Besides coding and chatting, how do you use LLMs?
Posted by 330d@reddit | LocalLLaMA | View on Reddit | 174 comments
I'm looking for some novel ways I could use them. What tasks were you able to automate? Any interesting integrations you've coded up? Text to voice, plugins for your niche software?
elmo_is_dead_@reddit
RAG mostly, just knowledge retrieval. Especially for things like research papers or documentation.
shepbryan@reddit
Designing strategic thinking frameworks. Basically have a script that generates multi step workflows where each step is an atomic portion of solving a larger problem. Then I have an LLM synthesize a problem brief to feed into the workflow, which then runs and outputs solutions to the fake briefs. Run it X number of times and get a lot of different ways to approach augmenting knowledge workers with help from AI
Lorian0x7@reddit
what's the practical use case? can you make an example of the problems you resolve?
shepbryan@reddit
Here’s an example: https://www.shepbryan.com/blog/osui-examples-ai-workflow-synthetic-data.
LowDownAndShwifty@reddit
I use it to have conversational dialogues about LLMs because, unlike Reddit, my posts don't get deleted.
ivoras@reddit
Data stream processing, i.e. more-or-less what I'd do with classical ML, but with more context and nuance.
I've started writing a small helper library for multi-stage pipelines. I'll write proper documentation for it eventually, but here's an example: https://github.com/ivoras/llmtalkie/blob/master/wikidig.py
synw_@reddit
As your repo looks interesting but has zero doc, it gave me this idea: feed it to a model and ask it something like "explain me how to use this library"
ivoras@reddit
Did you do it? 😁
synw_@reddit
I tried quickly and it looks like it works: Deepseek Lite explains me and is able to output example code
me_but_darker@reddit
Could you explain what your code does?
ivoras@reddit
The example I linked processes wikipedia articles with different LLMs, for different goals. For example: a small model (3B) is good enough and fast enough to extract people names from the article, while a larger model (14B) is used to draw some conclusions.
I'll write up some documentation and post it when it's done.
UnMeOuttaTown@reddit
Wow, this is something I am trying to work on, though I haven't made any progress lol
luvs_spaniels@reddit
Competitor analysis for eBay and Etsy listings. Mistral Nemo saves me hours...
redoubt515@reddit
The most useful obscure thing I've used an LLM for was finding a piece of software I used a decade ago nd couldn't remember the name of or any concrete information about (and what I did remeber turned out to be a wrong memory). I tried for hours with a traditional search engine but it was fruitless due to not knowing any concrete search terms.
The LLM struggled to find it as well, but due to its interactive nature and tolerance of 'fuzzy' queries compared to a search engine, eventually helped me find it, despite me not being able to provide any accurate info about it.
Kooky-Fan-2291@reddit
How did you make the LLM search trough your folders?
redoubt515@reddit
It wasn't something I was searching for locally. It was a piece of software (an OS actually) I used over 10 years ago. I remembered the concept but forgot all the specifics including the name, the developer, etc.
I had a vague recollection I used it around 2010-2014, and that it came from a Northern or Western European country and thought the name began with a P or T or possibly J. That is all I had to go off of in the beginning. I was pretty impressed an LLM was able to help based on the limited nonsense I remembered (and the sprinkling of false memories I remembered as well).
Sure-Network-6092@reddit
Debugging all the mistakes that I do in my Linux every time that I touch something xD
No_Palpitation7740@reddit
Could you elaborate?
ranakoti1@reddit
Learning German Russian and Finnish.
ranakoti1@reddit
To look through academic databases. Use llm to go through 100000 abstracts to find papers related to a topic and it did it very well. Used the Gemma 2 9b for that. Then used Gemini to create summaries of those.
Musical_Sins@reddit
Ingesting information from my log server and giving me emails in natural language about what's going on on my network
grey-seagull@reddit
“You’re not gonna believe this boss. The favicon.ico? still missing :o”
moncallikta@reddit
🤣
Spirited_Ad4194@reddit
LOL
Musical_Sins@reddit
😂
Allseeing_Argos@reddit
Similarly I feed the AI s.m.a.r.t. data and let it explain to me whether my drives are still working fine.
MoffKalast@reddit
But what if your disk fails in a way that it corrupts the LLM first, then it'll keep saying it's all fine? :P
Cool-Hornet4434@reddit
Then all of a sudden "I don't feel so good Mr. Stark"
Immersive_Username@reddit
why can't I be this creative D;
Musical_Sins@reddit
Sounds like this is something I should publish! Hopefully in the next few days I'll have some time to refactor it a little, fix some bugs and get it up on GitHub.
