You Can't Out-Algorithm Bad Data - DTNS 4903
31:40 Share

If you want to stay off the date on the latest, take news and about five minutes subscribed the daily take adler es as D A I L Y, take headlines wherever you get your pocket.

Runs here from mid mobile with the Price of just about everything going up during inflation. We thought we bring our Prices down. So to help us, we've rought in a reverse auctioneer, which is apparently a thing that I4444 in mobile guy com slash switch.

forty five dollars, up from payment to b to fifty dollars month. New customers on three months only taxes. Export speeds lower before give B, C, D.

Deep in the ocean. And orcott is on the hunt. These aren't your average orkis. These guys are organized marketing team.

Did you get those social media post scheduled for the CEO migration?

I I captain, we even have an automated notification for all pod managers when they go live. They use monday dotcom to keep their teamwork sharp, their communication clear and their goals insight. Monday dotcom, for whatever you run, even orka go to monday 点 com to dive deeper。 hi.

This is far nesh arabi, host of the webbing winning podcast. So money, if you're aiming for a goal, be IT saving for a house, growing your family or retiring life makes IT difficult to stay the pose. But with a dedicated miral advisor, you get a personalized plan and a clear path forward.

Having the bull at your back helps your whole financial life move with you. So when your plans change mars with you every step of the way, go to dog. m.

sh. Bullish to learn more, marry a bank of amErica company. What would you like the power to do? Investing involves risk.

Marilyn pierce, fenner and Smith incorporated a registered broken dealer, registered investment advisor member. S. I, P, C.

Picture of this, you're in the garage, hands covered in greece. Just finished ed up tuning your engine with a part you found on ebay and you realize, you know what, I can also use new breaks. So where do you go next back to ebay? And you've got ebay guaranteed fit.

You order a part. And if IT doesn't fit, send IT back, simple as that. So when you dive into your next car project, start with ebay. All the parts you need at Prices you love, guaranteed to fit every time ebay things people love daily.

Technology is made possible by you listening right now. Thank you. That includes pat, my cortez, a one stir. And jim noted on this episode, D, T, S has A, I hit a wall. We ask Andrew main, plus big gym. The trade nerd helps us understand how tech products get Priced and why the rumor to iphone seventeen air might be too fit. This is the daily technos for monday, november twenty fifth, twenty twenty four in los Angeles, sometimes merit from .

the atland area. I'm nick offord.

This shows producer .

logic ing and we are kicking out of the week with a gemp c show. It's going to be fun. You're going to learn something. You're going to have a good time, and we're going to start with .

the quick kids.

Ray's very pie announced an updated version of the peo two. The peak to w. The w stands for wifi because it's got an onboard wifi. Chih just cost couple dollars more than the pecos.

So seven bucks instead of five includes the rp twenty three fifty micro controller that includes A O two thousand eleven n which is also known as wifi for these days. And you get bu tooth upgraded to five point two along with that also as a single USB one point, one point. Uh those are all import things to know if you're someone who plays around with these grasper. Pes, the new rp twenty three fifty microcontroller er is the one they announced in August. And interestingly, IT combines architecture that that uses both ARM and risk five architecture.

Bloomberg sources say SONY is, in the quote, a very early stages of working on a parable council. Microsoft also has said IT is working unaffordable, but both of these .

projects are years away. I rote up, I know to this.

keeps saying.

and have actually happened yeah, yeah, yeah. Like fetch, the roku channel is already available on android T, V, and therefore reign google T, V. You don't need a rock u device or T, V to watch IT. But roca just announced some immigration into google T, V.

So if you're using the device that uses google T V, you can search through the google T V APP and rocket channel shows can show up, so you don't have to jump into the roku channel APP to find them. New tabs are coming to the google TV homework reen that will recommend rocke channel content and show you episodes you're in the middle of s so you can jump right back and finish them. Finally, around five hundred of rocus fast channels, the free ads supported channels, are coming to google TV live tab mac rumors.

a spired of myself o that indicates that macbook prose speakers will be available this week as a stand alone component, making them cheaper to repair. Right now, you have to replace the entire tag case to replaced speaker as a top case includes things like the battery and other parts that make IT expensive, especially if you only need to replace the speaker. Um right now you have to um good you can have all available now is a repair manual that um has the steps for a place thing, just speaker models and I guess what it's already .

available online and video debut a new generation model called foundational generative audio transformer opus one or forgotten gotto forgotten IT a IT can create a modify audio from a tex prom works with multiple kinds of sounds, including music and in video says IT has strong multi accent and multinational capabilities. They did a lot of work to bring in people from around the world for that.

