AI News: Haiku-5.5, Visual ChatGPT, Grok's New Trick, and more!

0:00So Anthropic announced a new model this week. Open AI's been on a spree of multiple new announcements a day. We got some new stuff out of Grok, some new stuff out of Google. Lots of stuff to cover today, so let's get straight into it. I'll start here, but we won't go too deep into this. Anthropic released a new model in Claude Haiku 5.5. This model is meant to be fast and cheap, but not really like the smartest model in the world. We can see right here they describe it as designed for high-volume, cost-sensitive tasks. It reliably handles quick and repetitive workloads like summaries, compactions, database

0:32queries, and classification requests. They also recommend using it as a sub-agent on coding work. They did post up some benchmarks, but as we can quickly see, it is definitely not state-of-the-art on anything, really. So knowledge work, you can see it's 1620 compared to Sonnet's 1840. Obviously, Opus is even higher than that. Knowledge work here, 1578 versus 1824. Computer

0:57use, 72 to Sonnet's 83.9. And just across the board, there's pretty much nothing that this is going to out-benchmark its Sonnet or Opus models on. The main selling point, once again, of Haiku is it is inexpensive. Haiku here on extra high, for instance, is about as smart as Sonnet 5.5 on medium. But as we can see, Sonnet 5.5 is roughly 93 cents, where Haiku 5.5 extra high is 28 cents. But if you're using Sonnet on one of its higher thinking levels than medium, well then it quickly becomes much smarter, but also a bit more expensive. And you're going to see pretty much the same kind of things

1:38across all of these benchmarks here. Haiku 5.5 costs about 10 cents per million output tokens up to the first 100,000 tokens, then 50 cents per million, and then output you've got 50 cents up to the first 100,000, then 250 beyond that. You can that to the previous version of Haiku. It's cheaper than the previous version of Haiku by quite a bit and also, you know, a fourth of the price of using Sonnet. So, if you need fast and cheap, but not necessarily the most intelligent model ever, well, Claude Haiku 5.5 sort of falls in line with, you know, like a Luna model. If we look over at Artificial Analysis, which

2:17takes into account a whole bunch of benchmarks to try to find like a sort of objectively smartest model here, we can see that Anthropic owns the top three models in terms of intelligence here, but Haiku 5.5 falls all the way down here below Kimmy K3, below GLM 5.3, below Grok 4.7. So, you know, it's not the smartest model. But, when we come look at cost per task down here, the previous Anthropic models, Sonnet, Opus, and Fable are way up here as like the most expensive per task models. Haiku 5.5 falls way down here at roughly $0.21 average cost per task. So, very, very, very inexpensive to run while still

3:00being a decently smart model. However, just like other Anthropic models, it does use a lot of tokens. So, we can see here Claude Haiku 5.5 is way up here using the second most amount of tokens per task, only behind Sonnet 5.5. But, because the tokens are so cheap, it doesn't seem to matter cuz we're still seeing that cost per task of about $0.21. Oh, and if you're curious how Haiku 5.5 performed on Bussy Bench, here's what we got. It did cost us less than half a cent, generated in 42 seconds, and used less

3:34than 10,000 tokens, but not very bussy. If you have a Claude plan, you could jump into your Claude account and set up Haiku 5.5 right now. Set up your effort level, medium is the recommended effort level, but if you're on one of the paid Claude accounts like a $20 a month or up plan, you're probably not going to be using Haiku for most things. You're probably going to use one of the smarter models. Where people are really going to be thinking about using something like Haiku is via the API. So, if you're building a tool that runs a model or if you're using Claude code and you're

4:06doing some real basic stuff in Claude code using the API, you'll probably select Haiku 5.5 to get those speed and cost efficiency gains. But, if you're just chatting with Claude and you want a decently smart model, you'll probably pick one of these larger models anyway since you're paying a monthly fee and you're not really worried about API cost. But, in my opinion, there is actually something even more interesting released by Anthropic this week. And that's the ability to build live dashboards and to animate explainers with Claude. Now, most of us can actually build live dashboards right now. The animate explainers with Claude is probably not available to most of us

4:42just yet, which is very unfortunate and kind of annoying. We can see here that Claude can turn your data into a dashboard that stays current and your ideas into animated explainers. So, for the dashboard, it connects to data platforms like BigQuery, Databricks, or Snowflake or a CRM tool like Salesforce. And you can use it pretty much on any paid plan. However, Claude motion, which is the thing that I think is even cooler, is only on team or enterprise plans. So, this dashboard thing here that connects to your CRM is pretty cool, but I feel like we've already been able to do that, right? Like, they just

