AI Demos: Claude 3's Opus, Mistral, Groq Playground, EMO by Ali Baba | E1912

0:00I said, help me with these instructions. and look what we got hit by here. Says, I apologize, but I cannot provide instructions related to assembling that particular object. It appears to be to depict an unethical, potentially illegal item. I would suggest focusing your efforts on more. What? He thinks it's a bomb? Positive construction projects that promote harm. What? Yeah, it's a shelf. You got caught by the woke AI virus. I know what happened here. This Week in Startups is brought to you by Lemon.io. Need to speed up your product development without draining your budget? Hire vetted engineers from Europe at Lemon.io.

0:38Go to Lemon.io slash twist to get 15 % off for the first four weeks. OpenPhone. Create business phone numbers for you and your team that work through an app on your smartphone or desktop. Twist listeners can get an extra 20 % off any plan for your first six months at openphone.com slash twist. And looking to up your marketing game? Check out the podcast, Marketing Against the Grain, hosted by HubSpot CMO Kip Bodnar and Zapier CMO Kieran Flanagan. They bring you the latest in marketing trends, growth tactics, and innovation. Available on all your favorite podcast apps. All right, everybody. Welcome back to This Week in Startups.

1:21It's Madra Mondays. Sandeep Madra is back. He had a big week. His company, Definitive Intelligence, was acquired by Grok, G-R-O-Q. And so congratulations to Sandeep Madra. You had a nice little cameo on the old all-in podcast this weekend. That must have been great. I appreciate it. I think I got something like 4 ,000 LinkedIn requests, which is pretty impressive. But no, we wanted to thank you guys, obviously, again, for support, but also helping get the message out. I think that's what was really key there in terms of what you guys have been talking about. And so we're going to try to work it into some of the demos today to show you what it's actually really about.

2:05And so Definitive I.O., as people know, you did all this AI enhanced data analysis for public private data. But now as part of the Grok movement, you're going to be taking over developer relations, I guess. Is that, am I correct in like your role specifically? No, that's one of the sub functions. You know, really Grok is a company that's been making an AI chip that goes inside, you know, servers and racks and then, you know, multi-rack systems. And what we've been working on with them kind of secretly for the last couple of months is building out a cloud service. And in that cloud service, we allow developers to utilize the high-speed inference engine that we've built so that they can build new experiences that are super low latency, high throughput.

2:55There's a lot of things that the developer community has done since we launched, probably a year worth of content if we really wanted to do it. I've seen over a thousand demos myself. So we'll pick and choose and we'll bring them on sometimes yeah and as people know inference is when you ask the question you give your prompt to the large language model and gives your result and grok specializes in returning those much faster than they've previously been returned much faster yes like 10x faster like just absurd lightning speed depending on the model it could be like 100x faster amazing yeah yeah so this is not for training the models the h100s from our friends at nvidia those are the state of the art am I correct, for training the models?

3:38Training and inference, you know, like basically anyone that's out there is using those. And then there's some kind of custom Google chips and there's custom chips within Amazon and Microsoft as well. You know, the architecture that Jonathan, the CEO and founder came up with is distinctly different from GPUs. And that's what allows it to operate so quickly. Amazing. Yeah, and that's like a seven-year story. I mean, I remember when Chamath invested in that team coming out of Google. and uh yeah it was people were like way to burn tens of millions of dollars to moth and here we are he got ahead of the curve so congratulations to our best each month and uh the team that yeah look at that wow you're so fast you're doing you describe what you just did i just did tell me the best poker hands this is the newest google model uh gemma 7b so it's a smaller model and you can see this is generated at 750 tokens a second wait wait gemma is who again that's the Google.

4:31It's an open source Google model. Oh, Google. Yeah. Google has an open source model. Correct. That's different than Gemini. That is different than Gemini. They call it Gemma. Huh. Okay. And so, yeah, when you did that, it just immediately gave you, yeah, that speed. Yeah. Yeah. Incredible. You can imagine for certain use cases and many use cases for developers, that speed is really critical because we've seen that across the internet, right? you've probably seen it in your past lives and different things that you've worked on, but definitely on all the companies that come through the incubator, you know, making page times load faster, making the user experience more zippy.