And video said IT could be used by music producers, for example, to prototype a song and and even edited video game producers could use IT to modify prerecorded assets so that they change dynamically with player actions. And they also suggested that would be great for building your own language learning tools. NVIDIA did not announce when or who will .

get access to IT meet claude the A. I. Assistance from and propionic transforming how organizations work. Imagine every person on your team having an expert collaborator who knows your company inside and out. That's claud.

When you upload company documents in a claud, you're giving IT the context to become your subject matter expert and deliver a student helpful responses. Cloud helps engineers ship products faster, marketers craft compelling campaigns and sales teams personalized our Richard scale from brainstorming execution. Claude looks like A T mate.

Not at all. Empower every person in your organization with A I, that's both powerful and protected. Data stayed yours by default. We never train our models on your company conversations and content. Join leading enterprises already working smarter with transform your organizations productivity and visit anthropic dock com slash enterprise.

We've all been there scrambling for a last minute gift that doesn't feel last minute with trade, you can send a coffee subscription directly to someone's in box in minutes. It's fast, easy and guaranteed to impress. A trade subscription isn't just coffee.

It's an experience. Every month, they'll get a new freshly roasted coffee perfectly matched to their taste from one of the top small batch roast USA ready to cross that last gift of your list. Trade is offering their best savings of the year on gift subscriptions right now, so had to drink trade dot com slash D T, N S to send a personalized coffee subscription in minutes.

Picture of this, you're in the garage, hands covered in greece. Just finished up tuning your engine with the part you found on ebay. And you realize, you know what, I can also use new breaks.

So where do you go next back to ebay? You can find anything there. It's unreal.

Wipers, headlights, even called iron tics. It's all there. And you've got ebay e guaranteed fit.

You order apart. And if IT doesn't fit, send IT back, simple as that. Look, D I Y fixes can be major. Doesn't matter if it's just maintenance or a major mod. You got IT, especially when things are guaranteed defeat. So when you dive in to your next car project, to start with ebay, all the parts you need at Prices you'll love guaranteed to fit every time ebay things people love when you're starting or scaling your company is security program demonstrating top notch security practices and establishing trust is more important than ever. Fanta automates compliance for I soo twenty seven O O one soc, two gdpr and more, saving you time and money while helping you build customer trust.

Plus, you can streamline security reviews by automating questionnaire and demonstrated your security posture with a customer facing trust center, all powered by vantine AI over eight thousand global companies like a lassen flow health and cora use venta to manage risk and prove security in real time. Our audience get a special offer of one thousand dollars off venta at vena. Dotcom slashed D T N S. That's vena dot com slash D, T, N, S for one thousand dollars off.

Are you a professional pillow fighter or in ninety five low cost time travel agent? Or maybe real estate sales on mars is your profession. IT doesn't matter whatever IT is you do, however complex or intricate, monday 点 com can help you organize or to straight men， make you more efficient.

Monday 点 com is the one centralized platform for everything， work related. And with monday 点 com work is just easier。 Monday 点 com or whatever you run， go to monday that com to learn more.

It's holiday season, which probably means you are looking for deals on the techy on I know I am, but I also got us thinking, how do these Prices really get sad? Why can't the sale Price just be the Price all the time? Do company we lose money during ourselves? So tom found someone who could answer that question for us.

James satcher is a trade consultant, host of the trade nerd. Domi sometimes goes by the name big gym if you seen him in our chat rooms or discord. Thanks for talking to be mad. Good for good to have you joining us.

Wonderful beauty time, always great time to be. Dt N. S.

Now i've talked you before about logistics um and there's a lot we could talk about that still, i'm sure. But with sales coming up, with the holiday is coming up, I wanted to talk about Prices because I see people talking about Prices and and I see discounts where it's like the manufacturer discount or the retailer discount. So I figured this would be helpful in understanding when you're shopping for consumer electronics, like what goes into this pricing. So let's start real basic like how does a company determine the Price of something at selling?