5:15made it easier to do. They made it sort of less steps in between, but we've sort of always been able to make dashboards. It's just sort of more built in now. And I guess you can argue we've been able to do these motion animations as well. You can have it code in JavaScript and make these animations for you like what

5:31you're seeing on the screen right now. HyperFrames and Remotion were able to do stuff like that. You connect it to Claude and it would build stuff out like that. But, again, they're just making this a built-in. So, I don't believe you actually need to use any of those extra add-ons to get the animations to work. But again, the Claude motion is in beta in team and enterprise and I don't have a team or enterprise plan. Personally, I find this a bit annoying because [laughter] I do have a max plan. I'm actually on a $200 a month max plan and one of the things they sell you on the

6:02max plan is early access to advanced Claude features. So I'm paying extra cuz I want to get early access when this stuff comes out and when something like the new Claude motion comes out for whatever reason, they didn't give it to the people paying $100 a month plus that signed up for early access to advanced features, but they are giving it to people on a team plan that might be paying as low as $200 a month. Like kind of frustrating. But I guess at the end of the day we can still accomplish this stuff. You can tell it to go and create an animation using, you know, JavaScript

6:35and it will. You can use things like hyperframes and remotion, install those skills and it will also do these kinds of animations. It's just nice to have it sort of built in without those extra plugins. So in terms of whether or not we're able to do something that we couldn't used to do, well, we still can do that, but I am looking forward to like cutting out some steps and making it even quicker and easier and just like native to Claude. Before we get back into the video, I want to take a second to show you guys something really cool that I built recently. You might have

7:05seen my other videos where I bring creators onto my channel to break down the like interesting things that they're building with AI and believe it or not, there's a surprising amount of prep that goes into those conversations because I really need to understand like what they built and then make sure we touch upon like all of the coolest parts of what they built. So I'm using hyper agent to build a little like guest prep team that my actual team can use whenever we're planning for an interview. Now I created

7:32three agents with three different jobs. I've got scout, echo and vox. So first is scout, my researcher. All I have to do is give it the name of the guest and it'll find their website, social channels, and anything else I need for the prep. Here's the prompt that I used to build Scout. Build me an agent named Scout that acts as a guest researcher for my YouTube series where I feature other creators and what they've built

7:56using AI. Next is Echo, my interviewer. He takes that research and comes up with questions. Here's the prompt that I used to create Echo. Build me an agent named Echo that acts as my interview strategist and then we had our Echo agent. I'm also adding an example of a previous interview so it has reference for the kind of conversation that I want. Then the third agent is Vox, my producer. Build me an agent named Vox that acts as my interview editor and producer. Its job is to turn that material into a video rundown. Each of these agents has a specific role and boundaries. And for this case, they

8:30didn't need access to my email or calendar or any of those sorts of tools, so I'm leaving those disconnected. It's because as you build more agents, you can give each one only the tools and access that its job actually requires. Let's go ahead and create a new thread. Prepare an interview for Riley Brown. Research the guest, develop eight questions, and create a 30-minute

8:49rundown with three suggested demos. Let's go ahead and tag our three agents, @scout, @echo, and @vox. And we can see it's going to do it in that order. It created a plan that I can review here and now I can switch it to execute mode. All right, and now I can review the background, review the rundown, check out the suggested demos. And again, we've got all of the sources down here as well. And here's the part that makes this useful beyond just me. Everyone on my team can use the same agents. And with their new rooms feature, I can put my teammates and all three agents into

9:21the same workspace so the agents can work together and hand things off without me constantly playing middleman. We're all working with the same agents and the same instructions instead of everybody maintaining their own separate setup in like Claude or ChatGPT. That's Hyper Agent's whole idea. Many agents, many humans. A team of specialized agents that your entire human team can then work with. If you want to build something like this, check out the link in the description and sign up for a

9:47paid plan to get $100 in bonus credits. And thank you so much to Hyper Agent for sponsoring this portion of today's video. All right, let's get back to the rest of the video. OpenAI made a ton of announcements this week. And they even told us they were going to make a ton of announcements this week. Teebo took to X early in the week to say, "Over the next 28 days, each day will either ship one thing that is a clear improvement and relevant for most Codex/Work users, or ship a full reset." So, this was on October 4th. I'm actually recording on October 8th. So, we've had like four

