Editor’s Note: Retraction of article containing fabricated quotations

We are reinforcing our editorial standards following this incident.

See full article...

 

[Marcus Andreus]

i rewatched The Man From Earth last night. i still think it's a great film, even if it does assign a little too much importance to the religious aspect. but what it really made me wonder is: do people really still watch films on DVD? i only have a DVD-quality copy of it because i can't find a BD, and even for a film that relies so little on visual fidelity, it was really very distracting how low quality everything was.

the only reason i ask this is because i've seen a couple of posts recently where someone said "i don't have the DVD, so i watched it on streaming". is "DVD" just a synonym for any physical media now? or are people really still watching films on DVD on their 100" TVs?

I have a 75" 4K TV and SD doesn't really bother me a whole lot.

Most things I watch are HD or better, disc or streamed, but I own a few things on DVD. The whole series of both M*A*S*H and China Beach. A few random movies. Notably I just 'bought' Blade Runner from Apple TV (I have the Blu-Ray... and the DVD, but was interested in the HDR presentation) and Dangerous Days, the "making of" documentary for it is still only in SD.

Encoding quality is probably nearly as relevant to me as actual resolution. My personally ripped and encoded version of Dangerous Days looks better than the Apple version, for example. On the other hand, I was re-watching some of season 1 of Altered Carbon recently and I kept getting distracted by massive banding and macro-blocking in the backgrounds.

But yes, some people also just use "DVD" as shorthand for "a physical disc."

 

[AI_Skeptic]

I'm just curious, what type of AI related articles would Ars members want to read?

 

[vonduck]

so... did anyone read about that police chief who told uk parliament no ai was used in something and that someone googled the info and it turns out it's all ai slop (or something like that)?... which basically sounds like someone googled some shit and just took the ai slop from the top.

not only can you not actively use ai when writing articles, you can't even look at the ai slop at the top of search engines.

oh and reading both the retraction article and the retracted article tells me absolutely nothing.

 

[mygeek911]

I'm just curious, what type of AI related articles would Ars members want to read?

The bubble popping. AI bros taken to account for real harms they’ve caused. AI removed from making medical, legal, and governmental decisions.

 

[SomewhereAroundBarstow]

I'm just curious, what type of AI related articles would Ars members want to read?

The snarky reply that popped into my head was "Ones written by Molly White," but honestly that's not far from the truth. I value distanced skepticism over cozy access, and I am far more interested in the broader societal effects of any technology than the touted benefits of it in specific uses.

 

[adespoton]

I thought there already was an attempt at a second Pony Thread, back in the end days of the old forum when the number of pages of replies was causing Moonshark errors.

So this would be Pony Thread III: Hay Hard With a Vengeance.

Well, that and, Ponies I is still going strong since the architecture fix. Maybe this thread should be reserved for Don Knotts?

 

[Aurich]

I took Lee's advice:

Enough is enough—I dumped Google’s worsening search for Kagi

 

[anguisette]

I'm just curious, what type of AI related articles would Ars members want to read?

i would like to read stories about "i tested AI for this real task i do frequently, and here's how it went".

notably, "real task i do frequently" does not include "make a shitpost on Memory Alpha about Janeway murdering Tuvix". yes, we're all sad she wasn't prosecuted for war crimes, but that is not a test of AI, it's a meme shitpost by itself. it doesn't test AI in actually useful tasks. also, if you give the AI 7/10 when it couldn't even do what you asked it to do, that's not a serious test, it's just boosting.

what i'd like to read is "i am a tech journalist and i spend all day in the copy mines writing prose for a well-known online journal whose name rhymes with 'Arse'. i tried replacing myself with an LLM for a week. here's how it went". let's see the prose you wrote vs the prose the LLM wrote. let's see what you explained that the LLM didn't understand needed to be explained.

i'm not even prejudiced about the results. if someone did that and found the LLM was just as good as them... well, that's pretty interesting by itself. but the point is the reporting should be in-depth, detailed, and relevant to the real world. that's what i want from Ars.

i'd also like to see some reporting about how the #1 Best Mars Moon Colonisation Programme is run by a company whose AI is mostly notable for generating child porn, but that's probably more than we can hope for at this point.

more generally, i think what i want is to know how LLMs will affect me, personally, and people like me. i don't get this from Ars. like, i enjoyed the story about OpenAI's huge wafer of confabulation, even if the comments provided more details than the story, and i don't want Ars to stop covering things like that. but it's really hard to find any non-credulous reporting about AI these days, and that's a niche Ars could easily fill, if they wanted to.

you know how Techdirt is really great at covering government policy bullshit and tells it straight? i want Ars to be the Techdirt of AI.