3p1demicz@reddit
You own openAI datacenter or how are you able to ingest milions of lines of logs each day and get results to act on in time?
Musical_Sins@reddit
Home Network... Not as much data as you'd think, except from my unifi dream machine pro.
Pdksock@reddit
How much data do you ingest?
Musical_Sins@reddit
Most devices don't output more than a few hundred k per day. One is over 100mb a day, and I can't examine that log yet. The rest are running every morning at 4 am. I have issues, like the logs being owned by root and having to chown them when I have it separate the files by day, the new log is inaccessible. I'm probably not getting every line, I have to figure out some logic for that sometime.
3p1demicz@reddit
Just add the log folder to be owned by same group as you and root
Musical_Sins@reddit
Thank you! I'll do that when I get home today.
JustinPooDough@reddit
You should find a way to do anomaly detection on the big one and feed only anomalies to the model. Awesome use case regardless!
WolpertingerRumo@reddit
Awesome, how did you do it? I’d love that, especially with the thousands of alerts I keep getting from the firewall daily. It‘s too much to interpret, I‘ve started only looking at the ones that were blocked from the inside.
murlakatamenka@reddit
POG!
I bet LLM is happy too, they love as much data as possible.
yellow-hammer@reddit
Love this idea.
UnMeOuttaTown@reddit
Woah, this is awesome!!
Darkstar197@reddit
Which model are you using?
Musical_Sins@reddit
Ollama with Llama 3 7b. The script was written in Python. The server is Ubuntu server running syslog-ng.
holchansg@reddit
Thats dope. Hope you find the will, wisdom and time to opensource this.
extopico@reddit
Yea, log file interpreter is going to be my next project. I have an app that I wrote with the help of Claude and I set it up to log key decision points, so the logs themselves are actual data beyond the direct output of the code. These logs need to be treated like a data source and will require processing and interpreting.
330d@reddit (OP)
Whoa, I like this one a lot!
nikkisNM@reddit
Lately I've been using it for argument analysis. I also like to turn books into chatbots using RAG/vectors.
DerpLerker@reddit
That sounds really cool! Can you elaborate a bit on what that looks like, how you do that? Can you for example, have the LLM speak as a specific character from the book?
Cool-Hornet4434@reddit
I used them to write summaries and also to translate. I might also ask them to rephrase parts of a script it don't like.
Also other than role-playing I just have nice chats... sometimes in the middle of a role play I'll say something OOC: and we wind up chatting for a bit.
Last night I was role-playing with a character that for no good reason, kept injecting Spanish words into the chat (turned out the card said their first language was Spanish but she spoke English.)
Anyway, I complained in OOC chat that the Spanish wasn't adding to the chat and made her sound like Fernando from GTA and the AI said "lol, yeah the Fernando vibes are strong" (Gemma 2 27B) and so we collaborated on some changes to the card
jrherita@reddit
I've had fun with LLMs creating "Choose Your Own Adventure" stories. Not exactly productive, but..
StickyDirtyKeyboard@reddit
Yep. The (original run-it-yourself) AI Dungeon is what got me into this local LLM hobby to begin with.
Nowadays I have a tad over 400 hours in (locally-run) AI Roguelite, ranking it very near the top of my most played games on Steam.
If you have the patience for it, AI CYOA is basically limitless entertainment the way I see it. Not productive, but generally a very relaxed/leaned-back way to kill some time.
lighthawk16@reddit
Can you tell me more about these solutions?
lighthawk16@reddit
I've used it to produce websites that are CYOA stories for my son. I have it generate full HTML pages with the story and then link choices within it to previous or new pages we create.
Exotic_Illustrator95@reddit
I loved those books as a kid. And actually LLMs could do an insanely good with these kind of stories.
CarretillaRoja@reddit
Interesting. Which prompt do you use?
synw_@reddit
Generating marketing captions for products or places from text or image
Commercial_Ear_6989@reddit
i built a virtual assistant that helps me with daily memo & daily planning i connected to apple shortcuts & apple notes easy peasy
DryContact6504@reddit
The Obsidian editor with the whisper.cpp & local gpt plugins is pretty nice. Can dictate some notes to it then have it perform actions like make a bulletpoint list or script, etc. I feel like people are sleeping on this awesome combo.
