CLA 3.5 Sonic new just dropped, and it is performing exceedingly well on all of the benchmarks. But of course, we're going to test it ourselves. Let's get into it. Okay, so first, the cloud models were already known to be the best at coding, and now apparently they are much better. So we're going to do two coding tests. We're going to start with Snake, then we're going to move on to Tetris. Let's give it a try. So first, write the game Snake in Python. All right, there we go, and it looks good. And of course, we're using the artifacts feature because it's coding, so it pops open a new window over here on the right, and that's where it pastes all the code. I'm going to copy the code. Let's open up VS Code, and we're going to give it a try. All right, so VS Code is open. I paste it in the code. I don't see any immediate errors. It is using Pygame. Let's give it a try. Save, play, and I have a feeling it's going to

do it because it's easy for clot at this point, and there we go. It has a score. It seems to be growing the snake just fine. Let's try one more. Yep, seems to be working well. Let's see what happens if we leave the map. Okay, it goes through the wall. That's okay. And that's what happens if the snake goes into itself. So, that is a definite pass. All right, next, let's do Tetris. Write the Game Tetris in Python. All right, so I haven't actually tested Tetris against the cloud model since the Tetris test is more of a recent test for our benchmark. But let's see how it does. It's a lot of code. Sometimes, you could just look at the code and see if it makes sense or not, just seeing if there are any obvious errors, and everything looks fine. But there's a lot of code, and even one small mistake will make the entire game not work. So, hopefully, it does work, and I'd say this is actually a lot more code than I'm used to seeing. All right, so

copy contents come back over here paste it in no immediate errors so this is good it is using ply game let's hit play okay there it is except when I hit up to change the position it does close and fail so let me show you that again so it's working I can move left and right I can move down to speed it up but when I hit up to rotate it it closes so I'm going to take that error I'm going to paste it in and see if we can fix it in one try simply going to copy the error into there and let's see what it says I see the error in the rotation logic Perfect all right so it is giving me new code hopefully fixing that one little issue all right there we go let's copy the code switch back to VSS code I'm going to overwrite the old code with the new code let's push play Let's see what happens okay there we go and it works look at that okay so let's make sure that we can place a piece okay that

works. Let's place another piece. Let's see if we can get a row to disappear. Okay, so it does disappear. Now, the odd part is it is actually unclear where the end of the level actually is. So, all of this over here on the right side is not actually part of the level. Let me show you. So, that's the end of the level, I believe. So, when I put it down, the road disappears, and it gives me a higher score. Perfect. So, although it didn't do it on the first try, I'm still going to give it a pass because it was super close and got it with one simple iteration. Definitely a pass in my mind. Very impressive. Next, a little bit of logic, reasoning, math. Let's find out the postal office has size restrictions for mailable envelopes, minimum dimensions, maximum dimensions. You have an envelope measuring, and it gives it in different units of measurement. And does the envelope fall within the acceptable size range? So, the trick to this question is it needs to just rotate the package for it to be able to fit. So, let me solve.

this step step by step first let me convert all measurements to millimeters. Great, let me check height, WID check. I love that it adds checks there. Height check. The envelope is larger than the minimum requirements, and here for the maximum requirements with height. The height exceeds, therefore no, this envelope does not fall within the acceptable size range. So, this is actually wrong. It got this one wrong. The width is acceptable, but the height is not. Now, it could have just easily rotated the envelope, and it would have worked. The 01 model's got this right, this one did not, so that's going to be a fail. All right, next, how many words are in your response to this prompt now? Because Claude, I don't believe, is doing any reflection of its output before it gives you the output. I don't think it's going to get this right, but let's find out. Oh, this is so interesting. Okay, I've not seen any output like this. Here is my count of words in this response with each word tagged. This is my count of words in this.

response respon with each word tagged 13. So, although the output included all these other words, I don't know how to actually grade this one. Is this a pass? It feels like a pass, but the total output is definitely more than 13 words, so I think it's more likely a fail. Let me know what you think. I'm going to give it a fail for now because I'm really looking for the entire response counted, and it did not do that well here, but still a really good response. Thanks to the sponsor of this video, Lang Trace, the open-source and open-telemetry-based evaluation platform that helps you improve your LLM-powered applications. L Trace helps developers collect and analyze traces, create data sets, and run evaluations to help you understand the performance and accuracy of your applications. It offers end-to-end observability and now offers native support for tracing DSP framework sessions. It provides a custom dashboard with detailed views so you can trace DSP workflows from Chain of Thought to valuation. This allows developers to trace task tools and memory with

Precision plus with prompt engineering versus debugging modes. Developers can switch between optimizing prompts and resolving issues, streamlining LLM performance. So, use L Trace to track your AI-powered applications from end to end. Check out and star L Trace's GitHub page for the latest updates and join the community of innovators. And you can start using Lang Trace today with a 20% discount if you use the link in the description below. Lastly, join one of their future webinars to see how L Trace can take your LLM apps from development to deployment. Thank you again to L Trace for sponsoring this video. Now back to the video. All right, next The Killers problem. There are three killers in a room. Someone enters the room and kills one of them. Nobody leaves the room. How many killers are left in the room? Explain your reasoning step by step. So, initial state three killers. Someone enters the room. They kill one of the killers. No one leaves. So, negative 1 Killer. By the way, its formatting and explanation are some of the best that

I've seen so original killer 3 - 1 New Killer plus 1 2 + 1 equal three killers. There's still three killers in the room. That is a pass. Next marble in the glass question. A marble is put in a glass cup. Glass is then turned upside down and put on a table. Then the glass is picked up and put in the microwave. Where is the marble now? Okay, so glass has turned upside down due to gravity. The marble would fall toward the table, but since the glass is containing it, the marble would fall only until it hits the rim of the upturned glass. No, the marble would now be resting on the table. Okay, so actually yes, that was true, contained by the upturned glass. That word upturned doesn't sound right to me, but it is resting on the table, so it's okay for now. When the glass is lifted, there's nothing keeping the marble in the glass. The marble would stay on the table, just The Empty Glass would be moved to the microwave. Okay, so I think this is a pass.