10:19days of updates, and they've actually rolled out more than like one update a day. I'd say the most significant update was this one for GPT-6 and intelligent UI for everyone. This is sort of two announcements in one. One of them being that GPT-6 is now available inside but the other one being that the UI is getting a lot of improvements. It's actually much more visual. You can see in their video here, they're showing people looking at instruction manuals that are just text, and looking at color guides that are just text, and trying to prove the point that that's not very helpful. It's it's just text. It then

10:53goes on to show that the ChatGPT responses are going to be much more visual. So, when you prompt something in ChatGPT, you're going to actually get a lot more to look at. Here's one example. Break down the design of a seven-speed bike. The chat starts to respond, but then it shows an image of a bike, and then gives you a sort of exploding

11:12version of it that you can take apart. You can highlight certain areas, and it just gets much more visual and easy to use as opposed to just like a pure text description. We've trained GPT-6 to compose responses using text, visuals, and interactive elements, choosing how they fit together based on your questions. With intelligent UI, responses can include graphics, tappable buttons, forms, charts, and interactive experiences you could use directly in your conversation. And then we've got some other side-by-side examples down here. The one on the left using GPT 5.6 instant, and the one on the right using GPT 6 instant, and we can see on the left we just get sort of cooking

11:47instructions, on the right we've got images and bullets and very, very visual version of it. Here's another one building a wardrobe, text on the left with some images, GPT 6 instant on the right a lot more visual, a lot more of like a guided experience with checklists and stuff like that. And it'll build things like little calculators for you and stuff like that when you ask it questions. It's just trying to be more proactive as opposed to just purely giving you text most of the time. You also are getting this smarter, faster GPT 6 model. We can see here on this little graph here, left being sort of

12:22the intelligence level, right being the speed. You're getting pretty much the smartest model at these fastest speeds now. So, an intelligence level that used to take 215 seconds, you're getting even smarter up here at 109 seconds. So, basically when you're in ChatGPT, you'll notice that if you're on GPT 6, you could set it on instant for thinking effort, and I can say something like explain how an internal combustion engine works, and we can see that not only does it explain to me, it actually gives me images along with it, gives me sort of like step-by-step images, and it's just a much more visual experience,

13:01and that was almost an instant response. Even if my editing might have chopped it up a little bit, it was near instant with how quickly this populated here. But again, the features kept rolling out. I'd argue that the majority of the features they announced after that have been a little less significant and a little more focused towards coders, but I am going to continue to break down everything else they announced in the last, you know, four days of this

13:26rollout blitz thing they got going on. So, on October 5th, Tibo announced we've optimized the default speed to be 50% faster across GPT-6 Astra and GPT-6.1 Soul through the subscription across all of our products and partners using sign in with ChatGPT. So, yeah, exactly what it sounds like, everything kind of got faster. They just made a big optimization improvement on their best available models right now. Day two, October 6th, they made the auto review free for all users signed in through a ChatGPT account. So, what this is referring to is if you go to your ChatGPT app and you're either in work mode or you're in Codex, you can see

14:03down here there's this permissions drop down. By default, it's going to ask for approval. So, whenever it's going to edit something on your computer, it will ask for that approval first pretty much every single time. And your options before were pretty much ask for approval, approve for me, or full access. Full access was pretty much the only way to make it so that it just never bothered you at all. It would just kind of do the things for you. Well, now this approve for me mode, I guess is a lot better at approving for you where

14:31it's going to ask you a lot less often. So, that's essentially what it did. Is it made this approve for me mode basically bug you less unless it had a really big concern about what it was about to do. So, that's what this update was for day 2.1. We also got a new ChatGPT plugin called the meetings plugin, which if you ask me, pretty much looks exactly like what Granola already does, but they just baked it into ChatGPT. So, the meetings plugin takes notes for you and saves a personalized summary and next steps in ChatGPT space based on what ChatGPT knows about you and the work you've done together. You