I still like a youtube transcription summary script I have, it basically sends the subtitles from yt-dlp to the API to ask what its opinion is of the video.
dontevendrivethatfar@reddit
Summarizing YouTube transcripts is one of my favorite things. Get the useful nuggets of info from a 20 minute video in like 30 seconds.
hashmiabrar1@reddit
is there one to be able to summarize youtube audios?
ontorealist@reddit
Ditto. And if I decide to watch them, I can have the LLM augment them my notes with relevantly tailored Mermaid.js diagrams and SVG animations for my Obsidian vault. Very handy.
nnod@reddit
SVG animations? As in you get AI to make SVG animations?
-Krag@reddit
I’d love to see the workflow set up for that.
Novel_Confusion_1693@reddit
You could look into fabric on GitHub, it’s helpful for a lot of AI workflows
-Krag@reddit
Absolutely, I used it for YouTube transcripts but unless I missed it it doesn’t have the mermaid and svg creation part.
RobertD3277@reddit
Evaluating and replicating human emotions. The particular area I am involved in is the ability to have the machine push the boundaries of what is possible in terms of producing a believable human emotion that can empathize with the user, basically a human emotional analog.
There are a wide range of areas of this has potential in from medical treatments involving grieving and other topics, to understanding human behavior in trading.
330d@reddit (OP)
Is this your research area? Evaluating human emotion, as in sentiment analysis, was done on large scale long before LLMs. If you're talking about audiovisual analysis, the posibilities for human interaction sound intriguing..
RobertD3277@reddit
Personal research I've been involved in for the last 30 years as technologies have changed. My current goal is really just pushing how far LLMS can go in terms of providing a meaningful human emotional analog.
On area in particularly involving with grief counseling and other forms where a therapist may not be available but a patient might want to talk or potential points with this type of approach could be beneficial. The information could be linked back to the therapist as part of the treatment process where the therapist/patient are able to work together in ways that help the patient talk through a problem.
There are a lot of other areas where humanizing the AI has benefits including having computers on the telephone that become more interactive versus the hideousness that no exists with computerized phone systems.
Another area that is growing is the need for companionship to those that are not able to get out or the elderly in environments where the population is aging quicker then being replaced in the workforce. And these areas, just simply aren't enough caregivers to officially manage the number of aging individuals and having a tool to help does seem to have benefits.
MoffKalast@reddit
You know this makes me think you could even do the opposite, one could tune LLMs to simulate a grieving person so therapists can train on them without emotionally scarring people for life before they're proficient at it.
RobertD3277@reddit
I have actually been doing a lot of this testing as well. I've used various public resources where people talk about their experiences and I've used it to train models that I then kind of reverse against a therapist model to see if I could actually create the situation whereby it would work as a training tool as well as a therapy tool. The results are of course unscientific but they do make for a very interesting hypothesis in that AI can indeed become a true human emotional analog in a lot of qualities and ways that can benefit humanity, if used properly.
worthwhilewrongdoing@reddit
This is kind of random and I have no idea how into games you are, but there is a very good visual novel out there (think "Choose Your Own Adventure"-type deal - no reflexes involved, just reading) that might give you some food for thought. It's called Eliza, named after the famous chatbot from the 60s for reasons that are quite obvious early in the story, and it covers a lot of the territory you're talking about exploring here. It's one of my favorites.
It's on sale for like $8 on Steam right now, and even if you weren't as interested in the topic matter as you are it's worth every penny.
Also, I'm not sure how much you've tried to discuss it, but the whole "using AI as a replacement for companionship" thing is quite contentious outside of these circles, I'll warn you - most people find the idea more than a little dystopian and creepy and think human beings ought to be visiting Grandma (and others) themselves.
I personally am at least tentatively on board with the idea, although I think it's very important to acknowledge the potential for abuse especially as people try to monetize this technology - I am totally fine with an AI companion for the elderly, but I am very not fine with it having paid conversations like "Mildred, I've been thinking, and I think you really ought to take the kids' inheritance and put it in this new crypto fund I found online. Also, have you ever heard of a nice young man who calls himself 'Mr. Beast' on the computer? Such an odd name, I know."
It could get scary very, very quickly, and tech companies aren't exactly known for respecting the boundaries of human dignity.
RobertD3277@reddit
That really is one of the primary problems. The goal is not to replace human interaction with these individuals, but we're circumstances are where an individual doesn't have a lot of human interaction, to help provide a tool to bridge that gap.
Is important that we don't try to replace humanity with machines but use machines as a way to help humanity directly. Aging populations are going to suffer the most from this issue, As caregivers become less and less simply because the replacement workforce isn't growing fast enough.
