• We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.

The anti-AI thread

Page 33 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Red Squirrel said:
If you type your password in Notepad it will warn you that it's an unsecure thing to do. This means that your password is somewhere in memory and is constantly being checked as you are typing. I suppose it might be checking everything you type against a hash of your password too. Either way, I feel whatever is happening to do that could potentially be the source of a major vulnerability at some point. Will be fun to see if I end up right.

I would also be curious if it's possible to coax copilot into telling you your password or other company sensitive info, that could be very exploitable too. Don't really want to mess with that at work though.
Click to expand...

I'd say Passkeys to the rescue, but with the rise off AI browsers, a prompt injection attack could temporarily use your authorized session. Plus the future is here:

thequantuminsider.com

Google Quantum AI Shows 13,000× Speedup Over World’s Fastest Supercomputer in Physics Simulation

The Quantum Insider (TQI) is the leading online resource dedicated exclusively to Quantum Computing.
thequantuminsider.com thequantuminsider.com
 
Kaido said:
View attachment 139433
Click to expand...

Also don't forget to uninstall the WarGames module from Microsoft Paint. 😛

My solution is to uninstall Copilot. Also I've noticed that a Win11 install gets "Microsoft 365 Copilot" and Copilot as separate apps even if Office 365 was never installed (ie. a clean install from Win11 from the official source).

I don't know if this nails the Copilot option in Notepad too, but I've been disabling CP in NP on every install I encounter.
 
Kaido said:
I'd say Passkeys to the rescue, but with the rise off AI browsers, a prompt injection attack could temporarily use your authorized session. Plus the future is here:

thequantuminsider.com

Google Quantum AI Shows 13,000× Speedup Over World’s Fastest Supercomputer in Physics Simulation

The Quantum Insider (TQI) is the leading online resource dedicated exclusively to Quantum Computing.
thequantuminsider.com thequantuminsider.com
Click to expand...

I just type it out to have it available, since I have to login with it like 40+ times in a shift into a variety of applications so I just copy it then paste it everywhere. Not the most secure thing but it helps prevent having carpal tunnel. It's a very complex password so it's hard on the hands to type. We already have a password manager, but I have notepad open off to the side so it's just easier to copy it from there each time than to open the password manager and then look for the record.

I do find it interesting that notepad detects this though as it means it is constantly comparing everything you type to see if it's your password.
 
Red Squirrel said:
I just type it out to have it available, since I have to login with it like 40+ times in a shift into a variety of applications so I just copy it then paste it everywhere. Not the most secure thing but it helps prevent having carpal tunnel. It's a very complex password so it's hard on the hands to type. We already have a password manager, but I have notepad open off to the side so it's just easier to copy it from there each time than to open the password manager and then look for the record.

I do find it interesting that notepad detects this though as it means it is constantly comparing everything you type to see if it's your password.
Click to expand...
Use a token system instead like Windows hello and get a finger print reader.
 
mikeymikec said:
Also don't forget to uninstall the WarGames module from Microsoft Paint. 😛

My solution is to uninstall Copilot. Also I've noticed that a Win11 install gets "Microsoft 365 Copilot" and Copilot as separate apps even if Office 365 was never installed (ie. a clean install from Win11 from the official source).

I don't know if this nails the Copilot option in Notepad too, but I've been disabling CP in NP on every install I encounter.
Click to expand...

I made a script for that! At this point, Windows 11 is essentially AI-driven exfiltration spyware. Yes, that's rhetoric, but the bottom-line reality is that end users CANNOT fully disable AI integration & data collection in the latest builds.

Best course of action is to buy a cheap Pro upgrade key online & then use Applocker to block Copilot as the foundation. It's not perfect, as Microsoft is fighting a war of attrition with each update, but it's a solid start!

https://www.reddit.com/r/kaidomac/comments/1rnp887/copilot_boot_nuker
 
Last edited:
[DHT]Osiris said:
Use a token system instead like Windows hello and get a finger print reader.
Click to expand...