 

[graylshaped]

It cannot “help track down quotes” without creating content. Selecting the quotes you want to use IS AUTHORSHIP. If I cite a case and want a quote from it, selecting which quotation best benefits my argumen is part of my specialized skillset. That quote is part of the brief I turn in or what I say at oral argument. It is, while a quote, also a fundamental piece of MY content. I stake my credibility to what I turn in, including the portions I choose to quote.


I don’t need to pretend that marketing BS makes what some dorkass middle manager thinks of as AI comparable to what some other person thinks of as AI. I don’t need a poll. I don’t need a study. None of that matters to what I said. Common sense (and common law) matters.

If you are writing something that contains quotes, then the quotes you select ARE PART OF WHAT YOU WROTE. If you allow a computer to select those quotes, you have just used generative slop machines to dictate your content. The machine wrote part of what you are putting your name on. If there’s a policy against doing that, you’ve violated the rule.

The reason you can put together a book of recipes and have it be copyright protected even though a single recipe cannot be is because selection and arrangement are, in fact, a form of authorship.

It’s not wholesale letting the slop machine write the article, no. It’s the equivalent of, say, using slop machines to generate the opening lines, or the title and blurb. But Ars policy would require disclosing any of that because it’s actually part of the content you turn in to them and hold out as something you alone authored.

We're putting it differently, but there isn't a gap between us, other than you seeming to assume when I say "track down a quote" that means "and stick in the final draft."

Crafting policy is hard. Have you ever done that? Written policy for a company and worked through drafts, word by word, with HR reps and union bosses and lawyers and senior leadership, including business leaders in other parts of the company who may or may not be affected by what your part of it is doing? That's a partial list, by the way.

 

[RetroTrailRunner]

Shambaugh has posted another blog and he mentions wanting the original article and comment sections restored.

https://theshamblog.com/an-ai-agent-published-a-hit-piece-on-me-part-3/

 

[Zionyx]

I just got a report from this thread complaining that a user was "posting too much", and man, if that was a bannable offense a lot of you would have to find something better to do with your day.

Carry on!

 

[LeeF]

I'm just curious, what type of AI related articles would Ars members want to read?

I can't speak for Ars members in general - judging by the comments in any article that touches on AI, there's a large percentage who are vehemently against it. Speaking for myself though, I'd be interested in a mix something like this:
-What's this Agentic stuff all about that everyone is banging on about? Is any of it useful?
-AI ethics
-Introduction to locally-hosted LLMs, and examples of useful applications for them
-Technical deep dives in the vein of Hannibal's articles about the Pentium microprocessor architecture

[Edits: typo and author misattribution, mistake my own fault]

 

[Alhireth-Hotep]

Context matters. According to Benj Edwards website, he is the Senior AI reporter. What does this mean?

1. Out of all the journalists on Ars' staff, he is the expert on "AI". He should know that LLMs, like Claude and ChatGPT, are probabilistic word generators . In fact, according to his website, he's the one who coined the term "confabulation".

"Confabulation" comes from psychology, and it was the word I was immediately using in 2021 because it's the obvious analogy. He did not "coin the term". That is a lie.

 

[graylshaped]

I'm just curious, what type of AI related articles would Ars members want to read?

Putting this behind a spoiler:

Spoiler: Too Soon?

More “stupid shit people do because they think AI is competent” stories we can share with people who have not yet been bitten by it are always timely.

 

[Bad Monkey!]

I'm just curious, what type of AI related articles would Ars members want to read?

Sam Altman interviewed while he contemplates on the Tree of Woe.

 

[PsychoArs]

I'm just curious, what type of AI related articles would Ars members want to read?

I work at an MSP.
What are the risks to my clients?
How to talk to my clients about those risks.
How to detect those risks.
How to mitigate those risks.

I have family.
What are the risks to my relatives?
How to talk to my relatives about those risks.
How to detect those risks.
How to mitigate those risks.

