Why OpenAI Has Literally Been Fighting Goblins - DTNS 5259

0:00This episode is brought to you by Capital One. Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called Chat Concierge and it's simplifying car shopping. Using self-reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trade in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One. Mother's Day is coming and Benihana just made it easier to celebrate in style. Buy a$50 gift card and get$10 back.

0:37Buy$200 and get$60 back. This deal won't last long, so don't sleep on it. Treat mom to hibachi, sushi, handcrafted cocktails, and a dinner table show she'll be talking about all year. Make the night unforgettable. Gift cards are available now at Benihana.com. That's Benihana.com.

1:02Huyen Tue Dao: Of course, I say that right after you've already pressed it. Give me one second. I didn't put this in my prompter.

1:18Huyen Tue Dao: This is the Daily Tech News for Thursday, April 30th. We're all out of April 2026. We tell you what you need to know, give you the important context, and help each other understand. Today, why goblins infested OpenAI's codecs, how it got rid of them, and what it means for understanding model training. I'm Tom Merritt. I'm Wintwet now. Let's start with goblins. So last week, OpenAI released source code on GitHub for the Codex Command Line Interface, or CLI. This revealed that the system prompt for the CLI, system prompt is something that reigns in its behaviors no matter what you type in. So it included instructions, and I quote, never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query.

2:20Huyen Tue Dao: That prohibition is repeated twice in the 5 ,000 words of the system prompt. Other elements of the system prompt are probably more understandable, like never use destructive commands like git reset hard or git checkout unless the user is clearly asked for that operation. That's a good one to have in there. Also, I think a lot of people will like that there is a system prompt that says not to use emojis or em dashes unless explicitly instructed. I think we all like that one too. But up until now, I don't think there's been a whole lot of discussion. There has been some, but not mainstream discussion of chatbot goblin talk or raccoons.

3:01Huyen Tue Dao: Certainly not as much as there has been about em dashes. And the instruction was not present in earlier models. Well, OpenAI, in a fit of transparency, posted on its blog that the penchant for mentioning goblins and gremlins showed up starting in GPT 5.1. And the company did a lot to kind of track down why this was happening. It believes the behavior was a side effect of developing the nerdy personality. For a while, you could choose to let ChatGPT respond to you with a nerdy personality. The reward system was scoring outputs with goblins and gremlins much higher. They were like, oh, yeah, that's very nerdy.

3:40Huyen Tue Dao: And so that encouraged the chatbot to use those more often. And interestingly, the way the reward system and the reinforcement learning system works, it then spread outside of the nerdy personality. OpenAI explained the feedback loop this way. So that playful style is rewarded. Some rewarded examples contain a distinctive lexical tick like goblins. That tick appears more often in rollouts. The model-generated rollouts are used for supervised fine-tuning. That's how it leaks from the nerdy personality because it shows up in stuff that is then used for fine-tuning training. And the model gets even more comfortable producing the tick.

4:24Huyen Tue Dao: Now, that nerdy personality was retired in March along with the introduction of GPT 5.4. And at that point, goblin reward signals were filtered out of the training data. They're like, we don't need to be rewarding the use of the word goblin or gremlin or any of that stuff. However, GPT 5.5 had already started training before that. So some of the goblin behavior creeped in to 5.5 and they decided they needed to filter out. Now, once this was discovered and went a little mainstream and people started talking about it, humans being humans have now created plugins, forks and skills that specifically remove or override the prohibition against goblin talk.

5:07Huyen Tue Dao: There's even some that encourage it. And OpenAI has embraced that as well. Nick Pash said they might add a goblin mode toggle to codex if enough people want it. And they did, in their post explaining how this all happened, give you some cut and paste CLI text if you want to remove the instructions prohibiting the Goblin Talk. Goblin Talk. Welcome to Goblin Talk. There's a lot of stuff here. So, number one, of course, nerdy style was rewarded previously as well as playful style. I do think this is really interesting. I kind of I've been thinking about this stuff a lot and it's it's so interesting because of the way that you know models work and with it was training and things like that that a lot of these fixes tend to feel very reactive right.

