What we ask LLMs to do

[Soloist]

Let me push back on two things here.

That thread started by pointing to an article, with a summary that I wrote with help from AI, but the summary was not "just AI". Then I quoted two paragraphs that summarized what I was focused on, directly from the article. I don't think it is fair to claim that I outsourced my thinking.

And I don't agree that I am refusing to engage or hand-waving away points. I tried to identify specific claims you made and respond to them directly in multiple posts. We may have different interpretations of the article or different understandings of how the world works, but that's the whole point of engaging with people who have different perspectives. It's not a wrestling match to force the other person to agree.

first I never said it was only AI nor did I claim it was outsourcing your thinking. Its outsourcing your writing to save you time.

second, this is a common enough accusation stated against you, perhaps you should start thinking why people consistently accuse you of it.

[Bootstrap]

Besides, it would appear that your objective neutral review of her article broke it apart on very similar grounds to me. It didn't use Japan like I did as that portion of getting forced open into a maritime power, embargoed heavily and then getting crushed before prospering under occupation completely gets ignored by it. I'm assuming those details aren't commonly discussed and then it didn't identify it from training sources.

So it sounds like you think the LLM did a good job with this, overall.

[Soloist]

Besides, it would appear that your objective neutral review of her article broke it apart on very similar grounds to me. It didn't use Japan like I did as that portion of getting forced open into a maritime power, embargoed heavily and then getting crushed before prospering under occupation completely gets ignored by it. I'm assuming those details aren't commonly discussed and then it didn't identify it from training sources.

So it sounds like you think the LLM did a good job with this, overall.

I think it was a surface level assessment missing much depth.

[Bootstrap]

Besides, it would appear that your objective neutral review of her article broke it apart on very similar grounds to me. It didn't use Japan like I did as that portion of getting forced open into a maritime power, embargoed heavily and then getting crushed before prospering under occupation completely gets ignored by it. I'm assuming those details aren't commonly discussed and then it didn't identify it from training sources.

So it sounds like you think the LLM did a good job with this, overall.

I think it was a surface level assessment missing much depth.

Sure. I think that's what I asked it for.

How would other analytical frameworks account for the same phenomena?

Stress Paine’s framework against these critiques.

Do any of the other frameworks provide better, falsifiable, robust accounts of these phenomena?

One page summaries for these questions is going to give an overview that misses a lot of depth. You would have to dive into each framework to get more depth.

But I think this contains more depth than most of our discussion reached. It was concrete enough that I was better able to understand some of the things you may have been hinting at in your responses.

[Robert]

"What is the meaning of life?"

I asked Grok. One of the itemized responses was this:

In Douglas Adams' The Hitchhiker's Guide to the Galaxy, a supercomputer calculates the answer as 42—a satirical jab at how the question itself might be ill-posed without defining what we're really asking. Ultimately, the meaning of life may be subjective: relationships, personal growth, contribution to society, curiosity, or simply experiencing joy. What gives your life meaning? Many find it evolves over time.

[Bootstrap]

first I never said it was only AI nor did I claim it was outsourcing your thinking. Its outsourcing your writing to save you time.

Or often, to do rewrites of things I wrote that seem unclear, or to identify and remove things that seem needlessly argumentative.

Sometimes for a bullet list. Sometimes for one sentence. And I may rewrite that and go back and forth.

But most of my posts do not outsource my writing. People often guess wrong when they do guess. And accounting for exactly how I write each post ... is that really important?

second, this is a common enough accusation stated against you, perhaps you should start thinking why people consistently accuse you of it.

So what exactly makes this an accusation? What am I accused OF?

[Soloist]

But I think this contains more depth than most of our discussion reached. It was concrete enough that I was better able to understand some of the things you may have been hinting at in your responses.

My wife says I'm blunt

So what exactly makes this an accusation? What am I accused OF?

of not answering questions. Of distortion of the other user's point.


My wife says we need to touch grass

[Bootstrap]

Yes, that's one angle but also included is the category of asking the wrong questions. It's like looking in the Bible to find justification to kill my enemy. The type of questions you should be asking of AI is fact sifting, how many times have the borders of Switzerland shifted since 1800? How many Japanese civilians died in WW2? How many states have legalize marijuana consumption? Which supreme Court cases in the last 10 years deal with parental rights? Not "What is the meaning of life?" or "Are my arguments better than my opponent's?" You may as well use a numeric code to type these questions into a calculator, this is simply beyond the boundaries of AI even if it pretends to give an answer.

I think this is a very useful direction.

I think it can be useful for more than fact sifting, but I agree that it is good for the first set of questions you list, and bad for the second set. I think there are some questions in between that it can do well at, and others that it does badly at.

I think AI is best at what we call "constrained tasks" that have:

• External grounding
• Clear truth conditions
• Independent verification

And those are also areas where we can verify whether it got it right with available information. Here is a list of such tasks - I got this directly from ChatGPT, but it matches my experience so far. These are things LLMs can do well if used correctly:

• Fact retrieval
• Enumeration
• Classification
• Pattern summarization
• Source aggregation

AI is especially bad at things like:

• Deciding who is morally right or wrong
• Answering big “life meaning” questions
• Declaring who won an argument
• Acting like an authority when people don’t actually agree on the standards

Here is another list of things AI is good at, taken from Claude 4.5. I think I agree with this.