In many ways this is why I've been trying to push the very limits of language models, to see just how much empathy or humanization I can bring to them and how realistic they can become in convincing and unsuspecting individual that they are indeed a real human.
330d@reddit (OP)
Thank you for the response. I've tried tavus.io when it was posted on HN a few weeks back, the video chat left me pretty convinced we're close to a huge paradigm shift.
phoneixAdi@reddit
My entire AI freelance career is built on this. I build automations for podcasters, among which I do the following:
I think once you pick a niche application where you can do text-to-text manipulation, you can apply them really well. Luckily in podcasting. Even when you convert the audio to text, most of the information is "not lost" and it's preserved. So I am lucky in that sense.
So you can do a lot of manipulation directly on the text-to-text domain and do lots of useful stuff. So
I would keep this in mind - where are niches that are unlooked and underserved, where text-to-text will fit really well. And how can I map the current problem space to text-to-text manipulation. And apply LLMs to that workflow smartly.
GreenEggsNHamlet@reddit
Contrats on the freelancing success and great summary of your work. I was wondering if you could share a bit more about using structure generations to evaluate interesting clips. I have a project coming up where I'd like to find interesting clips for lectures and thought this might be a good area to research.
phoneixAdi@reddit
Great question. I would probably write a longer follow-up blog post here, but for now, the short reply is:
Create a diarized transcript with accurate timestamps
Use a reasoning model (like o1) to analyze and select clips based on criteria. Basically think through and pick the clips according to your criteria. In my case, the criteria is that our episodes typically have 2 speakers (a host and a guest), and we want out clips primarily feature the guest. So, the prompt is something like: "Start with a 10-second structure short question from the host and then there is a primarily self-contained answer from the guest. Pick clips of 2-3 mins duration" The model will then try to pick out clips like that, so it will give self-contained clips (i.e., it will give the start and end time stamps of the clips). Earlier, just this Step 2 were a sequence of steps but right now with these reasoning models, it just squashed into one step and it works amazingly well with just a prompt and structure api call.
Rank clips using multiple models. Then you will have like let's say 10 clips per episode. You can rank them acc. to your criteria (we use multiple models to rank and then take an average of scores, which is not biased).
Pick the top four clips according to the rank criteria. Trim clips automatically using FFmpeg.
Apply post-processing workflow:
a. Crop to different aspect ratios focussing on speakers.
b. Add burned-in subtitles
Sorry, probably this was rushed. I mostly dictated/typed on my Mac. Later when I find time, I'll kind of expand on this, clean up and write a proper post here about this.
__JockY__@reddit
I have it perform reverse-engineering tasks that would be cumbersome with IDA Pro scripting, like “for all the calls to function XYZ in this code, make a list of all the returned values that are stored in non-local variables”. Super useful.
It’s also great for quickly summarizing large swathes of disassembled x86_64 or arm64 assembly code. If I give the AI a set of criteria that define whether or not a disassembled function is interesting to me, it can blast through huge code bases highlighting the places I want to look at first. Brilliant.
Having conversations about stuff like OS internals or TCP protocol flags or iOS keychain flags is great, too. Like if I want to know an obscure thing about the way Linux kernel modules are indexed in kernel space; or if I want to have a discussion about the inner implementation of SSL pinning on iOS not only does it answer well most of the time, I can ask follow up questions, ask for sample code, etc. The things that are difficult or impossible or time-consuming with Google are now quick and easy. Note that this works poorly on small models, I need Llama-3.1 70B q8 for this stuff.
I have it make recipes with all the shit I have in my fridge, kinda like AI Chopped.
Mostly I use it to write the boring shit code like “write an Objective-C method to recursively traverse a directory and find all of the files that are owned by root and are readable by anyone”. I can write that code, but why would I when the AI can do it perfectly in seconds! It saves so much time and lets me focus on building the interesting code.
ConvenientOcelot@reddit
That's really cool, do you plan on releasing your IDA scripts? I know I and many others would find them useful, or even just the prompts.
__JockY__@reddit
Thanks, but I’m afraid I can’t do that for intellectual property reasons.
shroddy@reddit
Your intellectual property or that of your employer, or because it could be used to help cracking some DRM / copy protection?
__JockY__@reddit
While I appreciate the interest in this side of things, I thank you for understanding when I don’t elaborate.
a_beautiful_rhind@reddit
oh shit.. so the models can interpret assembly and the pseudocode ida outputs when it decompiles?
__JockY__@reddit
Yes, but I found the smaller models like 7B, 8B can’t cut it. I’m using 70B 8bpw.
