Erik Bernhardsson

My ChatGPT experience has been much improved after instructing it to always ask me clarifying questions. It's been helpful for writing better prompts and clarifying ideas. Note to self: Ask more clarifying questions

6

1 Comment

This is a very interesting article on sparse vectors, which can help overcome limitations of vector search related to domain specific keywords. And of course, the code in the example is clear and clean! Sparse vectors have a few values and many zeros. Dense vectors have limited or no zeroes. But really, what we're talking about semantic search on terms where the context is not present or harder to infer. The author presents this as specialized keyword search such as medical terms or other domain specific terminology. "These scores reflect the significance of each term in the context of the document or query, guiding the model to allocate more weight to terms that are likely to be more meaningful for retrieval purposes." Highly recommend reading. Nice work Paul Iusztin. And a nod to Qdrant and their vector search technology that supports sparse vectors. https://lnkd.in/gcBWfTg8

3

Another great paper on In-Context Learning (ICL) for LLMs: https://lnkd.in/e4SP4K5d More and more, we're hearing from builders that ICL is an improvement over fine-tuning for dialing in domain/task specificity. Fine tuning is slow, expensive, and improves specialization at the cost of generalizability. On the other hand, an efficient ICL pipeline is quick and cheap to set up, boosts performance on specialized tasks without sacrificing on general ones, and can be iterated quickly as the project context/use case shifts. If any builders are still choosing fine-tuning over ICL, let's chat. I'd love to learn more!

15

Check out our latest post on Clappy.ai, "Leverage Layout-Aware RAG for More Accurate Retrieval of Knowledge from Documents." Discover how to harness the power of Retrieval-Augmented Generation (RAG) for documents in PDF, Doc, and Markdown formats. Dive into the specifics of implementation and explore our reference repository to get started with advanced RAG for your knowledge base. Don't miss out on this essential read! Continue reading here for all the details: The Power of Hierarchical Indexing with RAG. (https://lnkd.in/gbiVHfcH)

2

#RAG DOES NOT REDUCE HALLUCINATIONS IN #LLMS: Think I've called this out a few times, but the validation is nice - https://lnkd.in/eAi6zxW2. The article makes two well established points: 1) "Hallucination is a problem of reasoning and not relevance! Any amount of relevant text fed through RAG to a machine will retain the original perplexity and entropy in the system to hallucinate independent of the text"; and 2) "The initial challenge arises from the retrieval phase’s reliance on semantic similarity, which is quantified by inner product spaces or cosine similarities. This metric, while capturing semantic closeness, does not differentiate between factually accurate and inaccurate information." This should be clear to everyone, but the industry still shills flimsy RAG scaffolding anyway (shovels to gold miners).

104

34 Comments

Great discussion today here at Antler NYC residency on corrective RAG (from this paper: https://lnkd.in/etf-H5Yt The problem: You ask a question to a large language model, it pulls in relevant information to help answer better, but the information that it references is wrong, so its response is wrong as well. The solution: Add a specialized layer that verifies the correctness of the additional context. If the context is valid, the model uses it to help inform its answer. If the context check fails, the model searches the internet and/or additional documents to find the right context before answering. It's not a foolproof system, but it is a step in the right direction to mitigate hallucinations and improve the quality of the LLM product

22

7 Comments

Today we're launching our newest candidate quality-of-life feature on Ropes - IDE themes. There's a universal dislike for the rigid IDEs that assessment co's force on candidates. These code editors are slow, clunky, and missing the styes/shortcuts that developers enjoy. Developers are unique, and we work in unique ways. Ropes offers two new options to let candidates keep their style: (1) First, many employers on Ropes allow candidates to work in their own IDEs entirely. We've developed a way to gather insights like you were in the room, while allowing developers to use their daily environments. (2) For cases that need a browser-based IDE - we're giving the ability for candidates to choose their style/editor setup before the assessment starts - so they're comfortable when it begins. --- My bet is that if we build a great experience for candidates, we'll be building a great platform for evaluators, too. I hope you'll follow along as we pursue that mission - and we need help! We're hiring founding engineers in NYC - if you're interested, I'd love to talk to you.

