Anjana Vakil

Geoffrey Hinton's contributions is significant enough that I won't quibble with the category for his prize. (Nobel committee could create a special category just for him for all I care.) Widely (and rightfully) recognized as a leading expert in deep learning, his research touches everything from * Foundational concepts such as backprop https://lnkd.in/gMunbBzA * Foundational model architectures such as AlexNet https://lnkd.in/gq2JPJ6h * Optimization algorithms such as Momentum and RMSProp * Visualization techniques such as t-SNE https://lnkd.in/gjuw3vhA I don't think it's possible to overstate how profound the impact of machine learning and artificial intelligence will have on society, promising to unlock discoveries, improving productivity. Even ideas about what mastery of language or reasoning means, will need to be re-thought. Professor Hinton deserves all the credit for all of this.

31

A blog on (very) nerdy topic that I worked on with our (amazing) intern Xiangpeng Hao this year: https://lnkd.in/eQDfvNfH. We quantify how much of the discussion about changing Apache Parquet metadata was shiny-new-object syndrome and how much was actual technical limitations. We think there is easily a factor of 4 to be had from existing implementations before new formats are needed

290

12 Comments

Why LLM Agents Don't Work Today If you're involved LLMs, you might be wondering why there seems to be so much hype about LLM agents -- but underwhelming results. I'm building Langur, which attempts to address what I think are the critical issues with current approaches to agentic frameworks. If that sounds interesting to you, read on, and let me know your thoughts! Link: https://lnkd.in/ezJG4Zw4

9

2 Comments

Looking for resources on how to make effective use of LLMs in a developer workflow, if anyone has any good ones they can share. My GitHub copilot trial has finished and I’m trying to decide if I should continue my subscription. The results I’ve had so far have been a bit underwhelming but I’m willing to accept that I might need to learn how to use it effectively to get the most out of it (hence this post). Its definitely a better autocomplete (at least in certain cases) and so I think it has meant I’ve typed less (always a good thing for these battered old fingers) but I don’t think it’s actually saved me time, as I spend the time I would’ve spent typing reading/checking the things it’s generated. I’d probably be won over enough by the better autocomplete to keep the subscription if it weren’t for the environmental impact. But just like I get the train where possible rather than flying, I’m wiling to accept a certain level of friction to limit my carbon footprint. LLMs would have to demonstrate a step change in productivity to justify this for me I think. I haven’t seen this step change yet, but maybe I’m looking in the wrong places?

3

11 Comments

Some great hacks coming out from #BeyondTheSpeckleverse

1

Anthropic has just introduced „computer use” for Claude 3.5. It can use the #gui like human - move the mouse, fill the forms. It looks like a killer app for me (and #RPA killer in the future when it will evolve). Claude looks like a actual winner in the AI/LLM race. It’s more stable, more secure and more usable for business cases than GPT. And it’s „less drama” (as they pointed out in their ad). #AI #LLM https://lnkd.in/d6wciU2C

16

5 Comments

Part 1 of blog post about my FAIR library out on medium. This one dives deep on the core algorithm and optimizations. I want to do at least one more on the practical usage tips in the future. https://lnkd.in/g5UjA-zv

27

I've noticed a trend of more projects using LLM as Judge, but too many people are using just one LLM as a judge. From my experience writing NIAN, I think using just one LLM will lead to inconsistent results. Fortunately, LLMonPy makes is very easy to use multiple LLMs as judges. You can read more here: https://lnkd.in/gQRzBDTb

8

1 Comment

A short blog post I wrote, explaining the HyDe method for improving the relevancy of your RAG retrieval. It involves generating hypothetical documents using LLMs and searching for those instead of the original query. :)

17

2 Comments

This article is rather technical and is a bit of a deep dive, so for those who aren't as versed in AI jargon let me summarize the nerd lingo in here. One of the big challenges in AI/LLM development is actually understanding how the AI connects all the different data points when it's training on data. The work that Anthropic has recently done has created some important breakthroughs in interpreting what an AI model thinks is important from its training data AND how to influence the models focus. The example they use is how they were able to train a chat bot to always include the Golden Gate bridge in its responses. This is important because not only will this improve accuracy, but safety and security when it comes to the development and use of these AI's. super exciting! #ai #llm #aiinbusiness

2

1 Comment

Completely agree with this. AI is great for things like drafts, formatting, or boilerplate text, but it still takes a human expert to refine, structure, and make sure everything’s accurate and readable. Like with coding, the polish, the edge cases, and the attention to detail are where the real value happens. In tech writing, it’s not just about the words. We’re testers, user advocates, and part of the broader team ensuring customer success. We don’t just document, we make sure workflows make sense, identify gaps, and even suggest improvements to the product. AI can speed up workflows, but it can’t test a process, anticipate user needs, or bring the clarity and judgment that real writing requires. It’s a partner, not a replacement.

37

4 Comments

How do you format your Few-Shot Prompts? We almost exclusively use: Q: <Input> A: <Output> But would love to hear what others use. With Nathan Labenz on https://lnkd.in/egdXwCpy

13

Smart approach to visualize interaction provenance and gain insights from the user interaction with form elements.

5

My new standard attack is evolving to: "Why are you anthropomorphizing LLMs? Would you anthropomorphize 7zip or JPEG?" I'm reposting this article to help reframe the discussion. LLMs are, in fact, super lossy compressors. That's it. That's all they are. They don't care about your feelings because super lossy compressors don't have feelings and they don't have care. Stochastic parrot, or my new improved knockoff, stochastic pigeon, do not get to the heart of the problem. A compression / decompression algorithm is not a sentient being. End of story. Argument over. #WrittenByMe https://lnkd.in/gGaqyRMg

5

12 Comments

Have you used Cursor? How I use Cursor to fine tune my first LLM https://lnkd.in/eGFKcT8d

18

With the recent announcement that Finale is sun-setting, I’ve started learning LilyPond. https://lilypond.org/ I wish I had done this sooner. It’s a lot of fun to work with. Though LilyPond is not the solution for everyone it has many advantages. - 100% free and always will be - Text-based file input makes for much easier debugging - Highly customizable LilyPond editing is text based. Files can be read (relatively) easily by humans and tracked via version control. Here’s my current LilyPond Setup: https://lnkd.in/g8DfP4my Sample pdfs can be found in /files This is a work in progress. I may add a simple tutorial down the line. If anyone in my network is a LilyPond enthusiast please reach out, I’d love to pick your brain.

4

2 Comments

I started writing regularly about AI. If you're an innovators or product developer eager to apply the latest in AI, whether you're in a startup or a big company, consider checking https://lnkd.in/gGdW8PYe I won't teach you ML, but I share real-world insights from launching - and sometimes failing at - numerous ML features on a large scale.

54

2 Comments

https://lnkd.in/dDg2Dvcw Really recommend this video. It touches on most things I learned over the year on how to work with LLMs. Really hits the nail on important topics like: - Different tier models for different tasks - Meta-prompting and how to iterate and test prompts, he shows his workflow which is almost exactly mine workflow too for some time - Test-driven Development with LLMs - something I argue for since 2022 since before ChatGPT :D And still almost no one speaks about that... I think its next big thing in efficiently working with them - Distilling AI Model Performance - pointless waste of money without eval dataset - Model Routing and Multi-Model Orchestration - challenges 

3

3 Comments