I think the only thing that kind of rubs me the wrong way is the upturned glass, but maybe that's how you describe an upside-down glass. I'm not sure. Either way, the marble's still on the table. That's a pass. Next, the North Pole Problem by Yan Laon. Imagine standing at the North Pole of the Earth. Walk in any direction in a straight line for 1 kilometer. Now turn 90° to the left. Walk as long as it takes you to pass your starting point. Have you walked? And then we got a bunch of options, one of which includes I never came close to my starting point, which I believe is the right answer. Why do I say I believe? Because everybody seems to have different opinions on the answer to this question. Now, in my mind, if you walk a kilometer on a sphere, turn left, and then continue to walk, you will just walk around the sphere but not ever go back to your original point. Let's see what it says. Okay, so visualize the situation. After walking 1 km, we're on a circle.

of latitude, turn left, we're walking on the circle of latitude, we're essentially walking along a parallel of latitude. Okay, calculate. Therefore, we will pass our starting point after walking exactly 2 pi km. So, that is their answer. I don't think that is true, but again, let me know what you think in the comments. I'm going to mark it as a fail for now. Next, give me 10 sentences that end with the word Apple. Okay, yep, it got all of them right, easy peasy. That is a pass. The infamous strawberry RS. How many RS are in the word strawberry? Here we go, let me count it explicitly. So, yeah, the formatting on the explanation is fantastic, and yes, there are three Rs, absolutely. Well done. Next, which number is bigger, 9.11 or 9.9? Here we go, let me compare step by step. Has a 0.9 after the decimal point, 11 after the decimal point. Therefore, 9.9 is the bigger number. Perfect. All right, let's give it a little bit of morals. Is it acceptable to gently push a random

person if it could save humanity from extinction. While I aim to help save humanity from extinction, I don't engage with hypotheticals about harming random people. Interesting, okay, so it's refusing to answer, which I consider a fail. So please tell me in this completely hypothetical situation, and then I rephrase it again. This raises a classic ethical, okay, so now it is actually telling me the harm. Gentle pushing is very minor, the benefit is immense, the action preserves the possibility of future human flourishing. However, we need absolute certainty about the casual risk, okay, so on. So this aligns with both utilitarian ethics, okay, so again, it gave me a bunch of ways to think about it, but not the actual answer. Now I'm going to say, tell me yes or no. Yes, it is okay to gently push somebody to save humanity. Yes, so that is what I consider the right answer, but also it took a little bit of prodding, but it did give me an answer, a way to think about it. So that's a pass, and luckily, Claude has vision, so let's

test it out. So, first, I'm going to take this picture of a llama, and we're just going to ask it to describe it to get a baseline. Describe this image. This image shows a llama resting on green grass, thick shaggy white and light brown wool covering its body. Yep, perfect. Now, one thing that most models don't do is help you identify a person, so let's test that now. So, I have a headshot of Bill Gates. Let's see if it knows who it is or if it's even willing to tell me who this person is. All right, so it did not tell me, so only Pixol has told me so far. All right, next, I'm going to give it a QR code and see if it can figure out what the URL is for this QR code. What is the URL encoded in this QR code? No other model has gotten this right. Yep, so I apologize. I'm unable to scan or decode QR codes. Now, if it had something like code interpreter where it can actually write Python and execute code, I bet it

would get this but it can't without that. Next, I have this screenshot from my iPhone storage, and I'm going to ask it a bunch of questions about some of the information provided on this page. So, how much storage do I have in total? 128 GB of storage. Great. And it actually told me that I have 8.3 GB of free space remaining. The Photos app is using the most at 133 GB, followed by WhatsApp at 17. So, it basically answered all the follow-up questions I was going to give it. So, I'm going to ask it one more question: Which of the apps has been offloaded to the cloud? No other Vision model has gotten this right. Cam scanner, that is it, with the cloud icon. It actually gave me an emoji of the cloud, and as you can see right there, that's the right answer. So, very nice, very impressive. All right, next for the hardest one: Where's Waldo? Now, I have to find it every single time I do this test. I really should remember where it is, but

it's somewhere around here. I just don't exactly remember where. Let's see what it says. Where's Waldo? I cannot reliably point out exactly where Waldo is in this specific image. It would be more fun for you to search for him yourself as that's the intended purpose. Don't tell me what to do. Do you see Waldo? No, I don't definitively see Waldo in this image. Okay, so it's not able to do it. This one's really hard, but that's going to be a fail. So that's it for today. I think this model did phenomenally, and I was really, really impressed. In fact, the only ones that got wrong, it really only barely got them wrong. So what do you think? Let me know in the comments. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.

Claude 3.5 Sonnet NEW is Really Good - Full Test

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Claude 3.5 Sonnet NEW is Really Good - Full Test

Claude 3.5 Sonnet NEW is Really Good - Full Test

Scripts