I asked the Forum Helper bot (gpt4o) the following question and got the wrong answer :grey_question: If it takes 30 minutes to dry 10 wet towels laid out in the sun, how long will it take to dry 15 wet towels? Forum bot answer: :x: “It will take 45 minutes to dry 15 wet towels under the same c…

Forum helper is tweaked to RAG with the existing forum content, but your question has nothing to do with it. In this case the “creative” persona, which is the vanilla LLM is a better fit.

Makes good sense.! .. thanks Creative Bot: “If there are more towels than can be laid out without overlapping or if the space becomes crowded, reducing the exposure to sunlight, then it might take longer. Assuming conditions remain optimal and similar to drying 10 towels, the drying time should re…

also worth noting, LLMs have an aspect of randomness which is set using temperature and top p. If you want more predictable / less random results, create a persona with decreased randomness. see: AI bot - Agents

I’m having good luck configuring various open and commercial LLMs that can be mixed and matched with standard and custom AI bot personas.. and the idea of having multiple ‘agents’ that each have their own RAG context, categories, groups, etc. way cool.. opens up a ton of use cases ! so.. wow ! th…

Yes. the earlier models are esp bad at even simple math / simple logic for instance with this simple puzzle: I live in an apartment building with a broken elevator. If I enter it and push the button for a floor with an even number, it goes to that floor but doesn’t open the door and then goes ba…

At the end though LLMs can not do math properly, if you give it access to JavaScript though… it is a totally different story: https://github.com/discourse/discourse-ai/pull/630

But that’s not the issue here, surely? The issue is with the reasoning. Giving the LLM access to a calculator certainly helps ( Chatbot has had that access for a long time) but does not make up for poor logic or reasoning: doing the wrong calculation “correctly” is arguably as bad as doing a wrong …

Math vs logic .. a mind-bending aspect of the “two different answers topic” IMO, thanks Please consider… I want to a buy a product online and I see three sellers offer the same product – all have customer reviews: The first has 10 reviews, all positive The second has 50 reviews, 48 positive The…

Footnote.. Not being one to leave well enough alone.. I added context on the Laplace theory of probabilities to the Discourse AI bot that got it ‘wrong’ ( Gemini based ) [image] It still got it wrong, even though it had learned about the theory from context. Then I asked it to apply it’s knowle…

They don’t learn. That’s true with OpenAI models, I don’t know others. A bot can or cannot use given information depending of tokens, algoritm and some other mystical things. But we can point it to right direction. And yet after five or so answers it has forgot that.

Two different answers from gpt4o - one right, one wrong.. !?

Support

Jagster (Jakke Flemming) May 16, 2024, 7:37am 6

And there is one universal truth too: language model is different thing than doing math and models are notoriously bad in math without external tools.

Topic		Replies	Views
Gpt3.5_bot user Support	2	579	February 3, 2024
Request: AI Bot Chat Threads Support (Threaded/Branched Conversations with AI) Feature ai	0	146	December 22, 2025
Answer feature Feature	1	689	April 1, 2023
[Ai Bot] Add user token tracking, custom AI personas, max context posts, document loading, custom API URLs, and localized chat titles Feature ai , ai-bot	1	751	March 22, 2024
Star common AI conversations Announcements ai	2	356	August 11, 2026

Two different answers from gpt4o - one right, one wrong.. !?

Related topics