5:59So the M dash thing I think is really fascinating because I was thinking that because usually these days people talk about well you can really tell when AI has written something because it has an M dash in it. normal people don't i mean i i don't know uh tom i i suppose you're someone who might have used an

6:13Huyen Tue Dao: mdash in your life caught myself removing m dashes from my writing today okay today okay wait do you so are you like chat gpt or or other bots do you like m dashes and and and overuse them or you just no no i don't overuse them but there were definitely like in this discussion that we just had when I was explaining what a system prompt was. I had originally put system prompt for the CLI, m -, the part that reigns in its behaviors, no matter what you type in, m -, and I actually changed it to commas. Oh my goodness. Okay. Well, fine. It's just so interesting because after hearing that, I was like, well, at some point they're just going to put in some kind of command to not do that.

6:56It's so interesting because again, I am not an AI person and I'm not especially not lms and i don't really understand very well how training works uh so mine might be being unfair but typically when it comes to software engineering doing things in this way where you kind of hard code slash react to things very specifically is i would like to say broadly is generally not not the best most stable most can most um kind of band-aid yeah yeah it's a band-aid um and so maybe that's unfair of me to say with a training model but traditionally i think it's software engineering, band-aids are not great.

7:29You can do band-aids if needed, but generally the band-aids need to kind of get replaced with stitches or, you know, a skin graft or something at some point. So it's so fascinating to me that it seems that with the way that LLMs work, maybe naturally all you can really do is band-aid or over time retrain them. Maybe that's where that comes in. And so it just makes me a little nervous. And it does make these kinds of things more reactive just like maybe less fun stories about you know less fun things that maybe grok has done and things like that they have to tell you exactly grok's uh grok's controversy was caused

8:06Huyen Tue Dao: they said by somebody uh putting something in the system prompt that they weren't supposed to yeah but as a as a prank as a prank yeah yeah and so you can you can make it not do stuff or you can make it do stuff you can you can make it come up with stuff that system prompt is basically just a prompt, right? It's just a prompt you don't see because it's setting the boundaries for it. If it makes you feel a little better, I'm not an expert on this either, but from what I do understand, what they do with system prompts is sort of a temporary fix. It is a band-aid. And you saw that in the way that they described this, which was 5.4, or I'm sorry, 5.4, yeah, 5.4 no longer had this problem.

8:50Huyen Tue Dao: We've been able to filter it out with the training. But 5.5 had been trained previously to this. So we had to put the Band-Aid on 5.5 until we could get back to having it have the same natural defense against it that 5.4 had. So it really is a Band-Aid in the truest sense of like, yeah, we put that on here, there until we can do proper treatment. Right. It gets you to that point. And the other thing to remember is these are just predictive models of what character should come next. So what you're training is weights. What is the probability that this should come next versus that? And so these filters are like, oh, you can't program.

9:34Huyen Tue Dao: I think that's where you and I have the same reaction of like, that's not a good solution. It's like you can't program it to do something. This isn't that solid. You have to kind of throw things at the weights and then see how it reacts. Sometimes most of the time it reacts the way you expect, but not always. Every once in a while just starts talking about goblins and gremlins and raccoons, in which case you have to go figure out, okay, how do we train that out of it? In the meantime, let's put a system prompt in to stop it from doing that behavior that we don't want it to do anymore. And I think that's so interesting.

10:06You kind of touched on something else. I was thinking talking about because these are LLMs are basically pattern matchers more or less that what is interesting and what someone told me a couple weeks ago is that actually it has been known that LLMs often are not good at understanding negation like literally putting no or not in the sentence um and there's a different reasons for that I literally just like you know kind of googled a few things but it as I understand the literature is out there that it is a known problem um among kind of like machine learning and like AIs that there are just certain like yes that's why sometimes I think I've definitely written prompts where all of a sudden it feels like it thought I said yes when I when I actually said no um and I think it has to do with the pattern matching and probably probabilistic you know um nature of LLMs and just like to reinforce it doesn't again it doesn't necessarily understand what's happening in a sense because it is a pattern matcher um it's just very very good at what it's doing and it's very good at sounding like it does I still treat it I still answer more more more I still anthropomorphize it on a daily basis even though i i know it really shouldn't so i guess to say that point also is like the best thing that we can do is tell it to not do things it's not really great at understanding negation so i guess to your point as what you said yeah the fact that you know llms are non-deterministic and that like given an input you cannot i guess my bad explanation of that um is that given an input you could have a different output depending on yeah nothing it's not deterministic like given the exact same conditions, exact same inputs, you could get a totally different response.