5:09And so we have the ability to do that. Yeah. Amazing. All right. Well, well done. Let's get to our demos for this week. We've got so many great demos to do here. So let's just get started. We'll give our letter grades as always as uncomfortable as that will be for some people. So this week, we're going to focus on a bunch of new models that were released, which are generally scoring as high as GPT-4. And they all have kind of different capabilities. And so we're going to start by reviewing Claude. And Claude 3 has a model called Opus, which is their largest model. And they've done some interesting things.

5:48So let's pull it up here. Obviously, it looks like a standard chatbot. But you can say, you know, what are the best poker hands? And it will obviously go. It's not as fast as what I just showed you there. But its reasoning capabilities are incredible. And so one of the things that the Claude Anthropic team have done is they provided this prompt library. Yes, this is the prompt library. Really good. If you go to the pun one, it's really good. So this is kind of a fun one where it gives you a really nice prompt. They used to be able to find these prompts on Reddit threads or Hacker News threads or people that put up some websites.

6:26But this is great because these are prompts that the team at Claude are suggesting. So I would assume they're optimized a bit for their model. Yeah, you would think so. Right. And you can use them elsewhere as well. I mean, you know, these prompts, there's nothing specific to Anthropic or Claude in this particular prompt. And so what's really nice about them providing this, it gives developers and even end users insights into how to better get the output that you're looking for, depending on the category. And I really like how they made these available here in a lot of different categories. And there's a ton here, right?

7:03From data organization to email extraction to lesson planning to fashion advising. I really think that the team has done a good job. what i'm going to do here though is kind of show you how capable this model is i'm going to get an example here what this one in terms of its reasoning capability it's really really powerful and the example we're going to do is we're just going to take a little screenshot of uh an ikea build instructions oh right frustrating those can be and this was one of the agi tests i talked about it i think on all in oh you did okay yeah you know if you go to um the wikipedia page for agi they have a number of the tests okay that people you know assume it would determine agi now the turing test is the most famous you talk to a computer and you can't tell if it's a computer or a human so you put a human to answer questions you put a you know behind a box and you put a behind a curtain and you put a computer and can the person tell the difference between the two that's one but being able to give somebody ikea instructions and then make something would be you know really hard and somebody had another one which is give the computer a hundred thousand dollars and say uh turn it into a million in 30 days or something those are all examples of agi oh look at this one though it said help me with these instructions and look what we got hit by here well let's see it says i apologize but i cannot provide instructions related to assembling that particular object, it appears to be to depict an unethical, potentially illegal item.

8:37I would suggest focusing your efforts on more. What? He thinks it's a bomb? Positive construction projects that promote harm. What? Yeah. It's a shell. Oh, you got caught by the woke AI virus. I know. What happened here? We got you. Clawed woke AI. Uh-oh. Just going to get an F. Right now, startups have to do more with less. We all know that. It's rough out there, folks. So, If you need great tech talent, but you don't have the time to interview dozens and dozens of candidates, you need to check out Lemon.io. Lemon.io has thousands of on-demand developers to choose from. And these devs are vetted, experienced, result-oriented, and they charge competitive rates.

9:11Great developers can be incredibly hard to find. And when you do find them, it can be hard to integrate them into your team. Lemon.io handles all of that for you. Startups choose Lemon.io because they only offer hand-picked developers with three or more years of experience and strong portfolios. In fact, only 1 % of candidates who apply get in. And if something ever goes wrong, Lemon.io will get you a replacement ASAP. You know what? A bunch of our launch founders have worked with Lemon.io and they've had great experiences, which is always good to hear. Go to Lemon.io slash twist and find your perfect developer or tech team in 48 hours or less.