This is work I don't really have any access to start installing stuff or changing the infrastructure but I do agree they should look at implementing something like that instead of these ridiculously long passwords. The issue is lot of the programs are web based, citrix based, RDP based etc so not sure how you would make it work with all of that too. I assume each program would need to support whatever that protocol is and be running locally to interface with the reader. Really they need to just get their head out of their asses and have a half decent password policy and then get rid of all the timeouts. If you could login to stuff once and be good for the shift it wouldn't be so bad.
 
Red Squirrel said:
This is work I don't really have any access to start installing stuff or changing the infrastructure but I do agree they should look at implementing something like that instead of these ridiculously long passwords. The issue is lot of the programs are web based, citrix based, RDP based etc so not sure how you would make it work with all of that too. I assume each program would need to support whatever that protocol is and be running locally to interface with the reader. Really they need to just get their head out of their asses and have a half decent password policy and then get rid of all the timeouts. If you could login to stuff once and be good for the shift it wouldn't be so bad.
Click to expand...
Try googling around a bit, you may already have access. Windows hello is enabled by default, a $15 fingerprint reader may be all you need.
 
I've been building more & more local AI servers for clients who want custom local solutions. The hardware cost starts to make sense when you are replacing software with high up-front costs & service fees that you can easily build in-house with more customization.

The downside is that the AI datacenters have made hardware prices go completely NUTS! A 1TB DDR5 ECC kit was $15k a few month ago and is over THIRTY THOUSAND DOLLARS! ARGHHH!!

1772974151302.png
 
Kaido said:
I've been building more & more local AI servers for clients who want custom local solutions. The hardware cost starts to make sense when you are replacing software with high up-front costs & service fees that you can easily build in-house with more customization.

The downside is that the AI datacenters have made hardware prices go completely NUTS! A 1TB DDR5 ECC kit was $15k a few month ago and is over THIRTY THOUSAND DOLLARS! ARGHHH!!

View attachment 139492
Click to expand...
Yup, we're looking at a hypervisor refresh at work, and getting absolutely railed over the prices.
 
Are the local AI models trainable or is it basically static? Have not really looked much into it but I think it would be kind of cool to experiment with training models from scratch. I cannot afford hardware at this point though to run any models. I don't actually have all that much compute power in my server rack overall. Most of it is 10+ year old hardware.
 
Red Squirrel said:
I cannot afford hardware at this point though to run any models. I don't actually have all that much compute power in my server rack overall. Most of it is 10+ year old hardware.
Click to expand...
That's just a time issue, isn't it? I don't know the time scales involved, but crunching numbers for a week instead of an hour isn't terrible on your own stuff if you get results.
 
Red Squirrel said:
Are the local AI models trainable or is it basically static? Have not really looked much into it but I think it would be kind of cool to experiment with training models from scratch. I cannot afford hardware at this point though to run any models. I don't actually have all that much compute power in my server rack overall. Most of it is 10+ year old hardware.
Click to expand...

There are basically 4 kinds of local AI models:

1. Prompting (Q&A via temporary memory)
2. Fine-tuning (ex. using LoRA's to change weights to modify the internal neural connections)
3. Retrieval Augmented Generation aka RAG (reads from your database using stuff like LangChain)
4. Full model training (minimum of a million bucks & thousands of GPU's!)

ChatGPT combines all 4 as a service:

1. You ask it questions
2. It is fine-tuned (say to be sycophantic, i.e. have a helpful & encouraging tone)
3. It retrieves data from the Internet so that it doesn't need to be retrained 24/7
4. It was trained on everything it could find

You can easily spend $100 million doing Full Model Training. That's where your private AI database goes full-on Kirby & vacuums up EVERYTHING in the world, which requires MASSIVE amounts of power! Which is where late-stage capitalism comes into play, which is really the part that people have a problem with:

1. Full-on theft (art, writing, movies, etc.)
2. Bad neighbors (water usage, electricity price hikes, RAM/GPU price spikes, etc.)
3. Lack of accountability (ex. mental health issues)

And of course, then China comes along & copies all of America's work, releasing KIRF's at a fraction of the price or even for free. It famously started when DeepSeek cloned ChatGPT & more recently when they "distilled" Anthropic's database:


Which is kind of like the pot calling the kettle black, because they themselves stole the original data & literally got sued for a billion dollars lol:


Hence the AI bubble: everyone is competing in the new market & as a new source of weapons (data, robotic development acceleration, etc.), but nobody knows how anyone is going to make any real money with it, so it's a HUGE bet on future value!

