JPeg

[JPeg @jpeg729]

@sudoingX @FrameworkPuter The hipfire project looks like a way to run llms at high speed on anything rocm compatible

[JPeg @jpeg729]

@Pirat_Nation Or use uBlock Origin Lite which works pretty well

[JPeg @jpeg729]

@Object_Zero_ If we can multiply the clock speed by 200, then maybe we can reduce the transistor count and still gain performance. This might be an interesting short term approach

[JPeg @jpeg729]

@rekokylnoeht @TeksEdge An moe model with expert offloading might run fairly nicely.

[JPeg @jpeg729]

@Aella_Girl A great question for biblical literalists is: was Adam created before or after the animals?

[JPeg @jpeg729]

@giveashitnature Put a much finer grill over it and the mosquitos sit on the grill to lay eggs through it, and the young can't get out. No need for chemicals.

[JPeg @jpeg729]

@ibuildthecloud If you add the contenteditable attribute you get a text editor

[JPeg @jpeg729]

@AiFreak_ @tmophoto @NVIDIAAI Until your coding agent decides to use subagents. Then a little parallelism goes a long way

[JPeg @jpeg729]

@haopeng_uiuc Given that the system prompt is simply tokens positioned between special tokens, does this explain why it is relatively easy to prompt inject a model to act contrary to its system prompt?

[JPeg @jpeg729]

@FUCORY @jarredsumner @wevm_dev @threadreaderapp unroll

[JPeg @jpeg729]

@Mike_Preston17 "it's written in C# and angular". Sound cool, right?

[JPeg @jpeg729]

@giveashitnature If there is a fine mesh a couple of centimetres above the water. They lay eggs through the mesh but the young mosquitos can't get out. No need for chemicals.

[JPeg @jpeg729]

@Mike_Preston17 So milk the tokens while you can, AND buy a spark while they are still relatively cheap

[JPeg @jpeg729]

@techpupparent @francoisfleuret Isn't the same true for images? An expanse of blue sky is far less interesting than a bird flying through it.

[JPeg @jpeg729]

@JGuaq @0xSero @huggingface If you run it yourself then the only tracking possible would be via tool calls, and someone would have noticed by now. Some censorship may be built in, but abliterated versions mostly fix that.

[JPeg @jpeg729]

Maybe RAG could help even then, if it retrieved content addresses instead of content chunks. E.g. file path + line number. Kind of like "this vaguely reminds me of chapter 2 of X", and then you choose whether to re-read X. The file path would hopefully be sufficient to decide whether to re-read the content.

[JPeg @jpeg729]

@svpino In one project I simply told it the source of truth is: 1. High level SPECS.md 2. The tests. I also said use TDD, and gpt-5.4 has been doing it pretty consistently.

[JPeg @jpeg729]

@mr_r0b0t @AgentArchetype @UnslothAI @IntelAI @NVIDIAAI Rocm maybe more mature but it still gets beat by vulkan on the same card.

[JPeg @jpeg729]

@fjzeit When I get time I intend to do create my own harness in c# and to include some of the custom tool parsing tricks from hermes to make it work well with local models.

[JPeg @jpeg729]

@neogoose_btw And if you ask for "books join tags" don't you get each book repeated as many times as it has tags?