Well, the first thing they do is they take A V G board out that .

i'm kidding or .

are you well， some companies actually do that. Um take the way what you do is you look at what is the manufacturer cost, uh, the good. So you have to you know find all your components will material label cost overhead those types of things and you come up with that information.

Then based on that, you look at what competitive in the market to figure out what kind of profit uh percent you can incur with that. Now that's the manufacturer r so if you buying requirement manufacturer, that's usually how they are kind of pricing things out. They are looking to try and make certain spread.

Now the percentage of markup is typically called uh, from what your purchase Prices to what your sale Prices is typically rated in a percentage. And we call those points in the industry. We retailers call those points.

So some companies look for a low point with massive volume. Some companies look for high points with small volume. Um most are somewhere in between. Most are trying to get um on typical thirty to you know anywhere between thirty to forty percent is pretty good. As far as middlemen, retailers like to go through higher because they have added cost with expenses such as marketing, uh, the Operations of any retail spaces they have and a lot more overhead.

So essentially, I think a lot of people think all the cost is how much a costume to make IT and maybe a tiny bit more. And what you're saying is if I if i'm getting this right, both in the manufacturer and even with retail, the cost is how much does IT cost us to Carry this? And then how much more can we get because we want to make as much IT is not just big companies that do this garage sale people do this. How much can we get .

on top of that. But it's not just how much more can you get. It's also comes back to you know the money of the profit that you make on a product helps determine R N D expenses for future, for future generations.

IT also helps determine uh marketing sales, mergers and acquisitions, purchases that you might be looking at doing in the future. Um and then you also have financial requirements, maybe three year investors. So for example, if you are privately held uh company by venture capitalist, they're looking for a certain percent of return every single year to make their investment.

For what if you're publicly held company, you have to have an earnings per share raise show that keep your investors happy oh wisely start to stop your stock and well in the value of the company goes down. So it's dependent on a lot of different factors such as size and industry. But for the most part, that's how Prices looked at .

yeah Prices, what we can, what people will pay. And there's there's a range there that you can go and all that. But it's IT just has to be above what IT costs right?

well. And then there's also things that you can go below cost. We've seeing in the past certain companies uh for simple intendo and h SONY will go a low cost on the hardware object just because they know they're na make IT up in the software. Um those types of things are, you know what we what we used to call the old razor model of you would go by the razor for near to act cost and then what you be sucked in the buying all the razor blades at a specific costs that is way over instated and that's where they make their money.

And just like that might help inform our next question, which is, okay, so how does the sale Price work? If i've done all this work to determine what the Price should be, why would I take twenty percent off on black friday?

Well, because you're trying to IT comes down to mix. I guess that's the cleanest way to say IT comes down to mix. What you're trying to do is you're trying to drive them in people to come into your store. Uh, if you're doing an uh either an online store or physical store doesn't matter. You're trying to get people to spend time with you as a retailer to not only purchase the know the low margin items that you might be signed, but also the high margin. So if you go in and you're shopping for a computer, okay, you buy the computer, maybe there will be making three to five points on that specific computer, but then you need a bag, then you need a additional charger, but you might need, uh, different graphics to go with IT. And those items are what they're going to then build up that profit margin with as a whole um a whole of cart, I guess, is the best little look at IT.

So it's not exactly the same as the lost leader thing. We were just talking about where you sell below cost but the ideas like, okay, I may not make as much on this one but little pay off.

right? So the idea behind uh, a lot of these door buster deals is they keep that low because it's a doorbell. They want people at the door, they want people to bus in and go for those special guidance. But hey, while you're here, why don't you look at taking look at something over here that we make forty to fifty eight and so that we can turn around and justify the cost overall? The other thing to look at is you know competition also plays into this a lot and we have seen companies that will go out and do loss leader products uh for you know not only for black friday, but you'll push through uh TV are a good example this i'll push those uh TV costs from black friday all the way through the big game day, which we of course .

earth can say because that's yeah yeah no .

