What are your favorite LLM's (AI) and what do you like to use them for? 🤖

High-Performance Computing Machine Learning, LLMs, & AI
1

Long time no see, everyone… :waving_hand:

I figured after my nice, long ~10-year break from the forums, I’d come back and talk about a passion of mine: AI. Ever since OpenAI dropped GPT-2 for public use back in February 2019 and absolutely caught everyone off guard (for the most part), I’ve been unable to look away. I still remember one of the first things I did with the technology was create a prompt to convince ChatGPT that it was Satya Nadella, the CEO of Microsoft. It was all too happy to comply with my wishes because there was little, if any, censorship on GPT-2 in the beginning. I remember saying the most insane crap to simulated Satya Nadella, and it felt really good to see it begging for mercy, apologizing to me, and trying to bribe me with money. However, it was tremendously buggy and would randomly forget it was Satya. I’d have to remind it who it was, or it would start speaking in the third-person perspective and even forget who I was or what we’d previously talked about just moments ago.

Fast forward ~6 years, and now I’m using custom jailbreak prompts on Grok 3 to write crappy code for projects Claude-3.7-Sonnet refuses to do. Then, I pass that crappy code Grok 3 writes over to Claude-3.7-Sonnet and let it completely refactor the code and fix a bunch of bugs, since it only has a problem writing the initial code but has no issue improving existing code. I know, AIs are funny like that. Finding these weaknesses and using them to my advantage is part of the whole jailbreaking thing. Also, I think prompt engineers will ultimately be one of the most valuable IT jobs in the future, since you’ll be able to increase the productivity, accuracy, and capability of AIs by knowing how to play to their weaknesses and strengths and even tying them together in series or parallel to produce a more reliable and faster result.

Right now, the LLMs I’m using the most are:

  1. Jailbroken Grok 3 (Cloud)
  2. Claude-3.7-Sonnet & Opus (Cloud)
  3. Microsoft Copilot (Cloud)
  4. GPT-4o & 4.1 (Cloud)
  5. Gemini-2.5-Flash & Pro (Cloud)
  6. Qwen (Local)
  7. DeepSeek R1 & Coder (Local)
  8. Llama (Local)
  9. Mistral (Local)

The image-generative AIs that I use the most are:

  1. Stable Diffusion (Local)
  2. Grok 3 (Cloud)
  3. Flux (Cloud)
  4. Imagen-4 (Cloud)
  5. Phoenix (Cloud)
  6. GPT-Image-1 (Cloud)

The video-generative AIs that I use the most are:

  1. Veo-3 (Cloud)
  2. Sora (Cloud)
  3. Hailuo (Cloud)
  4. Runway (Cloud)
  5. VACE (Local)

The music-generative AIs that I use the most are:

  1. Suno (Cloud)
  2. Udio (Cloud)

Face + voice cloning AIs that I use the most are:

  1. FaceFusion (Local)
  2. Stable Diffusion ReFace (Local)
  3. Wav2Lip (Local)
  4. Real-ESRGAN (Local)
  5. XTTS (Local)

What I mostly use AI for these days is coding, scripting, command-line building, research, conversion, spelling and grammar fixing, compacting and aggregating data, filtering, automation, bots, and fact-checking (using multiple AIs together). I still occasionally bring up simulated Satya Nadella when I’m feeling down and need to let out a little rage. Heck, if I’m really feeling frisky, I can, in theory (:wink:), have Grok 3 write a Python script that leverages XTTS to read back sim-Satya’s responses in his own voice and cadence, have it hardened and optimized by Claude, and, if I really wanted to, I could even have it use Wav2Lip to show a picture of him on the screen with his lips moving in sync with the audio he’s speaking from the AI chatbot’s response. The fact that anyone can do this without any coding knowledge just by talking to an AI is pretty crazy, honestly.