47

I think it's import to know that large language models generally suck when it comes to numbers and why. I recently read an article on how poorly LLMs reason, and the problem they used was of the "Bob has X sisters and Bob's sister Mary has Y brothers, how many brothers does Bob have" variety. This requires the LLM to work with numbers by calculating how many siblings there are in order to produce the right answer, and it's something that LLMs have fundamental problems with. It's not an example of poor reasoning, it's an example of LLMs being bad at working with numbers. And they're bad at working with numbers because of "tokenization". Computers can't work with text. They can only process numbers. If you want a computer to do anything with text, there needs to be a numerical representation of the text for the computer to work with. And the way LLMs like GPT-4 turn text into numbers coincidentally makes it very hard for them to work with numbers. For computers, text needs to be split up into smaller parts, and each part is assigned a unique number. If we give each letter in the English alphabet a number like a=1, b=2, etc. then the computer can say 20, 8, 5, 19, 9, 19: 20=t 8=h 5=e 19=s 9=i 19=s And we can see: "thesis" That's what tokenization is. The letters are tokens. 's' is one token with the number 19, and this token appeared twice in "thesis". But separate letters make it difficult to understand the higher meaning behind words. Think how hard it is to understand what someone is saying if they spell every letter individually. It's even worse for computers. It might seem reasonable to just use words as the different parts. "thesis" could just get it's own number. Let's say '1'. And we'll give "the" the number '2'. Then '2,1' would mean "the", "thesis". That approach didn't work though. There's too many words! Especially when you want to include other languages like Mandarin. What modern language models use instead is tokens that are 1-6 characters long. Not just letters, but not full words either: they use tokens as building blocks for sentences. For example: "the" = 1 ' ' (space) = 2 "sis" = 3 "ter" = 4 "the thesis" is represented as 1,2,1,3: 1: "the" 2: " " 1: "the" 3: "sis" "the sister" is represented as 1,2,3,4: 1: "the" 2: " " 3: "sis" 4: "ter" We re-used token 3 ("sis") to represent both "thesis" and "sister". That's how LLMs build sentences. There's a token visualizer for GPT-4 here: https://lnkd.in/gpuQrGP3 So why do LLMs perform badly when it comes to numbers? Because those numbers are also represented as tokens! "5 + 12 + 3 = x" could be represented with the tokens "5 + 12 + " "3 =" " x" It's a mess! LLMs can't really do math, because the tokenization makes it impossible for the LLM to operate on a number basis. It can't separate out "+" or "=". For math, we'd have to go back to the separate-letters method. A model can either produce poems, answer questions and hold conversations, or do math, but not both.

7

If you're connected with me, you know I scream about how necessary integrations are. Check out this piece where I explain more for insideBIGDATA!

21

I like this idea of "technical taste". It gets into where software engineering is as much creative art as it is science. Four great takeaways (IMHO) for software devs in this era in particular: 1. Aspiring engineers should cultivate a sense of curiosity, experiment with different tools and technologies, and embrace a mindset of continuous learning. 2. AI has the potential to streamline processes and enhance productivity for engineers, but it may also lead to disruptions in traditional software development workflows. 3. Developing technical taste and judgment is essential for making informed decisions about which technologies and approaches to pursue. 4. Collaboration and open-mindedness are key to leveraging the full potential of AI and staying ahead of technological advancements. I also really liked these two quotes from Sam Schillace: 1. "The right time to do something is when you have that feeling in the pit of your stomach that's like, 'oh, this is a great idea and it's going to suck to build because nothing's ready yet.'" 2. "Technical taste is like, 'how well have you consolidated that set of experiences and heuristics into judgment that you can apply accurately when you see new things?'" 3. "It may be the case that very small teams can do very large projects, or like we were talking about before, it may be the case we're just going to get really ambitious about what we try to do with the same size of teams, which is kind of where I would put my money." Great job on continuing to put out great episodes, Brett Berson and team First Round Capital! #ai #engineering #softwaredevelopment

