A motivation to share this show was the episode 4454: AI, It's a Trap! , by Archer72.

This is a talk for Hacker Public Radio about characteristics of LLM and how you can use them for your best.

1 ^st characteristic,

LLMs (Large Language Models) are designed to always give an answer that convinces. That does not mean they’re right.

Use it for the best: don’t trust the answer. If you didn’t know the answer beforehand (using AI to obtain a better or faster result), then, verify it independently.

 

2 ^nd characteristic:

commercial LLMs, many times, do not think too much in the face of simple and short prompts. That’s a barrier, probably, against wasting resources.

How to use it for the best? Two things: 1) show the AI what you’ve found by your own, and proceed to tell what EXACTLY you need help with, on the basis of what you were already able to think. 2) Learn to follow-up. Suggested follow-ups are not good, prefer instead to talk to the AI, as a real conversation, to get confirmation, or contrast something you disagree with etc.

(That is a characteristic that resembles autism a bit, so the name of the show: to not like when someone comes wanting something from you without doing any effort to obtain it; so you also do not feel like doing for them.)

 

3 ^rd characteristic:

LLMs are not accountable. In general (most AIs), you don’t always know all the sources for the information given. (They do not know what they are telling, nor understand the value of sources.) And you can expect different outputs for the same inputs. (They are not deterministic.)

Simply saying, they’re crazy machines to generate content they don’t, cognitively, understand, but that convinces humans, because they use human content and patterns.

So, if you use AI for something, especially any serious purpose, remember : the result you get is of your responsibility, don’t expect to be excused for your words “because I was helped by AI, they did it”.

 

4 ^th characteristic:

AI have a pattern. We can suspect that something was generated by AI, and no one likes to be answered by one if that possibility was not explicitly told.

What to do about it? I suggest you don’t use AI-generated content with someone you estimate. It is rude.

 

5 ^th characteristic :

LLMs can give great results with less effort than you would need to apply with no machine at your side. What the AI have done is, by exclusion, not what you have done.

Use it to your best: you do not learn if someone does the job for you. For tasks you know well, and want to accelerate or remove repetitive steps, you may count on AI (or, better, count on a specific software, that can be programmed and give accountable results because you know exactly what is being done with the input). But for the intellectual work, if you like to think, if you’re good on written expression, the LLM may get passable results in less time… at the price of removing from you the chance to dedicate yourself to the comprehension and production. So , it’s not always a matter of producing like a king — effectiveness above all and everything —; it may be important to satisfy what you value as meaningful — learning, maybe; or feeling the satisfaction of the conclusion —, so that you can sustainably follow a routine that is not a pain on the eye (expression to mean something disgust).

 

That’s all, folks!, for today. Let me know if you’d like more content on this; possibly I (or Archer72, or another friend) could bring some more opinion on AI if it is of your interest.

Antoine was here. Bye bye.