search like a Ctrl F for the internet and don't want it to be too clever.
Yeah, gonna be honest. I ridiculed people for it before, but now I use Google's AI a lot. I haven't used Google in over a decade otherwise and I still don't. I use Brave for search but Google's AI is better than anyone else's for what I do. You heard it guys, I was wrong. I admit it.
Google's search AI caused a panic in my household just several nights ago. We were looking up an issue with our pet and the very first thing my partner saw searching google was an AI answer saying you need to panic, this is an emergency, get your animal to an emergency room asap before they die. That of course kicked her into Oh Shit mode. But aside from that answer none of the actual search results either of us were finding backed up the AI.
We had to decide whether or not to drag our pet to an after hours emergency vet, with all the associated stress and cost, or ignore the AI result and go off everything else we were reading. It's one thing to dismiss AI answers that seem wrong when it's a domain I know well or the stakes are low, but this was not that type of scenario.
In the end we opted to ignore google's AI and fortunately it was absolutely the right call. So, thanks google.
I think that's foolish engagement with the technology though. Never panic, always double check. I don't take it as factual. It's just a statistical summary of something it read and it must always be treated as such. In your case, why not just ask 5 other AI's and take a poll?
Google's AI is solid, but a lot of its benefit comes from the fact that it hasn't been enshittified into the ground like their search. The plain incontrovertible benefit is that it has the potential to reduce clicks from the user - instead of sending the UX through clicking into sites and then back to the main search pane the information is immediately available. That click through ends up being expensive for three main reasons 1 - the site is sometimes potato though google "solved" this with amp, 2 - it takes a while to scroll to the site entry passed all the sponsored links (a problem google created for themselves) and 3 - paywalls and "Please subscribe to our newsletter" popups are legion, but while this was always a little bit of a problem, it is a much more pronounced problem due to 1, the insertion of amp that threatens financial ruin on being crawled by google.
So, at the end of the day Google's AI is better than ye olde Google Search, and ye olde Google Search is now very difficult to accomplish because of how much Google has poisoned the well. Kagi is excellent but most normal people can't stomach paying for search when there's a "free" alternative so your options are Google's AI, Google's Search or and alternative search - most normal folks don't realize non-Google search alternatives exist (outside like Bing) so Google's AI ends up capturing a lot of usage.
My non-tech savvy mother started reading the stuff the Google Search AI answers to for some searches, and she's already fed up with it saying whatever. To her it doesn't matter that the "AI can make mistakes" because (in her own wording) "if it's faulty, don't answer".
There's a difference between "linking to a source that may be incorrect" and "you providing the text that's blatantly wrong", and Google seems too big to care about it.
Google only cares when something gains traction (and thus potentially hurts its bottom line". For example, it was answering "How many p's in Google" with "There are two p's in Google"[1] for long enough for me to get non-technical family members making a joke out of it. Google fixed it, only for it then to briefly tell you there were no p's in 'Alphabet' either.
Those particular bits on nonsense seem to have been stopped for now, but let's not fall prey to Gell-Mann amnesia. The only problem Google has fixed is "our LLM was hurting our reputation in this specific case". They have not, and likely cannot, fix the underlying problem.
I fail to understand why people think LLMs working in tokens (and thus having a hard time counting letters) is such a huge gotcha. Ask it to write you a bash script that counts the amount of p's in a word you input it'll do it fine. Then ask it to write that script in C, JavaScript, Lua, and a few other languages. Then ask your family members to do the same. The gotcha can go both ways.
Because it proves they are not, in fact, intelligent and do not, in fact, have any understanding of what they are doing. It's really obvious what people think it's a huge gotcha: because it is.
That difference can be pretty small when most of the web is also consumed by low effort automatically generated slop. You can't escape it. I know there is a good web out there, but the search engines refuse to give it to me. Probably their algorithm. preferring recent content over good content.
I suspect the only real answer is an economic one, Something like Kagi, where hopefully by paying for results you change from being the product to a customer and this is incentive enough for them to provide good results.
I find that those "AI summaries" google tends to use by default, are hallucinating liars. I stopped wasting my time with this AI slop spam in general. Any "human" still using AI and targeting me, gets perma-banned without any further discussion. I kind of need ublock origin for EVERYTHING. (Ublock origin is great, but I need this on every level, blocking AI slop spam, blocking Nate's donation-daemon nag-widget for KDE and so forth - ok, the last one is easy to disable, just patching out the part where Nate thinks it is ok to harass people, but for AI slop spam from external sites I need something more effective than ublock origin, kind of like an ublock colossal shield.)