11:36And that's why, yeah.

11:37Huyen Tue Dao: To a coder is like, ah. Oh my God, guys. I can't, when I first heard that, like LLMs are non-deterministic, I'm like, excuse me? Yeah, you have to be careful with a lot of, you have to be comfortable with a lot of uncertainty. Yeah, and not to say that, I mean, not to say that software engineers are perfect at all, nor, neither are we deterministic beings, really. But that, I will say like, yeah, I understand that this is how it works. I think that there's a lot of benefits. And I mean, I've been studying AI actually since college, different flavors of it. And so I understand that there's a form of impreciseness or like there's a lot of like extra stuff going on a lot of times when you have these systems and AI and the way that machine learning works.

12:22And I'm probably not, it is not distinct totally from the way our brains work, but that as someone who writes software, it makes me nervous. Because, yeah.

12:31Huyen Tue Dao: One of my metaphors for it is riding a bike. When you first get on to ride a bike, you want to just sit on the bike and direct it, right? And you can't do that. You have to shift your weight. You have to balance. And that is a weird thing to learn because you're guiding the bike, not in the straight ahead logical way, right? It's more of an intuitive thing. And I feel like LLMs in a way are kind of like that. You have to shift your weight, right? It's not a thing where you can just put in a command and it will always act the same way to that command. I actually studied more evolutionary algorithms in school, and that's very similar.

13:09Even less interesting than that because you give it basically some kind of scoring metric. And all of these little programs basically exchange code randomly or not so randomly. They create baby programs from splices of theirs, and they just rate themselves against this metric. and even by that system yeah like that would be like riding a bike by trying to use your hands it's on the pedal you know like me it's even less interest it's even more crazy than your example sure at the same time though i think what's interesting yeah i think at the end what's interesting is metrics right and i think that's one thing that i think we're struggling with with llms and of course like which is i think is interesting it has an interesting appearance in this article where okay they had a certain thing they want to achieve right i totally get this they wanted open ai they want chat to be to be playful they want to have fun with it yeah they wanted it to seem like something that you would trust to code for example maybe making it nerdy or just they want to make it appealing uh probably for you know ux reasons but also money reasons because they want people to spend more time with them that's totally fair um but at the same time how do we measure this is always like the biggest the hardest thing even in just regular software engineering how do we how do we measure that people like our thing how do we measure that nerdiness of our models.

14:26And right now in my industry, we're like, how do we measure productivity from AI? And so I just think it's so interesting. There's all these different things in the story. I actually, I'm very charmed by this story, to be honest. I don't find it that alarming, but it is like a very friendly way to digest a lot of different interesting things that are going on with AI.

14:43Huyen Tue Dao: Yeah, I agree. I think this is really a good instructive story if you want to understand how these things work to kind of dig in and say, oh, okay, I didn't realize that was the mechanics underneath it. And one of the reasons they're trying to give it personality and they put system prompts in there that say you need to be warm is if you remember a few years ago, these bots could get insulting and angry and hostile. And so I think they're trying to overcorrect to keep that from ever being the case. And sometimes it makes it sycophantic and sometimes it ends up with goblins and gremlins and records.

15:17Goblins and gremlins and ogres and trolls. Oh my. I'm very grateful for this story. And on that note, I am also very grateful for you guys, the listeners. DTNest is made possible by you. So let's take time to thank Jeffrey Zilks, Ailo Adam L., Philip Les, and Sun Bun.