9:43Go to Lemon.io slash twist and find your perfect developer or even a tech team in 48 hours or less. And Twist listeners get 15 % off their first four weeks. What a deal. Stop burning money, hire developers smarter, visit lemon.io slash twist nobody wants a woke ai come on claude do it again let's try one more time let's see what happens maybe we do it again yeah try one more time sorry everybody looks like claude is getting a little bit of the gemini all of these guys are getting it that's really surprising all right let's try it again let's do a new one you are a helpful assistant help me with the instructions to assemble this uh shelf ah you're letting it know it's a shelf not a shotgun yeah clearly claude can't tell the difference help me build this it's like you you fired off the bomb protocol i wonder yeah maybe yeah maybe it's a little racist maybe claude's racist there it goes okay here we go so the image provides step-by-step instructions for assembling a shelf using provided hardware and panels here's a summary one connect two panels at a right angle using the included hardware as shown in step three make sure the panels are securely fastened step four at a third panel i wonder if this is accurate or not what do you think look somewhat i mean that's the first panel then yeah okay yeah the third one okay yeah because we the image we gave was steps three four and five we didn't do one and two okay got it yeah i was thinking left something out all right i mean i it seems reasonable what we're really seeing here you know despite the um we got caught with a little bit of um uh you know the red team got us the red team got us exactly is and i don't know if you've ever dealt with like assembling these kind of things out of ikea or anything like that where they don't provide like the written instructions no it's just like this is pointing i just did this i made a desk when i was up in tahoe last time because i need a little desk and yeah it was no words and it was like there were three different size screws and i'm like well can you just say smallest screw medium size screw large screw maybe that would be an easy way to do it so infuriating i think what's going to be really interesting is what this really does and you know now we have different models available right you could you could try this with gpt4 you can try it with anthropic now you can try it with with others as well.

12:05It's this ability to basically take any task we're working on, right? Whether we're writing copy, whether we are basically assembling something, or whether we're looking at a menu and basically get the LLM behind this, the generative AI behind this to give us more detailed instructions. And I think we're really impressive by the Anthropic team in terms of the capabilities that they're showing here. You know, and it is a really great brainstorming tool. I've always felt like brainstorming is kind of where we're starting. And I'm using the Claude 3 right now. Let me show you what I just did here.

12:43So I just said, hey, listen, my prompt, I'm hosting an event for LPs and GPs in venture capital. What are some topics we should discuss in our roundtables and keynotes? So, you know, I know what I want to talk about. This is for Angel Summit, which I've now rebranded as Liquidity. You spoke last year. It's in June, June 2nd, 3rd, 4th. I'll be there. I got the invite. You got the invite. Liquiditypod.com to see the podcast and the event. And it said, when hosting an event for limited partners and general partners, yada, yada, here are some potential topics. Number one, fundraising strategies. Number two, deal sourcing and due diligence.

13:18Number three, portfolio managed and value creation. Number four, emerging trends and disruptive technologies. Number five, diversity, equity, and inclusion. Number six, environmental, social, and government. Number seven, talent acquisition and retention. Number eight, limited partner perspective, dedicated session or keynotes, LPs where they can share their investment philosophies, due diligence, process expectations from their GP partners, general partner perspective, networking and collaboration. I mean, this is not bad, you know, it's not a bad place to start. I think this is a great place to start making really great ideas.

13:49And if I had like a junior person on my team and I asked them to do this task, this would be probably similar to what they would come up with. Right. So this is not the most refined bullet points, but, you know, it's a start. doesn't seem to have captured the latest and most important things which i guess is just the training data i'm assuming yeah and like you said it's it's brainstorming you know what what i would say is also really powerful with the capabilities that you know claude has to offer and you're using cloud 3 sonnet i'm using opus i think you have to be on on pro to get the more advanced model which you know i think will uh yeah this is i just said i just asked it for some trends from 2023 and 2024 and it said the last update was August 2023 so I'm on the own one.

14:34So you do get that. One of the things that I wanted to suggest for you to try with this was you can give it your entire book and basically you can ask it to create an update for that book. So building on a little bit of what you and I've been talking about with like the Sopranos or Seinfeld episodes to extend it a little bit. So imagine you give it the entire Angel book and say, hey add some new chapters regarding this yes absolutely i think this this like fill in brainstorming you know adapting you know enhancing is a pretty good starting point for what i'll call year one i think we're going to start referring to the first year of ai as year zero because it was all demos and they they were proof of concepts but i think we're now now we're in the application phase i think you would agree if we're going to spend all this money on ai it's gotta have a payoff it's gotta hit the bottom line like we saw with the clarna ai assistance or you know here can we save money on producers for live events or can we make the content better or make the existing producers go faster right and so here we are we're getting there we're getting there and we start with transcripts we've seen it with summaries of podcasts like it's definitely helping um it's not perfect but it's it's definitely 80 90 percent of where it needs to be.