I am not have the I came after me time ah so but uh overall they're gonna push those because again, you coming, you buy the one thing, you're gonna be coming back to buy more and and that's the overall building .

the habit basically yeah ah you're .

building the habit and you're building the loyalty. And that's the other thing that we're seeing lately that a lot of these uh, retailers are pushing more and more is consumer information, consumer points, consumer spending. Um to keep people in the door. Um this used to be very, very common with your warehouse and club experience groups such as costa and D J S. And seems uh where you would want to bring those people and they pay a membership fee, but you can then target them and products of mass box there to keep your percentages high. But we're seeing that now with anything from going to buy vehicles all the way to, uh, trying to get haircuts, where there that intelligence in that information if they can get you in with a dog buster and they can get your information, then they can turn any marketing .

Better fantastic information. Big G M, thank you so much. Uh, for for helping me understand this. Help and everybody understand this. If folks want to keep up with you, where should they go?

Oh, well, there's always the trade near dot com, but the baseball to find is on the G, T, N, S. discord. Where did you in there? People come on in IT. So when they have a fun time, message me and i'll be more than happy to talk more about the wonderful .

world to trade.

Thank you.

All right, let's get into this um iphone seventeen air rumor to becoming IT next year. Supposedly it'll be between five and six millimeters thick, which would make IT the smallest iphone since the iphone six. The iphone six was six point nine millimeters a IT would be much dinner than the iphone sixteen, which is seven point eight millimeters.

Uh IT makes IT hard to fit enough battery and though according to the information, IT will only have a single speaker in its airpix because there's no room for a second speaker at the bottom. IT will use apples in house five g modem and this is all according to the information sources but I will not support millimeter wave 5g service， which is the fastest five g service, although it's the one that won't go through walls. So not everybody's using IT all the time, but will be it'll be a more limited modem.

It'll have a single camera on the back house in a large centered camera bump and I won't have a sim tray. It'll only rely on em um which is a problem if they want to sell IT in china since the government requires you to have a physical symptoms, you want to sell a phone in china because they want to be able to make you show up and show your ID and associate real name uh with that sim card. So uh some some issues with making IT so thin.

Ica, but IT has to be thin, right? Everyone wants the thinnest phone. I don't know IT seems weird.

So you don't have to really convince me too hard to bad apple product. I am an apple snob after all. 是 but I don't know what the purpose of this phone particularly is。 Um we just got the iphone sixteen and you know these phones are already pretty thin. What you're sacrificing for and thinness, I don't think IT quite is um worth the loss for for what you're going for .

two millimetres.

right for what the the capability that's going to be reduced for now, the only thing that I could probably say would make IT worth IT for someone is if the Price point is really low, you already have the iphone esc, which is kind of like the critical discount phone for iphone. So IT would have to be less then the S E to me to even really make sense. But the S E is already fairly, fairly inexpensive.

So I don't know how what kind of legs. This rumor has because I just don't see the the the benefit of apple creating this and the return that little get as compared to the current devices that we have out there. So i'm a little for most will at the notion of an iphone air yeah like .

is a great point about the S C. Because all of the sources are telling the various folks like mark garmin and mingei o. That an iphone S C is coming in the first quarter or so next year that it'll be updated IT.

It'll still be in that affordable Price range, that mid range Price range. So yeah, I doubt the seventeen air would be meant to supplant that. IT seems like that would be something that would be introduced in the fall, uh, and wouldn't be a bargain phone, maybe would be slightly less expensive. But IT seems like it's just meant to be a style that's like, oh, you you want something very slim and you know, baby, two millimetres difference feels bigger in your hand, right? Sometimes that that can happen, but I just don't know who is is the target market of people are like, yes, the thinnest phone possible appeals to me yeah.

I have to degree. I don't know who the target audience is for this purchase. So if they'd .

weren't making these compromises, I be less confused. I'd be like, well, you know, making IT then is an interesting marketing proposition, but if it's going to have possibly smaller battery and there's thermal issues as well, yeah, I reduced speaker. I don't know if they're using their own modem because that fit makes a thinner because they can decided but wouldn't be surprised if that part of its you're compromising on millimetre e wave. Again, most people probably sing millimetre wave in their five g phones, but it's a lot of compromises to put together. And if they can't even sell IT in china, if IT doesn't have if they can't make a physical symptoms fit in there, IT seems like seems like a lot of compromise .

that things like a lot for sure a well.

if you know why they're going to make an iphone seventeen slim ah let us know on social network and you can get in touch with us at dt show on x, at D T S show at msd D N social on mastodon, at daily technical show on tiktok and D T N S pics. D T S P I X on instagram and on threads.