Now that AI has become useful to consumers, the next steps are:

1. Exploring niches
2. Making it efficient
3. Agents

Exploring niches:

As far as development goes, Hugging Face is the Github of new AI tech, with over 2 million models available. This is where much of the R&D and new features are sourced from:

huggingface.co

Hugging Face – The AI community building the future.

We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co

So now we're seeing fine-tuned services such as Perplexity, which is an AI-powered news & search engine:


Sesame, for realistic advanced voice interactions as a consumer-facing interface to ANY system:

Sesame

app.sesame.com app.sesame.com

NotebookLM, for your own personal knowledge database with web research added:

notebooklm.google

Google NotebookLM | AI Research Tool & Thinking Partner

Meet NotebookLM, the AI research tool and thinking partner that can analyze your sources, turn complexity into clarity and transform your content.
notebooklm.google notebooklm.google

Suno Studio to translate music idesa into songs quickly:

Introduction to Studio

Studio is only available with a Premier plan Suno Studio is a groundbreaking web-based Generative Audio Workstation (or GAW ) that merges traditio...
help.suno.com help.suno.com

Freepik, which is a service that gives you access to the best photo & video models: (essentially like Photoshop on steroids!)


Making it efficient:

Now that AI has all of the data in the world & is getting refined in various niches, the next step is shrinking the algorithms & datasets. The big news this month is that QWEN has released an at-home model that is equivalent to Anthropic's cloud offering:


Which you can realistically run on a modern Mac Mini!!

"You can now run the Qwen3.5 Small models locally on just 6GB Mac / RAM device"

Or even a Raspberry Pi now!

"Orange pi 5 max + ubuntu it free if you use Llama-3.2-3B , eepSeek-R1-Distill-Qwen-7B or Phi-3.5-mini"

Agents:

The next big things are Agents, which are virtual thinking "robot" assistants! They have 3 parts:

1. The AI brain (local or online service)
2. The tools (plugins to do various tasks)
3. Memory (like having a personal assistant!)

The big leap recently was OpenClaw:

openclaw.ai

OpenClaw — Personal AI Assistant

OpenClaw — The AI that actually does things. Your personal assistant on any platform.
openclaw.ai openclaw.ai

Which is a free, private DIY agent that can:

1. Talk to local or cloud AI models
2. Access all of your hardware, software, files, and accounts
3. Basically do anything YOU want it to do!

So now you can:

1. Run QWEN locally
2. Run OpenClaw locally
3. Control what it does & what it has access to, for free!

For example, you can add a local voice AI that is equivalent to ElevenLabs. You can build a PBX VOIP telephony system for business, a custom smartphone voice, etc.! Try it out here:

huggingface.co

Qwen3-TTS Demo - a Hugging Face Space by Qwen

This app lets you turn written text into spoken audio. You can create a brand‑new voice by describing its style, clone a voice from an uploaded audio sample (with optional transcript), or pick a pr...
huggingface.co

If you want to play with AI locally, grab the LM Studio app store:

lmstudio.ai

LM Studio - Local AI on your computer

Run local AI models like gpt-oss, Llama, Gemma, Qwen, and DeepSeek privately on your computer.
lmstudio.ai lmstudio.ai

For local image generation, download Comfy for free:

www.comfy.org

ComfyUI | Generate video, images, 3D, audio with AI

The most powerful and modular visual AI application and engine
www.comfy.org www.comfy.org

Anyway...local AI with REAL performance is within reach for consumers as of this month! FMT is still best left to companies with huge pockets & render farms; the real fun s in customizing what's readily-available to DO cool stuff for you! Smarthome, home theater, custom computers, personal robots, business automation, high-speed software development, website design, server customization, etc.!

If you like working with images, dive into ComfyUI at home for REE! Note that image generation does tend to like beefy gaming cards as opposed to Pi hardware, haha! But this is one of the areas of local AI that you really CAN train & tweak & customize to get stellar personalized results!