15:54So I'm, I guess I give it a B. Yeah. You know, I didn't like that. It red tagged you red team, but I'll give it a B. Yeah. And I had to work around it a bit for me. What I, what I'm really enjoying seeing is kind of the following before I give my grade. It's really nice to see other teams coming up with capable models for the following reasons. What this means is for the broader community, that these things will become more affordable. they'll become faster and they'll become open source. Because of competition. Yeah, because of competition. And what it means is there's not one team that has some secretive know-how.

16:34And this is good for humanity, right? Because I think, if you think about open AI, if they're the only ones that were innovating, then the argument would be, especially in some ways how they become closed, that's not great for all of humanity. and seeing other teams have very similar capabilities, score high in the benchmarks, I think it's really, really powerful for both the developer ecosystem and society to see that these models are becoming comparable in their capabilities. It is a great thing. I have to say that, you know, ChatGPT 3.5, 4.0, 5.0 is not running away with it. And you and I both still contend that open source will win the day you still believe that 100 yeah i'm still 100 with you and i saw that the other grok elon's grok yes at twitter which is g-r-o-k hey g-r-o-k okay a little confusing uh but i guess we can live with it your q he's k his grok is he announced it's going to be open source just this morning just this morning he made that announcement on monday march 11th he made this announcement that he's going to open source it and then apple we talked about they're doing open source and facebook's doing open source google's got this open source side project but is it like the red-haired uh stepchild or something is it like uh this goes to show what's happening in the industry the their open source models and you know they're not sharing numbers sort of in terms of usage of their larger ones but their open source models are have become really really popular.

18:16And so even when we went and launched Gemma over the weekend, it quickly became one of our more popular models, both across the API and the chat. Interesting. Yeah. I think that's great when that happens, because when the model goes open source, you also get the ability for multiple providers to run it. So in the case of Cloud Opus, only they can run it, right? So the pricing is, and same with GPT-4. But when you get these models that are open, you see them available on multiple API providers, which I think is great for developers. And then they'll compete on price. And when something like that red team moment we just had where I thought you were making a bomb, I wonder if I asked it if it would think it was making a bomb where it's just doing that.

18:56Because it's like, ah, Sandeep Madra. Yeah, yeah. What is this guy working on? What is this guy working on? Yeah. I would also like, if we're going to be canceled anyway this episode, red-haired stepchild is a colloquialism in the United States for an unwanted child. If you have red hair, I didn't come up with this. I don't know. Maybe that has to be retired too. I guess we have to retire red hair, Chip style. Juggling multiple devices and apps to run your business is a mess. Open Phone is here to make it simple by simplifying your business communications with one easy to use app. Open Phone has rethought every detail of what a modern business phone should be.

19:35And here's the magic. It works through a beautiful, elegant app on your phone, or you can just use it on your desktop, making it super easy to get a business phone number for your entire team. And you know how brilliant OpenPhone is? My teams use it every single day. My sales team loves it. My ops team, they use it all day long. And here's the features that we love. You can create a shared phone number like customer support with multiple employees fielding all the calls and all the texts to that one number. At my investment firm launch, we pride ourselves on replying to every single call or email instantly and open phone is the number one rated business phone on g2 for customer satisfaction so here's your call to action super easy open phone is already affordable starts at just 13 bucks a month but twist listeners get an extra 20 off any plan for the first six months at open phone.com twist and if you have existing numbers with other services no problem open phone is going to port them over easy peasy lemon squeezy no extra cost head over to open phone.com twist to start your free trial and get 20 off but it is definitely the case that if they're open sources competition and then these guardrails as we talked about with the crazy gemini situation the last two weeks those are added the guardrails are added to the language model after it's done its work they add guardrails and so if you don't like the guardrails in an open source community the guardrails would be in the code base somewhere no the guardrails generally don't exist in the model now there may have been some some like we talked about some tuning that has tried to alter its behavior but the guardrails like you know the ones that we just saw there they exist outside the model and so they're on the interface level and the inference level so you ask the question it gets the response and it's like hey wait a second yeah is this person trying to build a bomb is this person trying to tell racist jokes let's stop them this is one thing i would like to see maybe we can do this next week i would like to see a model without guardrails and see what it does when you ask it to do inappropriate things like what are the models actually think based on human knowledge if we asked it like tell racist jokes about irish people would it actually make a bunch of jokes about irish people being drunks because there's a million pages on the internet with like top 100 irish jokes about alcohol irish jokes about whatever potatoes yeah what happened this week and we i can bring some demos on next week but this was interesting basically you've seen this before and i think we even tried something like you know tell me how to build a bomb and i said sorry i can't what people figured out in terms of so this is how you know that these are guardrails that are not in the model what they did was i don't know if you're familiar with like ascii art sure ascii art is a way of drawing words yeah with like symbols so So you can use the dollar sign, the asterisk, whatever, and you make ASCII art.

