←
Marco Fahmi2,639 posts
Marco Fahmi avatar

Marco Fahmi

@dataronin

Mostly lurking.

◷ Joined January 2011▧ 44 Photos and videos
LSE Impact Blog avatar
⟳ Marco Fahmi retweeted
LSE Impact Blog@LSEImpactBlog·26 May 2024

✍️Dorothea Strecker, Heinz Pampel, Rouven Schabinger and Nina Leonie Weisweiler, explore how common data repository shutdowns are and suggest what can be done to ensure data preservation in the long-term. #OpenData wp.me/p4m9em-cPR

Jeni Tennison avatar
⟳ Marco Fahmi retweeted
Jeni Tennison@JeniT·26 May 2024

I am really troubled by the proposal from the Tony Blair Institute for a "National Data Trust" (NDT) for health data, particularly by the idea that the next government might actually go for it. institute.global/insights/po…

NHMRC CRE on Achieving the Tobacco Endgame avatar
⟳ Marco Fahmi retweeted
NHMRC CRE on Achieving the Tobacco Endgame@CREtobacco·15 Dec 2023

Congratulations 🎉 to CRE researcher Dr Carmen Lim @Cwernlim who has been awarded $660,000 by the @nhmrc to develop a program to understand how pro-vaping campaigns on social media influence young peoples' attitudes towards the use of e-cigarettes.

UQ News@UQ_News15 Dec 2023

#UQ researcher Dr Carmen Lim is analysing the social media influences behind young people taking up vaping - and developing online content to support them to quit. More here: tinyurl.com/yt5ffavz @Cwernlim @UQHealth @ncysur @nhmrc 📸Adobe

Post attachment
Kevin Gee avatar
⟳ Marco Fahmi retweeted
Kevin Gee@kevg1412·17 Apr 2023

"He's not even a 10x engineer. He's like, 100x, or 1,000x engineer" Gmail creator Paul Buchheit on how Bret Taylor once rewrote Google Maps in a single weekend:

Post attachment
Axios avatar
⟳ Marco Fahmi retweeted
Axios@axios·24 Aug 2023

As adoption of generative AI grows, providers are hoping that greater transparency about how they do and don't use customers' data will increase those clients' trust in the technology. trib.al/tCr3DVd

Rachel Woods avatar
⟳ Marco Fahmi retweeted
Rachel Woods@rachel_l_woods·18 Aug 2023

There's a resurgence of interest in fine tuning LLMs I've yet to see a successful public use case where fine tuning > prompting. But here's where I see fine tuning *mattering*: First, fine tuning is for teaching an LLM specific tasks or behaviors Not teaching an LLM new knowledge. For new knowledge, use Retrieval (store your data in an outside database and strategically pull the right chunks in to give the LLM context to your question) But even in teaching LLMs specific tasks or behaviors - here's the catch... LLMs are remarkably good at picking up tasks and behaviors from just a good prompt THIS is what makes LLMs mind blowing after all So that begs the question. Where is fine tuning actually helpful? Some use cases I could see developing are teaching LLMs tasks that are exceptionally difficult to describe, or fit into ~10 examples you can add to a prompt. One way to think about this: if it would take someone a few weeks doing a task to 'master it' instead of being able to read training materials and get the picture... That *may* be a use case for fine tuning But proceed with caution To truly teach an LLM a new behavior or task, you'll need to treat this like a machine learning project, not just throwing examples in and getting magic in return (which it still blows my mind that ChatGPT does this so well for us). Things like: - Dataset design - Training and test data - Overfitting + more as the tooling around fine tuning gets more sophisticated The other obvious use case is cost. If you can get a super small language model to do a task instead of GPT-4, there's meaningful cost savings there. And if you're using a language model to do large scale tasks like triaging your customer support inbox, or analyzing public data for insights The costs can add up. But if you're wondering where the heck to invest in fine tuning... My answer at the moment for most businesses is still: Make sure you can't do it with prompts.