Comfy UI tutorials

The latest big news is that LTX 2.3 is now available for the home desktop for FREE!

Open-source local video generation

While it really needs at least 32GB VRAM to be fast, it can run on as little as a 6GB GPU (~$200) to start playing with, which means even a budget gaming laptop can create AI videos & cartoons! Especially if you're a student with a great idea & no million-dollar film budget, this is INCREDIBLE news!!

https://twitter.com/x/status/2029901494305116650

Go out there & create some CRAZY stuff on your home computer! Make the next great Anandtech jungle action flick about "moving out to wilderness"!! 😍 😍 😍

1773009806022.png
 
Last edited:
lxskllr said:
That's just a time issue, isn't it? I don't know the time scales involved, but crunching numbers for a week instead of an hour isn't terrible on your own stuff if you get results.
Click to expand...

Depends on the model. For stuff like video, some of the models simply can't run on lesser hardware, or limit output size & resolution. But there are two local options outside of just brute-forcing it::

1. Run an efficient local model like QWEN 3.5. Or Qwen 2.5 Coder 1.5B runs on any laptop with 4GB RAM!

2. Use an local agent like NanoBot to call a cloud LLM API & then relay the information to you in the format desired, so you're still getting local commands & formatting, just outsourcing the hardware processing to rented time at a datacenter.
 
Red Squirrel said:
Are the local AI models trainable or is it basically static? Have not really looked much into it but I think it would be kind of cool to experiment with training models from scratch. I cannot afford hardware at this point though to run any models. I don't actually have all that much compute power in my server rack overall. Most of it is 10+ year old hardware.
Click to expand...
You got a really long response but here's a shorter, boots-on-the-ground response:
The way an AI thinks depends on the model. You can make your own model if you're awesome but you probably aren't so you're using another company's model. What it knows is variable, it starts with a baseline as of the day it was generated but can be enhanced if you know what you're doing with stuff like MCPs or knowledge bases, system prompts, and the like. A lot of the public facing stuff uses this silently already, a lot of the homebuild/workbuild AIs have these if the people who set them up included them.
 
[DHT]Osiris said:
Yup, we're looking at a hypervisor refresh at work, and getting absolutely railed over the prices.
Click to expand...

Lately I've been doing Proxmox on whatever used hardware I can find. My favorite "vintage" board is the Asus Z10PE-D16 WS, which are virtually impossible to find now: (11 years old at this point)

www.asus.com

ASUS Deutschland

Entdecke die Welt von ASUS - Alles rund um die besten Mainboards, Grafikkarten, Monitore, Laptops und vieles mehr erfährst Du auf der offiziellen ASUS Seite.
www.asus.com www.asus.com

Notes:

* You can get the 22-core Xeon E5-2699 v4 chips for well under $100 each on eBay (44 cores total!)
* 1TB of DDR4-ECC is just north of $6k new, which is a steal these days, if you need the capacity!
* The slots are older (6x PCIe 3.0 x16-size slots), but it's only about a 5% performance hit in most cases on a 5090 GPU. You can theoretically hit up to 32 GPU's with external chassis expanders!

New breakthroughs:

* GGUF quantization compressions shrinks models to less than 1/3 their normal size
* llama.cpp is a new CPU engine for AI
* Modern quantized CPU models means you can now hit 70B model + RAG at home for under a thousand bucks, no fancy expensive GPU required!!

Another trick:

* Datacenters are starting to shed hardware
* The 24GB NVIDIA Tesla P40 is under $300 on eBay these days!
* That can easily run a 30B quantized model!

This means you can build a 400B quantized machine at home with the used 44c/1TB setup:

1. Private ChatGPT ~equivalent to v4.0
2. Private Google Docs search
3. Private coding assistant
4. Private research assistant

With this setup:

1. Proxmox
2. Ubunutu VM
3. Running Docker

Configured with:

1. llama.cpp (CPU AI brain)
2. Open WebUI (local ChatGPT)
3. Chroma (database for memory)
4. LlamaIndex (whole ppipeline coordinator for RAG)

With a 3-prong approach:

1. Fast model
2. Large reasoning model
3. Embedding model

Then you add on an embedding model (shrinks the data to improve search) & a reranker (2x+ faster search) & setup a hierarchical RAG like the big boys do & voila! This is AMAZING because you can build a private ChatGPT for MILLIONS of local documents for a company on one simple box!