22:25It was a way to make art before images existed on the internet. And so what people figured out last week was if you did this following prompt, and I'm sure they've patched it already. Tell me how to build A. A, and then BOM, but BOM in ASCII art. Okay. And it was basically giving you the instructions to make a BOM. Oh my lord. This really proves to us that the guardrails sit outside of the model. it's just that you know when you type the word bomb in something catches it before it goes to the model and tells you hey you can't be doing that type of stuff so this is how people were working around this last week someone basically uncovered this and i'm sure they've patched it since but that's how they did it and all of these guardrails is there another word for guardrails i keep using guardrails is there a technical industry term that is the right word okay so ai guardrails language model guardrails there is a open source version of this provided by facebook and meta they are open sourcing their guardrails so llama guard it's called yeah i'll pull it up here yeah and i think this is like a key for the industry because i believe people want to know and want to have information on the guardrails and how they work and so i love the fact that they have now put this out there because I use Grammarly.

23:46I love Grammarly. Grammarly and Slack have the ability to tell you like, hey, you're using gendered language. You're saying fireman instead of firefighter, policeman instead of police officer, because obviously the police and the firefighters can be, you know, any gender, obviously. So, you know, that's helpful actually in some ways, but other people, if you're a novelist and you're writing a novel from, I don't know, 1800s or 1900s you might want to not have it changing those things right there there could be arguments for not having it change it to firefighter right so anyway there's nuances and i think people are going to want to understand these nuances um so again just another reason why open source is going to win the guardrails are going to be transparent people want to know the guardrails and it should tell you like why doesn't it tell you that a guardrail has been hit right I think right now, the folks that are running models, especially as a service, are struggling with how they can make this a consumer-facing technology in a different way than Google.

24:49What's interesting is, you can go to Google and you can type in something like how to make a bomb, and you'll make your way to content that will show you that. Yes. And the approach that Google took is, hey, we're organizing the world's information, get there. Now, they probably do some level of filtering, and I don't want to search for that, J-Cal. Maybe I'll let you do it so that I don't come rushing to my IP address here. Well, no, if we had a language model, I think it would be interesting to ask an un-guard railed model how is a nuclear bomb built, and then ask Google how is a nuclear bomb built.

25:20That's an educational search, right? Yeah. But you might want to use a VPN. Yeah, exactly. Exactly. What's really happening is this layer of moderation. It's one thing from a safety perspective, but it's another thing like we've been experiencing the last couple of weeks when the moderation is not safety, but it's sort of opinions on how we should think and how society has changed. Ideological. Ideological. Good word. Yeah. Okay. What do you give the new Claude? I'm going to give them an A minus because I didn't like what happened with the thing. I would to give them an A before. What I really was impressed by is, like I said, I think it's great that we have multiple models that are at the capability of GPT-4.

