amrb

It's a red flag that the 1.2bil model has to fit in gpu memory, happy to be provided wrong when the code drops

Can't say I mastered the concept either, I'm waiting for the code [0] to be release so I can run some head-to-head tests.

[0] https://github.com/Aleph-Alpha/trigrams

An alternative approache to BPE tokenization https://arxiv.org/abs/2406.19223

Can check out their project at https://github.com/bigscience-workshop/petals

Speaking of quantized vectors https://huggingface.co/papers/2309.14717

Great project and I'm happy to see it expand to more models!

What's the new reddit to try?

Anything can end up in logs, then it depends on getting access to hosted splunk via employee creds, for a hypothetical breach.

There is a salary requirement, as not to under cut local works. Of course if you working over 40 hours a week maybe the company gets it's pound of flesh!

Good talk on the paper https://www.youtube.com/watch?v=ut5kp56wW_4

So we just recreated all of the previous SQL injection security issues in LLM's, fun times

The "one weird trick" to squeeze limes for extra juice

Group names are 10/10

Sorry to hear you had this experience, I would say its worth giving another go maybe you can check out the IRC first to check the vibe before committing.

I appreciate they say they need to learn more about 'exec' when asking GPT4, it also plays well into some of the reading strategies I've seen to get a high-level understanding then read the documentation with more general idea going in, could also lower frustration for new engineers so they push through to the end!

Sounds cool but I'm not seeing how this discovers IOC's or reverse engineers malware.

I'd like to see a yearly benchmark for models, could be logic puzzles or a suit of tasks but as it stands there is not good way to measure the ability of models.

Can some one test fizzbuzz, sounds silly tho a lot of models fail on the combination check in my tests.

You should look up for hackerspace's in you city, it's a space with tools to do projects.

We have seen hair works as a Nvidia only tech.. Compatibly could be why the licence theory was rejected.

What is chatgpt-turbo api pricing like .0001 pre 1k tokens, your paying more in workers salary at the moment.

A type of battery the small round one.. not the year

I could see Nvidia licensing the tech to game publishers but they didn't get much uptake, so open source it is.

feel more likely to rip game assets and build in a supported engine like Unreal.

The of the widescreen fixes for the Unreal games was the addition of a dll file in the game folder.

https://en.m.wikipedia.org/wiki/Strange_loop

Seems the fine tuned models I.e. gpt4all and alpaca are trained as LoRa's. Best advice is jump in and try the demos on hugging space!

Silly question, but would it be more impactful to use grey or open models to achieve the goals? End of the day the finally model may need to run outside a datacenter so if people don't fine-tune models this could be a limiting factor.

LoRa has been pretty popular and untill the llama leak was not aware of it, maybe will see something cool out of the open assistant project, we have a lot of English and Spanish prompts and was crazy to see people doing an massive open source project for ML.

Would like to see a yearly benchmark's for models like this!