Even more precisely, human writing contains unpredictability that is either more or less intention (what might be called authors intent), as well as much more subconsciously added (what we might call quirks or imprinted behavior).
The first requires intention, something that as far as we know, LLMs simply cannot truly have or express. The second is something that can be approximated. Perhaps very well, but a mass of people using the same models with the same approximationa still lead to loss of distinction.
Perhaps LLMs that were fully individually trained could sufficiently replicate a person's quirks (I dunno), but that's hardly a scalable process.
Chiplets are different physical chips, potentially from different nodes or processes. They're physically connected prepackaging. It's beneficial compared to sticking them on a board because then you get the shortest interconnects, and you aren't limited by pinout density. This should let you get the best bandwidth to power ratio, and better latency.
Wow, he admits to using two AI tools: He used Claude Code, which failed because the blog was intentionally set up to refuse AI crawlers, so he pasted the page into ChatGPT. Then he blames ChatGPT for paraphrasing the hallucinated quotes.
He makes the claim that he was just using AI to help him put together an outline for his article, when the evidence clearly shows that he used the AI's verbatim output.
There's no federal entitlement to being paid if you're sick, so companies come up with their own policies.
So companies often have a strange concept of "sick days", a specific number of days a year you're allowed to be sick. If you're sick more than that you have to use your vacation days, or unpaid leave when you're sick.
(And of course American companies often have weirdness around vacations too. More so in companies where there is allegedly "unlimited time off". But that's kinda off-topic now.)
During COVID my company had mandatory days off (I think 14) if you reported any COVID symptoms. Those days were unpaid of course. The cherry on top is the people paid the lowest were the ones who couldn't work from home and were most likely to get COVID. This was pretty common at other places too.
Depends entirely on the workplace and the individual. You can tell people not to work when they're sick, but it's not like they're not aware of deadlines for things that, in some cases, only they can reasonably do.
As someone who has deadlines, and is occasionally sick, if I have a high fever I am not working. Nor would my manager thank me for it if I did. If you have a high fever, you’re mentally impaired and shouldn’t be doing anything important if it can possibly be avoided.
I don’t totally agree with this. There’s a gap in his story that most journalists wouldn’t leave out like he did. According to his post, the order of events was:
1.) He tried use Claude to generate a list of citations. Claude refused because the article talked about harassment and this breaks its content policy.
2.) He wanted to understand why so he pasted the text into ChatGPT.
3.) ChatGPT generated quotes; he did not verify they were actual quotes.
I don’t see any sign that he actually read the source article. He had an excellent lead in to that - he had Covid and mentioned a lack of sleep so brain fog would have been a valid excuse. He could have said something as simple as ‘I was sick, extremely tired and the brain fog was so deep that I couldn’t remember what I read or even details of the original author’s voice.’ And that would have been more than enough. But there’s nothing.
That’s an odd thing for a journalist to leave out. They’re skilled at crafting narratives that will both explain and persuade and yet the most important part of this whole thing didn’t even warrant a mention.
As a basic rule, if a journalist is covering something that happened via blog posts, you should be able to expect the journalist to read the posts. I’d like to give this writer the benefit of the doubt but it’s hard.
I think there's still something missing here. This is a strange place for ChatGPT to confabulate quotes: extracting short quotes from a short text blog post is about easy as it gets these days. GPT-5.2 Pro can handle tens of thousands of words for me before I start to notice any small omissions or confabulations, and this was confabulating all that at just 1.5k words?
So since he says he was sick and his recollection cannot be trusted (I don't blame him, the second-to-last time I had COVID-19, I can barely remember anything about the worst day - which was Christmas Day), something seems to be missing. He may not have pasted in the blog post like he remembers. Or perhaps he got routed to a cheap model; it wouldn't surprise me if he was using a free tier, that accounts for a lot of these stories where GPT-5 underperforms and would explain a lot of stupidity by the GPT. Or didn't use GPT at all, who knows.
The DOW is at 50,000! 50,000! If you get this reference (and even if you don’t), there are many alternative actions he could have taken, including not acknowledging this at all.
Yeah, but the problem is that by not making it clear that additional actions may be coming, they're barely restoring credibility at all, because the current course of action (pulling the article and saying sorry) is like the bare minimal required to avoid being outright liars - a far cry from being credible journalists. All they've done is leave piles of readers (including Ars subscribers) going "wtf".
If they felt the need to post something in a hurry on the weekend, then the message should acknowledge that, and acknowledge that "investigation continues" or something like that
You don't announce that you're firing people or putting them on a PIP or something. Not only is it gauche but it makes it seem like you're not taking any accountability and putting it all in the employees involved. I assume their AI policy is fine and that the issue was it wasn't implemented/enforced, and I'm not sure what they can do about that other than discipline the people involved and reiterate the policy to everyone else.
They just needed to expand "At this time, this appears to be an isolated incident." into "We are still investigating, however at this time, this appears to be an isolated incident". No additional details required.