15:34Huyen Tue Dao: Yay! Thank you, every single one of you. I'll be honest. Nothing tests my patience quite like sitting on hold, listening to the same four bars of music on repeat. Turns out, all that time adds up. The average person spends over 40 days listening to that dreaded hold music. 40 days! This episode is sponsored by Parloa, the AI platform built to make that a thing of the past. Parloa's AI agents are available 24 hours a day, 7 days a week, fluent in any language, and smart enough to remember every customer across every interaction, because no one should have to explain themselves twice. From payment processing, to roadside assistance, flight rebookings, to appointment scheduling, Parloa delivers the kind of customer experience that turns a one-time buyer into a customer for life.

16:22Huyen Tue Dao: See Parloa's AI agents in action at parloa.com. That's P-A-R-L-O-A dot com. This episode is brought to you by Capital One. Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called Chat Concierge and it's simplifying car shopping. Using self-reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trade-in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One.

17:03Huyen Tue Dao: There's more we need to know today. Let's get to the briefs. All right, well, let's give you some hardware chaser. After all those AI goblins, Motorola had a Razer family announcement on Wednesday. So Motorola announced pre-orders will start for the Motorola Razer Fold on May 14th for$1 ,899 USD, shipping May 21st. The new Motorola Razer family goes on sale May 14th, starting at$800 for the base model,$1 ,100 for the Plus, and$1 ,500 for the Razer Ultra. The Moto G87 mid-range phone gets a 200 megapixel camera coming to Europe soon for€399. Motorola also announced the Moto Buds 2 Plus True Wireless Stereo or TWS earphones.

17:52These buds were announced at MWC in March. The Moto Buds 2 Plus are available in the U.S. now for$149. They go on sale in India May 8th at 12 p.m. for 5 ,999 rupees, which is about$65. You can also get the brilliant edition of the buds with 12, count them, 12 Swarovski crystals encrusted on each earbud and an undeniably, very sexy, more eye-catching Pantone violet indigo paint job. The sexy part was my editorializing because I'm a purple girl. There's a similarly crystal encrusted version of the Motorola signature phone. They're available worldwide with the exception of North America.

18:35Huyen Tue Dao: Yeah, I'm very curious why North America isn't getting the Swarovski crystal encrusted Motorola devices. But OK, yeah, things are weird right now. Must be a tariff on Swarovski. They pay by the difficulty to pronounce. Yeah, great family line. In fact, I think the G87 might be the coolest one of there just because you get if you're into cameras, which I'm not, but I know so many people are. You get a really good camera. The front camera got an upgrade to for 399 euros. Not bad. Not bad at all. Motorola is always hitting that that middle that middle range price point pretty well. And also they have some of the the most like fun, interesting colorways.

19:18And yeah, like, OK, the Motorola Signature phone, it's extra, but it's got this like quilt. It's not only Pantone Violet Indigo, but it's got this quilted pattern where this Swarovski crystal.

19:30Huyen Tue Dao: Like it's a pillow for your crystals to lay upon. It's so extra, but at the same time, I feel like, especially with smartphones, the specs on smartphones minus foldables are kind of just logarithmically going up. Anything like Swarovski crystals and a brilliant purple to distinguish things is a win in my book. Yeah, I think Motorola just is solidifying its place as the RC Cola, right? Coke and Pepsi can go have their day and battle each other and do taste tests. But if you want a good, reliable, affordable phone, get Motorola. rolling. If you don't get the RC Cola reference, just go look it up on Google.

20:12Huyen Tue Dao: It's third place, whatever your favorite third place thing is. That's what we're talking about. It was a huge day of tech company earnings on Wednesday as five of the biggest tech companies announced double digit revenue gains on the year and increased capital expenditures. They're making more money so they can spend more money. Obviously, not everybody wants to hear all the numbers. You You can go look it up if you're interested in a particular company. But these are the numbers worth paying attention to as a consumer, as somebody who's like, I just want to know if these companies are still going to be around, if they're going to be cutting services.

