Daniel Bourke

📢RAFT: Sailing Llama towards better domain-specific RAG 1️⃣ Cheaper - Meta Llama 2 7B language model 2️⃣ Faster then RAG - the model answers queries like a student writing a closed-book exam 3️⃣ Easier - Azure AI studio simplifies the process of fine-tuning, testing, and deploying. Learn more 👉https://lnkd.in/durJv7Jk Well done Tianjun Zhang and Cedric Vidal - Blown away by this new approach💓 #ai #meta #microsoft #chatgpt #programming #development

9

1 Comment

A completely new neural network architecture just dropped - the Kolmogorov-Arnold Network (KAN). I don’t mean a new type of neural network model, but a foundation change in the core architecture of NNs. Multi-layer perceptrons have been core to all deep-learning frameworks. At its core, they are simple units that sum the weighted input from other neurons and apply some form of non-linear activation (most of the time RELUs). At a foundational level, it relies on the theory of universal approximation. KANs on the other hand rely on The Kolmogorov–Arnold representation theorem, which states that any continuous multivariate function can be represented as a composition of a small number of univariate functions and a small number of operations on those functions In practice, KANs do not have any edge weights, instead, they have a trainable univariate function parameterized by a B-spline. This seemingly small but foundation change makes KANs much better than MLPs in terms of speed of training, model complexity, and explainability. I have seen some incredible results coming out in the short time since this was published. The paper comes from one of my favourite scientists Max Tegmark and you can read it here (Liu et al., 2024, https://lnkd.in/d8iSTAGb). Fair warning, while the concept is simple the paper is quite verbose and detailed.

46

Verdict is out, and it didn’t even take me a week to decide: Windsurf SOARS over Cursor for coding editors right now. Anyone looking to take the leap, here’s what’s awesome: 1) the interface is better. It’s more like vscode (and cursor was already very similar) 2) there’s a nice “write” vs “chat” toggle to change between workflow types 3) CONTEXT is way better: I don’t need to “select files” to make it understand my Codebase- it can see and intelligently navigate it to know where to look (or at least where to start looking) 4) Yup. Three’s a magic number. Give a like if you found this helpful 😉

3

I stumbled across a fascinating blog about how Mantle tackled a common dev nightmare: converting a prototype from R to their production stack (Golang and ReactJS). Instead of spending months on it, they used AI (Gemini 1.0 Pro) to cut the workload by two-thirds! The trick? Let AI handle the repetitive stuff, so their engineers can focus on what really matters. If you’re into AI or software development, this one’s super interesting! https://lnkd.in/gFQKkAcY

7

Thinking about building your own AI model? Why stop at theory when you can create your own? Building a Large Language Model from scratch isn't just a tech challenge; it's a gateway to revolutionizing how we interact with technology. Here’s why it’s a game-changer: 🔎 Customization: Tailor the model to specific needs or tasks. 🔎 Innovation: Drive forward the boundaries of what AI can achieve. 🔎 Knowledge: Deepen your understanding of machine learning and AI. Sebastian Raschka,’s book, "Build a Large Language Model (From Scratch)" and its accompanying GitHub repository provide all the tools you need. This resource is a treasure trove of knowledge, offering practical insights into coding, pretraining, and fine-tuning a GPT-like model. Here’s what you’ll find inside: Step-by-step guides to code your own AI Tips for pretraining to ensure your model understands well Strategies for fine-tuning to achieve specific tasks GitHub Link: https://lnkd.in/gThNJcqm ♻️ Repost this if you agree. #AI #MachineLearning #Technology #Innovation #DataScience

4

Douglas Hunley: On CVEs, BoMs, and Releases 🚀 New Post Alert: Navigating the PostgreSQL Galaxy with Douglas Hunley 🚀 Imagine this: You’re sailing the vast cosmos of PostgreSQL with just a Bill of Materials as your trusty star map. Sounds like a geek’s dream, right? Well, welcome to the world of Douglas Hunley! In his latest piece, "On CVEs, BoMs, and Releases," Douglas takes us on a journey through the PostgreSQL automation universe. His mission? To ship tested-together versions of components that make up the CPA product. It's like getting the band back together, but for databases! Why is this a big deal, you ask? Because in the database realm, having all your ducks—or bytes—in a row can transform your day from chaos to cosmos. And Douglas? He’s like the Yoda of automation, guiding us with his wisdom. BoMs aren't just for grocery lists anymore, folks. They're the backbone of ensuring every piece of software code is strutting down the release runway in harmony. So, next time someone mentions PostgreSQL and BoMs, pretend you’re on a cosmic adventure—because with Douglas's insights, you really are! Kudos to Douglas Hunley for making database automation sound like a space odyssey! 🌌 #PostgreSQL #Automation #DatabaseGalaxy #DouglasHunley You can find the full article here: https://postgr.es/p/6I4