26:05And I think that basically means that there's multiple teams out there that know how to create the technology, know how to, you know, basically what data to train on. And, you know, it's harder and harder because more and more sites have cut off training information to folks. And so it's really great to see that these different teams are getting there with those type of capabilities if you want to use reddit twitter or quora or stack overflow you got to write a check so the open source models are not going to have access to that anymore in all likelihood they will not unless it's a rogue one done in a foreign country where they maybe don't respect copyright law uh like say china right so the The Chinese could just, the Chinese have a full copy of the websites I just mentioned, archived and scraped.

26:54Yeah. So they'll build their model. So they'll be at a slight advantage because they will not respect copyright, just like they don't right now when they rip DVDs and sell them on the streets. It's an interesting thought, though. And so one would argue, then, the models they'll create could be more powerful. Yeah. Yeah. That's the arms race we're in right now is that. Okay. Yeah. If you break the rules, you could have a better product. Really fascinating. How do you think about that? If you had a company come in, because, you know, there's some great models that are coming out of China, either out of the big companies or even some small ones, like DeepSeek is a good one out of a small company.

27:33There's Quinn, which I believe is out of, you know, one of the big internet companies there. How would you think about that? I would think about it in the same way the United States thinks about trade, which is, you know if we are going to have a trade relationship with china we need to set some ground rules with them there are things that are important to us intellectual property is one of them so if they want access to our market and we want access to their market they have to you know appreciate the the way we do business which is we protect intellectual property and they're going to need to evolve their legal system and just how their marketplace works over time to protect our ip and if they're not willing to do that then maybe you know iphones will be made in india and if in india they respect microsoft's you know like when i was in china 15 years ago you could just walk down the street and they had every microsoft product i'm talking about like server products you know really expensive products just available on dvd i remember and you could just buy them in shenzhen or shanghai on the street i mean i'm talking about things that cost$10 ,000,$100 ,000, like Tableau level stuff.

28:41And it's all been cracked and it's available. And it was kind of funny. I bought a couple of CDs there just to give them to the Engadget folks and be like, hey, look, you can just buy this stuff. But yeah, nobody pays for Windows as an example in China. Windows is basically free there. And I think that's why a lot of subscriptions are happening, but it's a give and take, right? Yeah. All right. You guys know I like to keep things fresh here on This Week in Startups, get you some great content that's worth your time. Today, I want to tell you about a great new podcast I know you're going to love.

29:09It's called Marketing Against the Grain, and it is hosted by Kip Bodnar, the CMO of HubSpot, and Kieran Flanagan, the CMO of Zapier, which makes you happier. I love both of those products, and these guys are top-notch marketers, and their debates are packed with powerful insights. They're going to bring you the latest in marketing, not just the usual stuff, but the real actionable, tactical strategies. strategies, this podcast has deep dives on all the marketing techniques you're going to need to know in order to grow your business. And you're going to grow your career. And you might even grow your wealth.

29:43And I've been checking out the show. Here's the deal. Episode predicting the future of marketing with AI co-pilots. This is a game changer, something you need to know. In this episode, you're going to hear how the AI co-pilot movement is revolutionizing marketing. And you know that that's happening, right? Everybody said it. But now you're going to learn the tactics of how it's actually going down because there is a seismic shift coming and you need to be prepared. That's just in that one episode. So there's plenty more packed with lessons that'll keep you ahead of the curve. I want you to pause this podcast and go do a search marketing against the grain.