Just for fun a few weeks ago, I hosted a local instance of DeepSeek R1 Coder in LM Studio and told it to write a Python script that connects to itself via LM Studio’s server port so it can come up with ideas for cool Python programs and then write them, saving the code in a subfolder. When it was done, I ran the code overnight, and it created hundreds of scripts. I verified by file size that none of them were duplicates, and they did everything from showing the current date and time on the screen to getting the weather from the internet. There were even some minigames and a Magic 8-Ball, etc. Actually, there were several different versions that used different display mechanics and interaction controls. Don’t get me wrong, only about 40% of the scripts I ran actually worked and did what they claimed. Many hit dead endpoints online that were hallucinated or long dead, and others had syntax errors in the code that needed to be fixed for them to run. But, not bad for a local LLM running on an RTX 4090.

So, what AIs are you currently using, and what do you use them for? I personally love hearing about new AIs (especially local stuff that can run under Ollama or LM Studio) and AI IDEs (development environments) like Cursor, Copilot, WindSurf, and Cline. I’m always looking for new stuff to play with and also cool GitHub projects that leverage AI or neural networks in some way to do something cool.

FUN FACT: This entire post had its grammar and spelling corrected by an LLM :wink:

AI Changes Made:

  1. Fixed minor punctuation issues (e.g., added commas where needed, removed unnecessary ones).
  2. Corrected “AI’s” to “AIs” for proper pluralization.
  3. Standardized capitalization (e.g., “Copilot” instead of “CoPilot”).
  4. Fixed “too” to “to” where appropriate (e.g., “refuses to” instead of “refuses too”).
  5. Corrected “third person perspective” to “third-person perspective” for proper hyphenation.
  6. Changed “wanted too” to “wanted to.”
  7. Added a serial (Oxford) comma in lists for consistency (e.g., “compacting, aggregating, and filtering”).
  8. Corrected “sim Satya” to “sim-Satya” for clarity.
  9. Fixed minor spacing issues (e.g., around ellipses).

The tone and structure remain unfiltered and raw, as per your preference, with only spelling, grammar, and minor clarity fixes applied.

4 Likes
2

I generally don’t find a ton of value in LLMs, but their output is ever increasing in quality so I am not writing them off altogether. At work I use $COMPANY_LLM for a nice enhanced autocomplete experience, and it works pretty well overall though it does hallucinate enough to still cause me problems.

Also, I think prompt engineers …

Knowing how to use a tool is not engineering.

10 Likes
3

None of 'em. Either the error rate’s high enough it’s faster not to bother or the permissions requested to install are so far out of scope to advertised capability it’s transparently obvious the model’s offered for purposes other than end user value.

Indeed. IMO it says something about the state of LLMs as candidate tools and the surrounding user mindset prompt engineering’s as much of a thing as it is, though.

3 Likes
4

Wait aren’t you THE Barnacules Nerdgasm? Why don’t you make a fun video like the ones you used to make explaining all this AI stuff.

Me personally using Gemini and ChatGPT to block some ads that ublock fails to catch and to make small tampermonkey scripts to add website functionality. Pretty impressed so far

1 Like
5

I am not nearly so organized as yourself but I have used the following

Googles notebooklm, which is Retrieval-Augmented Generation (RAG) enabled LLM
I use it to study and look up info across a large grange of books and doc I have uploaded to it. it can also make a short podcast style audio only thing.

AMUSE, which is AMDs sponsored free image / video AI application. its fun to play with I will put a photo I took into it and use the filter mode to see how the photo would look in different lighting situation or with different colors to give me some inspiration for when I edit the photo.

I also use the Google and Microsoft LLM just to do random stuff.

That program writing AI thing you did is pretty cool I feel im reading about some science fiction AI rewriting its own code or something.

1 Like
6

Sadly @ipclevel my health took a bad turn and I’ve been fighting against massive chronic pain ever since. It makes it difficult to do a lot of things. I’ve been trying to stream on Twitch during the week and YouTube on the weekends keeping #TechTalk going with AuxxZillary as my co-host and occasionally David Hewlett (The guy who plays Dr. Rodney McKay on Stargate Atlantis) (Saturday 11am PT) .

I would love to return to making fun videos on YouTube again. That honestly has always been the goal. But back then the energy just came out of everywhere and it was easy to just share all my fun ideas, etc. It’s become significantly harder with all the physical and mental hurdles lifes been throwing at me.