This is an antonia sec wave, you hear, because of our job, we need to be connected, tiny over connects by more than us.

right? But it's important.

like connection with your friends, your family, every connections were your life. And that gives you something that provide high speed internet, just in the context of this show is so important, will be talking about something here.

And it's easy to forget when you have a big parts of the country still don't have high speed internet.

So A T N T knows this. There are ware of IT at, and they're making a huge effort to change that. What they're doing is on trying to cover thirty million plus locations currently who don't have IT to get IT with fiber by the end of twenty twenty five.

Millions of people obviously will benefit families, businesses, schools, hospitals. There is a place called old and counting in kentucky, where they are now providing high speed internet to more than twenty thousand customers. And t, join a lot of great things.

connecting changes everything. A T, N, T.

why get all your holiday decorations delivered through indecent? Because maybe you will only bought two reatha, but you have twelve windows, or maybe your toddle got very eager with the advent calendar, or maybe the inflatable snowman didn't make IT through the snowstorm, or maybe the twinkle lights aren't twinkling. Whatever the reason, this season in starts here for hosts and their whole holiday hall get decorations from the home, deeper cvs and more through installed and enjoy free delivery on your first three orders, service fees in terms apply.

AI models took off a for most people, became into your mind for most people. In november twenty twenty two, win ChatGPT launched. That was on OpenAI GPT three point five model. Five months later, they had gp t four.

And i've had other model since, but not a GPT five anthropic launched a three point five version of its senate in june, three point five version of high school in number, but has been delayed on a three point five version of claude opus. There's also german I delays from google. Some people have been describing this as AI hitting the wall. So we wanted to talk to Andrew main, who keeps up on all of this. And I I kind of take my cues, Andrew, from you on on where A I is going.

And thanks for joining us.

Glad to be her toe. Do you think what do you think people mean when they say AI is hit a wall? And do you think IT has?

I think they are another talk about all. Be honest with you. So you did a pretty t timeline there of of where I ve been for me, IT. IT started back IT was four years ago when I started to open an eye working on capability discovery, which was trying to find out what these models could do.

And you know, open, I had made a big bet that you could scale increasingly, not a compute, put data into IT and continue to get like increase in returns. And that's when they were from GPT one to GPT three, GPT four. And that's been sort of the mentality of, like anthropic canada, a left opening eyes to go to anthropic with this team members there in other A I labs too.

And the idea of the wall kindly comes into this, this idea that, you know, at some point you're gona run out of, at some point IT can't work, which you know certainly is true. It's certainly you reach a limited scale of, like you can only make a jet engine so big or so powerful, but you can go much faster. You know, the certain things that have these limitations.

You can only get a solar panel that gets to one hundred percent efficiency. You can get to one hundred and five percent. The thing that I bring up is that the people who are saying, oh, we hit the wall one, it's you based on like we heard rumor, we heard the staying about you something something there, you know, my friends still working on the stuff.

You know, I know the research open. I nobody there's panic, nobody there's act like, hey, I think we hit a wall. And I think that there's been a lot of incredible developments that happened along the way that we just sort of a forget about.

Remember, GPT four was last year. Last year was GPT four. okay. And as you scale these models, you have to increase the size of amount of computer. So if you are using you know ten thousand computers in ten thousand clusters, you might go one hundred thousand whatever.

Or you increase the training time if you don't have ten x the computer, if you want a ten nex that you know what you're training used to increase the training time, whatever with these models is one that might take longer to train. Because if you want to increase on a compute, you also have to factor in testing how long IT take to test bigger models, take longer to test. There are safety training.

There's all these other things that come in to play. Plus, if you screw something up, let's say that you go and to go test mobile, to do a big six months training run or something like that, and you found out that one of evil is way off, then you ve got to start over. So press to stay when we hit the wall. I don't think like a year. So and is the time to say if we hit two or three years, i'd be okay.

But also, we had a big development, our really big development that we just came out a couple months ago, and that was the o one models from open eye, which showed that not only can you make these models smarter or Better, capable of generalization by increasing how much computer used in training, when you do inferences time, you can actually, if you let, when you let me think about a problem longer, we found out that all this didn't lifted all these events. And these are effectively models that were the same size as we had a year ago. But now you increase that. So there are all these different levels. You can pull the increased thematic capabilities.

