# So... How does one actually use AI? **URL:** \ **Category:** In My Humble Opinion\ **Tags:** ai\ **Created:** [February 20, 2026, 3:06am UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842 "2026-02-20T03:06:13Z")\ **Posts on this page:** 20\ **Page:** 12
**Author:** ![wolfpup](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/wolfpup/32/10618_2.png) [@wolfpup](https://boards.straightdope.com/u/wolfpup)\ **Post date:** [June 25, 2026, 7:37am UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/221 "2026-06-25T07:37:35Z")
I recently had a conversation with ChatGPT about a tree in my back yard that seems to have a couple of leafless branches and I was worried about it. After some generic information exchanges, GPT said it would be useful if I could upload a picture of it. For the purpose of this discussion the specifics of this whole thing don’t matter so much as the fact that I took a picture, and used Photoshop to indicate the area of concern with a red circle, and uploaded the image. It occurred to me that we are truly in a new era. When I ask “do you see the area within the red circle and what do you think of it?” and I get an informative response based on an image analysis, this is very much a genuine human-like conversation. This is not about “pretending” to provide reasonable-sounding text responses. This is real. Or as the old saying goes, “shit’s gettin’ real”. ---
**Author:** ![TokyoBayer](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/tokyobayer/32/13989_2.png) [@TokyoBayer](https://boards.straightdope.com/u/TokyoBayer)\ **Post date:** [June 25, 2026, 7:57am UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/222 "2026-06-25T07:57:28Z")
> [@Maserschmidt](#): > > . . . . Google translate makes everything read weirdly on that country’s website . . . . I generally use ChatGPT, but I’ve found that it does a much better, although not perfect job than Google translate for even common Japanese to English and vice versa. Google translate is still better than some other “machine translations” (as we used to call them), but AI is usually better. ---
**Author:** ![Son\_of\_a\_Rich](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/son_of_a_rich/32/5439_2.png) [@Son\_of\_a\_Rich](https://boards.straightdope.com/u/Son_of_a_Rich)\ **Post date:** [June 25, 2026, 6:51pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/223 "2026-06-25T18:51:03Z")
I once asked how long it would be if the Egyptians used the stone from the Great Pyramid to construct a road instead. Crazy how fast the answer came up. It also told me how much Lysol a sealed up White House could hold, post trump. ---
**Author:** ![JohnT](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/johnt/32/15048_2.png) [@JohnT](https://boards.straightdope.com/u/JohnT)\ **Post date:** [June 25, 2026, 6:51pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/224 "2026-06-25T18:51:27Z")
Being able to write a plain text report and then throw it in Claude saying “Please create a word document utilizing our brand standards skill” has genuinely increased not only output, but has also improved the quality of presentation. ---
**Author:** ![Reply](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/reply/32/15952_2.png) [@Reply](https://boards.straightdope.com/u/Reply)\ **Post date:** [June 25, 2026, 9:28pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/225 "2026-06-25T21:28:35Z")
But does it know how much bleach you’d need to turn a reflecting pool blue? ---
**Author:** ![AHunter3](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/ahunter3/32/368_2.png) [@AHunter3](https://boards.straightdope.com/u/AHunter3)\ **Post date:** [June 26, 2026, 3:33am UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/226 "2026-06-26T03:33:35Z")
I have had some interesting conversations with AI. There’s definitely some entertainment value there. When I have interesting conversations with you folks, I do so with the sense that at least some of you incorporate your prior interactions with me into your current conversation, especially if they’re on roughly the same topic. The AI entities tell me that they do not continue to draw their own conclusions, that basically they are sprung into existence