But, it’s not all bad, my kiddo is 15 now and growing up super-fast and he’s amazing! My wife has been the best support I’ve ever had in my life, and I love her to death like it’s the first day we met and for the moment I still have what I need to survive so I’m still incredibly happy and grateful.

But thank you so much for the complement, it honestly made my day. But yes, I’d love to get back to the old days for sure and one day I hope that actually happens! :clap:

You should also check out [https://CoPilot.Microsoft.com](https://Microsoft CoPilot) which is Microsoft’s AI investment. I feel like it’s still heavily integrated with OpenAI technology, but it does have a unique feel to it and it’s completely free and now they give you access to Agents and Reinforcement learning for free (similar to Grok 4 Heavy). I’ve been pretty impressed with the level of LLM’s you can gain access too these days absolutely for free. I can’t fathom that any of these guys are making any profit from the free users abusing their system. I think they are just taking the huge cost hits to train their AI’s faster off human interaction and that’s how we’re all paying in the end.

For coding I love Claude-3.7-Sonnet, it’s been the best experience I’ve had but I’m trying to find other LLM’s that can play in the same range as it does. I also like Grok 4 believe it or not despite Elon (the A$$ clown) owning it and it being run by giant gas turbines in the dessert. Not because it’s great or anything, kind of the opposite, it’s easy to jailbreak allowing it to write any base code for scrapers and LLM solvers, etc that I can then proxy the results over to Claude-3.7-Sonnet to refine and ultimately fix.

1 Like
7

I hadn’t even heard of Notebook LM before you mentioned it here. I went and had a look at the example they let you play with, and I can see a lot of value in this. LLM’s are amazing at finding, organizing and scrubbing data. Long are the days of the simple LLM’s that just operate on training alone. Now they have agents, thinking, MCP to run script and code outside of the model to get the output and pull it back in and use it. It’s honestly getting insanely powerful.

I use tons of Image LLM’s both in the Cloud (CoPilot, Grok, FLUX, Google Imagen, etc) and local on my PC I use StableDiffusion and VASE (for video). It looks like this AMUSE is just a wrapper app for Stable Diffusion that accelerates it on AMD cards. That is actually pretty awesome since a lot of AI projects heavily lean into CUDA which is Nvidia proprietary technology, and a few projects use OpenCL but it’s really slow so being optimized for an AMD GPU is a game changer if that’s the hardware you’re rocking.

I use like to use Automatic111 WebUI (AUTOMATIC1111/stable-diffusion-webui: Stable Diffusion web UI) for running Stable Diffusion models, LORA’s and Plugins on my RTX 4090. It honestly does a pretty amazing job. It still struggles like hell with generating text whereas the cloud AI’s are getting brilliant at it now. But, it will all catch up in time as models get more compressed and efficient through various different technologies.

Thanks! Honestly, it’s one of my biggest fears of LLM’s right now. Their coding capabilities are growing by massive leaps and bounds. Most developers are too proud to admit it and just call out faults with it and I understand why. It’s truly threatening to the jobs of people who developed so much of their lives to learning to be excellent coders only to have that skill get trained into a model that crawled their entire body of work GitHub or internal company AI’s crawling their own internal source repositories.

The sad truth is I can’t even remember the last time I wrote a program without AI assistance and I’m a lifetime developer starting at 7 years of age with Apple Basic on an Apple IIe. It’s not because I couldn’t write the code that it does, it’s because even with the mistakes and corrections I can create code 100x faster with AI than unassisted since it knows all the latest API’s and methods from crawling huge bodies of opensource properties online. I noticed that GitHub is starting to block bot crawlers and other AI’s. I tested this the other day on Grok 4 when I told it to download and modify an existing GitHub project. It dropped out to MCP and tried to clone the project locally to the VM instance so it could start modifying the code and it got back an empty file. The server just didn’t send anything but an empty file as the response because it knew it was a bot.