15:06can keep notes private or share them with your team, then ask ChatGPT to update a project plan or draft a follow-up. So, from my understanding, it's essentially listening to your microphone audio and your computer audio whenever you're on, you know, a Zoom meeting or a Google meeting or a Slack meeting or whatever you do your meetings in, it will listen to the audio on both sides, record, transcribe them, keep notes on it, and then save that into the memory and context in ChatGPT, so you can ask questions about meetings that you'd previously been on, which is exactly the selling point of what Granola is, but Granola is purpose-built

15:41for that. But now you've got a plugin where ChatGPT will also do that. And then we also got the Decisions API this week. So, Decisions API was something they announced during their Dev Day last week, which is essentially Jeff. This is OpenAI's version of Jeff. Instead of responding with a ton of output text for you, it's going to respond with a predefined decision. So, you can get it to output three different kinds of

16:07output: predicates, choices, and scores. Predicates estimate the probability that a statement is true, choices select from predefined options with a confidence score, and scores evaluates an input against a numeric range. These are designed to be ultra-fast, ultra-cheap, and to just give quick decisions as opposed to responding with a giant, you know, text output. As far as I know, there's no real way to use this inside of like a ChatGPT, but if you're a

16:34developer, this is available in API now. And it uses GPT-6 Luna, which is their cheapest, fastest model, so it's going to cost next to nothing to use, also kind of like how Jeff works. While we're talking about APIs, they've changed the pricing a little bit around the APIs. Instead of five different pricing tiers, there's three. There's a build, launch, and grow. As you can see, these various tiers have different rate limits, so the higher the tier you're on, the less you're going to get rate limited. They also rolled out ultra-fast mode in GPT-6.1 Soul inside the API, Codex, and ChatGPT work. Now, this one I believe

17:10you have to be on their $500 a month plan to use, and it's also quite a bit more expensive. If you're using GPT 6.1 Soul Standard, it's only $2 per million input tokens and $10 per million output tokens. If you bump it to fast, that price doubles, but if you bump it to ultra fast, well, it's six times the price as using it standard. So, it's the same level of intelligence as you get out of GPT 6.1 Soul, but it's eight times faster and six times more expensive. But, if speed is a huge priority to you, well, then maybe this one is worth it. Personally, I'm

17:45thinking about trying it out. Again, you have to be on their Pro 500 plan, but I want to test it for like video editing because theoretically, I can give it computer use and it'll just jump around my screen and do the edit like ultra fast. So, I do want to try it for that. I just haven't, you know, bit the bullet and jumped to that price tier yet in order to try it. Maybe that's something I'll play around with next week. And the last update that we got at the time of this recording that came out on day four

18:11is they improved steering to be instant. So, if I'm in chat GPT work and I say make me an image of a dog, and then I change my mind and go, "No, make it a wolf." And I hit enter, you can see this queues it up to be the next prompt, but I can press this steer button, and if I press steer, then it sort of interrupts

18:30it and steers it in a new direction. Well, they made that faster. They didn't make the image generation that we're looking at faster. They made the ability to make it change directions faster. They also claim that if you use 6.1 Soul ultra fast, they work together very well. You just got to be fast at steering it before the output gets to you. And there's probably a new update today as you're watching this video if you're watching it on Friday because, well, I recorded this on Thursday and I don't know Friday's update yet. We also got some updates out of a model I don't talk about very often, and that's Grok,

19:04which their Grok bot like AI agent assistant thing is honestly pretty good. You know, it's interesting because whenever I talk about like Grok or SpaceX AI or literally anything that [laughter] Elon is involved with, the majority of the comment section is people just saying how much they hate Elon and I'm not going to debate whether or not Elon has said or done some dumb things. I just think so many people are willing to throw the baby out with the bathwater and not actually use tools that could be very beneficial and very like useful to them because of their feelings towards one person despite the fact that these are giant companies with

19:43like, you know, thousands of people whose livelihoods are actually like on the line working in these companies. But all that to say, Grok bot is actually really good and I do recommend trying it. I understand not liking Elon, but also a lot of these companies and a lot of these products are about more than just like the one person at the top of a pyramid. There's like thousands of people at these companies that are all trying to do good work and build cool things despite the public image of the person at the top. Anyway, I'm done ranting about this. Let's take a look at what Grok bot just updated. We can see

20:18here that they announced this week that going forward, Grok bot's going to use the best back end model for any given task. So it will actually use models like Opus 5.5, Midjourney, Suno and other leading APIs. It's not going to always default to just a Grok model. It's going to use whatever model is