I find it takes me more time... sometimes it has the answer, often it has pure bullshit. I need to verify everything that comes out of it myself before repeating at anyway.
If people are just naked copy-pasting that field... which, ugh, of course they are... they are doing themselves and others a disservice.
Pretty wild to see this in a technology forum. The whole idea of technology is to do more with less. It inherently reduces the set of people who are still economically valuable. Have you ever wondered how many livelihoods computer programmers have destroyed? Now that programming itself is on the chopping block, suddenly some moral line has been crossed?
> The whole idea of technology is to do more with less. It inherently reduces the set of people who are still economically valuable
I don't think either of these sentences is true.
> Now that programming itself is on the chopping block, suddenly some moral line has been crossed?
I didn't say anything about a moral line, I just said that there are a lot of programmers who are very excited to remove themselves from being employable. I didn't even say whether I thought that was good or bad!
>The whole idea of technology is to do more with less
No, that's not the whole idea of technology. Technology exists to, and this is a non-exhaustive list: fostering creativity, play, enabling autonomy and self sufficiency, connecting people to each other, collaboration, building skills and enhancing people's natural capacities. That's a smalls et of examples of what one could imagine technology to do.
Are these AI tools connecting people, genuinely enhancing their capacity to think, increasing or diminishing their creativity and self-sufficiency? The answer looks pretty obvious to me.
People who think technology is something to be used to "do more" and then discarded have no actual interest in technology as an object of study, or as a set of pro-social tools. That's not the mentality of a hacker. Why would it surprise you that people on a website called hackernews don't have the same attitude as a guy in an MBA class?
It absolutely is. Technology is what lets you farm more with less resources. Manufacture more with less materials. Get more things done in less time. Fostering creativity and play? Yeah, by getting more things done in less time, so as to free up that time for all the fun stuff you actually want to do. Of course, to others this "fun stuff" is just yet another inefficiency to be eliminated.
We might love technology and think it has intrinsic value, but the rest of the world could not care less. It's about maximizing profits and minimizing costs. That includes the costs of human labor. No hacker here thinks of the amazon mechanical turk employees that AI will certainly get rid of, for example. I bet the hackers here will be at their most self-satisfied when they finally manage to replace doctors.
So it is indeed pretty rich to read complaints on so called Hacker News about hacking itself being automated away by AI models. It's a very arrogant perspective. Did hackers think they were above this? Did they think they were irreplaceable? Hilarious.
I think that a lot of those overly enthusiastic engineer AI fanboys are just playing a rational game driven by their perception that AI is a effective substitute for them, and that the only way to survive the comming culling is by being seen on the market as something an AI thought leader.
Basically, signalling that they are going to be cooperative subjects for the enemy's occupation of the land.
"I, for one, welcome our new giant insect overlords" is, IMHO, the operative meme here.
Others are just addicted, the cycle of fast interaction and reward in coding agents is not very different from gambling or crack cocaine.
I think the prevailing mindset amongst developers who use LLMs is that actually LLMs are more of an effective augmentation of programming tools in the same way that an IDE is, and the marketing angle comes from perceived demand for that augmented skill set.
Many developers even seem to predict an increase in demand in the medium to long term as AI written systems increasingly begin to need human attention.
I think the hyper enthusiastic ones are more vocal, but there's a quieter and larger group who are somewhat more measured about it.
Yes, I am addicted to food and shelter and to being able to work in the profession that I specialized in and spent years learning to do well. Without these things I would literally die.
When I was fresh out of college in my 20s I watched two engineers get fired in front of me. They were hired to be DBAs. When I finished rewriting the software into a proper mvc their entire job became redundant since anyone could now make the database edits needed from an admin webpage. And I was better at the DBA part of my full stack role than they were. This is no different.
People need to be independently rich, or be extremelly frugal to not care about the hypothetical obsolescence of their jobs.
What the fuck do you expect? That people just cheer a brave new world of diminishing salaries and disappearing jobs along with some vague promises that every thing will be alright?
Yes, I absolutely do expect computer programmers to cheer as the brave new world they helped create is ushered in.
How many people here got rich by automating away the jobs of others? I mean, what is this? Others are fair game, but programming is sacred? That's quite simply the peak of absurdity.
Every single step in this saga has actually created millions of jobs.
It's pretty funny seeing this play out in the HN comments because I never got any consideration before when I had to learn new frameworks and languages. But suddenly machine learning is the line one should not cross? Playing the world's smallest violin right now.