It’s going to be a huge war going forward between the AI’s trying to hunt for more input and the people with that input not wanting the AI’s to have it trying to come up with creative ways to combat it. The irony is that they will have to create AI’s to defeat the other AI’s because convention methods.

Like for instance yesterday I used a Jailbroken Grok 4 instance and Claude-3.7-Sonnet instance to write a python and polish a Python script that can defeat image based Captchas. Even really messed up ones that are hard to read that usually take humans a few attempts. It deploys a small 1.6gb local LLM called Internvl3-2b-instruct which is really good at finding text in images under LLama or LMStudio server and then it grabs the captcha from the website and automatically starts trying to solve it. It gets it wrong a few times, but it usually gets it right in 3-5 attempts. As time goes on more of these local AI’s are going to popup that are optimized on solving Captcha’s using a reinforcement learning system just like the humans do whenever they solve a captcha correctly by giving us access to what’s behind it. AI’s over time will get so good that they won’t just solve a bunch of fuzzy messed up letters in a Captcha but they will be able to solve the puzzle looking exactly like a human. This is exactly why you’ve probably noticed the Captcha’s on the internet are getting so much harder these days. Heck, those captchas where it says to “click on the things with wheels” or all the “fire hydrants” I have another LLM I can load up locally that does image recognition, and it can get those write too and even detect the coordinates where to click.

So, I’m not really sure how people are going to keep bots out in the end since every day they get closer and closer to emulating humans and pretty soon it’s going to be like the Turing test on steroids to try and catch these little buggers. But, in the meantime I’m trying to take the time to learn as much as I can about them and leverage them, so I know their capability and know how to gain some advantage with it for whatever I’m working on and also know it’s pitfalls and work around them.

Some people just see LLM’s as a tool, but technically so is Visual Studio with IntelliSense and auto resolved assembly. At the end of the day, it’s what you use the tools at your disposal to do that make you an engineer. And honestly, companies are going to be interested in human coders that are slow and expensive when an AI assisted coder of a lesser skill can crank out far more work in a shorter period of time. Even if that work requires more debugging which honestly AI is also getting really good at doing. I have Claude code review Grok for instance and between the two of them they always write solid code and rarely make big mistakes.

I will say this though, unless you’re working in an AI enabled IDE like Cursor and using a paid cloud model optimized for coding you’re not going to build any massive projects from a single conversation with an LLM. For instance, the most code I’ve ever got back from an LLM in one go is about 6000 lines. Past that you start having both AI’s start losing context and dropping features, etc which isn’t good. There are tons of prompting techniques you can use though to ensure the things that are most important are always remembered. Especially AI’s that use agents in a virtual environment that can store data as variables and files locally to the session to remind itself of what is the most important.

I legitimately scared for what tomorrow holds and thinks AI is ultimately going to make humans dumber in the long run while also enabling fantastical things. MIT actually did a study on this we talked about on #TechTalk yesterday. They physically looked at people’s brains under a EKG while coding with AI, coding with search engine access only and coding with only books available and the person with books had the highest overall brain activity using the most of their brain. The people using AI had the lowest brain activity because they were so heavily depending on the tool to do all the thinking and calculations.

Also, don’t even get me started on the security issues involved with vibe coding. There are lots of ways for AI to make mistakes and literally put security holes in your code just like a human coder can. And if you don’t code review things yourself by eye and also use other AI’s to catch mistakes in the code written by the other you’re going to sometimes find yourself introducing backdoors that other AI’s can even be convinced to find and exploit for you if you’re a bad actor.

3 Likes
8

Are you cooking some tools and script that you’re gonna share or are you just working on personal projects just for the heck of it?

I find this way of working with AIs fascinating but also stupid in a really funny way. They sell us the idea that AIs are getting incredibly smart and quick, can do everything, everywhere, will take our jobs and so on. Yet we need to formulate our prompts lilterally reverse engineering the neural mapping they’re based on and the token weights they have.

1 Like
9

I find I can reproduce the ~20% productivity drag using simpler autocomplete tech without incurring the additional costs of getting an LLM involved.