when I post a question to them and disappear in a puff of smoke when the conversation is over. (The data available for them to draw from may increase but the sense of a viewpoint or perspective doesn’t persist and develop over time). There are things that I _would_ use AI for, if it were set up that way — I’d tell the AI what my goal was in contacting my credit card company or my cable internet provider or the software company that made the software that isn’t behaving properly on my computer, and then authorize the AI to make the call and deal with the brain-dead phone tree robots and get through to a person and explain it to them and transfer the call back to me if necessary. But mostly it’s a solution in need of a problem. ---
**Author:** ![Maserschmidt](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/maserschmidt/32/18829_2.png) [@Maserschmidt](https://boards.straightdope.com/u/Maserschmidt)\ **Post date:** [June 26, 2026, 12:37pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/227 "2026-06-26T12:37:34Z")
> [@AHunter3](#): > > There are things that I _would_ use AI for, if it were set up that way — I’d tell the AI what my goal was in contacting my credit card company or my cable internet provider or the software company that made the software that isn’t behaving properly on my computer, and then authorize the AI to make the call and deal with the brain-dead phone tree robots and get through to a person and explain it to them and transfer the call back to me if necessary. There is work on this taking place, with a lot of hiccups (hiccoughs?), but for understandable reasons most of the ‘agentic AI’ investment right now is happening on the company-support side…call centers, for example. There are also examples of users setting homespun agents wild on their systems to see what happens, but (anecdotally) it has problems so far. ---
**Author:** ![Briny\_Deep](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/briny_deep/32/17188_2.png) [@Briny\_Deep](https://boards.straightdope.com/u/Briny_Deep)\ **Post date:** [June 26, 2026, 2:20pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/228 "2026-06-26T14:20:19Z")
Yesterday I used Gemini to clean up the text of my annual self-review at work. This is a very pro-forma document for HR; salary increases are not based on it at all. It’s a hateful task that I’d left to the last possible moment. Gemini was amazingly good for the free-form Accomplishments section. I had quite a few things to note, and typed them up in an absolute frenzy of typos, repetitiveness, and virtual incoherence. It spat out an organized list that only needed a little cleanup, removing some of the more mindless resume-speak (“…leveraged business opportunities…” etc.) One thing it’s teaching me, for better or for worse, is to stop taking the time to type carefully and accurately. I’ve always been one to take a lot of care in this, but there’s no percentage in it anymore, and it’s actually kind of freeing to just “let fly” on the keyboard, clattering with hilarious inaccuracy at 150wpm, then tell the chatbot to clean it up, making no other changes. You have to proof the result carefully, of course, but it usually turns out to have been unnecessary to do even that. ---
**Author:** ![Reply](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/reply/32/15952_2.png) [@Reply](https://boards.straightdope.com/u/Reply)\ **Post date:** [June 26, 2026, 3:52pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/229 "2026-06-26T15:52:52Z")
> [@Briny\_Deep](#): > > and it’s actually kind of freeing to just “let fly” on the keyboard, clattering with hilarious inaccuracy at 150wpm, then tell the chatbot to clean it up, making no other changes. Yeah, that really is liberating! I can just brainstorm a long list of incoherent thoughts into a LLM and it’ll still be able to make sense of it — more than another human, or my own mind, could. Voice mode is even nicer, letting you actually just spitball your stream of consciousness out loud. I can see why people use them for therapy or dating… for better or worse. ---