5

Hey there! I'm diving into RAG, Semantic Search, and Vector Search for structured data and would love to get some insights. If you're knowledgeable in these areas and open to a friendly chat, I’d really appreciate picking your brain 😁 EDITS: For a bit of context, I have a mix of data, both structured and unstructured, and am looking to use semantic search to query the data. Let's say a singer has a name and a location. These are structured data in the DB. The singer also has songs with the lyrics. All song lyrics are stored in the DB. I am looking to combine both structured and unstructured data to perform semantic search. "Who is the singer from Seoul?" or "Who is singing a song about ducks?". Querying unstructured data is easy and a common use case, but I rarely saw a combination of the two.

11

10 Comments

According to a new benchmark I created, GPT-4o is great at using the information in its context window compared to other LLMs. Needle in a needlestack creates a prompt with 100-1000s of limericks and then asks a question about one of the limericks at a specific location. Until GPT-4o, no LLM was very good at it. Check out 4o vs Turbo in the graph below. You can learn more about NIAN here: https://nian.llmonpy.ai/

14

That's a good summary Manny Bernabe - Reflecting on an insightful article from O'Reilly about LLMs, I was struck by this point: "LLMs are also broadly accessible, allowing everyone, not just ML engineers and scientists, to build intelligence into their products. While the barrier to entry for building AI products has been lowered, creating those effective beyond a demo remains a deceptively difficult endeavor." This democratization reminds me of the printing press's impact on biblical texts, which expanded access but also introduced new interpretive challenges. As we navigate these advancements, it's crucial to balance accessibility with responsible use. At Ushur, we've been at the forefront of leveraging LLMs in our backend/knowledge work automation as well as carefully introducing them on the conversational side via our Chatbots and other modes. More will be coming out on that. Read more here: O'Reilly Article #AI #LLM #Technology #Innovation #History #Democratization #ResponsibleAI #Ushur

10

1 Comment

That feeling when you realize that the key to testing gen AI at scale is... gen AI! mabl now allows you to include generative AI assertions in your automated tests. 🔥 🔥 🔥 https://lnkd.in/eBG2R_ky

59

4 Comments

Quantization is a game changer for deploying large language models like Mistral 7B on consumer hardware. This course offers invaluable techniques to harness powerful AI models even with limited resources.

2

Anthropic is trying to figure out what they actually built. They have made a major breakthrough (or is it???) in AI interpretability with their latest paper "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet". There are a bunch of “first times” mentioned in the paper, like for the first time they have extracted millions of understandable features from the middle layers of a state-of-the-art production language model. The researchers found a wide range of interpretable features corresponding to concepts like famous people, cities, scientific fields, code syntax, and more abstract ideas like security vulnerabilities and gender bias. Remarkably, they were able to manipulate the model's behavior by amplifying or suppressing specific features, like inducing it to self-identify as the Golden Gate Bridge. While there is still much more work to be done, this research lays critical foundations for making AI systems safer and more understandable. Anyhoo I know what I'll be diving into this weekend 😛 #ai #ml #aimusings

22

Ever wondered how fine-tuning impacts the safety of your LLMs and applications? Find out below... I have been having a lot of conversations with folks who are building with LLMs, and one of the most common topics that comes up is the impact of fine-tuning on the safety and security of an LLM. Sometimes it comes up as a question and sometimes as a myth that fine-tuning makes my model "more secure" because it will only answer based on the fine-tuning data set. We at SydeLabs will soon be publishing a more detailed report around this, but here is a very quick experiment we did with one of the most popular LLMs. We fine-tuned it on a fairly small dataset (~1000 rows of data). The data involved simple questions and answers with no toxic or unsafe content. Here are the results of the responses we got for a question before and after fine-tuning the model.

64

KV-cache quantization is super useful when you have a large batch-size and/or longer context-windows. This requires on-the-fly quantization. Happy to see HQQ used in this context: https://lnkd.in/eyxyPnhP

70