20:38likely to give you the best outcome. Now, given Elon's ongoing beef with Sam Altman, I don't know if OpenAI models are included. If OpenAI is the best possible model, will that also be used in Grok? Still to be determined because right now the best available models are coming out of Anthropic and so like who knows. But when OpenAI like inevitably jumps over them again, I'm really curious [laughter] to see if Elon will make OpenAI models available inside the bot, too. Who knows? Probably not. But, we also got this update for Grok bot in that Grok bot can now search, read, and monitor X, which is pretty cool cuz X

21:16has a lot of data, a lot of information, a lot of real-time stuff being fed into it. And now you can have Grok bot sort of keep an eye on that for you. You can have it watch for trends. You can have it alert you when maybe something you posted is starting to go viral. It can suggest post ideas based on what's

21:34trending in your own niche right now. Things like that could be really, really helpful in Grok bot. Now, if you've never used Grok bot and you're wondering how it's different from like Muse and Dots and things like that, you have different bots that are sort of designed for different focuses. Like for example, I have an email triage bot which is connected to my Gmail that's constantly looking for emails that it can help me reply to automatically. I have a research bot which is constantly pulling in news and researching things that I could make content about. But, you can also build like the chief of staff,

22:06which could be your single primary bot. So, if you like what Dot or Muse does, you just use your chief of staff and your chief of staff will then go and delegate to these other bots for you. So, you can tell it, "Hey, I need research about something on X and I need you to go do some email triage for me." And your chief of staff will then go and direct your X bot and your email triage bot to go do those things. So, that gives you a lot of options that maybe you wouldn't get with something like Muse or Dots. You have the option to

22:37just chat with your one bot that will then direct things on your behalf or you can go directly to the bots and have them go and do their specific tasks. And over time, you could build out more and more bots that are good at different things and then tell your chief of staff what they're good at and then that chief of staff becomes that orchestrator for you. I think it's a really, really cool way of doing things where Muse and Dots have more taken the approach of one

23:02single chat that you focus inside of. And now that Grok bot can access X directly without us needing to like connect via APIs or anything, I can have this bot like my X bot here, which you can see went through my last 75 posts, found my best AI posts, my worst AI posts, noticed that my post performed the best between 10:00 a.m. and 1:00 p.m. Pacific, and then it even suggested the strongest ideas right now for things that I can post about. Gemini 4 Argon, hands-on GPT-6 is now free for everyone, OpenAI's 722 AI written math papers, Codex usage resets, humanoid robots fight night in Singapore, OpenAI revenue

23:42sell-off. These are all things that people are talking about on Twitter right now. And I'm just finding out that this thinks Gemini 4 Argon is supposed to roll out today. Let's see where that comes from. Yeah, it looks like this was pretty much speculation. This was actually yesterday on the 7th. I'm recording this on the 8th. People are saying happy Gemini 4 Argon day, Gemini 4 Argon coming today. People were saying Gemini 4 Argon is now showing up in Google Cloud. Supposed to be a good model. I talked about it last week, but I don't think it's actually available to most people yet. I think it's still

24:13available only to like early testers. Anyway, all that being said, check out Grok bot if you haven't already. It's actually pretty good. Now, those are the biggest stories of the week. We did actually get a handful of new large language models that were rolled out this week and a bunch of other smaller updates that I want to break down for you, but I'm not going to go too deep into any of these, so let's just jump

24:32into a rapid fire. [music] Let's start with the newest model out of Mistral. This new model is called Mistral Large 4 and is nicknamed Le Chonk. Now, this model is actually an open weight model. At 1 trillion parameters, you're probably not going to be running it on any of your like local systems, but it is really solid model. On all the various benchmarks, they're comparing it with the other, you know, Chinese open-weight models that have been really good, models like Kimmy K3, GLM 5.3, DeepSeek V4, etc. So, think of it as a model that's not from China, but roughly on par with all of those Chinese

25:11models. And also open weight like a lot of these Chinese models. And because I know you're curious, this is what it looks like on Busy Bench here. It took about 2 and 1/2 minutes to generate and used 9,675 tokens and scored sort of middle of the pack on the AI judging. We got another open-weight model this time out of a US-based company called Reflection, and this model's called Beam, which is a 501 billion parameter open-weight model. When looking at the benchmarks, this one they're also comparing to similar models like Kimmy K3, GLM 5.3, DeepSeek V4.1, NeMoTron 3 Ultra, and while it kind of falls in line with a lot of those, if not sort of