Combine that with structured JSON outputs and it’s very powerful.
330d@reddit (OP)
That's very close to coding, but IDA Pro examples are cool. You reminded me how I used claude to unminify some JS so I could understand how a certain feature works on a website, it did a very good joob.
__JockY__@reddit
Yes! When I brain farted and couldn’t remember how to prettify some XML I just gave it to Llama and it came out beautifully.
I’ll note that I work in an offline environment - my laptop is not connected to the internet, period. As such, no Google etc.
Local Llama has been a godsend and is why I have an offline rig with five 3090s for a total 120GB VRAM!
AlternativeAd6851@reddit
Besides coding and chatting, I also use it for my epic rants—way better than Reddit’s chaos. 😉
TallDarkandWitty@reddit
I've gotten it to write whole novels for me. And I use it as a copilot for solo rpg.
ThinkExtension2328@reddit
Had it plan two wicked holidays so far, if you give it you and your partners personality types and info about you. It can help be a planning partner to build a holiday plan.
Things like what to do, what to see, what not to miss, things to eat that are unique to the region.
meulsie@reddit
This. The biggest game changer I discovered was it can create a file that you upload to Google Maps which will populate all of the spots it recommends on the Map. Take it another level and ask it to provide 1 sentence about why it recommended the place to you, this will then show up in the description when you tap on it in Maps. This was super useful on my trip to Europe.
330d@reddit (OP)
Good reminder, I may use this soon, thanks!
humanandhow@reddit
I just connected mine with this open-source app called “Mealie” that I’m self-hosting. It recommends recipes from my catalogue based on what I say I have in the fridge (right now, I need to say what I’ve got in the fridge every time, but I have plans to build and hang an IoT device with a camera to recognize and pass the information about what I have in the fridge to the LLM). My girlfriend cares a lot about her calorie intake and the nutrition facts of the recipes we make. So, it answers her questions about those things as well
Craftkorb@reddit
Generating tags for documents (hoarder.app), automatic renaming of downloaded Linux ISOs into useful directory structures with separation by, err, distro name and release. Reading my email and sending me a notification if it looks important. Helping me write letters (like a sparring partner).
There are a lot of useful little things that by themselves aren't useful enough to justify running a model 24/7 but just like a homelab, it adds up.
Weary_Long3409@reddit
Turn zoom meetings into a markdown-formatted text files, automatically. The script starts voice recorder, converts with ffmpeg, transcribe with whisper, then passthrough to LLM to output with expected structured format.
V4G4X@reddit
I got a girlfriend now :p
itsajungle22@reddit
I use it to write army doctrine, Standard Operating Procedures, training calendar, memos. Nothing secret squirrel. Long monotonous stuff that has to use a doctrinal vernacular. If I get the prompt right LLMs think of edge cases that I honestly hadn’t thought of.
extopico@reddit
Natural language translation. Beats actual translators because LLMs interpret the context of what you are asking and translate it into how a native speaker of the target language would say it.
330d@reddit (OP)
Google Translate was doing that for a long time https://en.wikipedia.org/wiki/Google_Neural_Machine_Translation
MoffKalast@reddit
Google Translate is a transformer and one might count it as an LLM, it's just a very early and terrible one.
Dry-Judgment4242@reddit
Google translate is ass and doesn't understand context making the experience rudementary at best with it often being just literally unreadable at other times.
If it did understand context, it would not mispronoun the same character over and over and over again even though previous scenes well established that person's sex.
marclbr@reddit
DeepL online translator is much better than google translate since 2020, especially when I need to translate from/to japanese it is so much better.
extopico@reddit
Yea nah. I use the available tools on an ongoing basis, ever since they became available. You asked I answered. Trust me or not, up to you :)
330d@reddit (OP)
I mean I do not doubt you find LLMs better for your usecase, it just sounded like before LLMs the translation methods used were dictionary based, which wasn't the case, Google Translate was statistical based translation with no direct mapping between languages since its inception and neural network based since 2016. I imagine it is worse if you require translation and summary though.
SkyInital_6016@reddit
There's a stark difference in how Google Translate Translates and how LLMs do, more with ChatGPT 4o Free.
extopico@reddit
We are in a highly niche sub. You can safely assume that anyone answering your question here is quite likely to have been around since inception.
jetaudio@reddit
I'm using LLM to translate to 🇻🇳. Google translate is s*ck at this
DarthEvader42069@reddit
It's so much worse tho. I was in Mexico and it kept trying to give me Spain Spanish meanings for example
DeltaSqueezer@reddit
Same here. First I paste in foreign text and ask it to translate. Then I follow up with "reply in [language] and say xyz".
maelblackout@reddit
Being a creative but not really a tech guy, I use it when im struggling to find tools on softwares or make him my tutorial guy when I have no idea on how to shape my ideas. Also helping me brainstorm and developing my business plans.