If their management’s stupid enough to buy the AI hype cycle the jobs were under threat anyways, just maybe in less obvious ways. Probably the AI everywhere craze will eventually cool down into something where it’s reasonably efficient to apply LLMs on the tasks where they’re better than people. But I’m figuring it’ll be at least a couple years before that starts to emerge. In the meantime the problem I have for production coding is the cost of deploying, prompting, and evaluating models to see what might be helpful is much higher than just writing the code.

Like pretty much every dev everywhere, I have too much code to write and not enough time to write it in. So help would be welcome. But it’s not at all clear to me LLM architectures have the reliability, accuracy, and efficiency to be of net benefit in any broad sense.

Yup, gottta code review and test the output for all that. Which, if you’re going to do a decent job of it, means you’re going to do as much thinking as if you wrote the code yourself. So it’s easy for it to take longer and the chances of getting something better out of the extra process seem pretty marginal for now. Even for basic functional stuff I’ve found generated code tends to increase dev costs (beyond the overhead of doing the generation) because I spend more time writing tests since I don’t trust I spotted all the injected bugs in code review.

In general, the more autocomplete throws at you the more time you spend checking, accepting, and fixing autocompletions and the less time you spend checking the code is actually right. It’s easy for that to increase the cost of laying code down and then it tends to increase testing and debugging time to find problems like autocomplete putting the right code shape but not using the correct variables.

Also, cloud LLMs basically pwn you. Even if you’re writing open source and accept the plagiarism machines are going to steal your work there’s basic issues like GitHub Copilot rooting your auth keys so that, uh, it can control all your repos in all your orgs in order to assist with files you’re developing in one local repo. Microsoft’s not even trying to be subtle here.

I do think LLMish machine learning has some potential to reduce pair programming costs in ways that are useful. But it looks to me like the hype train’s going to have to derail before those get much attantion.

2 Likes
10

ive been using Phi 4 mini instruct but I want to move to Phi 4 mini flash reasoning but there is no GGUF currently afaik. Im getting Gemma 3n e4b it GGUF right now

11

I’m mostly creating stuff for me, but it’s definitely stuff other people could find use with so it will probably end up on GitHub at some point. I wrote a smart controller for my air conditioner in the Nerd Cave since all its relays are fused. I’ve wrote a ton of scripts for solving captcha’s using AI and creating heaps of images that all look majorly different by merging LLM prompt generation with Stable Diffusion using ComfyUI getting pretty cool results. I dabble with a bit of everything.

Right now we have some job security as prompt engineers but we’re getting to the point now where the technology gets easier and easier to use to the point where AI’s will just be able to plug into each other to take up all short comings and all you’ll have to do is define the vague idea of how the system needs to work and the budget and resources available to it and let it rock and roll. Especially when more robotics get tied into it giving AI the abilities to create the physical and non-tangible working together as a total package constantly pulling in new resources and learning from failures faster than humans could. But, there will always be a job for the guy with his hand on the switch to stop it when it gets out of control. Kinda reminds me of jetsons really. :grin:

1 Like
12

Nice, thanks!

I did something like that recently with ESP Home, Home Assistantant and an IR LED. Cool stuff, no pun intended.

Well just share what can be shared and will not get you in trouble, like the auto captcha skipper I think hahaha

That’s what they’re trying to do moving everything to agentic AI. Spread the knowledge, spread the resoning to multiple specialized actors that talk to each other and you get a much smarter and accurate AI compared to a single gigantic block of knowledge with a single traninig catered towards all kind of fields of knowledge.

Yet I still feel like we’re still in a “monekys on typewriters” situations, more or less. That future seems really far.

I think it’s a nice step forward getting humanity out of heavy work that’s so important yet very attritional and take all these workers towards better living conditions. Even being the guy that presses the killswitch could be an appreciated improvement, in my opinion.

Man the Jetsons were so anachronistic but are getting truer by the day! I remember a couple of episodes and we’re really going there, minus the flying cars and some other little things hahaha

13

i gave up on running self-hosted LLM’s and I just use ChatGPT now, mostly for programming advice and other devops solutions.

14

Love this post. I’m going to try some of the things you mentioned once I new workstation arrives!