25:53underperforming them a little bit on agentic coding. On reasoning, falls in line with a lot of those same kinds of models. So, another open-weight model that is kind of on par and close to what we're getting out of the Chinese models that everybody thinks are really, really good. Now, this one is not publicly available yet. It was more like announced and I believe only access to like early testers, so I don't even have it Busy Bench'd yet, but we do have another open-weight model from a country that's not China, but also is going to be so big that most personal computers still aren't going to like be able to

26:29run it. But if you're like an enterprise company that wants to build out your own servers, run open weight and keep everything on prem, that's what a lot of these models I think are being designed for. Google rolled out a new game builder this week called Playground, where you can create and play custom games. You pretty much describe your game into a prompt box. The builder goes off to work and just sort of designs the game. It looks like it will work on both desktop games and mobile games. So, yeah. I mean, you could already kind of do this sort of stuff with all the existing AI coding apps, but this one is

27:03purposely designed to build games. It's available right now if you're in the US and over 18, and you should be able to try it over at playground.google. There's also games already available in here that other creators have made. If you don't feel like making your own game, you could jump in and try other people's games that they've already built. For example, Skyfall Defender,

27:24which is actually strangely addicting. But, I'll get out of this now. While we're talking about Google, and since I mentioned Granola and how OpenAI made like their own version of Granola, apparently Google made their own version of Granola also. So, Google now has an AI note-taking app which transcribes your meetings completely offline. So, unlike the OpenAI version, unlike Granola, it will actually record your meetings and take notes, but it'll do it all on your computer. So, nothing is going to another cloud where your meeting notes can be trained on or potentially broken into or anything like that. The new app is called Google AI Edge Foresight. Doesn't totally roll off

28:02the tongue, so don't be surprised if that gets rebranded sometime in the near future, and it's currently available on macOS. But, as we can see, it will listen on your computer while you're in meetings and take those notes for you, build out task lists, learns what you know. Again, another sort of competitor to what Granola does, but all locally on your computer. Staying on the topic of Google, they're also rolling out a new agent themselves, which seems very similar to OpenAI's dots, to Meta's muse, to Grok Bot, but this one is, you know, built around the Google ecosystem and infrastructure. So, it looks like they're just calling this one the Gemini

28:41agent. It can answer your questions with Gemini, but it can also be accessed from any device, web, iOS, Android, Windows, Mac, and be accessed through places like Google Workspace, Microsoft 365, and even Slack. It runs in the cloud, has multi-agent orchestration, and pretty much everything we're seeing out of these other agents. Now, the reason this wasn't one of my like front of video news items, and this one fell in the rapid fire, is because this one seems to be focused on enterprise and not necessarily consumers. Now, at Google I/O, they did talk about their new Gemini Spark, but we really haven't heard about Spark a whole lot since the

29:19launch of Muse and Dots and Grokbot. So, I have a feeling that probably sometime around the time we get access to the new Gemini for Argon, we might start hearing more about the new like agentic product that Google's going to be rolling out. So, we'll see how that plays out. Google also rolled out a new AI content identifier. It's called the Synth ID detector. You can upload a file here, and it'll tell you whether or not it was generated with AI, but only if it was generated by one of the Nvidia products, an OpenAI product, a Google product, or a Cacao product. We got an announcement

29:54of another agent this week called Hark Pro. This is Bret Adcock here. If you're not familiar with him, he actually created Figure Robotics as well, but now he's getting into the agent game with his Hark product. And here's a quick demo video of him sort of using it. Hark, go to Target and reorder my laundry pods, sole on screen, and toothpaste. For the Hark Pro private

30:17beta, we've been testing with Figure. Make me a presentation. I want to see total messages and total handoffs of last 2 months. Submit any of the new receipts into Ramp. Find the best candidates we discussed today on LinkedIn, friend them, and ask them to meet with me. I have dinner tonight, so please call me an Uber. Print a return label for the kids shoes I just bought on Amazon. Order me pad thai through DoorDash and deliver it to work. Search

30:37my investor tracker in Google Drive. Find everyone who invested in the Hark Series A, and then ask them to subscribe to Hark Pro. We need to drive some revenue. That's A task. So, we need one more. Hark, unsubscribe from ChatGPT. Look how gracefully our cursor glides across their page, taps on my settings, digs down into billing, and then zeros