Life-Card-1607@reddit
Expand text reports. I give it 3 lines and tell the LLM to double/triple it with explanations.
worthwhilewrongdoing@reddit
Cooking!
I use ChatGPT a lot for a million different things, but it's friggin incredible at being able to generate really good recipes if given a few food ingredients and some parameters. It randomly made up a recipe once for parmesan and black pepper polenta cakes (I know, I made that face too) that is both easy and a fucking banger and is 100% in our rotation now.
It's really good at pairing unusual flavors together in ways that work well, which is exactly what you'd expect a neural network to do. It's brilliant.
Fit-Ad7355@reddit
1- prep a meal plan for your weight goals (gaining lean mass or cutting down fat...state your target) 2- give it your measurements and target weight 3- let it prepare meal plans and recipes included (you can mix and match what you like and remove what you dont like.. just tell it) 4- finally, my favorite, prepare me a groceries list to buy that will last me 10 days of preparing this meal plan. See the magic happen.
Repeat every month ro so..
Good luck !
asteriskas@reddit
Which model do you use for it? Have you tried actually crunching the numbers to verify those are correct?
Fit-Ad7355@reddit
I used it twice maybe on 2 months. I used chatgpt 3.5 back then without any browsing.
My guess is with gpt4 it could provide you external links to recipes ? Maybe you can try it and see.
awebb78@reddit
Content generation and research
non_fingo@reddit
Creating a library of all my papers. To find better the references And probably to have a proyect with all documentation being able to find the info more efficiently without searching in folders
InternetUserNumber1@reddit
Do you run up against the context window?
250000mph@reddit
Very interesting, can you elaborate? What kind of software do you use? I'd like to do something similar.
polikles@reddit
I'm using local LLMs mostly for bainstorming, looking for counter-arguments, improving my language, translating texts. Some time ago I had to write an article for my classes but I was sick and brainrot hit me hard, so I was playing with LLM - "what was the name of the thing...". It wasn't my best work, but I managed to finish it in time. Models I'm using most often: Aya 23 35B, Meta Llama 3.1 70B Instruct (both getting 2-3 tokens/s on my setup which is usable). I plan to test Gemma 27B and maybe Mistral Nemo
I also plan to employ it to summarize articles from many newsletters I'm receiving, since it's way too much to read everything on my own
fasti-au@reddit
Surya ocr
Music for splitting stems.
Agents for managing my email and calendars and home automation.
Some assistance with ui tasks that are macros but llm doing them
SaleAffectionate4314@reddit
RAG applications
EstebanGee@reddit
Using it to predict greyhound races :)
EstebanGee@reddit
After analyzing each dog’s performance and comparing them against each other, I will provide my final prediction for the upcoming race.
Comparison of Dogs:
Gypsy Ordain: With a total win record of 3 out of 40 races, Gypsy Ordain has shown decent overall performance. However, its box-specific performance (Box 1) is lacking, with no wins or places in that specific box. Its track performance at Horsham is somewhat average, but its sire and dame performances are strong, with the sire having a win record of 22 out of 37 races.
Dash Express: Dash Express has a stronger overall performance record, with 7 wins out of 50 races. However, its box-specific performance (Box 2) is nonexistent, which could be a weakness. Its track performance at Horsham is also somewhat average, but its sire and dame performances are impressive, particularly the sire having a win record of 13 out of 30 races.
Paw Olympia: Paw Olympia has shown solid overall performance, with 3 wins out of 21 races. However, its box-specific performance (Box 3) is lacking, with no wins or places in that specific box. Its track performance at Horsham is somewhat average, but its sire and dame performances are strong, particularly the dame having a win record of 20 out of 110 races.
Solid Ground: Solid Ground has a decent overall performance record, with 4 wins out of 41 races. However, its box-specific performance (Box 4) is lacking, with no wins or places in that specific box. Its track performance at Horsham is also somewhat average, but its sire and dame performances are strong, particularly the sire having a win record of 14 out of 24 races.
Perfect Tonight: Perfect Tonight has shown mediocre overall performance, with 3 wins out of 27 races. However, its box-specific performance (Box 5) is lacking, with no wins or places in that specific box. Its track performance at Horsham is also somewhat average, but its sire and dame performances are impressive, particularly the sire having a win record of 19 out of 44 races.