I operate a computer.
What are the risks to me.
How to talk to myself about those risks.
How to detect those risks.
How to mitigate those risks.

I don't need articles about what AI "can" do. That's trivial to come by. There are adult cinemas with cleaner floors than the regular press when it comes to AI.

What I need is a dispassionate, accurate, detailed examination of the constantly-evolving technology, to keep me abreast of the pitfalls. I need a skeptical and informed eye telling me what the cult members don't want to talk about. The failure modes. The mistakes. The flaws.

The things swept under the rug.

And you know... if a product/service is released and it's legitimately fit-for-purpose, it's okay to come to that conclusion. But I'd rather read "we tried to find flaw with X and couldn't" than "brand new X is awesome!"

Yes. I'm asking for things that reinforce my confirmation-bias. But I promise I'm an adult and will pay attention where flaws aren't found. It's just that the consequence of issues in this sector so dramatically outweigh the benefits that we need to start there.

In the infancy of online banking, we needed articles like: "don't do it unless your bank uses HTTPS" and "don't do it unless your bank supports MFA" and "don't do it on someone else's computer". We didn't need articles like "online banking lets you pay bills" or "now you can transfer funds anywhere with two clicks" or "take out loans, fast!" Because what you could do was super, super secondary to what you shouldn't do.

That's my $0.02, but I'm Canadian, so that's like $0.15 USD.

 

[jnv11]

I'm just curious, what type of AI related articles would Ars members want to read?

Real life scientific advancement like DeepMind’s AlphaFold AI algorithm that solves medical and biological problems that were previously impossible to solve without AI would be a good candidate.

 

[AdrianS]

There it is folks: tacit permission to make this Pony Thread 2: Derpy Boogaloo.

Ponies vs Battleships - the conflict begins.

 

[AI_Skeptic]

Not sure what else he could have done? You can say "he shouldn't have used AI tools in the first place" but he was feverish. It's very much like being drunk. Your ability to make sensible decisions is...not grea

When a person is sick with a fever, the last thing they will do is to deploy a new tool in their workflow, and when the new tool doesn't work, attempt to troubleshoot that tool. Instead, they would say "forget this, I'm too tired to learn, let me use the way I am familiar with".

Plus, the name of the tool was never given :O

 

[Dark Jaguar]

this is a real problem, and not just with generative AI. it's like, you can build a tactical nuclear weapon to take out the skunk living in your back yard, but once you do that, aren't you going to be tempted to deploy the next one on your neighbour's yappy dog? and there aren't even any consequences to doing that. some weirdos are like "i don't think everyone should own tactical nuclear weapons", but let's be real, the market has spoken, and nukes are the next serious investment opportunity.

at this point, it's simply unrealistic to oppose private ownership of tactical nuclear weapons. the conversation has moved on, and there's no point re-litigating ancient history. we might have to endure a short period of mass destruction and people dying of horrifyingly painful radiation poisoning, but once we get past that initial hurdle, you'll see that private ownership of tactical nuclear weapons just makes everyone so much nicer, and more polite. who's going to wolf-whistle you on the street when you might have a nuke in your pocket?

anyway, as a practicing journalist, i sometimes use tactical nuclear weapons to prepare stories (i NEVER use them to generate prose) and no one has ever complained that i'm doing a bad job. well, except my old boss. but i have a new boss now, and we get on great.

Davy Crocket approves this message. So does Senator Armstrong. This brilliant prose was created with nothing more than a find/replace function.

As far as the point, I've said it before... the pithy words "you can't put the genie back in the bottle" aren't strictly true, and it's worth the effort to try considering the sheer damage that's been done. Also... and this is just a personal gripe... genies traditionally ALWAYS go back in their bottles, eventually. (And, toothpaste can go back in the tube, I mean a machine put it in there in the first place.)

 

[Dark Jaguar]

I'm just curious, what type of AI related articles would Ars members want to read?

Articles on the science behind it, on developments, on analysis of it's accuracy and what effects it has on people, oh and of course, stories of when it goes wrong. I just... would make sure the articles on the tech itself don't read like an advertisement and include criticism including notes on what claims they've made have actually been demonstrated.

 

[AI_Skeptic]

Let's forgive, learn, and move on.