Computer used to be a job for a cadre of women who did math. That job went away when computers, the physical machine, got enough software that those women were no longer needed.
I've noticed a lot of weaselly statistics associated with AI adoption.
Company replaces phone support with AI chatbot, then says "Call center interactions dropped 80%! People must really love our AI bot," even though they were given no other choice.
AI features are popular in the sense that people are using it. I think the popularity lessens when asked if people want these features.
My point being that people who visit chatgpt.com/claude.com/etc by their own free will are not the same as people who now have to use AI summaries on Google because they are just showing up there and making the ten blue links harder to find.
That’s pretty much it. I have ai in my tracker, ai in my search engine, ai in my team chat, ai in the os (work computer),… I’ve never asked for it, but I bet I’m being counted as one of those users.
I don't think you understand that there's a difference between a user who wants an AI chatbot and a user who wants to perform a web search, and even if they're the exact same user, they expect for a web search to operate like a web search and not like a chatbot.
I don't think anyone who works in product management at any company in 2026 understands this, so you're not alone.
Bluesky is insufferable. Literally any issue in any program is labelled "#vibecoded" by these clearly genius-level engineers who've never shipped a bug.
Remember the scene in "You got mail" where there are bunch of people who protest against big retailer and support the local bookstore but at the end of the day, they have no extra sales.
All these anti-Google, anti-facebook, anti-Instagram, anti-OpenAI, anti-Claude stories are exactly that. Provide copium and feel good for a handful of people for a few days.
Indeed, but justice requires that we recursively continue all the way to the base case, until all 32-bit integers are products of 16-bit integers, all 16-bit integers are products of 8-bit integers, all 8-bit integers are products of 4-bit integers, all 4-bit integers are products of 2-bit integers, and all 2-bit integers are products of 1-bit integers. Only when we have reach all the way down that list to the very, very smallest of the numbers around us and brought justice to them will the future be able to arrive. I literally can not wait for that day.
That's some primo-grade flamebait even by The Register's standards; wrapping a fairly common, perhaps mundane at this point, observation about how what answers the AI favors will affect what answers people tend to be given if AI use increases around the religion flame flash point is definitely going to get some views.
Look for the followup article where they wrap the mundane observation around the politics drumbeat (people who use AIs may be exposed to political beliefs you don't believe in) once they notice the article's view count.
I think it's mostly emergent. I can even point to a plausible mechanism, which is that if you think of an organization as a network of people and how they are connected together, you can think of "responsibility" as something that arcs through an organization like electricity and burns out whatever it courses through, prompting the creation of alternatives to avoid getting zapped the same way in the future.
It isn't completely inevitable, I think it's possible for relatively strong leadership to understand that the processing of responsibility through an organization is a necessary feature and people handling it without external forces conspiring to make it even more like that it will "burn out" a part of the org is a necessity and a good thing. But it's really easy for an organization to just default to burning out the path and evolving ways to avoid it in the future, and it is very motivated to make it happen.
Thinking that teachers (or AIs) are always right is the anti-intellectual way to go through life. Between a teacher or an AI just going off of its training data is a close call; between a teacher and an AI doing a web search it's no contest, it's the AI.
It is arguably a characteristic of any intelligent system, that at least some part of it must be opaque to itself, but the previous sentence is more defensible than a generalized claim.
If you don't understand what that means, tell me from your own metacognitive insight what parts of your brain are being used to read this. Not because of learned knowledge about what parts of the brain do what, through your own insight in your own functioning. You can't, because you don't have any.
This isn't just that human rationalize a lot. This is below that. This is that even if you notice yourself rationalizing, which is something you can train yourself to do, you have no access to the underlying computations/processes of the rationalization itself, or the process of noticing you are rationalizing.
There is arguably still a sense that we experience in which we humans could reasonably say "No, I'm pretty sure I used addition-with-carry to answer you", so that is perhaps not the easiest example to think about the experience of. But there will always be some question of "how did you do that" to which you can give no answer because the answer is in the firing of the neural net itself and you, who is in one way or another the product of that firing, do not have access to that. How did you quickly catch that ball that someone unexpectedly threw at you? You just did, as far your neural net is concerned.
(Also, while I've expressed this in terms of your conscious experience, this doesn't have anything to do with "consciousness". Neural nets in general do not get this feedback and do not and can not have arbitrary metacognition about their own functioning. This is an artifact of my writing text to address conscious beings.)
reply