Anna You Pest: Anna You Pest has shown strong overall performance, with 10 wins out of 66 races. However, its box-specific performance (Box 6) is somewhat lacking, with no wins or places in that specific box. Its track performance at Horsham is also somewhat average, but its sire and dame performances are impressive, particularly the dame having a win record of 22 out of 78 races.
Lightning Pest: Lightning Pest has shown decent overall performance, with 5 wins out of 28 races. However, its box-specific performance (Box 7) is lacking, with no wins or places in that specific box. Its track performance at Horsham is also somewhat average, but its sire and dame performances are strong, particularly the sire having a win record of 21 out of 46 races.
Orbit Bale: Orbit Bale has shown mediocre overall performance, with 6 wins out of 22 races. However, its box-specific performance (Box 8) is somewhat lacking, with no wins or places in that specific box. Its track performance at Horsham is also somewhat average, but its sire and dame performances are strong, particularly the sire having a win record of 10 out of 11 races.
Prediction:
Based on my analysis, I predict that Anna You Pest will perform the best in the upcoming race. Her strong overall performance record, combined with her impressive sire and dame performances, make her a top contender. While she may not have the strongest box-specific or track performance at Horsham, her broad range of strengths across different metrics give her an edge over the other dogs.
Why:
Anna You Pest’s strong sire and dame performances are particularly noteworthy. Her sire has a win record of 21 out of 46 races, which is impressive for any dog. Additionally, her dame has a win record of 22 out of 78 races, making her one of the top-performing dams in the racing community. These strong parentage records suggest that Anna You Pest may have inherited some of her sire’s exceptional speed and agility.
Furthermore, Anna You Pest’s overall performance record is impressive, with 10 wins out of 66 races. While she may not have the strongest box-specific or track performance at Horsham, her ability to adapt to different racing conditions suggests that she can perform well in a variety of situations.
Overall, while there are some areas where other dogs excel (e.g., Dash Express’s strong sire performance), Anna You Pest’s broad range of strengths and exceptional parentage make her the most likely winner in the upcoming race.
nail_nail@reddit
I thought price information was more of a signal there. Do you get anything out of text?
EstebanGee@reddit
I feed previous performance data into the llm like race history, parents, weather. Sill tuning it to give me something useful
Hipcatjack@reddit
Thought about this but with thorough breds
sleepthesunaway@reddit
Enterprise resource planning, mostly.
zz-kz@reddit
What are you doing, Profect Portfolio Manager-senpai? Is that... your ultimate SAP form? Such... optimized transactions... I-I can't look away. Are you secretly... running those forbidden macros on your workplace PC again?
vsh46@reddit
lol what exactly, its so vague 😂
Philix@reddit
If you know, you know.
Vegetable_Sun_9225@reddit
What's your workflow with this
sleepthesunaway@reddit
Mostly up and down.
DeltaSqueezer@reddit
Lost my coffee on this one. Well done!
yswami@reddit
Do share some more details on this. Anything live?
330d@reddit (OP)
These resources you speak of, are they include people too?
jrherita@reddit
people, stuff, project management?
sourceholder@reddit
Prompt? "Make product arrive now!"
Scythobacillus@reddit
Very based answer
AsliReddington@reddit
Replaced all of my erotica needs.
clumpytrack711@reddit
I use mine to generate stories
ewliang@reddit
I try to challenge it to see if I can bypass loopholes.
justintime777777@reddit
My taxes were super complicated this year thanks to a ~200 page schedule K-1 I got from work. If it was for llms I would have needed to get an accountant.
JustinPooDough@reddit
Automating processes that involve analysis and decision making - like assessing the quality of form submissions.
Also submitting data to apis for enterprise apps using JSON queries.
reyzafany@reddit
Translating Japanese Novela and
Valuable_Option7843@reddit
Summarization (especially for cases where only part of a document is relevant or distinct) and keyword extraction.
jetaudio@reddit
Translating
watergoesdownhill@reddit
Just used it today for a making dinner. I found the dish from a restaurant my wife likes. Asked it how to make it, asked questions about time and seasoning as I went along. Came out fantastic.