20:45Huyen Tue Dao: On the money-making side, Alphabet revenue up 22 % and Cloud, which is important for those of you who use Google Docs and stuff, up 63%. So Alphabet booming. Gemini paid subscriptions grew 40 % on the quarter. So that is definitely a growing business that Google will continue to pay attention to. Microsoft revenue up 18%, not bad. Cloud up 40%, very good for them. AI business up 123%. I don't think a lot of people saw that company. But gamers beware, Xbox hardware fell 33 % and game services fell 5%. So they do have some work to do there. Meta revenue up 33%. This is going to get lost in the headlines because Reality Labs lost$4.03 billion.

21:34Huyen Tue Dao: They're still spending a lot on Reality Labs intentionally with the belief that it will pay off someday. They're treating it like a startup. But revenue is up for Meta. However, this is concerning. Daily active people dropped. That means they have fewer people to sell advertising to. They still have$3.56 billion. So they're not going away tomorrow, but something to keep an eye on. Amazon revenue was up 17 % and Amazon Web Services, AWS up 28%. That's its fastest in 15 quarters. So the overall story here is if you're doing cloud, you're doing very well. And if you're someone who's like, is there really demand for these cloud services?

22:10Huyen Tue Dao: Yes. Is there really demand for these AI services? Yes. That was true for Microsoft and it was true for Alphabet as well. Samsung, in a little bit of a different situation, revenue up 69%. And that's entirely, well, not entirely, but that is largely due to RAM. They are selling RAM at the highest level and the highest price that they have in a long time. All right, so what are they going to do with all that money? Let's talk about the spending side. Amazon, Meta, Microsoft, and Alphabet expect to spend 77 % more in capital expenditures than the record$410 billion combined those companies spent last year.

22:50Huyen Tue Dao: Alphabet's raised their expenditures to$190 billion. Amazon plans to spend$200 billion. Meta raised its projected spending from$115 to$135 to$125 to$145 billion. And Microsoft is on track to spend$190 billion in CapEx this year. Side note, Europe's software companies, SAP, Capgemini, Dassault Systems, beat estimates with positive earnings as well. So that's a good upbeat for software, especially in Europe. And what about Samsung? Or for that matter, SK Hynix and Micron. Samsung said demand would continue to outstrip supply into 2027. That's not good news for us, but it's good news for their bottom line.

23:32Huyen Tue Dao: However, none of these companies are talking about CapEx spending to build capacity. So the cloud folks are spending a lot of money to build data centers because they've got a lot of money going in. And they're saying the only thing that kept our numbers from being even bigger was capacity. A couple other factors to keep track of, the spending done now will spread out over years, and that will impact profitability for years to come. So you spend$190 billion on CapEx. That doesn't come out of your bottom line this year. You spread it out with depreciation over several years. So the big revenues now are a bit of a bulwark against the fact that these costs are going to continue to hit their bottom line for a decade or so.

24:14Wow. Yeah, I mean, it's just as just a cog in the wheel or wait, a cog in the system, a cog in the wheel, a cog in the wheel, a cog in the wheel. um i these things always kind of like i i never quite understand um how how these things work and how like different court i just i just don't understand the strategy sometimes but it is

24:36Huyen Tue Dao: would you like me to try to explain it i'm certainly not a business professional but the strategy is uh we are getting so much demand for our cloud services look at our revenue okay if we build more capacity we could sell more of them so we're going to spend a lot of money with the idea that we will make even more money. Okay. Is there any, like, I don't know, is there a point where, or is there any, like, legislation that, I don't know, it sounds like a gamble, which I guess is... That's every business ever. I mean, that's fair. It's perfectly fine. The reason a small business takes a small business loan is exactly the same principle.

25:16Huyen Tue Dao: Okay. You're saying, I don't have the money now, but I'm going to spend it because I'm going to get the money. Now, in these cases, they have the money now because they have big piles of cash that they can use. But they're saying, I'm going to spend it because I'm going to make it. And I think that's the thing to pull out of these earnings reports is if you're like, yeah, but are they really making it or is it really a gamble? They're making it. They're making the money now. And they know that we're maxed out. We could sell more. Let's make more. The interesting thing is the hardware makers not doing that.