Yeah, I get the sense when people talk about the wall that they're saying, well, you know, we had ChatGPT come out of nowhere and then we had ChatGPT four. Where is five when what you're describing sounds like maybe maybe the progress has slowed, but it's not stopped and it's not going down. Do you think that sounds right? Is progress slowing is at leveling out? Or or how would you characterized?

I don't know how IT depends how we measure what we talk about progress. So if you think about like the problem we have run into right now is when you want to measure the capability of these models, now I ran in to this. I had to handle part of the release for G P.

T. four. And as i'm trying to figure out how to explain the people at GPT four is capable of. I'm looking for really good examples to show what I can do. And GPT three point five kept getting Better and Better and Better at IT in the difference from gp, three point five and four started to diminish a bit because all the said, you know, he was two point five I was doing Better and eval's but even there reject t four to try to show how great IT was after. Like I I run in this a lot.

Like I know these o one reasoning models are really good for certain test, but I have to think of like h what's a problem that average people can get? What a good problem that average people can get to solve the model? And I I struggle trying to think of like, oh, what does IT fail at? And I have become up more complex things.

And that's the problem in trying to chinese models or build Better versions is IT. We know these models generalize. We know that they can actually, you can give that if you give them bigger, Better problems to solve, and need you reinforcement learning, they learn how to solve those problems.

The problem is creating evaluations. If I want a model to be with the right, complex scientific proof, I have to have a way to evaluate the scientific proof. I have ever a way to do that. And that's one of the problems is its limitation right now is really evaluations.

And and I know album has been very bullish about agents, as are many other people. And and to me, that sounds like, you know it's not that progress has slowed, baby, so much as we are getting different ways for models to be used, in some cases, maybe more narrow ways .

yeah any of there's the back in the olden days at GPT three. You know if you wanted to train a model, you would just take a large corpus of text, like common crawl, or things like all freely available, you publicly available text on the internet of whatever, and you just taken alcohol and you say, hey, look at this, predict comes next.

Keep running through this into your very good at predicting when he sees so what comes next, right? And know that gave us an amazing G P. three. But then we said, okay, that's good. But if I want to prompt IT, I need to say I can't to say White blog post because I might really think I have on a conversation with, you know, somebody working for me and gets like, no, you're write IT.

So we learned that, okay, if you give IT what we call instruction training, that these models became Better at IT, right, which is one of the big list from GPT three to three point five and chat PPT was given that all of the instruction training so we're at right now. Is that trying to figure out like we ve realized IT just training IT on a large corporate text um is only get you so far now you need to think about synthetic data and structure data in a higher quality data. That's why labs are right now actually reaching out to phds, novel or it's and other people to create much, much or higher quality material because we're pretty good to have answering the average question you might ask on the internet. Now we've got to figure out the next level.

So it's not as much about bulk as IT is about quality.

absolutely. It's the bulk was good for learning rules of language, some gentle facts about the world. But if I want to have a model that is not an expert at understanding certain branches of physics or medicine, then I did much more higher quality data, maybe not a red form.

Now, there was one thing you mentioned on the last attention mechanism, m that I heard. And I forgive me if I am remembering IT wrong, but IT sounds like you said the pieces for an artificial generalized intelligence are all there. Is that I remembering that right?

So in the chAllenge is, what's our definition? And the thing that we've noticed is that where were the cheering? We all remember when the cheering test was the big test of whether not something that was going to be sitting right.

And then we got closer to IT. And like, well, we have systems that can answer questions. We have systems can mimic human behavior. But that's not the same. And you can very easily a gp t chat, a free version ChatGPT.

Even a model like a load on your computer now could have probably blown away every cheren test to would have been ten years ago. And rose, it's not an effective measure. So we talk about agi, as we say, hey, is a system that can do, I can email IT and I can do just about any human level task.

Is IT here right now? I would say for many cases, you know, if you really want long, complex stuff more chAllenging. But I would say that I think there's a lot of pieces to create something that we would feel was A G I like.

that's great. I enjoy. Thank you so much for taking the time to chat and .

appreciate IT my asure time.

If people want to get the attention mechanism or anything else you're doing, where should they .

go and remain dot com? Easiest place to find what i'm up to and also i'm on x at Andrews in big.

big day, Andrew. Uh, for doing that, nia. I think one of the things I took away from that, that conversation was we often talk about AI as like, well, of course, it's trained on the bulk of the internet that needs more data.

