Activity
Mon
Wed
Fri
Sun
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Jan
Feb
Mar
What is this?
Less
More
Memberships
Chase AI Community
47.2k members • Free
Clief Notes
4.6k members • Free
4 contributions to Clief Notes
6h •
Is This Actually An Opportunity?
The AI Automation Agency (AAA) Business Model is very popular amongst a sector of YouTube creators such as Nate Herk, Liam Ottley, etc. They offer courses to learn Claude Code as well as learn how to start your own AAA. These courses are marketed to enable you to make your first $10k with automations and so on. After doing research, there is a common proposition from these creators: Offer 24/7 lead capture (chatbot/AI receptionist), an SMS booking system, a social media DM bot, or a “speed to lead” system (a workflow that responds to customers right away). These are the 4 core offers, find a low tech industry and sell one of these offers. My opinion here is that those 4 offers are not very high value and are most likely offered by many SaaS companies or even website hosts. Which brings me to the question: Using Claude Code, are there opportunities within small to medium businesses to automate high value workflows? Everyone wants to start their own AAA but I am not seeing any high value offers. It’s all just basic add-ons that are already offered. Maybe I am not seeing something here, that’s why I want to get some community input on it. Do you think Claude Code workflow automations are a feasible business?
0 likes • 3m
I agree automating human communication is low value. I don't even see how sms booking is better than an interface that allows me to see availabilities and just pick one. The true value I see is automating how domain knowledge context is collected. The high value solutions require specific domain knowledge and an interface that actually helps the business and workers do their job. So automate how work procedures and business/client specific data is collected and stored. Then design the interface that helps a human use it.
2d •
Bare metal inference
Just to show off how big of a glutton for punishment i am. the network drivers along took 3 days of active work. but i can run models from usb. no os, no kernel.
1 like • 23h
This is nuts I have been thinking about how you could route inference to any gpus in a home or office environment. This bypasses all os configuration? The usb can be matched to the hardware config only. Just boot to this and send requests over network. An application with some inference router can orchestrate different agent hierarchies across all available local compute!
1 like • 20h
@David Herrera Totally fair I cannot even imagine what most of that stuff involves and while it is very cool likely not enough practical benefit over dual boot or regular os install.
2d •
Hi just joined, my Obsidian+Claude-Code experience
I am a student and recently admitted to a new program to study intelligent systems engineering. I have been a self learner for the past year just pursuing full stack development and Linux basics, there seems to always be more to know. But, I have been loosely on this AI 2nd brain trail for a while now and now its possible with current models and gaining momentum like in this community. First checked out the Youtube and resonated with this idea of what is the new layer of abstraction, what's worth building and investing into. A little bit ago a document rag processing pipeline seemed to me the best thing you could do and now its all native. In the past few days Gemini now has multimodal rag embeddings too. Structured context and instructions in markdown is the layer where it is both human readable and AI readable. As the models improve you can't take away the need for well structured context, in fact it only makes it more valuable. So what is the best way to do this. The most straightforward thing for me was Obsidian + Claude-Code, Claude-Code runs in my vault root and Obsidian is a great markdown frontend with plugins, notably obsidian CLI and local rest api. First symlink the .claude folder to a folder like CLAUDE so it is not hidden in obsidian. Then the CLAUDE.md tells it how to interact with your vault. Ultimately the native tools Read, Edit, Write, Glob, and Grep are often best but GET /active allows it to see what's open via the rest api. Obsidian CLI has some commands that can allow more capability faster in fewer tokens, but not all are better than the native md editing. My initial strategy is building light commands that instruct it how to use these obsidian commands and also bash to get more deterministic outputs in fewer tokens. Make common tasks repeatable and maybe transition to run it on local ai requiring less inference at some point. However, there are some disadvantages still. It feels like there is not total agreeance between Claude-Code standards and obsidian standards. I gave up on using an Obsidian MCP because there is too many of them and its unclear the tradeoffs between them. Maybe I should be making my own? It also needs deeper integration with an IDE, sandbox dev container environments, a knowledge graph it can traverse not just the obsidian tags and links it has to grep. Another thing is how do I want to interact with it from my phone, a mobile interface focused on different actions and views not just obsidian mobile, although it may be versatile with proper setup.
1 like • 2d
@David Herrera Yes I took a look at teamide I definitely will be checking it out, its about the same line of thinking. You're right I don't want to lock myself into something but it will be good to continue to experiment.
2d •
local AI build recommendations
Hi everyone, I am glad to join the community of like minded people building and working with AI. I had a quick question.... I am currently on intel 10600K 24gb 2666mhz ram RTX 3060 TI I am going to upgrade to Ryzen 7 9700X x870E motherboard (haven't decided which one yet) 64 GB of ddr5 ram ( just waiting for markets to cool) and I wanted to know what do you all recommend for GPU's to run ? I plan on running dual GPU s but I want to know if single will be better to run a local AI and automation business ? please let me know below based on the poll which GPU would be best suited and if its better to run single or dual GPU. thank you all in advance.
Poll
7 members have voted
0 likes • 2d
To me it just comes down to memory bandwidth and how much / what kind of work it is doing. Comparing: 5090 32GB at 1792 GB/s 575W TDP 4090 24GB at 1008 GB/s 450W TDP 3090 24GB at 936 GB/s 350W TDP what i have for comparison: 5060 ti 16GB at 448 GB/s 180W The 4090 is quicker but ultimately model capability and context window is most important. Its not worth paying much more for the 4090 as its mostly the same as the 3090. Technically by whatever calculation the 5090 is more power efficient even at the higher TDP, if that matters to you. Although the PSU required will not be cheap but maybe you are undervolting anyway. The other thing I've learned when it comes to dual GPUs is while motherboard bifurcation is ideal it only affects it in a few ways depending on what you are doing. It was quite different than what I expected using ollama and lmstudio sharding. pcie 5.0x16 = 64GB/s pcie 5.0x8 = 32GB/s pcie 4.0x16 = 32GB/s pcie 4.0x8 = 16GB/s 1. Model Load time: The model must load off the ssd over pcie each time it is loaded into memory. This is bottlenecked by the ssd usually. A gen 5 is ~14GB/s and gen 4 is ~7.5GB/s both under the pcie 4.0x8 speed (16 GB/s). However this could be a pcie bottleneck without bifurcation, still if you are not changing the model often doesen't matter only delay on startup. 2. Prompt Refill: The pcie does actually control a minor delay at higher context windows each time the prompt is submitted. The context window 64k, 128k, 256k translated to GB of KV cache and quantization needs to be loaded onto the gpu over pcie. This is more for context switching though not chat continuation. So if 4gb-8gb of context window kv cache changes it will load .125-.25 sec for 32 GB/s and .25-.5 sec for 16GB/s. This estimate is theoretical and this may be negligible anyway. 3. Model Training/Fine-tuning: Here pcie is a direct bottleneck, it becomes pretty apparent that training or fine tuning over multiple gpus is going to be slower than one. A 5090 would excel here but this is really what the dgx spark is for. 128GB at 273 GB/s, more memory than a single gpu and more memory bandwidth than pcie 5.0x16 64GB/s connection.
1-4 of 4
Online now
Joined Mar 11, 2026
Powered by