4

People rightly state that there is a lot of hype surrounding #AI. This reflects our obsession with applying new tools to *every* problem, fit for purpose or not. 😄 Here's the good part, though: when applied properly, AI is a huge force multiplier. Imagine a lever ^ 1000000. I've been sharing, any time I have the opportunity, the "good parts": areas in which AI can have real utility now. Not someday. Not maybe. Absolutely yes, and now. One area in particular is #RAG, Retrieval Augmented Generation. RAG == Turning an AI loose on YOUR DATA. Imagine all of those documents in which vital information is squirreled away, hard to search, even harder to correlate *across* those myriad documents. What a treasure trove if only you could leverage it! With RAG, you can...and pretty easily, too. Using AI and RAG, 1+1 pretty quickly adds up to 500. Or more. Even now, the power of your AI-enabled application is more limited by your imagination than the capabilities of the tools. If you want to know more, get just a taste of what you can do with AI using RAG, join me this Monday at #RAGHack. See below for more details and to save your spot! https://lnkd.in/gYAsiSP4

13

1 Comment

Here's all that we know about Meta Llama 3 so far > Trained on 15T tokens > 70B and 8B models released (along with instruction tuned) > 8K context length > 70B scores 82 on MMLU and 81.7 on Human eval > 128K vocab tokenizer - utilises 15% less tokens > Dense model architecture > Trained on 2x 24K GPU clusters > Both instruction tuned on a human-annotated instruction dataset > Open access, under Meta Llama license and terms of use What's next? > Scaling Llama 3 to 400B+ params (currently training) > Instruct scores more than 86.1 and 84.1 on MMLU and HumanEval > Would support multiple languages, higher context lengths. How do you use it? > Available in Transformers > Instruction version on Hugging Chat > Deploy via Inference Endpoints Time to get building! 💪

202

6 Comments

Your coding bootcamp certificate is cute. Now let's talk real education. Here's your real curriculum: 1. Debugging production fires at 2 AM 2. Deciphering legacy code written by Satan himself 3. Explaining to non-tech folks why the computer 'doesn't just do it' 4. Building something so complex it makes you question reality Your bootcamp taught you to code. Life will teach you to be a developer. Welcome to the school of hard knocks, where the tuition is paid in sleepless nights and the degree is respect from seniors. What's the harshest lesson the real world taught you? Share your war stories below. Bonus points for tears. 👇

143

33 Comments

Worth reposting here. 😅 https://lnkd.in/g4Kktqn6 I’m personally really excited for the soon approaching back half of the hype cycle where we can spend time adding some actual software engineering and scalability expertise to the decades of poorly written Python code from the ML world and start implementing features users want without slapping the widely abused phrase “AI” on everything. 🙌🏻 If you think it needs a chat bot, it doesn’t. If you think you’ll get more engagement or find market fit from a “✨ Try it with AI” button slapped on everything, it doesn't and you won't. The best AI-driven features are specific, solve real problems and are so seamless in your product that users don’t need a desperate “LOOK WE HAVE AI TOO” billboard to find them.

8

2 Comments

🚀 Claude Artifacts just got supercharged Inline code editing lands for Artifacts. Game-changer? Absolutely. Playing with it on a Python project. Artifact spat out a data processing function. Needed tweaking. Highlighted the line, typed my request. Bam. Optimised instantly. Why it rocks: • Maintain coding flow • Iterate at light speed • Pinpoint accuracy in edits Artifacts were already powerful. Now they're indispensable. What's your take on this update? Drop your thoughts below. #ClaudeArtifacts #AIAssistant #CodingProductivity

