Hey guys, welcome to the show. Today, we're checking out my Claude the AI and basically just an online instance of open Claude running in the cloud
24/7 available. So, if you find it too difficult set up at home, you can use this service. I'm showing you how much it costs, the token usage, the
different API model providers you can all different skills you can utilize the system. So, it's personal
AI system that everyone is checking out and it's like one of the fastest open projects that are out there at the moment. You can connect your instance
with WhatsApp, Telegram, Discord, web chat, all these kind of situations. So, use cases, code automation. So, let's test out some coding in this
situation. Smart file management, AI assistant, AI car purchasing negotiation. I I need this tool for
negotiating. So, you know, I always thought I'm getting ripped off. So, it's going to be probably good for that situation. So, let's log in and when
you log in, you get a pretty much open Claude just running in the cloud except check this out. We got access to Claude Opus 4.7. We got Claude Haiku,
we got Gemini GPTs, King K 2.5, and MiniMax. And what you want to check out is the AI models tab
because that tells you all about the prices. So, Claude Opus 4.7, it's around $22 for a million tokens. But, if you switch a route and go to MiniMax,
that's only $1 for a million tokens. Amazing. Now, you also have something called a backup model. So, if the main model isn't available, it'll switch
automatically to the backup model and Ooh, this is very good. You probably want to set this. Monthly
limit for this Claude. So, right now it's unlimited and I've only got 12 bucks in the account. So, I'll probably want to limit it to five bucks,
maybe. I'll put in a five bucks limit there just to see what's going on. I've already used 28 bucks. Oh, man. Okay, I'll put in 35 bucks. There you
go. We are set for this demonstration. All this looks nice. You got a little open Claude kind of like logo
over there to tell you it's vibing. So, we've got this feature called tool calls and tool results. So, show me the tool call that's making. It's
checking out bootstrap who am I reminding itself who it is cuz this open close situation has a memory. So, it has kind of like a living log file of
your interactions with it that you can update it depending on your experience with it. So, hey, I just came
online fresh workspace no memory yet. So, he wants to ask me some question. I'll say forget that. Can you tell me the latest papers on quantization?
So, I'm going to ask him now to do a research question. Let's just say we're doing a PhD and we want a research assistant to do some work for us. So,
it's parking the bootstrap and it's going to tell me the latest papers on quantization of LLMs. So,
it's making two tool calls. Let's click on the tool calls find out what it is. So, it's doing a web search for the latest patient papers in LLM
quantization. That's what I asked for and it went straight to archive. Archive, that's the probably the most popular site out there. And it's nice
that you can actually see the tool input and tool output. That's So, we've got one widening the gap
exploiting LLM quantization with outlier injection. High rate quantized matrix multiplication. Finer is better. Co quant. Subspace projection. They do
weight activation that means they run some queries through it check out the activations and find out what which ones get activated the most and I
guess they depending on which one get activated most they mix the position up. I say, "Are there any
stand out breakthroughs?" Be my analyst now. So, it's research the latest papers over here the top six and I'm asking it are there any breakthroughs
that I should implement in my inference engine. And there's actually two standout papers worth my time. Tier one implement now. It's It's just I don't
need to read anymore. I can just save my life's brain, make it focus on other stuff like watching TV
all day, and I can just get it to do the work for me. So, let's see. Lattice leech lattice vector quantization for efficient LLM compression. It just
came out on the 11th more scale quantization all the time middle. So, like apparently you can compress it down to two to three bits with workable
inference speed. So I'm regarding your number one suggestion, is it an inline technique where I can just
plug it in and it'll come up with the output or does it require a data to measure the activations? If anyone knows anything about quantization, the
data methods they require you to run inferences on the full fat model that doesn't fit in memory with uh your inferences and then it will pick the
levels to compress based on that. So that takes forever, especially if you're streaming it from this. So
I don't know if this is actually practical or if it's a Good question. Let me actually check the paper directly rather than guess. Short answer is
data free. But in practice you still want calibration data. So what is it? What is going on? So it just says it is actually data free and just messes
around with the actual data, so no activations needed. But where does it take the activations? So it's
benchmarked with some other stuff. So apparently this could be a winner cuz they're saying that it is data free. They're saying also you also probably
need data calibration to make it good. But the essence of it is data free. We're going to random topic, just picking our brains out, finding out some
stuff. You could have picked anything and just vibed with it around it. So they said do you want
release codes? Yes, check for codes. So unfortunately you cannot find any code for this implementation. So right now we're going to go into paper and
figure out some stuff or maybe we can actually get it to write it. But speaking of which, speak since it can do some web searches and it can do some
sort of research for you, I want to make this into the money-making agent. So the idea is can you
write me a skill or job that will run this query every week to let me know the latest paper and deeply analyze on its effectiveness. So this is where
the cookie will crumble or not. I've kind of vibed with it. I gave it a task. I reasoned through that task to find out if it's actually worthwhile or
not. Now I'm going to ask it can it make an automatic job that it will run once a week for me to get
the latest papers and find out is it feasible and get the code that is to implement it. So, let's just see if we can figure this out. Good idea the
cronus approach. It's going to make a weekly cron job that spawns in an isolated agent turn. It does a search deep and reports back to the chat. No
skill file needed or anything. I'll say, "Yeah." So, we didn't need to make skill file. It's just going
to make a cron job, which is just a job that just runs on a schedule. So, it's going to schedule it for once a week. So, once a week I'm going to know
the latest happenings in the world of LLMs and that is pretty cool. Don't need any more assistance for me. I've got this tool doing it for me. So,
with weekly paper scan, this is the message it gave me. Review step one. So, it pretty much just put
the skill or them in the message. So, this is the message it's going to do and it scheduled it to run once a week. So, it's done that job there and
it's made the directory and it's added it to my memory. So, that memory file is going to check the job and the job is going to run once a week. So,
every Monday at 9:00 it's going to do an isolated session. It won't blow our chat history. Going to be
separate out of this. Going to fetch the last 7 days of papers and it's going to triage them, deep dive and report back 400-700 word opinionated and
that is awesome. So, you test the job now so we can run it and see if it works and that will be our first report. You know what I'm going to do that.
Yes, test the job. So, it's running the job right now. It's doing it in the background so it should
be a few minutes and it's going to stay out of the way and in the meantime here I'm going to click on the cron jobs tab to see if it actually had
added it to the weekly panel. So, we have yes, enabled job. One job has been enabled and there it is. This is the job that we've just made. So, we
pretty much got a weekly employee doing this work for us. Now, in case you're interested, there are a
plethora of different skills you can use. You can plug it into your Google workspace, quick capture, self-improving agent. Yo, this is where we're
here already. AI is happening. We got um peekaboo. Now, this is something great for your kids as well. Peekaboo. Can do that one as well. AI PPT, add
nano banana, obsidian. All this kind of stuff is all built in for you. So just as easy as that we use
my claw AI running in the cloud to connect to your WhatsApp, connect to your Slack, connect to all this stuff, gets to do jobs for you. We made it do
an automated schedule tool to do our job for us every week. It's going to do my job. What kind of jobs would you get running on your system? What do
you guys think? Hope you guys found the video useful and enjoyed the show.