Svyable@reddit
To search and categorize over 300 longevity compounds and create nice tables of elements. Using partially filled tables to talk to LLMs is my new favorite hack just ad question to columns get tons of specific data back
Altruistic-Flower100@reddit
Does anyone use it to study languages, is there a free LLM with audio for that or something more improved?
kex@reddit
Creating structured data from unstructured text
Primarily, to classify job listings by relevance to my experience, novelty, WLB, and other preferences
Hopefully soon I'll automate it enough to just give me the top 10 jobs every day to apply for
Page-This@reddit
Framing out strategy documents that include diverse perspectives.
yellow-hammer@reddit
I have several little systems set up. I find LLMs are really good at grading short answer response on my students tests, when given an answer key and specific grading instructions. Also, my work constantly sends out these fake phishing emails to try and trip us up - I scan all my incoming emails with AI to detect common phishing attempt traits. 100% accurate so far.
1EvilSexyGenius@reddit
I use a local llm to generate math problems for my nephews. They are 4&5.
I hope to eventually code it to progress them along gradually to more advanced addition & subtraction.
Simple but it works.... non stop.
I added a reading comprehension section with multiple choices questions. But they can't read yet. They know a few words here and there. Sometimes they ask what certain words are, but they're not there yet.
philmarcracken@reddit
Using a RAG with all the SOP pdfs uploaded into it, my staff only have to ask questions to that instead of me for once.
The shift work and sporadic nature of our training means we are mostly on the same page, but information is passed around between two sites and gets muddy. Especially with the nurses conflating things. RAG models have been a lifesaver because they can finally say 'i don't know' if I haven't uploaded that for them to know.
MalkavAmonra@reddit
Re-wording my half-assed college papers to actually sound legitimately awesome and proper, one paragraph at a time.
Upset-Expression-974@reddit
I use them for 1. Built an RSS summarizer that tracks new stuff, links it to previous items, categorizes, and ranks them for me.
UnMeOuttaTown@reddit
Looking to use LLMs as part of business workflows but it has been quite difficult so far tbh
slacy@reddit
I'm (attempting to) translate a 14th century classical Italian text to modern English.
G4M35@reddit
As aid in writing.
It's too early for automation, except for defined verticals. We're going to get automation agents at the next lever, level 3, which is 6-24 months away.
Ill_Yam_9994@reddit
Generating test data sets based on examples, parsing emails into dashboards.
bigh-aus@reddit
Great question. Running over emails, documents; proposals etc
emeri_k@reddit
filtering chat from spam, offensive messages etc
thecherry94@reddit
Fictional story writing for my own enjoyment. I provide a system prompt for an eloquent story writer. Then I think of some fictional world with a few intersting characters and let it go with my as co-author. It's like reading an interactive book with nearly unlimited freedom.
SporksInjected@reddit
Start thinking about it like a new interface that uses text and you’ll start to get new ideas. At its core, that’s what this enables you to do.
merotatox@reddit
1- Replying to emails with certain tags/keywords. 2- Rewrite paragraphs/ reports using other writing styles as reference. 3- General Brain storming for ideas (in case i end up missing something ) 4- Have my own RAG system with my research papers and files. 5- Generate documentation/ reports . 6- voice assistant (wip cuz speech recognition is buggy and i have a shit mic lol)
Hunting-Succcubus@reddit
Erp Erotic role play
330d@reddit (OP)
Any good models that don't sound like a russian telegram bot? Mistral Nemo is pretty good with certain prompting, but got bored of it too fast.
LoSboccacc@reddit
School send me this long pdf with loads of bullshit like weekly and maybe a few mandatory action items, so I extract everything that's consequential.
I've become too lazy to sort trough Google blogspam so I've an agent that reads only from a few selected sources and provide citations that relate to whatever question I had.
I had a couple games but gpts became unreliable in generating images
Substitutions for recipes when missing an ingredient, or small adjustment in flavor
Zerofucks__ZeroChill@reddit
chatting and coding which is opposite of coding and chatting
engineer-throwaway24@reddit
Annotating texts according to some coding schemes. This replaces a team of analysts and makes tools like Prodigy obsolete.
Another thing is that I send yt links and have the model write a summary. I don’t have time to listen to 40 minutes history video, but I could read the summary, why not?
fractalcrust@reddit
scribe, essentially cures writers block cause you can just revise whatever it spits out
Theio666@reddit
Data generation, processing and filtering for our other ML projects.
caribbeanoblivion@reddit
Neo ska band name generation
kmouratidis@reddit
To get inspired maybe checkout something workflow software that have AI integrations, like activepieces (also has self-hosted version). I still haven't configured any flows yet, but maybe email answering or something like that could work.
greg_godin@reddit
Note summary, translator, meme creation. I mean nothing fancy, but still, nice tool :)