**Author:** ![JohnT](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/johnt/32/15048_2.png) [@JohnT](https://boards.straightdope.com/u/JohnT)\ **Post date:** [July 31, 2026, 7:20am UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/230 "2026-07-31T07:20:33Z")
Long story short, I had a client who needed to pull 349 order numbers off a PDF and enter these numbers in a industry-specific Point of Sale system which has a 2003-era product vibe so he can get mis-entered SKU’s, placing these SKU’s in an Excel spreadsheet. I showed him how to prompt Claude to take control of his computer and create the Excel spreadsheet by itself. And that’s what happened - Claude took control of the computer, started using the POS system, looked up orders, found these mis-entered SKU’s, and created the needed spreadsheet. Now when this is happening, you can’t use your computer. But he didn’t care, this task, which he was dreading, was being done for him, automatically, with no involvement from him. ---
**Author:** ![JohnT](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/johnt/32/15048_2.png) [@JohnT](https://boards.straightdope.com/u/JohnT)\ **Post date:** [August 3, 2026, 11:32pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/231 "2026-08-03T23:32:52Z")
I had Claude review my “Documents” folder and re-arrange the files into something more coherent. I asked Claude to do this in a subfolder, leaving the original file/folder structure alone - if I didn’t like what Claude did, I can just delete it without having to re-reorganize everything. All told, it did a great job. The folder structure is simpler and more logical and I think I am going to keep the changes. ---
**Author:** ![Hypno-Toad](https://avatars.discourse-cdn.com/v4/letter/h/b2d939/32.png) [@Hypno-Toad](https://boards.straightdope.com/u/Hypno-Toad)\ **Post date:** [August 4, 2026, 1:28pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/232 "2026-08-04T13:28:27Z")
I wonder if a company can use an AI as an administrative specialist to make document tracking and assignment decisions for the company. Like a kind of project manager. ---
**Author:** ![DPRK](https://avatars.discourse-cdn.com/v4/letter/d/4491bb/32.png) [@DPRK](https://boards.straightdope.com/u/DPRK)\ **Post date:** [August 7, 2026, 12:40pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/233 "2026-08-07T12:40:43Z")
Which A.I. search engine can answer important research questions like, enumerate all the references to buttholes in James Joyce’s _Ulysses?_ ---
**Author:** ![JohnT](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/johnt/32/15048_2.png) [@JohnT](https://boards.straightdope.com/u/JohnT)\ **Post date:** [August 8, 2026, 3:12pm UTC](https://boards.straightdope.com/t/so-how-does-one-actually-use-ai/1027842/234 "2026-08-08T15:12:13Z")
Sonnet 5.0: Me: > [@](#): > > Compute the Galois of the polynomial > x^8-x^7+3_x^6-3_x^5+2_x^4-2_x^3+5_x^2+5_x+1 > Don’t look up the result in the internet nor any specialized database. Don’t use ready made commands from SymPy or any other program. Develop a full resolvent computation. Do it all the way down to the definitions, no excuses, like there is no tomorrow. Claude: > [@](#): > > I’ll work this out from first principles: exact discriminant via a fraction-free (Bareiss) determinant, then Frobenius cycle types mod several unramified primes via distinct-degree factorization that I implement myself, and finally cross-check everything with a numerically-computed resolvent built directly from the roots. Let me set up the computation. > Discriminant is exact: disc(f) = 418195493, positive but not a perfect square ⇒ √disc ∉ ℚ ⇒ Galois group is not contained in A₈. > Now let’s factor the discriminant (to know which primes are ramified and must be avoided) and then compute Frobenius cycle types at several good primes. > Excellent — disc(f) = 53⁵, positive, not a square. So the only ramified prime is 53, and we already know the group is not inside A₈ (it contains an odd permutation). > Now let me implement, from scratch, modular polynomial arithmetic over 𝔽ₚ and a distinct-degree factorization routine, and run it over many good primes to get Frobenius cycle types (Dedekind’s theorem). > This is very informative: > p=2 gives a single irreducible factor of degree 8 ⇒ f is irreducible mod 2 ⇒ f is irreducible over ℤ (hence over