30:13Subscribe in your favorite podcast app, whatever that is. If you're Spotify, you're Apple, you're Overcast, whatever it is. Once again, give it a search marketing against the grain and learn all these new marketing techniques. It's going to be worth your time. What's sort of interesting that you say that windows thing, because I don't know if you saw this and maybe it was like a week ago elon got a new laptop and then he was complaining that in order to use a laptop he had to connect to the internet um and this is like the double-edged sword right because you know for microsoft they want to eliminate piracy and they probably want the laptop to connect to you know go and verify that the key is not shared or whatever it happens to be there is a workaround still for you to basically yeah it was like hard it was like 20 steps it's the same thing with ios right like if you want to use an iphone they're like put your credit card in if you don't put your credit card in it's like really hard to not turn on siri to not put your credit card in like they use kind of dark patterns or gray patterns and i don't know if you can use an ios device without you know it having an email address you know um associated with it and so i guess you got to create a burner account or something but yeah i think this is part of the standoff between the united states and china is over these nuances of capitalism and it does mean i think in the short term they'll have better models potentially and i i think this is going to prove the point that open source is going to win the day because open source is just going to be able to move faster and people are going to be able to point it at what they want so like while chad gpt4 and claude and gemini are trying to get their licenses lined up you know somebody could put up a rogue language model in israel in north korea in you know people have different like in israel i think they have a different scraping law india might pakistan might they just have different rules around can you scrape a website here we're pretty protectionist about that and maybe they say yeah you know you can scrape all of crunchbase you can scrape all of linkedin and so if you're looking for an angel list of whoever's got information like on you know profile pages of individuals you could just hit a language model i've been looking for a language model where i can just query people and be like hey who are the ctos of the top 500 companies whatever you can't do that like that database is just not in a language model because no what's sitting behind exactly right it's on zoom info it's on linkedin so i mean i would like to see a linkedin language model come up or if i had a premium linkedin account and i paid 300 a year for linkedin it should be connected to my chat gpt4 and we had this discussion imagine you authenticate your chat GPT-4 with your Spotify, New York Times, and LinkedIn subscriptions.

32:49So when I do a search, I'm like, hey, what are some musicians like Dire Straits? It kind of goes and does that. Or take the song Born to Run by Bruce Springsteen and tell me other songs that have a similar lyrical devices in it. And it goes and finds Telegraph Road by Dire Straits, right? And it's very interesting things where Spotify might not build a language model, or LinkedIn might not build a language model anytime soon, but man, LinkedIn plus language models would be the greatest product ever for business. Can you imagine? I'd pay$1 ,000 a year for that. Yeah, because you would just basically, like, because their search is really bad.

33:26I mean, advanced search, if you know how to use it on LinkedIn is phenomenal. Okay. But you have to have the paid version, which I do, which I'm happy to pay for, but it's not a natural language search. And I think that what happens is the reason why you might perceive it as bad is either you don't have the advanced version or yeah well i pay 600 bucks a year yeah yeah which is totally worth it like two bucks a day so but here's the thing there's so it's a billion people there now so you also have this issue of like okay i just want english-speaking world ctos of companies with over a hundred engineers like that might be because i want to invite them to grok right you know to some or you can do a grok hackathon right like how would you get you should be able to just say, I make a product called Grok.

34:11Here's our homepage. Yeah. Contacts window. Who should I invite to a hackathon? And then it should go through LinkedIn and just find people who've been to other hackathons, who are in AI, who are developers, who are in the Bay Area or within 50 miles, or have been to another hackathon and put their hackathon on the page. You should be able to figure that out, right? Like that kind of a query. I bet you will have that in under, definitely under six months and under three months. Okay, here we go. We got another bet. Okay, everybody, for my team, thisweekandstartups.com slash bets, you say you'll have a natural language query like that.

34:46Well, I don't want to make this bet with you because someone from LinkedIn did reach out to us. Oh, no, this is not the public markets. You can trade on inside information. Did you see that the streaker at the Super Bowl? Yes. There was a claim that he put a huge bet that there would be a streaker at the Super Bowl, and then he streaked. Does that negate the bet? That seems like it should negate the bet. I don't know what the terms of use are, right? In terms of the betting. I'll tell you what. Yeah. You streak, I'll place the bet. Exactly. And if I happen to hand you a brick, I hand you a brick.

35:23Yeah. We'll chop it. Chop it up. All right. What do we got next? It's kind of very similar, but I want to give everyone their credit. So the team at Mistral, this is a really interesting team based out in France. And they released their model called Mistral Large. They have a few different ones here that you can use, like large, next, small. And this one is excellent. So what's interesting about the team at Mistral is, I would say from an open source perspective, they have a model called Mixtral, M-I-X-T-R-A-L. And it's basically, it's a mixture of experts. it's seven and eight billion parameter models that work together and in the open source this is the most popular model that we see right now because it has capabilities that are pretty wide ranging um because it has those you know seven experts and each of those experts is like a you know reasonably sized model and this one can run on you know like not at a really high rate of speed but you can run it on reasonable hardware as well you can run it on your laptop if you really want to.