recently, i spent some time working on cross-lingual alignment for LLMs via encoder injection! treating languages as modalities is a compute-efficient way to extend understanding of low-resource languages without extending pre-training or the tokenizer

hypothetically, if one wanted to research at a frontier lab in 1.5 years (hypothetically winter 2027) and wanted to know how to develop the necessary skills and credentials to do so, what would you suggest to them (asking for a friend)

holy guacomole Alice in borderland season 3 is out

we built PokeOS. Poke now - interacts with your iMessage proactively - fills forms/searches on your local browser - can write/debug code directly in your IDE

built with @_rajanagarwal in under 2hrs at cursor hackathon

cursor can now read your terminal like we do! we built observe, a way for cursor to monitor your terminal in real-time and execute fixes based on logs before the program even finishes

thus, it is the well-timed frog, just like cursor that acts on your live terminal, that ends up the farthest

but leaping too often indubitably tires out the legs

a frog must leap before the lily pad sinks

i built shadow: a powerful open-source background coding agent with a real-time interface. for async, parallel work on long-running coding tasks! made in a few weeks with @_rajanagarwal @elijahkurien

Just working on my poetry

Too much bread got me sour on this dough

🤯 Gemma 3 is available on Ollama! Multimodal is here for Gemma. 1B (text-only): ollama run gemma3:1b 4B: ollama run gemma3:4b 12B: ollama run gemma3:12b 27B: ollama run gemma3:27b

sike,

https://t.co/HOUfWcLzUk

frontend joshua just dropped

Introducing Outlook: A novel ML indicator that consistently outperforms both individual stocks and major indices. Leverages weighted future returns and custom decay functions to identify market trends across all conditions. Thread below

Day 13: ChatGPT Wrapped End 2024 with another wrapper. Built in 48 hours with @sdand https://t.co/xKbCzVahIY

we're turning @uwaterloo into a supercomputer! arceus is a cross-device distributed compute network for training large models, using model/tensor/pipeline parallelism. you can train anything, from deep neural networks to language models on the network. oss & deployment soon!

We're turning @UWaterloo into a supercomputer 💻 Arceus is a distributed training marketplace for renting unused MacBook compute to a global network, giving developers cheap compute to train large models on consumer hardware clusters! With @_rajanagarwal @josh1yan @simerusm