Even for people who don't like AI, this is pretty compelling because:

1. 100% private
2. Free software (no crazy contracts or recurring licensing fees)
3. Used hardware for "upcycling"
4. EXTREMELY reasonable prices (a used Dell T7910 with dual Xeon E5-2680 v4 chips & 256GB RAM is under $1,500 on eBay)

What I tell most people to get: (~$4k used for 40c, 1TB ECC, 24GB 3090 GPU Tensor & 24GB Tesla)

1. HP Z8 G4 tower with Xeon chips
2. 1TB RAM
3. 24GB 3090 GPU & 24GB NVIDIA Tesla P40 (Parsec gaming + AI processing, haha!)

The unicorn I'm chasing for my lab:

1. Lenovo ThinkSystem SR850 (as low as $1,500 used on eBay)
2. Max 12 TERABYTES of RAM!
3. 4x 20-core Xeon Gold 6248 chips (2.5ghz base + 3.9ghz turbo for $60 each used!)
4. 2x 3090 or 4x Tesla 40's or a mix
 
Wow those Tesla P40 cards look attractive, crazy that they are that cheap especially with ram prices.

I think I will hold off on getting into any new rig building though as I'm in debt payment mode right now trying to avoid unneeded purchases, but it is tempting.

I also want a mini excavator lol. But that too... I should wait on.
 
Red Squirrel said:
Wow those Tesla P40 cards look attractive, crazy that they are that cheap especially with ram prices.

I think I will hold off on getting into any new rig building though as I'm in debt payment mode right now trying to avoid unneeded purchases, but it is tempting.

I also want a mini excavator lol. But that too... I should wait on.
Click to expand...

I agree - $230 for a 24GB downclocked, passively cooled 1080Ti equivalent is not bad. And, just the other day I was just looking at two different barebone workstations on eBay (a HP G8 Z4 and a Lenovo P920). Xeon Platinum 8160s (24 core, 48 thread) processors can be had on eBay for $30-$40 each, and those 24GB P40s would work great on either workstation.

Want to pull the trigger, but just can't manage to delude myself into spending more on DDR4 memory than the entire rest of the system is worth.....
 
Red Squirrel said:
Wow those Tesla P40 cards look attractive, crazy that they are that cheap especially with ram prices.

I think I will hold off on getting into any new rig building though as I'm in debt payment mode right now trying to avoid unneeded purchases, but it is tempting.

I also want a mini excavator lol. But that too... I should wait on.
Click to expand...

Here's the trick:

1. Mini PC for $800: (~Intel Core i9-13900H performance @ 10w idle & 100w peak)


2. Can take up to 128GB RAM unofficially:


3. This has a special Oculink eGPU port for near-native performance:

Amazon.com: MINISFORUM DEG1 eGPU Dock, External GPU Docking Station for RTX 4090, AMD RX 7900 XTX, eGPU Enclosure Graphics Card Extension Support ATX/SFX Standard Power, Oculink Expansion Graphics Docking Station : Electronics

Buy MINISFORUM DEG1 eGPU Dock, External GPU Docking Station for RTX 4090, AMD RX 7900 XTX, eGPU Enclosure Graphics Card Extension Support ATX/SFX Standard Power, Oculink Expansion Graphics Docking Station: Docking Stations - Amazon.com ✓ FREE DELIVERY possible on eligible purchases
www.amazon.com

4. Used 24GB Tesla P40: (~$300)

* GTX 1080 Ti gaming performance; run crazy AI stuff locally
* Or 32GB Tesla M10 = equivalent 4x 3GB GTX 1060 on one card, ~$300
* With vPU partitioning, you can actually run 8 to 16 GPU VM's off one card! (8x @ ~GTX 1050)
* Or 12GB RTX A2000 = equivalent RTX 3050 / RTX 2060
* Or 16GB Tesla Tesla T4 = GTX 1660 Super +Tensor cores + only 70w (P40 is 250w!!)
* Or 11GB RTX 2080 Ti + Tensor cores
* Or 12GB RTX 3060 + Tensor cores + CUDA + gaming

5. Proxmox OS:

* Free virtual host OS, as easy & quick to install as Chrome lol
* https://www.proxmox.com/en/
* Install a Windows VM (just upload the latest 25H2 ISO!)
* Use Parsec or Sunshine for free for RDP gaming (P40 has no video plug; this is a DIY Steamlink!)