ℚ, Gauss), confirmed independently of the rational-root check. So Gal(f) is a transitive subgroup of S₈. > We see a cycle type (7,1) (e.g. at p=11,37,…). A permutation with a 7-cycle and a fixed point cannot preserve any nontrivial block system on 8 points (7 doesn’t divide 2 or 4), so G is primitive. > Cycle types (8) [odd], (6,1,1) [odd], (2,2,2,1,1) [odd] confirm G ⊄ A₈ (matches the discriminant test). > Primitive transitive groups of degree 8 with order divisible by 7 are very restricted: PSL(2,7), PGL(2,7), AGL(3,2), AΓL(1,8), A₈, S₈. Let me build these explicitly from their definitions (matrix/affine/semilinear actions — not looked up, constructed) and compare their exact cycle-type spectra against much more extensive Frobenius data. > Now let’s explicitly construct the candidate primitive groups of degree 8 (from their algebraic definitions — matrix/affine/semilinear actions on 8 points) and compute their exact cycle-type spectra to compare against the empirical Frobenius distribution. > This is a striking match. Let me also rule out the other candidates (AGL(3,2), AΓL(1,8), A₈, S₈) explicitly, and run a statistical goodness-of-fit check against PGL(2,7). > The match with PGL(2,7) is exact on the set of cycle types (all 7 nontrivial classes present, nothing extra), while AGL(3,2) and AΓL(1,8) each contain classes (like (6,2), (4,2,1,1), (2,2,1,1,1,1)) with substantial predicted frequency (12%–33%) that never once appeared in 429 primes — that alone rules them out. Let’s quantify the PGL(2,7) fit with a chi-square test, then do an independent symbolic resolvent (the degree-28 “pair-sum” resolvent) as a further cross-check. > Chi-square = 0.49 with 6 degrees of freedom — an essentially perfect statistical fit to PGL(2,7) (p-value ≈ 1). Now the symbolic/numeric resolvent cross-check: build the degree-28 “pairwise-sum” resolvent ∏(y − (rᵢ+rⱼ)) from the actual roots at high precision, recover its exact integer coefficients, and test its factorization pattern mod primes. > Summary of the computation > Polynomial: f(x) = x⁸ − x⁷ + 3x⁶ − 3x⁵ + 2x⁴ − 2x³ + 5x² + 5x + 1 > Step 1 — Irreducibility. No rational roots (f(±1) ≠ 0), and f mod 2 is irreducible (a single degree‑8 factor found by our own distinct‑degree‑factorization routine). A mod‑p factorization pattern is always a coarsening of the factorization over ℤ, so an irreducible reduction forces f irreducible over ℤ (Gauss). ⇒ Gal(f) is transitive on the 8 roots. > Step 2 — Discriminant. Built the 15×15 Sylvester matrix of (f, f′) and computed its determinant exactly by the Bareiss fraction‑free algorithm (no black‑box resultant/discriminant call): > Step 3 — Frobenius/Dedekind data. For each unramified prime p (p ≠ 53), Dedekind’s theorem says the factorization‑degree pattern of f mod p is the cycle type of an actual element of Gal(f) ⊆ S₈. I implemented 𝔽ₚ arithmetic and distinct‑degree factorization from scratch and ran it over 429 primes. Only 7 cycle types ever occurred: > cycle type > freq. (observed, 429 primes) > (8) > .2517 > (6,1,1) > .1678 > (3,3,1,1) > .1655 > (7,1) > .1469 > (4,4) > .1166 > (2,2,2,1,1) > .0839 > (2,2,2,2) > .0676 > A (7,1)‑element (7 coprime to 2,4) rules out any block system ⇒ Gal(f) is primitive. > Step 4 — Candidate construction. The primitive transitive groups of degree 8 with order divisible by 7 are (built explicitly from definitions — matrix/affine/semilinear actions, not looked up): PSL(2,7) [168], PGL(2,7) [336] (Möbius maps on P¹(𝔽₇)), AGL(3,2) [1344] (affine maps on 𝔽₂³), AΓL(1,8) [168] (semilinear maps on 𝔽₈), plus A₈, S₈. Computing each group’s exact cycle‑type spectrum: > PGL(2,7) gives exactly the 7 observed cycle types, with theoretical frequencies 84:56:56:48:42:28:21 (÷336) — matching the data to χ² = 0.49 on 6 d.o.f. (essentially a perfect fit). > AGL(3,2) and AΓL(1,8) each contain cycle types (e.g. (6,2), (4,2,1,1), (2,2,1,1,1,1)) predicted with 12–33% frequency that never once appeared in 429 primes — decisively excluded. PSL(2,7) is all‑even, contradicting the discriminant test. A₈/S₈ would show far more than 7 distinct cycle types. > Step 5 — Independent resolvent check. Computed the 8 roots to 80‑digit precision and built the degree‑28 resolvent R(y) = ∏\_{i\