The Revolutionary Impact of Claude AI: A Game-Changer for Software Engineering

for the second time this year anthropic embarrassed open AI by releasing a new state-of-the-art large language model

that sweeps GPT 40 on every major Benchmark it's not even all that close and Claude is now King of the software

engineering Benchmark by a wide margin but I'm not here to Simply simp for Claud because yesterday they also

released what is perhaps the most dangerous AI feature ever handed over to the public the ability to open Excel and

do all your pointless spreadsheets at work the ability to fill out a patient's chart if you're a doctor the ability to

log into your Robin Hood account and YOLO your life savings and really anything else Humans Do by taking full

control over your mouse keyboard and monitor what could possibly go wrong in today's video we'll jump back on the AI

Doomer hype train and take this mind-blowing new technology for a test drive it is October 23rd 2024 and you're

watching the code report as someone with no biological friends my life essentially is my computer my wolf pack

consists of the Original Gangster GPT the base AF llama the desperately wants to be cool Gemini and of course the Apex

Alpha model Claude yesterday CL became even more powerful with an upgrade to the Sonet 3.5 model as you can see here

it beats GPT 40 on graduate level reasoning programming and visual Q&A it only loses to Gemini 1.5 on math but

that's comparing four shot to zero shot and it also sits on top of the software engineering Benchmark where it's able to

solve 49% of GitHub issues that it encounters one big caveat though is that it's comparing to GPT 40 and not the new

01 model which itself relies on the Chain of Thought technique to automatically reprompt itself thus

making compar comp Arison is difficult this upgrade is cool and all but the real game-changing new feature released

is something called computer use which is available to developers via the API now so I immediately put it to use and

started burning through tokens which cost $15 for a bundle of a million the first thing I asked it to do is

something I thought it would fail at which was to find the SVG code for the fireship logo but it actually succeeded

so let's break it down step by step because basically it prompts itself in an infinite Loop performing different

actions analyzing the results which lead to other actions until it solves the original problem in this case it takes a

screenshot of the desktop and notices that I have Firefox there that leads to another action to click on the Firefox

icon it now sees the address bar which leads to another action to move the mouse there and click on it which is

followed by another action to type out the URL now that it's on the website it finds the logo and right clicks on it it

then opens up the dev tools to inspect the HTML copies the code and returns it pretty amazing because we just did some

webs scraping entirely with natural language but it can use virtually any application on your computer like I also

asked it to build a net worth calculator in Excel or Libre office not only did it input all the data but also created the

formulas for the calculations I even prompted it to open up X paint and paint a picture of a horse and it created this

masterpiece it may not look like much but this image didn't come from some diffusion magic it was created with the

actual stroke of a pen like a real artist but this technology is far from perfect like at one point right in the

middle of a coding task it decided to go on the internet and browse photos of Yellowstone National Park and I do the

same thing when I get burned out from coding however there's a ton of potential for bad things to happen in

here like if you use this tool to manage your bank account it's only a matter of time before Claud drains it to invest it

all in Godus Maximus a shitcoin that's pumped entirely by Ai and now has a $500 million market cap but what you'll

notice here is that I'm not actually raw dogging this thing on my main computer it's actually running in a safe sandbox

with Docker in fact you can run it right now on your machine with one command you just need to have Docker installed and

have an anthropic API key but be warned this thing Burns through tokens extremely fast the good news is that it

mostly uses input tokens which are a lot cheaper and it needs all these tokens because like gp1 it takes the output of

one prompt and uses it as the input for the next prompt and we'll do that in a loop until it gets the result you asked

for or until it crashes which happens a lot but still clot is the best model out there when it comes to real world

computer environments at least according to the OS World Benchmark however it's nowhere near human level and will take a

lot of training and reinforcement to get there the main bottleneck here is that it requires a lot of compute time and

tokens to do simple things that us humans take for granted like all the prompts I've showed you have taken 5 to

10es minutes to complete but eventually that problem will be solved companies like Amazon Google and Microsoft are

investing in nuclear to power massive AI data centers but eventually Chain of Thought action models like this will

likely be baked into every computer and I'm not just talking about your personal computer but the brains of the robots

they're already building robots that can drive you around build you toys perform medical procedures and even put food in

your dish the future looks bright until one day you wake up and realize that the guy Claude is named after Claude Shannon

was right all along when he predicted that we will be to robots what dogs are to humans this has been the code report

thanks for watching and I will see you in the next one