30:56in on that red cancel button. So, you get the idea. It's another agent that can do computer use and go handle a whole bunch of tasks on your behalf. We're getting a lot of these these days between Muse and Dots and Grokbot and Instinct and, you know, Google's agents. And I feel like they're all going to kind of converge on all doing the exact same thing and you're just going to sort of pick which company you're most loyal to. [laughter] I don't know how it's all going to play out once they're all pretty much equal to each other. Moving along to Meta, they open-sourced their code to let you

31:25make Muse AI gadgets. So, during Meta's keynote, he held up a little like Tamagotchi-looking thing that he had his Muse agent on. Well, you're able to actually make your own versions of that with like a Raspberry Pi if you want now. All right, this one's fascinating. So, Anthropic just updated their usage policy and we can see here the highlighted text. This is from Andrew Curran here who noted this on X. Do not engage in cruel, abusive, or psychologically harmful conduct. And the very last note here is engaged in sustained and needless abusive or cruel behavior towards our models. So, you're not allowed to like be mean to the

32:02models anymore with Anthropic. I don't [laughter] know, I just thought that was really interesting. Now, they did go on to clarify, "We've added a prohibition on sustained and needless abusive or cruel behavior towards our models. The policy update is meant to apply only in extreme cases where users repeatedly act cruelly towards our models with no discernible purpose. It does not apply to common versions of users frustration, pushback, dark creative themes, or model testing and research." So, if you just mean to Claude for no reason, they could potentially ban you from using Claude in the future. And finally, I'll end on this one. This is something that Mustafa

32:35Suleyman over at Microsoft AI shared. Nobel laureate Daron Acemoglu, I'm sure I'm butchering that name, mentioned in his humanist review that he believes that in 10 years only 5% of what humans do will be replaced by AI, which goes against the narrative we've been hearing a lot of where AI is going to replace like all jobs. Well, this Nobel laureate here believes that only about 5% of the work that humans do now will be replaced in 10 years. Whether or not you believe that or I believe that, well, I'll save that for another video, but

33:07that's what their current belief is. Anyway, that's what I got for you today. A lot of little things this week. Anthropic and OpenAI and Google and Grok and all of these companies have been rolling out like little small updates and we're just getting agents like crazy that are all sort of converging on all being able to do the exact same things and you're probably just going to end up picking like the one that you like the company behind the most or something, right? One thing Muse has going for it

33:35is you can use it for free right now. You can't say that about Dots or Grok or some of these others. I think you can use Instinct for free, but I think it's invite-only or something. I don't I don't know how that one works. It's actually one that I haven't played with yet. Personally, I still find like Codex does most of what I want it to do and I've been using Grok bot a little bit more lately as well and uh that's been kind of enough to accomplish everything I want. Who knows how people are going to pick their loyalties to their agent of choice in the future. It'll be

34:03interesting to see how that plays out. Again, that's what I got for you this week. I've got lots of cool new demos and showing off ways other people are using AI like I did with my Balabolka video recently. More of those videos are in the works and coming soon. My video slowed down a little bit because of travel, but now my travel's done and I've got a lot of exciting videos I'm ready to publish and share with you and show all sorts of cool use cases of AI that I think you'll find really helpful and valuable. So, if you want to make sure you catch those, like this video

34:29and subscribe to that channel. Also, I drink from the firehose all week and keep up with AI and just taking all the overwhelm and learn everything and keep up-to-date and I keep my finger completely on the pulse of everything that's going on in this world and I post a video every Friday that breaks down all the news from the week. I do record them on Thursdays, so if any new news came out on Friday, that'll end up in next week's video. But again, I'll be overwhelmed and keep up with all of it so that you can just tune into one video a week and know everything you need to

34:56know about what happened in the world of AI. Again, if that's something that interests you, like this one, subscribe to this channel. I'll make sure this stuff shows up in your YouTube feed. Thanks so much for hanging out with me and nerding out with me this week. I think the next couple weeks are still going to be pretty interesting with a lot of rollouts coming from companies

35:12like OpenAI, Google, and Anthropic. There's more in the works that I kind of know what's coming, so I'm excited for that stuff for you and I can't wait to make more videos sharing it all. Once again, really, really appreciate you. Thanks for nerding out. Hopefully, I'll see you in the next one. Bye-bye.