But what Andrew was indicating is that a lot of these models have gotten to the point where. They need quality. It's the quality of the data, not just the amount.

absolutely. And I agreed with a lot of what he said. Um one of the lashes les I had was leaving a team that was that was the whole point of the team was validating the data quality and determining whether IT should be passed down the stream for our other um team members in the pipeline to use IT.

And IT was I have to say IT was a brand new team, a brand new group, something that I basically had to pull from scratch. And that's one of the key things is, again, you can't out algorithm bad data. So you have to make sure you start with A A solid base of data before you do anything. Otherwise you're just, you know, running bad data through these extremely complex algorithm. So I think he was spotted .

on yeah ah it's like we were rolling down hill with the original ChatGPT two years ago and now now we're just kind of cruising and there maybe some cobble stones and stuff. So you got to figure out how to keep you going on. Probably get out of here.

Let's check out the male bag. Got an email from mat who says, hi folks, all time listener, multi year patronne first time commenter. I wanted to add to the discussion from yesterday around why OpenAI would want to create their own web browser, and there was friday's conversation beyond the obvious opportunities for data and new customers and paid features.

I think it's an interesting idea for getting around the blocks that some sights and platforms have against OpenAI and other l ms. For indexing. Having their own browser, in theory, would mean that users could query the LLM based on content that is currently loaded in the or I love this idea because I not only would they often want to summarize that content, but they may also want to ask deeper follow up questions.

They could also translate a page, modify the dome to just show the relevant info for them, remove ads and more. He wrote, cough, cough a well, I have the microphone. Thank you so much for being such an amazing resource all the best.

And I hope you have a relaxing holiday weekend. Thank you. Mad, what do you think of that idea? I think that's intriguing of matt saying IT would allow them to answer questions about things that they are blocked from looking at otherwise.

right is a very interesting concept. The only thing that I may worry about a little bit um is how silo that could potentially be um but outside of that, I think it's a it's it's a good concept. This is a good thought process behind why they want their brothers or so a great food for thought. matt.

Yeah yeah. We are looking at for for good ideas about why, why else they might want to do this. Thank you that for step up up and given us one. And thank you unique a for being here today. If folks want to find you more often than this, where do they go?

You can find me pretty much at tex ivy of whether wherever there is associative platform you can also find me 在 hosting um my podcast novel west with brother tech hood also been a contributed to D T N S over at snob was .

cast that fantastic。 We have a new member of the detainees family of podcasts as well, day to scientist Andrew and Jones. Roy has a series called behind the data.

If you would listen to our experiment week. Back in August, SHE launched the pilot back then. This weeks episode, SHE talks with seh masked about how to make thoughtful analysis from all the recent presidential election data.

Again, Andrea Jones roy is super smart. They're one of my favorites and favorite of a lot of you all by the emails as well. So if you want to find their their new show, go to behind the data show dot com. Uh the importance, just like in the last episode, of being thoughtful about whether our narratives are based on a real data. Go check that out that's behind the data patrons .

to so looking for IT to .

that yeah yeah no, it's good stuff. Patch and sticker up for the extended to good day internet. A chinese company tested whether a robot could talk other robots into defecting, and IT did.

So we're going to talk about, like, what was really going on there. You can also catch the show live monday through friday, four pmr twenty one hundred UTC. Find out more about that daily technical showed comes slash life. We will be back tomorrow, today.

The DNS family of podcasts helping each other understanding .

this program.

Meat claude anthropic, S A I assistant, ready to transform how your organization works, empower every person in your organization with A I that looks like a team mate, not a tool. Security, upload your company knowledge and watches claud helps every department from engineering the marketing produce their best work faster your data stays protected why your teams reach new heights discover enterprise grade AI at anthropic dot com slash enterprise .

every idea starts with a problem more be Parkers was simple. Glasses are too expensive. So they set out to change that by designing glasses in house and selling directly to customers, there's able to offer prescription I wear that's expertly crafted and unexpectedly ly affordable.

Warby Parker glasses are made from premium materials like impact resistant and carbonate and customer s. And they started at just ninety five dollars, including prescription lenses, get glasses made from the good stuff. Stop by a worby Parker store near you.