Marcin Junczys-Dowmunt (Marian NMT)

Still no bookmarks?

Hi, the Microsoft Translator research team is looking for an intern for the summer. If you a PhD student in Machine Translation, Natural Language Processing, or related, check it out: aka.ms/mtintern

Just 10 days after o1's public debut, we’re thrilled to unveil the open-source version of the technique behind its success: scaling test-time compute

By giving models more "time to think," Llama 1B outperforms Llama 8B in math—beating a model 8x its size. The full recipe is open-source!

Rant: Apparently every vector-based sentence alignment tool insists on having an unusable file-based API.

It's more confusing than that. It does exist and seems to mean crucifix which had me even more confused. Suddenly very high stakes 😀

Rood?

Missing bookmarks are a much bigger deal for me. But I think it's funny that they didn't go for one of the most requested features. Seemed like an easy win.

Wrote up some notes on Microsoft's new Phi-4 LLM. They trained it on a LOT of synthetic data, and the details of how and why they did that are really interesting.
https://simonwillison.net/2024/Dec/15/phi-4-technical-report/

Chart of time vs: - number of cameras (exponentially increasing ), - giant squid footage (exponentially increasing ), - bigfoot footage (small and not increasing), and - good quality UFO footage (small and not increasing)

It's messier, but I think this one slaps the point home a bit stronger by adding the giant squid footage. I think unique weather, like lighting sprites, would make the point just as well.

Frontier Models are Capable of In-context Scheming Frontier models are increasingly trained and deployed as autonomous agent. One safety concern is that AI agents might covertly pursue misaligned goals, hiding their true capabilities and objectives - ...

the anthropomorphizing in this LLM scheming paper is through the roof and the interpretations are wild, but still a cute set of experiments and a fun skim, showing some interesting behaviors.

arxiv.org/abs/2412.04984

🚀 Introducing the Byte Latent Transformer (BLT) – A LLM architecture that scales better than Llama 3 using patches instead of tokens 🤯
Paper 📄 dl.fbaipublicfiles.com/blt/BLT__Pat...
Code 🛠️ github.com/facebookrese...

So... no edit button, huh?

Oh no no. VSCode is a an actual recommendation. My actual favorite piece of software that I didn't write.

Whispers in Microsoft: VSCode

This place needs bookmarks.

Leave links. I am old.

And they laughed at us when we pursued PhDs. Who's laughing now?

Ah, are these size-limited? And you guys continue with running numbering?

Oh. There you are. Where's that starter pack?

You convinced me 😉

Hi!