Keith @gnukeith

Small model but has the biggest personality out of the family. The rest are just very bland and boring, unfortunately. Thanks! I just work on stuff @brave and @obscuravpn then I try working on some research projects when I have time. Currently I'm figuring out how to massively reduce hallucination rates and tool call usage with LLMs using a Jev-style model. I somewhat did this with rei-v1 which is an experimental model that I created to make models search more relevant information. huggingface.co/gnukeith/rei-v1 I used to do a lot more, now I just kind of live in my own head and come up with theories and try to prove them.

For some reason a lot of you find this interesting so I'll try to explain this a little. To me LLMs are simply big text summarizers, I want them to be good at retrieving relevant information and what potentially would be relevant as well so I don't need to go back and forth with it. The only mode I'll let this slide with is Opus 5.5, it is the WORST frontier model when it comes to searching information. Grok 4.7 is the best one by far at this simply because it has access to X. GLM and Deepseek are surprisingly good at this too, Kimi is only good at this if you have Agent Swarm enabled. Qwen 3.8 27B and Nemotron 3.5 Lightning, I will stay on this hill. Nemotron 3.5 Lightning is the most underrated model from Nvidia ever. It is so quick and so smart that it is incredible. Everyone loves Qwen 3.8 27B, I don't think I need to explain much, capable, small enough to host locally but overthinks. It has an existential crisis every time you ask it a question. I love Kimi's personality, if Kimi ever makes a model that is 400B or less I will without a single doubt in my mind run it locally. We stan Kimi. Gemini. Fuck you. OpenAI models, over hated, Luna specifically is awesome. Astra is great but for some reason I can't use GPT 6.1 Sol there is something about that model that just doesn't sit with me. It doesn't vibe with me.

Keith @gnukeith

10 hours ago

My (personal) usefulness AI tier list.

35 6 688 33K 201

Nemotron is so absurdly fast that it feels like magic! For local stuff it's my go to for 80% of the tasks while Qwen 3.8 27B handles the rest. GPT 6.1 on low/med is way more capable in terms of intelligence compared to Luna but Luna feels like the only OpenAI model now with a personality, I think they went too hard on the personality with GPT 4o and GPT 4.5 and then hard pivoted to making it like a HR worker for HR, HR that owns all the other HR departments.

Keith @gnukeith

7 hours ago

@unlimwallworks Nemotron on a 5090 + Brave Search is so incredible, because Nemotron is really smart and really fast, so its like using something like Perplexity at 10x speed. I just look at LLMs as smart summarizers, you really don't need big models to search the internet anymore, even a 9B

0 0 2 422 0

It's so incredibly narrow minded that it probably had a tunnel for a mind. I do wish OpenAI models felt more "free" in their thinking because it just feels like they are trying to achieve the goal with as few steps possible. Which to their credit, probably saves a lot of tokens but I end up having to prompt it couple of times to get where I we want. Thank you ^^

Keith @gnukeith

7 hours ago

@Halios Significantly better at searching the internet for the right information, I'm not sure how to explain it but it expands really well. ChatGPT models in my experience have been very goal driven, you ask it to do something and it goes straight to it but nothing more, it doesn't

0 0 0 2K 0

My (personal) usefulness AI tier list.

Also I couldn't give less shit about your benchmark numbers, I'm not after benchmarks. I'm looking for models that are actually useful to me and I have different use cases and goals than you.

@Tim_The_Sandman Right?? It such a good model, I can't wait for Qwen 4 27B!

@sczy0x Yes! However I've been using it less due to testing Grok Bot and Snowpaw (my own thing) but it is really really good.

Significantly better at searching the internet for the right information, I'm not sure how to explain it but it expands really well. ChatGPT models in my experience have been very goal driven, you ask it to do something and it goes straight to it but nothing more, it doesn't expand well it is a very tunnel vision model.

Nemotron on a 5090 + Brave Search is so incredible, because Nemotron is really smart and really fast, so its like using something like Perplexity at 10x speed. I just look at LLMs as smart summarizers, you really don't need big models to search the internet anymore, even a 9B model is able to do it decently. For some reason Claude models are absolutely horrendous at searching new information, Grok is amazing at it.

@nickpio Very good idea!

@Programazing Mostly because AI's have been getting better at searching the internet so there isn't much to Perplexity anymore, their council mode is great though!

@kurufal I rate it based on usefulness, Astra was significantly more useful to me than GPT 6.1 Sol, maybe for you it is a B tier model but for me its staying in D.

@vivescene The reason why it is super useful to me at least is that it's really good at cyber, I don't have to worry about getting hit with "Hey this is bad and you shouldn't do this!!" It just does what I ask it to do and it does fine

@Moretti_ale90 Collected a lot of data and didn't even do anything good with it.

@0xcowsecurity Thank you, I knew you would understand

@Jush21e8 No one said wasn’t

This is the new GrapheneOS boot animation. It's generated frame-by-frame using a script. It has a minimum 3.5s total runtime for the intro including a small pause. The rest of the animation gets interrupted once the device is booted to avoid any delay so there can't be an outro.

No wonder people are concerned about data centers, the clouds are at their limits!