@DarioAmodei Committing to embedded safety evaluators the same month you cut @AISecurityInst out of a frontier release is a strange way to ask to be trusted on verification.
the thinking that "everyone will just vibe code their own software" was born from within a bubble
what's happening to software development right now has happened many, many times before to other things. you just need to look
web 2.0 allowed everyone to write blogs. did everyone write blogs?
instagram allowed everyone to share beautiful photos. did everyone share beautiful photos?
tiktok allowed everyone to make short videos. did everyone make short videos?
now vibe coding allows everyone to build software, and you think everyone will start building software? this time around, people are just different?
the mainstream is always a consumer, never a creator. go talk to some people outside of our tech bubble and you'll see - they literally don't give a f about vibe coding
@JFPuget@lifeofjer Don't shoot the messenger ๐ But tbh, some guardrails and best practices (even for startups) need to be in place before giving agents full control. Or they should be fine with taking the hit when prioritising speed over reliability.
@JFPuget@lifeofjer They do mention somewhere they wouldn't have stored their tokens if they had known it could be used for prod DBs too. Still begs the question of why give full reign to agents when startup environments are known to be not-standardised by definition.
๐๐จ๐๐จ
NN loss landscapes are full of permutation symmetries, ie. swap any 2 units in a hidden layer. What does this mean for SGD? Is this practically useful?
For the past 5 yrs these Qs have fascinated me. Today, I am ready to announce "Git Re-Basin"!
https://t.co/mRu5k3ONUm
This latest release for Transformers Interpret (v0.8.1) brings some features I'm really excited about particularly the new pairwise classification explainer. (1/8)
https://t.co/FZxgBQVeUj
What if I told you that you can *easily* control the behavior of GPT and change it in particular directions of your choice, with only a few simple and intuitive steps?
Meet โจLM-Debuggerโจ, an open-source interactive tool for inspection and intervention in transformer LMs ๐1/8
2021: we canโt reproduce big labโs research because we donโt have a 1000 GPUs
2022: we canโt reproduce this research because we donโt have a nuclear reactor.
@akilan27 Upto 3 vowels in the first guess is ok. But you have to filter consonants as well. Also adieu has barely any scope for greens except "i" maybe.
In 2021, I trained 693 models over 239 hours. I tuned 5 hyperparameters per model on average. Get insights on your model training with @wandb! #myyearwrapped
And I only started using WandB by late August.