Blak

pasionate about Technology , HVAC & Coding , Optimizing comfort with smart engineering

[Blak @blako87]

@rasbt first reasoning are traces for humans to be able to see what the llm will gave you ,for an response! an llm doesent win n intelligent with reasoning!

[Blak @blako87]

@0xSero i tried both ! but hermes bloated python files its a pain ! openclaw has to manny files like a rain forest ! these are experimental " tools" even with the most advance model they do not handle properly instruction!

[Blak @blako87]

@kyzoroX just for inference are ok ! but to train models on these cards its very painfull

[Blak @blako87]

@0xSero ornith is on 3 place ! agentic tool calls are running very food

[Blak @blako87]

@GoogleAIStudio i have tried but allways returns 402 api !!! so basicly paying for 402 requests im not intresting

[Blak @blako87]

@jun_song keep on going! stay focus! you will have succes

[Blak @blako87]

@TheAhmadOsman at first run inference is ok! fine tune rl etc.. every one say that! but almost nobody talks about set up properly and professionell training data pipeline for specific domains[ not codind!!! ] when i see from you a professionell training pipeline?

[Blak @blako87]

@jun_song why do you ned Premium +?

[Blak @blako87]

@ornith_ hey ! i would also apreciate when you guys release also some data sets! toolcalling and agentic behavior. training pipeline , training data pipeline ! this would contribute a lot to opensource comunity!

[Blak @blako87]

@babayagatwt the training pipeline & quality! this is the most important thing in ai! without it you will never have an model!! but nobody talks about it, beacause just a few well skilled devs knows how to do it!! think about it ! the model is the box , without the right tools is just a box!

[Blak @blako87]

@ivanfioravanti try @OpenRouter there is also available!

[Blak @blako87]

@sudoingX without the data set are this weights just weights! but with right data sets along them ?this gives you the power to improve your model

[Blak @blako87]

@analogalok please tell us how long it did to you to set up hermes to be able to do that? then do you have used frontier models for this? or do you have done manualy? why just posting the end result? why not the efort that you put it in it? cann tell you , you spend weeks on this!!!!

[Blak @blako87]

@sudoingX so in toolcalls when the context filling about 90% the models bekame repetitive !they arr looping with the same output! so the best context windows on qwen3.6 famili is 192k !

[Blak @blako87]

@jun_song the main issue are , that out there to many languages are. the llms are powerfull and cann write this! the software stack is to bloatet with so many different languages,

[Blak @blako87]

@analogalok be carefully with no quality loss! if you know math you will know what its hapens when you cut down the weights full byte

[Blak @blako87]

@jun_song i dont love hypes! we all know that nothing is one shot! engineering is a step by step ,hard thinking process, how to solve real world problems! so speaking the truth is allway the right way, this give the people the ability to think about the right things! we need more facts!

[Blak @blako87]

@Nabilkeano @jun_song every one thinks europe is behind! but the truth is a lot of EU companys have AI !but closed behind doors. we have Ai !working on real environment engineering! so yeah dont believe mass media!

[Blak @blako87]

@Hikari_07_jp im am very interested in your training data pipeline / format/ training pipeline! i saw that alone cann increase an llm performance very well ! so i will be glad to see some data sets on hugingface, from you. special for tool calls and agentic behavior. Nice work by the way!!

[Blak @blako87]

@HermesAgentTips plc trying onnlokal device