Post New

Here is the output of a chat where I ask DeepSeek(R1) to prepare a detailed financial report on ISRG based on fundamentals.
https://www.perplexity.ai/search/prepare-a-financi...
Earlier I had asked the same question of ChatGPT_o3mini and perplexity.ai, but I hadn't planned on doing a comparison so didn't save those chats.

I initially tried ChatGPT but it made obvious errors, so I then tried perplexity.ai, it too made obvious but different errors.

By the time I reached DeepSeek I was both frustrated (and amazed) how these "AI's" can so confidently spout garbage.

A big issue is that people will believe this stuff.

A bigger issue is that AI's are getting embedded in enterprise-wide software, so garbage analogous to this won't be very visible, but actions will be taken based on it.

As you can see in the chat, I was by this time very frustrated.

But it's more useful to always be very polite to a chatbot, you'll get better results (google it if you don't believe me).
This is due to a technical issue, not an anthropomorphic issue.

Your responses are entered in the AI's "context window" and help to prompt the AI responses based on its training data.
There are more instances of conversations in training data where problems get quickly resolved that include phrases like "please review this inconsistency ..." than conversations that include phrases like "you're incompetent, this part is inconsistent!"

Also the human-like phrases you see at the start of CoT (chain of thought) conversations, such as with DeepSeek R1, like "Ok, let's take this step by step" are not eye-candy intended to make the AI seem more friendly and doing some thinking. They are manually inserted by the wrapper surrounding the chatbot to get those phrases in the context window. It turns out this prompts more careful "thought" based on its training data than if those words weren't in the context window. So they are manually inserted.

Post New

AI Impacts FAQ

---

Contact Shrewd'm
Contact the developer of these message boards.