How to get started with AI?

W

wolffnx

Ok..I'am not a big fan of AI
but ,at least I want to learn, so
I dont expect a full howto from cero to %100 , but the guys who use it
What OS use for fully develop or get started? Can I do it in FreeBSD?
and if thereis a tutorial "AI for dummies"? .. I know that exists,but better listen it for experiences
Thanks
 
cracauer@

Thread 'Running a GPU-accelerated LLM on FreeBSD (2-line howto)'

Code: 
pkg install llama-cpp
llama-server \
        --host `hostname` \
        --port 8080 \
        --ctx-size $((64 * 1024)) \
        --temp 0.6 \
        --top-p 0.95 \
        --top-k 20 \
        -hf bartowski/Qwen_Qwen3.5-27B-GGUF:Q6_K_L

This is assuming you already have a working graphics card driver with Vulkan. You can test with `vulkaninfo --summary` from pkg vulkan-tools. Working Vulkan is in the binary NVidia drivers and the AMD drivers (including CPU-integrated GPU). Dunno about Intel GPUs.

This will install the llama.cpp LLM runner, which is an intermediate difficulty...
  • Like
  • Thanks
 
I use Claude AI in a browser, which means you can do it on any platform. I'm not a "professional user" of it though.
You can upload your files/projects and ask questions using normal English (and I believe, español también).
 
I recommend to ask a LLM like Claude Sonnet or Deepseek [1]. They will tell you. Ask "What are you for?" and "What can you do for me?". After that, ask them "What makes a good promt?". And so your journey starts.

If you have enough RAM and no dGPU like me (96GB RAM), you can run a local LLM (downloaded from hugginface) with some access to tools but not to internet (this needs other tools or agents - a LLM can help you to improve your tech-stack). I would recommend unsolth Qwen3.6-35B-A3B with a Quant >5. It gives ~24 tokens/sec which seems to be enough for basic tasks. I use my own, more recent version of llama-server as from pkg source (shameless self-reference). I can start a session in my browser after:
Code: 
GGML_VK_VISIBLE_DEVICES="" llama-server \
  -m "$model_path" \
  --host 127.0.0.1 --port 8081 \
  --ctx-size 262144 \
  --threads 8 \
  --numa isolate \
  --mlock \
  --batch-size 4096 \
  -ngl 0 \
  --tools all \
  --reasoning on

As far as I see the AI learning curve is low. You will learn some AI vocabulary / concepts like RAG and LoRA. But most importand: use LLM as a tool. Andrej Karpathy put it well: You can outsource your thinking but you cannot outsource your understanding.


[1] With Claude, you only have a limited number of credits available in the free version, but that’s more than enough to get you started. With Deepseek, I’ve never run out of credits. Avoid ChatGPT because it is too chatty - and I find it mostly useless for basic tasks. What makes ChatGPT usless for me is the endless loop of:
Code: 
user:      Please answer my question / solve this task.
ChatGPT:   here is the solution
user:      you are wrong at ...
ChatGPT:   thanks, now this is the correct solution
<start endless loop with "you are wrong at ...">
 
Some AI services are expected to be a jack of all trades, some are specialized.

Get an account with a web service? Sure, did that lots of times, now I have over a hundred passwords in a phone app.

Prepare to have a conversation on a specific topic? That's no different than walking into a shop and having a conversation with a specialist, face to face.

Not every service can easily determine that you're a complete moron who needs a lot of hand holding. Certainly not right away. But imagine you got questions about Win98, and you go to a chatbot whose specialty is FreeBSD. Even if the chatbot is polite, how long will the conversation go on before the chatbot recommends you go to a different chatbot, one who can actually answer questions about win98?

If you wanna train your own offline AI, good luck: AI Hallucination is a good term to know.

If you actually read Wikipedia's entry that I linked to, you'll learn about jaw-dropping stuff, like how AI-prepared stuff is not admissible as evidence in a court of law.

I think it's more about attitude than digital mechanics: If you treat AI as a tool that helps you think things through, you'll get better results than if you treat AI as free labor that will do your thinking for you.

An AI can produce an image that portrays something that is historically and factually inaccurate. Extending the same idea, an AI service can be used to make an animation to be posted on Youtube.

Has anyone heard of thispersondoesnotexist . com ? A web site that generates faces, supposedly using AI. A few years ago, when I was on Discord, I had to lecture someone on the dangers of taking that web site's results as gospel. A couple level-headed users helped me deliver that lecture, so I think that kid finally realized the importance of taking stupid shit like that site with a grain of salt, instead of accepting incomplete and potentially dangerous information with alarming alacrity.
 
Thanks to all , I have a more clear vision
In the past I worked in a proyect(network and aws part), but is was an AI proyect
they offer a service when recorded cameras footage(bussines,street,etc) to recognize attacks and violence against women,
the developer use a python or too..python modules to analize the videos,but the database has to be trained

the GPU resources was soooo expensive , now I know why
the main reason of this post is , I'am looking for a job, an AI is the leader word today..so , maybe is time to add resources to my brain
 
wolffnx said:
... now I know why
the main reason of this post is , I'am looking for a job, an AI is the leader word today..so , maybe is time to add resources to my brain
Click to expand...
I think it's extremely difficult to find a job in the AI field; there's a lot of competition. My son is a professional AI engineer (Promt Engineer). Much to my regret, he's not a programmer. But it's not all lost for him; he's not even 30 yet. Finding work is extremely difficult (he develops websites and chatbots - https://palmbit.media/ ). His saving grace is that he's an electronics engineer—he repairs computers (which is still a steady source of income). :-/
 
Yes, developing AI engines requires a strong mathematical background. There's less competition here, for obvious reasons (I clarified the wording in my previous post).
 
There should be a community attempt that achieves the same reults but doesn't depend on any proprietary/undocumented hardware. I don't believe the parallel processing GPU requirement to be crucial. A distributed p2p network should be able to reach the same performance by combining the resources of multiple nodes. A local GPU could be a shared resource, if that adds something. No need for a central corporate datacenter.
 
You must log in or register to reply here.
Back
Top