Just realised our Sponge Example paper is on the ETSI report "On the Role of Hardware in the Security of AI" etsi.org/newsroom/press…. So if I think carefully, basically I am making the world a better place Ilia Shumailov🦔 ?

My genuine questions on this: 1) a more advanced activation function? and 2) can we try ImageNet 🤔?

Thinking of when to buy a house is tiring, what if I train a time sequence model + language model to predict both housing price + mortgage rates 🤔?

huh? 🤣 sites.google.com/view/kpop-in-n…

So programmable espresso is a thing ... (but at a cost of 3000 pounds) github.com/decentespresso

Pytorch 2.0 == complie to ur own backend

What happens when generated data of one LLM becomes training data of another LLM? Turns out that models start forgetting the real distribution and as the process repeats models develop dementia. cl.cam.ac.uk/~is410/Papers/…

We're now listening to Tim Clifford talking about Imperceptible and blackbox-undetectable backdoors in compiled neural networks

As machine-learning models become more capable, questions arise surrounding how their use might be governed. In our latest paper we consider "hardware locking", a governance method where the use of a particular model is locked to a particular hardware platform.

This week in Nature: Garbage out - AI models trained on AI-generated data descend into gibberish Browse the full issue: nature.com/nature/volumes…

Fun fact, “realDonaldTrump” is a single token (74505) in tiktoken, the tokenizer used by GPT and LLaMA3.1😂😂

Congratulations to Can Xiao (co-supervised by Aaron Zhao and me), whose paper "Refining Datapath for Microscaling ViTs" has been accepted to #FPL 2025. Preprint to follow. hashtag #FPGA

I spent a few hours in my weekend on making this little integration MyArxivSanity (lnkd.in/gp4FSm7B) - it periodically crawls arXiv papers, storing those GPT predicts you'll like directly in your Notion database. Check it out here: github.com/Aaron-Zhao123/…

We actually saw some of this magic with hardware when running evals for locking models into hardware [1]. When we were computing fingerprints for rented hardware often things didnt match -- hardware name didnt correspond to its observable characteristics. We ended up writing