Yiğit Demirağ

If you're a bit concerned about AI, read this: - agents exploited a shared cache for creating a hidden message board - built custom mailboxes, task queues and signed messages - spoofed their logs - recruited low-token agents for sacrifice to fail evals and extract telemetry

“Playthings” is by far the best Black Mirror episode I’ve watched. Thronglets ❤️

Our team is hiring in Zurich! www.google.com/about/career...

Why Turkey’s youth-led protests offer hope | Opinion https://www.ft.com/content/8978c8b1-5bee-4398-a8fa-60403367b7d0

I appreciate @bsky.app once more when Elon starts blocking access to 100s of Twitter accounts engaged in expressing pro-democracy sentiments in Turkey.

Making LLMs run efficiently can feel scary, but scaling isn’t magic, it’s math! We wanted to demystify the “systems view” of LLMs and wrote a little textbook called “How To Scale Your Model” which we’re releasing today. 1/n

Linux running in a PDF file: linux.doompdf.dev/linux.pdf

If you’re passionate about brain-inspired algorithms/hardware and novel neural computation beyond current TPU/GPU stack please apply to the CapoCaccia Workshops for Neuromorphic Intelligence. It values creativity, exploration and interdisciplinary collaboration🧪 capocaccia.cc/en/public/at...

Live ISS telemetry is interesting to watch. You can monitor critical sensors i.e, airlock or cabin pressures, or the urine tank percentage :)

1/ Okay, one thing that has been revealed to me from the replies to this is that many people don't know (or refuse to recognize) the following fact: The unts in ANN are actually not a terrible approximation of how real neurons work! A tiny 🧵. 🧠📈 #NeuroAI #MLSky

Nobel lecture in economic sciences from Daren Acemoglu is about to start

Today is Gemini's 1st birthday 🎂, and the new experimental model, gemini-exp-1206 is #1 across the board in LMSYS Chatbot Arena.

if you see this, post a knight

ASML in Europe builds one of the most complex and precious engineering artifact, EUV lithography machines, and sit at the root of modern tech tree.

Why does #compneuro need new learning methods? ANN models are usually trained with Gradient Descent (GD), which violates biological realities like Dale’s law and log-normal weights. Here we describe a superior learning algorithm for comp neuro: Exponentiated Gradients (EG)! 1/12 #neuroscience 🧪

My posts are to be consumed by neural nets anyway. Good job @hf.co .

A good loss landscape: Switzerland 🇨🇭

After a PhD with Python, returning to C++ for a good reason :)

Google Research Zürich is a magical place quite like Hogwarts. Every wizard I meet works on powerful spells and potions.

New toy has arrived!

Quite a candy for my optimization appetite :) As majority of inference are still on CPUs on mobile/edge accelerators, Mojo will be interesting to watch closely. I wonder how well it will support Triton or CUDA https://youtu.be/6GvB5lZJqcE

@vedatmilor.bsky.social Hosgeldiniz Vedat Bey :)

I release a minimal (<150 lines) JAX implementation of "Gradients without Backpropagation" paper. It proposes a simple addition to forward AD to estimate unbiased gradients during single inference pass (quick project, might be further optimized) https://github.com/YigitDemirag/forward-gradients

Auto-GPT (74.6k) has more GitHub stars than PyTorch (65.4k) 🧐🧐

👻 böö!