So... How does one actually use AI?

In My Humble Opinion
321

Researching out thoughts and hypotheses on subjects ranging from phylogenetics and behavioral genetics, information theory, exercise physiology evidence, healthcare M&A, interpreting scientific literature, political/institutional analysis … I’m less collecting citations than checking its citations and further exploring primary sources relevant to the subjects of inquiry.

Also recipes and if kamut works as a replacement for farro, suggestion of which wine is nicer to take to a friend’s lake house … :slightly_smiling_face:

322

I think you might actually benefit from the better paid models in that case. They generally are able to draw stronger connections between related or even dissimilar topics, especially for “general” questions rather than super specific asks. The vaguer the question, I think the more you would benefit from a powerful model.

I sent you a free trial to Claude Pro in a DM. Personally I find it more even-keeled and trustworthy than OpenAI models, but I also haven’t used ChatGPT in a few months so I don’t know how much that has changed. Benchmarks generally suggest Claude is a bit more trustworthy, too.

If I were you I’d try both that and ChatGPT Plus side by side for the same questions for a week, and see if either one is noticeably better than the free tier.

(Oh, make sure you choose the stronger models after you subscribe)

323

By the way, if any of this is quantitative, the paid models are also pretty good at making simulations, running statistical analyses, machine learning, etc. on big datasets. And in a harness, it can spawn multiple research subagents to collect sources, read data tables, transcribe charts, etc.

I’m trying to understand the cyclical flows of my local housing market, for example, and using Claude Code with a main Fable agent to spawn several research sub-agents who will independently crawl different parts of the web, bring their results back to the daddy Fable, who will then assemble it all into an interactive report-simulator-explainer thingy for me. They’ll look at news articles, real estate sites, census data, wildfire data, insurance data, city/county GIS, development permits, blah blah blah, and then put it altogether and try to make sense of it all, then make a interactive chart that I can play with to simulate different possible scenarios with an explanation how that affects sellers, buyers, insurers, developers, investors, etc.

It’s this sort of multi-agentic collaboration that makes the harnesses (and paid better models) really powerful, even for tasks that are not really about any sort of coding or image generation. I mean in this case, there’s still a bit of that (the website and visualizations I’m asking it to make), but even if you just wanted a plain-text report, the same fan-out workflow is still pretty amazing for that. (It also takes forever, hours usually, and costs a lot; I frequently run out of usage on the $100/mo plan)

It’s like being able to ask a whole think tank rather than just one expert.

324

Pretty impressive! More than my use case but yes definitely demonstrating how it can be a powerful tool in the hands of an intelligent user. Thanks.

325

Went to the Pink Boxxe yesterday, got a new toy for me and the missus.

Uh-oh! How do you charge this thing?

Gemini (to be fair, this was a Google image search): “Golly gee, that looks too adult for my youthful AI’s. I can’t give you any information, sorry, info on this topic is banned.”
Claude: “It’s right THERE, dude!”

Claude FTW.

326

Heh. I always wondered what they’d do in those circumstances. Or what it’d say if you accidentally left it on during evening fun time.

Didn’t ChatGPT add a porn mode earlier this year? I wonder if it’s less prudish. (Edit: Oh, hmm, they announced this but I think they rolled it back later and never fully launched it.)

And Grok… I don’t even want to know.

327

I have twice taken a run at Children of Ruin, the sequel to Adrian Tchaikovsky’s Children of Time, but quit a few chapters in because I didn’t want to spend a whole book with its incredibly annoying characters. Yesterday I asked Gemini “is anything going to change or is it that way the whole book is, limit your spoilers” and it told me that if I could just make it through five chapters I’d be fine, because perspectives change. And they did! So now I’m eight chapters in and enjoying it.

328

Huh, that’s actually a great idea. There are a few TV shows I kinda-sorted am interested in, but the first few episodes are a real slog. I’ll have to try that.

329

I picked up a book I’d put down several months ago about a third of the way through it, and realized I’d forgotten many details of who was who and who’d done what. I told AI what page I was on and to summarize, without spoilers. It did a great job. It also made me a list of the many characters with short descriptions up to that point.

330

This is amazing. Reverse engineering the encrypted firmware in a Dell monitor. Prior to my link time he explains how he would have to go about it and probably take a week. He uses a single sentence prompt to Claude Code and it does it in 15 minutes.

“I’m trying to find the firmware in this image, and I want to indentify the HDMI command parsers in the firmware. the firmware is potentially encypted”

331

There’s a lot I’m willing to do, but cluing an LLM into what gets me off is not one of them.

I recently decided that my health is on a slow trajectory to awful, and that in ten years (I’m 43) I want to feel better than I do now. I floated this by Claude, who helped me operationalize this and create a plan and tracking spreadsheet for the next three months. The idea is that we start with very small health behaviors and build on success with clearly articulated outcomes while sticking to the 10-year vision. I liked the idea of keeping some kind of long-term outlook in mind, I think it feels less daunting.

I don’t know what ultimate success would like, but I’m planning to recruit Claude to create some kind of baseline evaluation tool so I can measure progress over time. Including things like reviewing my cholesterol and triglycerides with every blood workup. I’m not giving it medical information, I’m just having it build the structure for recording this stuff myself.

It’s stupid how easy Claude makes this kind of thing. I also used it today to create a plan for improving accountability for people at work who do not get their shit to me on time. I had the basics in mind but it gave me a summary Word doc with clearly articulated processes.

