Hacker News Reader

Comment by anon373839 | Hacker News Reader

They’re useful for tasks that don’t require correctness: brainstorming, exploratory research, sketching out the vague shape of a solution to a problem.

They’re mediocre to awful at consistently following instructions, unless you have the fortune of having a task and domain that are well represented in the post-training data.

Yesterday I needed to generate ~500 filenames given a (human written) summary of each document’s contents. This seemed to be the perfect task to throw an LLM into a for loop. Yet it took three hours of prompt engineering to get a passable set of results - yes, I could’ve done it by hand in that time. Each iteration on the prompt revealed new blind spots, new ways in which the models would latch onto irrelevant details or gloss over the main point.

View on HN

Sometimes it helps to have it outline the task first with a prompt like "Suggest good guidelines for [task]" and then "Follow those guidelines for [task]" once you check to make sure the guidelines are good. LLMs aren't good at inventing processes on the fly, but they are good at writing plausible-sounding and often even correct processes, which they then follow.

Applied to ChatGPT's 4o:

> "Suggest good guidelines for generating ~500 filenames given a (human written) summary of each document’s contents"

> "Pretend you have this data and create 10 example file names."

> "Financial_Report_Q1_2024.xlsx (Summary: "Quarterly financial performance for Q1 2024, including revenue and expenses.")"

> "Marketing_Strategy_SocialMedia_2024.docx (Summary: "Comprehensive marketing strategy focusing on social media growth for 2024.")"

> "HR_Employee_Handbook_v2.1.pdf (Summary: "Updated version of the company employee handbook with revised policies.")"

And so on. In general, you have to either create the process or have it create a process and refine it if it isn't quite there. You can even use standard processes and guidelines when they're available. Then it usually does a great job.

Interesting strategy— I’ll check it out. I get a bit annoyed that I have to essentially trick a tool into giving me useful answers. What’s the chance that a nontechnical user that was just sold a subscription to a Magic answer-generating machine would even realize they needed to figure out how to do things like that? LLM chatbots are amazing but they’re lagging behind the marketing inferences, and that’s a recipe for frustration that providers need to address. I think a lot of these problems need to be fixed with product design and communication.

It does seem like some of the messaging gets away from their actual documentation: https://platform.openai.com/docs/guides/reasoning-best-pract...

The reasoning models improve on this kind of thing, but that just means they do even better when provided with or asked to create and follow a process. o3-mini-high, since it shows its work, didn't just spit out a process. It showed the process of creating the process in its reasoning block. It considered the format above in there and removed and added things while considering different levels of detail before the final version before producing anything.

In my cognitive landscape, abstract reasoning is a huge spike for me while linear reasoning is closer to average. I just don’t think I need help with all of that exploratory stuff — I need something that’s going to root me in reality with hard facts, figures, and processes. I can see how folks that don’t exist primarily in “shower thoughts” land, like I do, and have better intuition with the concrete linear stuff might get more use out of it than me. Conversely, perhaps I get a bit more out of traditional search engine workflows than others?

That would definitely account for the difference in my perception of their utility vs other peoples’.

My experience has been that converting abstract ideas into a set of actionable steps is one of the areas where AI tooling is really useful. So for example, if I want to generate some music programmatically for a small demo, it gives me concrete suggestions of the existing tool landscape and it helps me generate a plan of action. Then I fill in the remaining gaps myself in order to execute.

Thanks— I’ll play around with that workflow.

Reply on news.ycombinator.com