Learn what exactly? Benj is a reporter with over 20 years of experience. In Journalism 101, on the first day, they teach students to always triple check quoted sources. The source Benj failed to quote accurately is from a website, which could have been verified using a copy-find-paste that would take 30 seconds. This is first day of journalism school lesson.

What possible things could Benj, a reporter of over 20 years, learn from this event?

 

[marsilies]

the only reason i ask this is because i've seen a couple of posts recently where someone said "i don't have the DVD, so i watched it on streaming". is "DVD" just a synonym for any physical media now? or are people really still watching films on DVD on their 100" TVs?

"DVD" as a term for any type of video disc, including Blu-rays and UHD BDs, has been a thing for over a decade now, coming up on two. A lot of people don't make a distinction between formats. It's like calling every video game console "A Nintendo."

That said, DVDs themselves are still very popular. Many people never bothered upgrading past their last DVD player, and DVD playback is more ubiquitous. Old laptops with DVD drives, older car entertainment systems, etc. There's a reason Blu-rays often packed in a DVD copy as well.

For the "RiffTrax Makes MST3K!" Kickstarter, the just sent out poll to see how many of the physical media reward tier want a DVD v Blu-ray, since they launched with only Blu-ray option, but received feedback from some backers that they'd like a DVD copy instead.

View: https://www.kickstarter.com/projects/rifftrax/rifftrax-makes-mst3k/posts/4603105

 

[Zukunftsweber]

Yeah, but do you get into that level of detail with customers?

Yes. Different field, but I've done presentations going into this level of detail explaining our release process. I've even put versions of it in our customer training.

 

[runswithjedi]

i would like to read stories about "i tested AI for this real task i do frequently, and here's how it went".

notably, "real task i do frequently" does not include "make a shitpost on Memory Alpha about Janeway murdering Tuvix". yes, we're all sad she wasn't prosecuted for war crimes, but that is not a test of AI, it's a meme shitpost by itself. it doesn't test AI in actually useful tasks. also, if you give the AI 7/10 when it couldn't even do what you asked it to do, that's not a serious test, it's just boosting.

what i'd like to read is "i am a tech journalist and i spend all day in the copy mines writing prose for a well-known online journal whose name rhymes with 'Arse'. i tried replacing myself with an LLM for a week. here's how it went". let's see the prose you wrote vs the prose the LLM wrote. let's see what you explained that the LLM didn't understand needed to be explained.

i'm not even prejudiced about the results. if someone did that and found the LLM was just as good as them... well, that's pretty interesting by itself. but the point is the reporting should be in-depth, detailed, and relevant to the real world. that's what i want from Ars.

i'd also like to see some reporting about how the #1 Best Mars Moon Colonisation Programme is run by a company whose AI is mostly notable for generating child porn, but that's probably more than we can hope for at this point.

more generally, i think what i want is to know how LLMs will affect me, personally, and people like me. i don't get this from Ars. like, i enjoyed the story about OpenAI's huge wafer of confabulation, even if the comments provided more details than the story, and i don't want Ars to stop covering things like that. but it's really hard to find any non-credulous reporting about AI these days, and that's a niche Ars could easily fill, if they wanted to.

you know how Techdirt is really great at covering government policy bullshit and tells it straight? i want Ars to be the Techdirt of AI.

404 Media is doing some amazing work, especially on the topic of NCII abuse. Sam Cole really digs into the topics and has thorough reporting.

Honestly, 404 Media is quickly becoming my favorite publication. It's incredible what owning their own news organization has allowed them to cover and, most importantly, how they cover it.

 

[Distraction]

There's nothing wrong about a journalist using a tool to extract quotes, even if that tool uses LLM technologies. However, said tool needs to have deterministic code that always confirms the quote is accurate and exists.

Even when quotes are accurate, a journalist should NOT ask an LLM-based tool directionless questions like "Provide the 5 most interesting quotes from this stack of articles." I don't need to know what Claude "thinks" of an article. I do not want every email, every piece of content I read to be curated by the same big three models.

But I have no problem with a journalist taking a stack of 100 articles and asking an LLM, "provide me with every quote that mentions something about topic X." You are guiding and directing it to perform a specific task. Yes, there is some "judgment" involved, but with proper articulation I think the risk of steering is outweighed by the fact you can find more relevant content more quickly.

Then the actual reporting needs to be done after reading the source material.

