Twitter Beğeni Hilesi

Oh so you agree companies taking the works and data of folks on the internet, without their consent to train GenAi models is not ok? Good to know 💖

you saw the super bowl ad and heard brad smith in front of congress. now you see what the company is actually doing. color me approximately 0% surprised.

More evidence that the scientific innovations we actually need from AI aren't the ones trained on copyrighted creative output without consent. GraphCast predicts the weather better and faster than what came before. It's trained, of course, on weather data. You can get the

Are there any CTFs involving RE of ML architectures? I want to see competitions where participants are given small quantized LLM’s with obfuscated inference code, partial architecture descriptions and the challenges are things like isolating specific activations, and tuning them

If I train a 124M model on listicle, markdown, small talk, do you think I could top LMSYS?

So incredibly proud of this work and hope to be able to share more tools and prototypes in the coming future! #futurelegaleducation #generativeAI #communicativeagents

if you haven't read the helen toner lunch with the ft yet, you really should: on.ft.com/3Yqg0jL

OCRonos-Vintage is also an example of an "open everything" model: trained with open code on open data under a permissible license, it is distributed under Apache 2.0 license. The training corpus includes exclusively content made available by digital libraries in the public domain

so happy to see other great orgs change the narrative. you can build useful systems that solve real problems without theft or trillions of parameters. it just takes careful, diligent work like Alexander Doria has been openly documenting

even🍓can't solve the root cause of this problem

Reupping this because it hasn't gotten traction. We need help testing a beta version of RECAP. Five minutes to install, then let us know if anything is broken. TY!

cursor is interesting and (another) example of an incumbent (copilot) getting leapfrogged...but, i'm not that impressed relative to "real" agents. e.g., if i want to add a fastapi dependency, don't add a 2023 version to pyproject.toml; run `poetry add fastapi`.

Ouch: Internet Archive losing its appeal. Will be terrible to cultural heritage in the US but also lots of potential consequences for AI training. Fair use is the main legal framework for using copyrighted content. x.com/PublishersWkly…

here's the 2nd cir's Hachette opinion from today, which serious implications for fair use analysis in the "AI copyright" cases. buckle up, gang. aleainstitute.ai/resources/Hach…

wouldn't it be wild if the best way to use cure cancer with AI was to build a model that speaks the language of cancer, not just rip off millions of artists and authors and journalists? who would have thought

the secret behind o1's red teaming % improvement - most of the improved performance is because the model simply times out before ending hidden token generation, which it also does 1-in-3 times anyway.

also, just to be clear, there are still two r's in strawberry if you even slightly vary the prompt.

so many of the problems in ai, both technical and legal/ethical, stem from poor metadata about our data. meanwhile, we've had the Dublin Core standard for a quarter-century. optimistic that adoption will finally accelerate with events like the october Dublin Core conf