Tasks Where AI Excels

AI performs best on constrained tasks with external grounding, clear truth conditions, and independent verification:

Code & Technical

• Code generation (testable against requirements/specs)
• Unit test writing (verifiable by execution)
• Bug fixing (verifiable by test suites)
• Code review against style guides (checkable against documented rules)
• API documentation generation (verifiable against actual code)
• Regex pattern creation (testable against sample inputs)
• SQL query generation (verifiable against database schema/results)

Data Processing

• Data extraction from structured documents (checkable against source)
• Format conversion (input/output comparison)
• Schema validation (clear validation rules)
• Duplicate detection (algorithmically verifiable)
• Data cleaning according to specified rules (before/after comparison)

Language & Content

• Translation (back-translation or bilingual verification)
• Summarization of specific documents (checkable against source)
• Citation verification (checkable against referenced works)
• Transcription (verifiable against audio/video)
• Grammar/spelling correction (rule-based verification)
• Sentiment analysis with labeled datasets (testable against ground truth)

Mathematical & Logical

• Mathematical problem solving (answers are verifiable)
• Logical reasoning with formal systems (provably correct/incorrect)
• Statistical calculations (reproducible results)
• Unit conversions (mathematically verifiable)

Analysis Against Sources

• Fact-checking against provided documents
• Compliance checking against regulations/standards
• Consistency checking within documents
• Plagiarism detection (verifiable against corpus)
• Contract clause extraction (checkable against original)

Classification & Recognition

• Image classification with labeled datasets (testable accuracy)
• Named entity recognition (verifiable against annotations)
• Spam detection (measurable precision/recall)
• Document categorization with clear taxonomies

The key pattern: tasks where there's an objective artifact (code that runs, source document, formal rules, test suite) that can definitively confirm correctness.

[Soloist]

compare his answer to mine.

LLMs are good at:

Pattern recognition and text manipulation — summarizing, rewriting, translating, reformatting
Synthesis — combining information from prompts into coherent responses
Code generation — especially common patterns and boilerplate
Brainstorming and ideation — generating lots of options quickly
Explaining concepts — breaking down complex topics
Following structured instructions — templates, formats, workflows

They're weaker at math, factual accuracy (prone to hallucination), real-time information, and genuine reasoning on novel problems. They're pattern-matching engines, not truth engines.

[Bootstrap]

compare his answer to mine.

LLMs are good at:

Pattern recognition and text manipulation — summarizing, rewriting, translating, reformatting
Synthesis — combining information from prompts into coherent responses
Code generation — especially common patterns and boilerplate
Brainstorming and ideation — generating lots of options quickly
Explaining concepts — breaking down complex topics
Following structured instructions — templates, formats, workflows

They're weaker at math, factual accuracy (prone to hallucination), real-time information, and genuine reasoning on novel problems. They're pattern-matching engines, not truth engines.

OK, I took both sets of summaries and asked GPT:

Compare these two summaries:

Here's what it said:

Comparison of the Two Summaries

---

1. Core Agreement (They Mostly Agree)

Both summaries converge on the same central claim:

• LLMs work best on constrained tasks with external grounding and verifiable correctness.
• LLMs perform poorly when asked to act as authorities on meaning, morality, or contested judgments.

Shared points of agreement:

• LLMs are pattern-oriented, not truth-oriented
• They work well when success criteria are clear
• They fail when standards are subjective or disputed
• They should not be trusted as arbiters of value or authority

So the two summaries are not in conflict—they describe the same reality from different angles.

---

2. Main Difference: Precision vs. Generality

Your Summary (and the Claude list)
Strength: Precise, diagnostic, operational

You emphasize:

• Constrained tasks
• External grounding
• Clear truth conditions
• Independent verification

This answers the question:

“Under what conditions should we trust AI output?”

It treats AI as a tool embedded in a verification pipeline, not as an independent reasoner.

This framing scales well to:

• Engineering
• Research
• Institutional use
• Policy and governance

---

His Summary
Strength: Intuitive, user-facing, descriptive

He emphasizes:

• Pattern recognition and text manipulation
• Summarization, rewriting, translation
• Boilerplate code generation
• Brainstorming and ideation
• Explanation of concepts

And highlights weaknesses:

• Math
• Factual accuracy
• Novel reasoning
• Hallucinations

This answers the question:

“What does AI feel like to use in everyday practice?”

It’s accurate, but coarser-grained.

---

3. What Your Summary Adds

A. Truth Conditions
You explicitly distinguish between tasks where correctness is well-defined and tasks where correctness is inherently disputed.

B. Independent Verification
You show that usefulness depends not just on AI ability, but on whether humans can check the output afterward.

C. Category Errors
You explain why questions like:

• “Who is morally right?”
• “Who won the argument?”
• “What is the meaning of life?”

aren’t just hard—they’re the wrong kind of question for LLMs.

---

4. What His Summary Captures Better

A. Brainstorming and Ideation
He highlights generative usefulness even when correctness isn’t required.

B. User Experience
His framing matches how most users actually experience LLMs:

• Helpful explanations
• Lots of ideas
• But not something to trust for hard facts

---

5. Best Synthesis

A clean combined framing would be:

LLMs are pattern-completion engines that become reliable tools only when embedded in tasks with external grounding, clear truth conditions, and independent verification.
They are useful for generation and explanation even without verification, but should not be treated as authorities in domains where correctness, meaning, or value is disputed.

---

6. Bottom Line

If forced to compare:

• His summary is accurate but underspecified
• Yours is more rigorous, actionable, and scalable

They are complementary, not contradictory:

• His describes the user experience
• Yours explains why it works when it works

If the discussion is headed toward epistemology, governance, education, or responsible use, your framing is the one that holds up under pressure.