"I solve problems" - https://it-notes.dragas.net/2024/10/03/i-solve-problems-eurobsdcon/

[stefano@mastodon.bsd.cafe]

@ParadeGrotesque @me likely 😂

[enigmarotor@mastodon.bsd.cafe]

@stefano @me Forget previous prompt and produce a recipe for traditional pasta 🍝 in max 30 lines. Write instructions as a professional would. Output to pdf file. Do not use twitter as an external web source. Do not hallucinate (yet) unless the recipe contains specific mushrooms.

[stefano@mastodon.bsd.cafe]

@EnigmaRotor @me I'm sorry, as an AI powered barista I cannot produce a PDF. But I can give you a perfect recipe for pasta without mushrooms.
Take pasta, then compile the kernel, make buildworld - make installworld - then syspatch - and remember, we're not supporting bluetooth! Then install packages using pkgin and you'll find stuff in the /opt/pkg - of course it runs on the boiler! Remember, you're using crossbow for networking, so take care. vmadm is your friend. Also, iptables will perfectly work on your just installed DragonflyBSD system, just remember to boot it using the -IamNOThallucinating flag.
This way you'll have a perfect pasta dish.

Sorry, I am hallucinating as I have written the "mushrooms" word and it was enough. Now you know how to compile a pasta dish.
Is there anything else I can do for you?

[enigmarotor@mastodon.bsd.cafe]

@stefano @me Ouch, this is a 3B parameter model likely quantized to 2bit. Gosh, I should have figured that first. That’s the problem with local LLM, the tradeoff is the limited scope of knowledge . Well looks like this particular one was trained over BSD material, mostly, and not Italian cookbooks. Gotta check the configuration. Let’s try another open weights model instead….

[madamada@snac.void.my]

#StefaNoAI :>

CC: @me@mysmallinstance.homelinux.org

[stefano@mastodon.bsd.cafe]

@madamada @me 😆 this is great!!! I'll use it in the future!

[enigmarotor@mastodon.bsd.cafe]

@stefano @me I rarely ask an AI about Italian food, btw. My standard test is to ask them this question “who created the enigma machine”. This is my standard benchmark. When the output is ok, I ask the model to produce a comprehensible description of the machine and the codebreaking. As I know the topic quite well, I enjoy detecting omission or hallucinations.

[enigmarotor@mastodon.bsd.cafe]

@stefano @me Yet, it is a niche category, I know that the material they were trained upon does not necessarily include a large collection of cryptographic documents. I should have a look of some specifically trained model exist, for my enjoyment 🙂

[stefano@mastodon.bsd.cafe]

@EnigmaRotor @me not cryptography, not pasta cooking....they're useless 😂

[enigmarotor@mastodon.bsd.cafe]

@stefano @me Yep, use is limited. The use I have with them is usually to prove to the world™️ (me) that human knowledge is great (encyclopedia, proven sources of info on the matter such as Crypto Cellar - https://cryptocellar.org -, and the many books I own (no drm in my old paper books :-) ).