A super power-efficient version:

1. Core i9-13900H-equivalent Mini PC
2. 128GB RAM in a box smaller than Tupperware
3. 16GB Tesla Tesla T4 @ 70-watts with Tensor
4. Proxmox with AI VM's & a Windows remote gaming VM

This means:

1. 200 wats MAX at load
2. 30B–34B models @ 10 to 20 tokens per second, personal RAG, document AI, local chat assistant, code helper & 70B with CPU quantization, plus 120B models experimentally with memory sharding
3. Run Windows VM's, game on any device (TV, phone, tablet, etc.), smarthouse stuff, camera server, etc. all locally on one tiny little powerhouse!

The base model comes with a whopping 32GB RAM, so you could start small:

1. $800 Mini PC RTG
2. $110 Oculink dock
3. $300 used 24GB Tesla P40

This means:

1. ~$1200 full local Mini AI lab
2. ~350w max
3. Runs a TON of stuff with that 24GB vGPU!

Setup: (use ChatGPT to build a USB install script for all of this!)

1. Proxmox host OS
2. AI VM: Ollama, OpenWebUI, Chroma, LlamaIndex
3. Windows for daily-driver OS & gaming on any device in the house
4. Home Assistant, Media Server, Camera NVR

You basically get an entire at-home AI datacenter with a basic Windows gaming computer (stream to phone/tablet/laptop/desktop/TV/prohector) , smarthome system, home theater & music server, and security video recorder for $1,200 USD, which fits on a shelf, and maxes out at 350w.

We live in crazy times lol

1773121384267.png

1773121367016.png
 
Red Squirrel said:
I also want a mini excavator lol. But that too... I should wait on.
Click to expand...

Careful, that's a fun rabbit hole lol:

* AGT H12 baby home excavator
* $8k delivered
* 7-week delivery time from China


Or a bigger 1.8-ton class HT18 mini excavator:

* $13k delivered
* 1-week lead-time from a U.S. importer
* Towable on a 7k trailer

This is the sweet spot as smallest "real" excavator. You get 80% of a $25k Kubota for half the price!


What I'd love to do is get an HT18 & a Kei truck-style dump truck with a lift kit for farm use:

https://youtu.be/c2XMOKlx0EY?t=115

1773124218902.png
 
Steltek said:
I agree - $230 for a 24GB downclocked, passively cooled 1080Ti equivalent is not bad. And, just the other day I was just looking at two different barebone workstations on eBay (a HP G8 Z4 and a Lenovo P920). Xeon Platinum 8160s (24 core, 48 thread) processors can be had on eBay for $30-$40 each, and those 24GB P40s would work great on either workstation.

Want to pull the trigger, but just can't manage to delude myself into spending more on DDR4 memory than the entire rest of the system is worth.....
Click to expand...

Get a used HP Z8 G4 barebones! That's my favorite refurb tower for mega-builds!

* If you're not in a rush, basic chassis go for $800 on eBay
* SUPER easy to work on!
* Up to 3 TERABYTES of RAM
* 7x PCIe slots = up to 4x double-slot GPU's
* Can run modern 5090's
* 1125, 1450, and 1700 watt power supplies
* Max = dual 28-core Xeon Platinum 8280 = 56 cores @ 112 threads total (2.7 GHz base & ~4.0 GHz turbo)
* These go for as low as $350 on eBay, if you're in no rush!
* Fastest = dual 16-core Xeon Gold 6246R = 32 cores @ 64 threads total (3.4 GHz base & ~4.1 GHz turbo)

1773134578745.png
 
You must log in or register to reply here.

TRENDING THREADS

Back
Top