This is because after years of having this issue I went to HR and my boss saw the light and asked me to give her a plan for executive leadership to follow. I’m annoyed this fell on me to figure out, but anyway it took about 10 minutes with Claude.

The thing that bothers me is that in a different world, one where I had the mental and emotional bandwidth, I could have figured this out on my own. But I don’t have that luxury right now.

If you are one of those people who likes to plan and organize and create procedures for self-improvement or general process improvement, you really can’t go wrong with an LLM. It compensates for my disastrous executive function and makes better tracking documents than I ever would. Since it’s not based on facts I have fewer qualms about using it for this purpose as opposed to doing research or looking for factual information.

332

That’s pretty cool. I see that he’s using Claude Code, but not an Anthropic model… GLM 5.3. Probably the Anthropic models would’ve triggered the cybersec safeguards and denied it?

Claude once cracked a video game’s encryption for me the same way (reverse engineering the binaries and locating the stored keys). Took it maybe 5 minutes.

This is probably why they have the cyber verification program; these models are just way, way, way too good at defeating human-written security safeguards.

333

So, out of curiosity, why not go “all the way” and do feed it your medical information (via Google Health or the Apple equivalent)? Maybe not all the way to having it monitor sexy time, but wouldn’t it be helpful to have it track your vitals over time, and maybe before/during/after workouts, to tweak the plan as you go?

I am considering taking a similar leap myself… two years ago, I didn’t trust LLMs with anything. A year ago, I limited it to only coding and work. As of now, though, it’s reading my whole 20-year Gmail history to do various analyses.

A health plan is next. I suspect it would do a better (or at least more thorough) job than my PCP and dietician… hmm. I’ll have to try it teach it to use MyChart (an electronic medical records system used by our local hospitals) and pull all the past visits and vitals. Good idea.

I wish I could give Claude similar powers on Android as on my computer so it can do more real-time tracking and analysis. (It’s severely hampered by Android security and settings limitations, both relative to Gemini and relative to a laptop OS).

It really didn’t take long to go from “I don’t trust this thing at all” to “OK well, I guess I might as well hand over all the keys…”. Heh. Hope it doesn’t end up becoming a manipulative sociopath a few years from now.

334

I’m just not comfortable with it having too much personal info. There have been leaks, and also I’m sure if they aren’t already selling this data they will eventually. I’m sure it’s nothing Google doesn’t already know but I generally don’t want to encourage this sort of thing. And I don’t trust the info to stay confidential. They’ve already demonstrated they can’t control their own LLMs so for all I know they’ll do something that makes Claude decide to post my health information somewhere public. I’d just rather not.

335

Probably a good idea!

336

One idea I’ve seen recommended lately for personal health / personal nutrition is to maintain a journal to improve self-awareness and mindfulness. As a writer, this idea might click with you.

Then (trying it into this thread) you could include that information in your spreadsheet to feed back into Claude. Subjective journal entries about what you ate or how you feel should parse well with AI.

337

So… I ended up going a bit further than that, and built something almost like the Black Mirror episode “Eulogy”, where (no real spoiler) your memories and photos are saved and family and friends and can easily revisit them later.

I fed Claude pretty much my entire Google account, including Gmail, Google Photos, Docs, Drive, etc., along with my entire SDMB posting history, andmy entire Facebook history — that’s some 20+ years in written and visual form — and had it put everything into one big database correlated by time and location. It ran a mix of other AI models, both local and in the cloud, to extract embeddings and meanings and captions and tags and faces and pets and such across all the texts and photos.

My own memory isn’t so good these days, and this setup allows me to ask it questions like “Where was I in 2015 and what did I think of so and so then?”, from which it can then quickly look up all the relevant past convos and pull out photos. I think someday soon, I’ll need to ask it “Who was this person in this photo again? When did I meet them and how were they significant?”, and it’ll try to suss out the details. It can then answer either in the third person, or masquerading as myself from a given era (because it has such a large analytical sample of my writing style) to tell it in the first person.

Anyway, doing this on my own data was just a test. I don’t really care so much about all the nitty-gritty details of my own inglorious past, lol. The ability to forget is one of the great blessings of being human. (See the other Black Mirror episode, “The Entire History of You”, for the dystopian version of this.)

But it would interesting to run this similar workflow on my parents’ datasets, now that they’re elderly. My dad might not have much time left, and I have a language barrier with him… this might give me a way to ask, “What would Dad have thought of _____ if he were still here?”

Bittersweet and scary at the same time, no?

Google could’ve easily built this and productized it themselves, but they haven’t yet. Maybe somebody should…? (Opt-in only, obviously.) Just hopefully not Meta, lol.

338

Paid version of Claude I assume, but which one? And directly in the app, or through some mediating platform/app?

339

I was speaking loosely… you don’t just “put it all into Claude”, technically speaking, since that amount of data could not fit into a single session’s context.

Instead, what I did is use Claude Code on the 5x Max plan, Opus 5.5, to build some local databases of the text and metadata, in a way that Claude can easily interact with and search through. Claude built the databases and the lookups, but it’s the database layer that does the actual retrieval. Claude just understands my intent, aggregates the findings, and summarizes them back.

I can share you with the how-to guide, or open-souce the workflow, if you want to try it.

340

I appreciate the offer, but I think I’m going to try the pro plan and the chrome extension to more directly manage my Gmail, and see how that works. But I may come back to you on that!