25:48Huyen Tue Dao: because they are worried that if they build the capacity, the demand will eventually stabilize and then they'll be left with too much capacity, which is going to keep your RAM price higher because they're not gambling, because they're not spending that money ahead of time and just raking in the profits. That's interesting. I mean, that is why hardware is hard because you can't scale it up and down in the same way you can with cloud and SaaS. Fascinating. All right. Well, Sony issued a statement in response to discoveries that a 30-day internet connection, a 30-day internet connection clock was showing up for PlayStation games.

26:22Sony wrote, Players can continue to access and play their purchased games as usual. A one-time online clock check is required after purchase to confirm the game's license, after which no further check-ins are needed. Sony did not address the issue of a console with a dead CMOS battery, which would be unable to verify the date and time. Sony did not address the 15-day window that was also theorized to mean you had to do an online check-in after 15 days in order to get an infinite renewal of the license.

26:53Huyen Tue Dao: Yeah, this was an unsatisfactory explanation. It did confirm that it exists, and it confirmed that it's not meant to be an ongoing every 30 days thing, which is very good. But yeah, what if you got that dead battery? That's a problem. And there were several people that noted that the first check didn't come until 15 days in. So what they're saying is either that that's not true or they're not pointing out that, yes, you have to do a one time online check in 15 days after purchase. And it would be good to know that most people won't run into this because they're connected to the Internet every time they log in on their PlayStation.

27:29Huyen Tue Dao: But, yeah, for that that that C-bomb issue, it would be nice to have more details on that. But at least they responded. it. Well, if you would like to respond to us, you can do so all kinds of ways, including social networks. We are at DTNS show on X, Instagram, threads, blue sky and Mastodon. And for TikTok and YouTube, spell it out the whole way. You can find us at Daily Tech News Show.

27:56This episode is brought to you by Capital One. Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called Chat Concierge and it's simplifying car shopping. Using self-reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trade in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One.

28:29Huyen Tue Dao: all right let's do some quick headlines that are going to make you go like yeah i knew that i heard it on dtns all right well thanks to motang who noted the story on the subreddit china's stellato s9 sedan developed by baic motorcorp and huawei can project a two megapixel image from its headlights letting you show a movie on an outdoor screen or wall it can also project navigation arrows on the ground. That's not going to be a problem if everybody got that, where everybody's arrows are in the intersection. But it's pretty cool, actually. I think that's pretty cool. Tokyo's less busy Haneda Airport, pro tip, fly into Haneda if you can.

29:08Huyen Tue Dao: It's closer. We'll start testing humanoid robot baggage handlers in May. Pro tip, fly to Haneda. You get to see the humanoid robot baggage handlers. Now, you won't see them on the ground right away. They're going to start with simulations to determine where they can best be used and how to actually implement them. But they will eventually join humans on the ground. And they say they possibly might have them in cabins, clean in airplanes. Haneda also has a lot of great restaurants, so like addendum to the pro tip. Nice. I guess fly to Haneda. Three pro tips. Three pro tips. All right. Well, in Japanese aviation, Japan's Air Kamui has developed a cardboard drone, the Air Kamui 150, for the Ministry of Defense.

29:48The idea was that they would be disposable, though still$2 ,000 to make.

29:52Huyen Tue Dao: I guess the actual reusable drones that fly back to you are even more expensive. Yeah, but these will just biodegrade. I guess it's not bad for$2 ,000, I presume. I mean, it's still a killing machine, but, you know, better than it biodegrades. Thanks to TechnoMensch, who noted on the subreddit that the ever-popular Notepad++ code editor has arrived as a native Mac OS app. It's even a universal binary. So you've got an old Intel machine. Amazing. 20 years in the making. I love this. I still love that we're excited about Notepad++. I know, me too. All right, well, even Reality's smart glasses announced Terminal Mode, which projects an icon in your field of view to let you know what your coding agent is doing.

30:38which is very appealing to me, but also, man, the brain fry is going to be so real. You'll never turn off.

