Anjum Sayed

🏁 Lead Data Scientist at @McLarenF1 🌱️ Building 🏅️ @Kaggle Competitions Grandmaster (datasaurus) 🎒️ Pacific Crest Trail Class of 2022

[Anjum Sayed @AnjumSayed]

I'm finally a Kaggle Competitions ✨Grandmaster✨

Wow, what an epic journey 💙. I couldn't have done it without my amazing teammates @Yamashita_136, @KinasRemek, @WojciechRosa & Alvor in the IceCube Neutrino Detection competition.

Feeling proud and grateful to finally get here

[Anjum Sayed @AnjumSayed]

@dieworkwear What we all need to see is how you organise all your inspo photos on your computer

[Anjum Sayed @AnjumSayed]

@chrisalbon Have you tried tuning your validation set?

[Anjum Sayed @AnjumSayed]

@mattpocockuk It’s really good now. My only complaint is that it’s greedy wrt to disk space, as it once it grabs a slice, it won’t give it back until you resize the partition via a bunch pwsh cmds. I’ve worked around this by just keeping all my files outside of WSL and mounting volumes

[Anjum Sayed @AnjumSayed]

@ucl @Bloomberg Ok to be continued tomorrow…

[Anjum Sayed @AnjumSayed]

Let's learn how to solve a Rubik's cube for the first time with @karpathy

[Anjum Sayed @AnjumSayed]

@ucl @Bloomberg I know it sounds contradictory but it’s getting easier?!?

[Anjum Sayed @AnjumSayed]

When they’re able to win medals using models trained on a GTX1080, then they’ll have my attention 😅

[OpenAI @OpenAI]

We’re releasing a new benchmark, MLE-bench, to measure how well AI agents perform at machine learning engineering. The benchmark consists of 75 machine learning engineering-related competitions sourced from Kaggle. https://t.co/zFoYr5pbiF

[Anjum Sayed @AnjumSayed]

@dannypostma I use RAG using an embedding model with a long context length - specifically Alibaba-NLP/gte-base-en-v1.5 which seems to be a good balance for my use case at scale. With a 8192 token context length, a long transcript can be broken into 5-10 chunks

[Anjum Sayed @AnjumSayed]

@arvidkahl Are you currently using distil-whisper? I haven’t seen any benchmarks yet but was going to check it out when I’m home next week for a full comparison. The discussion page in the repo suggests some teething issues with timestamps too

[Anjum Sayed @AnjumSayed]

@rajivayyangar I literally built a white noise machine (just for fun and to learn new things). I didn’t put it on PH because I thought it was too boring. Kicking myself now 😂 https://t.co/eTmjBNM17H

[Anjum Sayed @AnjumSayed]

@levelsio The amazing thing is that each one of those images started life as an image that looks just like this one 🤯

[Anjum Sayed @AnjumSayed]

@JamesIvings But yes totally agree, once the final technical distinctions are ironed out, that line is going to look very blurry

[Anjum Sayed @AnjumSayed]

@JamesIvings I guess as of today, it is still possible to draw a technical distinction by testing an LLMs ability to learn in-context (ICL). At the moment ICL is very weak (and poorly understood) and to get to AGI, we’ll need to figure that out

[Anjum Sayed @AnjumSayed]

@JamesIvings Ah ok yes I missed your original point