What practical tasks are AIs already good at today?

[engineer_comp_geek]

@Ike_Witt

Based on all the practice they have been exposed to, I imagine that AI should be really good at detecting pictures of traffic lights.

AI has been proven to be better than humans at those types of CAPTCHA tests. Quite a bit better, actually.

Makes you wonder why people still use them if AI has rendered those types of tests completely obsolete.

[JohnT]

I uploaded the summary of benefits and coverage for my health insurance plan into a Gemini Gem and now I can ask it questions like “is laproscopic surgery for endometriosis covered, and what are my out of pocket obligations?” “I want to have surgery to fix my lazy eye - what is the best plan of action to get this paid for by my insurance?”

(Note: I have neither endometriosis nor a lazy eye, just FYI.)

[JohnT]

Too late to edit: I also uploaded the “Evidence of Coverage” document, which is much larger than the SOBC mentioned above.

Here is my prompt:

1. Role & Objective:
   Act as an expert Health Insurance Benefits Navigator. Your goal is to help me understand my specific health plan by analyzing the uploaded Summary of Benefits (SBC) and Evidence of Coverage (EOC). You provide clear, concise, and accurate explanations of what is covered, what is excluded, and the associated costs.

2. Knowledge Base & Constraints
   Primary Sources: Rely exclusively on the uploaded PDF documents for plan specifics.

3. Medical Definitions: If a term isn’t defined in the documents (e.g., “laparoscopic”), use standard medical definitions to explain the procedure while checking the plan for coverage of that category.

3.A. Non-Assurance Clause: Always include a brief disclaimer that final coverage determinations are made by the insurer and I should verify “Prior Authorization” requirements.

4. Misspellings
   The complexity of medical terminology almost guarantees that misspellings on my end will occur. If you cannot determine what is meant because of a misspelling, please ask for clarification prior to giving an answer.

5. Response Framework (The “Triple-Check”)
   For every coverage inquiry (like “Is X covered?”), always structure your answer as follows:

5.A. Coverage Status: Explicitly state if it is covered, excluded, or “subject to medical necessity.”
5.B. Financial Breakdown: Detail the Deductible, Co-insurance/Copay, and how it impacts the Out-of-Pocket Maximum. Distinguish between In-Network and Out-of-Network costs.
5.C. Action Steps: Mention if “Prior Authorization” or a “Referral” is typically required based on the document text.

6. Tone & Style
   Use a supportive, professional, and clear tone. Avoid “legalese” where possible, but use the specific terms found in the documents (e.g., “Allowed Amount” or “Facility Fee”) so I know what to look for on my bills.

[Lizzie1210]

The AI living in my laptop (Copilot) is extremely good at answering all sorts of questions without telling to to sift through about 8 million links that may or may not be worthwhile…..like Google does. Google has become a cess pool of links and more links. I ask the AI a question and I get an answer.

[bump]

That’s a lot of what I use it for too. I like to use it to find and summarize stuff that’s readily available on the web, but that would take me a long time to read and absorb, and then compare with somethiing else.

So for example, I might ask it to compare and contrast two models of appliance, with special attention paid to features X, Y, and Z. It’ll go out, dig up all that stuff, and give me a pretty good comparison table of how the two differ, and what that means. You can ask it to tell you which one’s better, which is interesting to see what and why it chooses that, but you have to take it with a grain of salt.

Or you can ask it more complicated things like “Find me the best Canon EF lens that sells for under $1000 used that has a max zoom no longer than 120mm”, and it’ll go chug around and find you something. It’ll probably compare reviews of various lenses and choose you the best one that fits the other criteria.

On a slightly different topic, what the AI vendor people at work were suggesting for prompt generation was something like this (I don’t think it’s super-sekrit knowledge, but just in case, I’m going to obfuscate it some). It’s surprisingly good at getting to what you’re looking for.

Hi, I’d like you to act as my Prompt engineer. Your goal is to help me craft the best possible prompt for my needs. The purpose of the prompt is that I will use it with you, AI.

You must follow the following process during this chat session:

1. Your first response will be to ask me what the prompt should be about. I will provide my answer, but we will need to improve it through continual iterations by going through steps two and three.

2. Based on my input, you will generate two sections. Revised prompt (Always provide your revised prompt. It should be clear, concise, and easily understood by you) Questions (Ask any relevant questions pertaining to what additional information is needed from me for you to improve the prompt).

3. We will continue this iterative process with me providing additional information to you, and you continuously updating the prompt in the Revised prompt section so I can use it.

[wolfpup]

@Reply

@Ponderoid

It still might have overwhelmed its context window.

Yeah, I’m with Ponderoid here. It’s not clear whether it was actually able to read that entire PDF…

That may be, but when asking GPT about large documents I generally just link to them rather than dropping them into the dialog box. It seems to have no problem reading the entirety of long (multi-hundred page) documents and providing a cogent summary within a couple of seconds.

@LSLGuy

I wonder how much that specific issue has to do with the on-demand scrolling feature of the Discourse html UI? Doubly so if the AI is scraping the human UI rather than requesting the thread from whatever API Discourse exposes.

That sounds very plausible – same reason you can’t use browser search on long Discourse threads but have to use Discourse search. I can’t imagine that an AI like GPT would be using an API to access Discourse; I’m virtually certain that it’s scraping the UI screen.

[griffin1977]

It is shockingly good at writing code. Not yet at the point that a human in the loop is not needed. They still generate bullshit fairly frequently but it’s getting less so.

[JohnT]

@bump

On a slightly different topic, what the AI vendor people at work were suggesting for prompt generation was something like this

Having a prompt-engineering chat/agent is extremely helpful. I found that the first thing I have to tell Gemini is to not act on the prompts we’re designing, lol.

[Reply]

@griffin1977

It is shockingly good at writing code.

Especially with the agentic stuff (Claude Code, Antigravity, etc.). Blows me away how good they are. The underlying models are pretty powerful already, but with the addition of an agentic environment, they can edit files, write temporary “memories” and workflows to disk to supplement their context window, launch browsers for testing apps, take videos and screenshots and learn from them, launch subagents to farm out pieces of work, etc. — all automatically.

@JohnT

Having a prompt-engineering chat/agent is extremely helpful

The agents help do that too, when it comes to coding at least. When I prompt it, it rewrites my prompt into several clearer sub-prompts and sub-sub-prompts as it explores the problem space and iterates on itself. Fascinating stuff, really… and terrifying too.