The comments are trending towards being more critical as of my posting. A lot more asking what they're going to do about the authors, and what the hell happened.
> Greatly appreciate this direct statement clarifying your standards, and yet another reason that I hope Ars can remain a strong example of quality journalism in a world where that is becoming hard to find
> Kudos to ARS for catching this and very publicly stating it.
> Thank you for upholding your journalistic standards. And a note to our current administration in DC - this is what transparency looks like.
> Thank you for upholding the standards of journalism we appreciate at ars!
> Thank you for your clarity and integrity on your correction. I am a long time reader and ardent supporter of Ars for exactly these reasons. Trust is so rare but also the bedrock of civilization. Thank you for taking it seriously in the age of mass produced lies.
> I like the decisive editorial action. No BS, just high human standards of integrity. That's another reason to stick with ARS over news feeds.
There is some criticism, but there is also quite a lot of incredible glazing.
Yeah, the initial comments are pretty glazey, but go to the second and third pages of comments (ars default sorts by time). I'll pull some quotes:
> If there is a thread for redundant comments, I think this is the one. I, too, will want to see substantially more followup here, ideally this week. My subscription is at stake.
> I know Aurich said that a statement would be coming next week, due to the weekend and a public holiday, so I appreciate that a first statement came earlier. [...] Personally, I would expect Ars to not work with the authors in the future
> (from Jim Salter, a former writer at Ars) That's good to hear. But frankly, this is still the kind of "isolated incident" that should be considered an immediate firing offense.
> Echoing others that I’m waiting to see if Ars properly and publicly reckons with what happened here before I hit the “cancel subscription” button
No reason to trust that the comment section is any more genuine than the deleted fake article. If an Ars employee used genAI to astroturf these comments, they clearly would not be fired for it or even called out by name.
In 2003, Bush first tried gathering support for invasion in Sept 2002 at the UN. Congress granted authority for use of military force in October, and the troop build up started. Colin Powell tried one more major push for UN support in Jan 2002. The invasion was in March 2003.
That was the public facing attempts to gather support. Internally within the administration, they started working on invasion plans within a few months of 9/11. These plans continued to iterate up to being more or less locked in and approved (by Bush) in Jan 2003.
You can still do this in W11 notepad. Firstly, there's a global setting for having formatting/markdown being enabled at all, and secondly it only does the rendering for .md files. Finally, while formatting is enabled, and editting a markdown file, you have the option to toggle between formatted and "syntax" view (ie raw text).
"Except for vaccination for adults aged ≥65 years, ACIP makes no preferential recommendation for a specific vaccine when more than one licensed and recommended vaccine is available. Among adults aged ≥65 years any one of the following higher dose or adjuvanted influenza vaccines is preferentially recommended: HD-IIV3, RIV3, or aIIV3. If none of these three vaccines is available at an opportunity for vaccine administration, any other available age-appropriate influenza vaccine should be used (4,5)"
Trial designs are rarely perfect, and drug companies will try to stack the deck.
That being said, trial designs aren't made in isolation without consultation. Moderna's press release about this (https://feeds.issuerdirect.com/news-release.html?newsid=7346...) indicates that 2024 CDC said that performing an efficacy trial without using the higher dose would likely cause complications for recommending the mRNA vaccine for the >65 age group, but would not impact overall approval.
Firstly, in math, if you look at how the percentiles break down, it seems clear that while there an overall drop in performance from 2020-2023, it also seems clear to that the top end of doing relatively ok. For example, at the 75th percentile, the we are basically flat from 2008 to 2020 (before dropping in 2023) - 2008: 305, 2012: 309, 2020: 307, 2023: 301 (net -4). This contrasts with the median which went from 2008: 283, 2012: 287, 2020: 282, 2023: 274 (net -9).
This implies to me that whatever flaws are in the overall system (at least pre COVID), the top end was relatively durable.
Secondly, if you go the "student group scores" section and click through all of the different sub-groupings, the only group that looks to have an overall flat score at all is "Private: Catholic".
I think the combination of the upper end being pretty durable, as well as the higher scores in the only "self selecting" category in the dataset may support what a lot of people tend to grumble about - the distribution of domestic situations is not favorable.
The thing that popped out to me playing around with that is that when you look at most of these things from the late '70s through now, many show some modest growth or flatness through the '90s, then grew a little more up through the 2010s or so, and have pulled back a little in the last few years to roughly 2008ish levels. This is for both math and reading and for both 9 year olds and 13 year olds.
Depending on what exactly you are looking at the places that it grew or was flat or declines change, but overall the big picture look gives me more "keep an eye on it" vibes than "we've got a crisis" vibes that some people seem to think are justified.
The first requires intention, something that as far as we know, LLMs simply cannot truly have or express. The second is something that can be approximated. Perhaps very well, but a mass of people using the same models with the same approximationa still lead to loss of distinction.
Perhaps LLMs that were fully individually trained could sufficiently replicate a person's quirks (I dunno), but that's hardly a scalable process.
reply