In terms of Mr. Edwards putting Ars Technica at risk of libel suits (or whatever I am reading in every third post) ... that is ridiculous at least in terms of USA laws. The bigger risk of a lawsuit involves punishing an employee (W2 or 1099) prematurely due to the pitchfork wielding mobs demanding justice.

If y'all want to cancel subscriptions to make a point - excellent. That's how the system works. Personally I believe in grace & redemption, and while trust needs to be earned again, the idea that a single known lapse in judgment (that in of itself is largely inconsequential) should doom the rest of your career just to prove a point that LLMs are prone to abuse--well, that's pretty rough, and I hope nobody ever judges you that way.

Putting aside the fact that any results you get from asking an LLM to provide 'every quote from 100 articles that mentions something about topic X' would absolutely contain a lot of hallucinated bullshit, it is still selecting your arguments for you, whether you instructed it to or not.

You would also be left with a bunch of sentences stripped of any context and any conclusions you drew from them might very well be the opposite of what the authors actually wrote. Frankly, if you can't even be bothered to read the sources you're using you bolster your own argument, you shouldn't be using them at all.

In this case, the author was summarizing one short blog article. It would have been faster to do it the ethical way.

 

[AI_Skeptic]

Nobody's perfect and nobody's immune to illness so if it wasn't done with malice (and it didn't seem to be), why immediately call for the person to be fired?

Because Benj's excuse does not make sense.
1. He suffered from both COVID-19 brain fog (per his interview with Ed) and a bout of COVID-19 (per his BlueSky post). Therefore...
2. Why didn't he let his co-author, Kyle, know he's not feeling well and to review his work? I'm positive Kyle would have reviewed the quotes from the website if asked (even if it was "under the table, I shouldn't really do this..." type of thing)?

Or...
1. He used a new Claude based tool to extract the quotes, which didn't work and required time troubleshooting. However...
2. People with a high fever automatically revert to processes they are most familiar with, because they don't have the energy to learn and troubleshoot tools.

Or...
1.He used a new Claude based tool to extract the quotes, which didn't work and required time troubleshooting. However...
2. As the senior AI reporter, he should know that LLMs generate characters that are most likely to appear next, and that an AI tool that's able to extract text from articles without error deserves a article itself, explaining how the tool works while maintaining the AI company's stance of "we don't recite information, it's learned"?

Or...
1.He used a new Claude based tool to extract the quotes, which didn't work and required time troubleshooting. However...
2. When he asked ChatGPT to troubleshoot the tool, [insert magic here] and ChatGPT gave the quote from the website. He did not share the magic that happened that allowed ChatGPT to give the quote, because ChatGPT does not work like that.

I think Kyle's response of "no comment" is much safer.

 

[Jim Salter]

i actually have a question about this. earlier someone mentioned not wanting to use "PTO" to cover sick days. i thought this was just a terminology difference, but is it actually the case that in the US, you have to use vacation days to get paid sick leave? or is this something that varies by state? (in which case how does it work in NY, where i think Ars is based?)

PTO is separate from vacation, and covers everything from sick days to needing to go to a doctor's appointment or any errand you might need to run during working hours. It's typically expended in blocks of 0.5 days at a time, and you don't get a whole lot of it.

From what I understand of the work/life balance in most EU countries, you folks get considerably more vacation than our vacation + PTO put together.

 

[Jim Salter]

My experience of women going into men's rooms is that they couldn't wait for the line for the ladies' room and they're either using the urinal right next to the men or peeing in the sink, propriety be damned.

I've been into quite a few single-seat ladies' rooms on the same premise. There's only so long I'll wait for the dude taking a 40 minute steamer at the gas station to finish before I get aggravated and just go sit'n'pee in the ladies'.

 

[AI_Skeptic]

There's nothing wrong about a journalist using a tool to extract quotes, even if that tool uses LLM technologies. However, said tool needs to have deterministic code that always confirms the quote is accurate and exists.

Your first and second sentence contradicts each other. An LLM, by it's nature, can not confirm the quote is accurate and exist - and there's much more efficient ways to review quotes from websites as well (i.e. Python with Regex, for example). Can you walk me through how LLM can have deterministic code?

 

[Jim Salter]

I just got a report from this thread complaining that a user was "posting too much", and man, if that was a bannable offense a lot of you would have to find something better to do with your day.

Carry on!