30:45Huyen Tue Dao: YouTube announced that picture-in-picture mode is now rolling out globally to everybody. It's no longer reserved for the premium-paying users, and it's not US-only anymore. You can get it everywhere, eventually. Well, OpenAI has said it has signed contracts for 10 gigawatts of compute capacity, a goal it originally hoped to meet by 2029. And speaking of signing up all that capacity, OpenAI also rolling out a security-focused version of GPT 5.5 called, appropriately enough, GPT 5.5 Cyber that is very much like Anthropik's Mythos and may only be used by approved users. now I have been of the opinion that maybe Anthropic was limiting Mythos partly because of computing capacity so I wonder if that is also partly what's going on here but they do have 10 gigawatts on the way so that won't last for honestly the more that I've seen stories come out about open AI and stuff like that and also just okay that makes actually a lot of sense the compute capacity is the main issue who knows All right.

31:50Well, we end every episode of DGNAS with some shared perspective. Today, Howard has some further thoughts on the cursor agent that deleted a production environment for a rental car company.

31:59Huyen Tue Dao: Yeah. Did you hear the story earlier of the week? Oh, you know I did. I said stuff about it to somebody. Well, Howard wrote on Monday's episode, the discussion about whether cursor or railway, which was the platform provider, could be at fault, made me want to share the principle of least privilege. The general idea is that users, systems, or processes, agents in this case, are granted only the minimum access to systems, permissions, or data needed to do their job or task. The goal is to reduce the scope of a security breach. But I think in this situation, the agent may have had broad permission and access as it was acting on behalf of the human.

32:39Huyen Tue Dao: I have feelings, soapboxes, and rants, but I would encourage people to consider least privilege when granting access to AI agents rather than simply letting them act on your behalf. And he's got a link to a good article at Palo Alto Networks about that principle if you want to get that out of our show notes. um yeah Howard I it's kind of funny because last week I I finally kind of copped to downloading flawed code and letting it help me do certain tasks like doing prep for podcasts um and I also study Japanese so and I have a lot of like random you know vocabulary words that I want to get into like Anki which is like a flashcard deck builder and I want to do all these fancy things with it and at some point it started asking me for like hey can I can I access this can I control your browser can i do this and i'm like oh god i know i shouldn't but it and it but and and i guess to say as someone who knows that i and i i've heard uh i hope you have proud howard i have heard of like the principle of least privilege and i do understand it and i do uh see that it's very valuable and even as someone who has been aware and definitely um works at places where you know it's a pretty important principle to live by i still was like well maybe it's fine this time maybe it's okay it's because we're humans right i do the same thing sometimes where i'm like i know i probably shouldn't do this but i'm sure it'll be fine i'm sure it'll be fine as the bane of human existence yeah and i guess for me i'm like well it's just my personal stuff it should be fine which maybe is actually the opposite i should be thinking or at least it should be my own personal security should be as important as as as my job and it might be fine but it's that like yeah i don't want to go to the trouble of making sure it's fine yeah temptation that can get you and And also, oh my gosh, the cloud already did all these things for me.

34:21Let it just finish the thing. Let it just finish the thing. Yeah, yeah, yeah. Yeah, there's that pressure too. So I think it's a really good reminder that it is easy, but that we still need to put our noodle on it. All right, well, what are you thinking about? Do you have some tips, principles, of tips and principles that you want to share with us? We'd love to hear them. Feedback at dailytechnewsshow.com.

34:41Huyen Tue Dao: Big thanks to Howard for contributing to today's show. Thank you for being along for the Daily Tech News Show. everybody out there listening and watching you keep us in business so if you haven't already and i know some of you have this week so thank you for that become a patron right now patreon.com slash dtns the dtns family of podcasts helping each other understand diamond club hopes you have enjoyed this program.

35:15This episode is brought to you by Capital One. Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called Chat Concierge and it's simplifying car shopping. Using self-reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trade-in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One.

35:44Huyen Tue Dao: The savings are stacked at Mattress Warehouse. Right now, save up to$800 warehouse-wide on top brands you know and trust. Plus, buy one, get one free deals so you can upgrade sleep for your entire family and get a free adjustable base with select mattress purchases. With 0 % interest financing and free delivery available, getting better sleep has never been easier. And with our exclusive lifetime price guarantee. You'll always know you got the best deal. Save big today at Mattress Warehouse. We're more than a mattress store. We're Mattress Warehouse.