Don't tempt me to bring the Tarrbot back just to piss whoever that is off.

 

[Jim Salter]

Your first and second sentence contradicts each other. An LLM, by it's nature, can not confirm the quote is accurate and exist - and there's much more efficient ways to review quotes from websites as well (i.e. Python with Regex, for example). Can you walk me through how LLM can have deterministic code?

The LLM itself doesn't. But you can wrap all the deterministic tasks you'd like into an agent that can also invoke an LLM.

The workflow would look like this:

• agent triggers LLM to scan an article looking for interesting pull quotes
• LLM hands a selection of pull quotes to the agent
• agent triggers a grep on the source material for the exact text of each of the pull quotes the LLM offered
• any pull quote not found in its exact original form gets rejected
• LLM is invoked again, told it fucked up, asked for pull quotes that actually exist this time
• agent triggers a grep...

Lather, rinse, repeat until you've got the number of "interesting" quotes you asked for.

Now, keep in mind, in this context even if that tool works 100% correctly, even if every quote is 100% perfect and legitimate, it's still not okay. Selecting pull quotes is a very important part of writing the article. If you let the LLM select pull quotes for you, it's frankly just as bad as letting the LLM write the rest of the piece for you while you get the pull quotes. Either way, you're delegating not merely style quibbles but concrete, foundational tasks that determine the quality of the entire piece to the LLM.

 

[Snarky Robot]

In this case, the author was summarizing one short blog article. It would have been faster to do it the ethical way.

Yeah, that’s one of the many things that makes me wary of AI stuff. It’s scary how quickly people can replace a portion of their critical thinking with rote usage of it, even when it’s readily apparent the computer is doing a worse job or is taking longer. It’s like their brain atrophied from disuse.

Obviously, Benj would admit that using the AI tool made it worse given the outcome. But it would be fascinating to go back and watch him go through that process and see how he’d react to someone saying “that article will take you 5 minutes to read, max, why are you using AI?” because I‘ve seen the blank stares of people asked that question over much less vital things, and I can’t help but wonder if it went down like that. Where turning to AI is just so normalized that the alternative didn’t occur.

 

[Jim Salter]

Now that the conversation has shifted to who's using which bathrooms and who's posting too much, it's probably time to close this thread.

Have you considered simply not clicking into threads you no longer wish to participate in? Seems like an easier and more fair solution to your "problem."

 

[AI_Skeptic]

• agent triggers LLM to scan an article looking for interesting pull quotes
• LLM hands a selection of pull quotes to the agent
• agent triggers a grep on the source material for the exact text of each of the pull quotes the LLM offered
• any pull quote not found in its exact original form gets rejected
• LLM is invoked again, told it fucked up, asked for pull quotes that actually exist this time
• agent triggers a grep...

Why should we be using "agents" to call LLMs to call grep when Python and Regex can do the same thing - locally? Or use grep locally (I know it's built into MacOS, is it built into Windows)? The only advantage is that agents usually require Claude or OpenAI on the backend, allowing those companies to (try to) make a profit.

The way you describe it sounds like a Rube Goldberg device that's extremely inefficient.

Ninja Edit: That's on top of the whole "is it even ethical for a journalist to use a tool like that?" question. If I was a journalist, I would use a paper notepad, a pencil, and write down interesting quotes I see from a website or other sources. I'm a big fan of paper and pen/pencil myself.

 

[Aurich]

In the forum interface there's even a handy Ignore button to make a discussion vanish if seeing it in your threads irritates you.

 

[Jim Salter]

I always figured this was part of what's behind Ars' comment quoting policy (and yeah, I bent it a bit right here, by not including the entire quote myself).

No, you didn't. The rule is no modification of the quote you select, not that you must quote the entirety of the post.

That rule IS pretty draconian--it's been established that so much as bolding a word in the quote and adding (emphasis mine) just beneath the quote is a rule violation, as is [paraphrasing in square brackets]. But it does not prevent you from only quoting the bit you mean to reply to.

 

[Jim Salter]

Why should we be using "agents" to call LLMs to call grep when Python and Regex can do the same thing - locally?

I'm not entirely sure what you think an agent is. It's a framework of real code wrapped around the ability to call an LLM. Most commonly, the agent runs on local hardware and the LLM is invoked from the cloud via a personal API key for that user, but in some cases, the model and the agent both run locally.