1

How AI Should Really Work (and how it usually does) AI in Most Companies: - Doing the same old boring stuff over and over - Using data just a smidge to make decisions - AI projects that don’t chat with each other – like neighbours who don’t share the BBQ - Barely spending a dime on teaching people about AI - Chasing quick wins like a kid chasing an ice cream truck - Occasionally throwing the word ‘AI’ into their marketing to sound cool What Happens Then (based on what experts say in layman terms): [Kate Crawford, Andrew Ng, Yoshua Bengio, Thoughtworks, Forbes, PwC] - Not much money made from AI - Data all over the place, not connected - like socks after laundry - Missing out on cool new ideas - Spending heaps with not much to show for it - Projects dragging on longer than a cricket match Usually, the big problem is companies don't have a smart, connected plan for using AI. How AI Should Really Work: 1. Goals: - Set goals that match what you can do now, what the market needs, and what you can grow into. No pie in the sky dreams here. 2. Target Use Cases. - Find and pick the best tasks for AI, the ones that will really make a difference. Each task has different needs and benefits. Focus on those that fit your big plans - like finding the best wave to surf. 3. Data Strategy. Make sure you collect, clean, and organise data for AI. This means: - Making rules for data - Connecting data from all over the shop - Ensuring data is good and easy to access 4. AI Talent Development. Spend money on teaching and hiring people who know AI. This includes: - Teaching current employees new skills - Attracting new AI experts with good offers - Creating a culture where everyone wants to learn 5. Integration and Scalability. - Get AI talking to every part of the company, making sure AI solutions can grow as needed. No isolated projects; aim for a plan that matches your business strategy – like having a game plan for footy.. 6. Ethics and Governance. - Create rules and frameworks for using AI responsibly. Ensure AI is fair, transparent, and accountable. 7. Full-Lifecycle Management. Plan out AI projects from start to finish: - Planning and development: Clearly define what you want to achieve and how long it will take. - Deployment and monitoring: Put AI solutions in place and keep an eye on them to ensure they work well. - Maintenance and improvement: Regularly update AI models and algorithms with feedback and new data. Ensure AI systems communicate seamlessly across all departments, fostering collaboration and sharing insights. Avoid siloed projects; aim for a unified approach that aligns with your overall business strategy – like a well-coordinated team on a mission Keen to hear what you think: what’s the most important factor for successful AI integration in business? #AIBusinessStrategy #InnovativeAI #FutureOfWork Image credit: Brij kishore Pandey

1

I think, the most crucial skill in working with Large Language Models isn't prompt engineering nor fine-tuning or knowing how to implement a good RAG system - it's knowing when to and when NOT to use them at all and if using, knowing which one to use. I recently spoke with a company about their quality control process. They had implemented a LLM based method to detect anomalies in their work streams. While they were able to get it quickly working it had some issues and was little more than a prototype. - 1-2 second latency per analysis - multiple thousand $ monthly API costs - Privacy concerns with their data going to an LLM provider - Inconsistent detection patterns In our discussion, we figured out, that a traditional classifier (isolation forest in this case) was a better solution. - significantly reduced latency - On-premise deployment - Full IP protection - higher accuracy - Clear statistical basis for each detection, including confidences(!) - Fraction of the ongoing costs In the end, also a local hosted LLM finetuned to the use cases helps to create human readable explanations of the anomaly - but only if needed. With all the talk about O1 (o3!) don’t forget: LLMs are remarkable tools and perfect for quick prototyping, as well as some more involved tasks but they're not the answer to every problem. Sometimes, the best innovation is choosing the right tool for the job - even if it's not the shiniest new technology. I’m interested in other examples on this others have encountered - and also feel free to reach out if you want to learn more about this. #AI #MachineLearning #Innovation #TechnologyStrategy #Engineering

31

3 Comments

Been playing around with GPT-4o some more and it's really quite good. Hard to justify using anything else for agents unless the task is really simple. That OpenAI API account is getting a good topping up next pay cheque. Meta, can you quickly whip up a multi-modal open source Llama 4 for me so I can go back to using this fancy graphics card? #GPT4o, #Llama #OpenAI #Meta

2

I stared at a bug for hours. This 15-minute hack solved it: 1. Set a timer for 15 minutes 2. Walk away from your desk 3. Do something completely unrelated to coding 4. Return with fresh eyes The solution wasn't more coding. It was stepping away. 90% of the time, the solution becomes obvious. Your brain needs breaks to problem-solve effectively. What's your go-to method for tackling stubborn bugs?

312

94 Comments

I want message limit credit when Claude makes a mistake. I have been using prompts inside Claude.ai to create a no-code web app. The experience has been enjoyable since I liken it to programming with a junior developer. For example, when I add new functionality via a prompt, Claude makes a mistake. I then tell Claude what's not working, and it has to go through and fix the error. It doesn't always fix it and sometimes makes the problem worse. I used to deal with this earlier in my career and had to jump in many times and make the changes myself. In this case, I'm just prompting the AI in hopes of helping it understand what it did wrong. I am okay with this process, except it takes away from my message/usage limit, and I must wait hours before continuing. For now, it's forcing me to be more concise with my prompts; tell Claude to mark that particular iteration (see my screenshot where it references Point D) so I can quickly go back to a working version. You can also step through different versions in the interface, but once you reach your message limit, you must wait. Now I'm stuck waiting until 9 PM Pacific time. At least it gave me time to write this post. Anthropic #claude #claudeai #nocode #lowcode #AI Kevin Witt Mark Levin Benjamin Elliott #usage #prompts #aiprompts Emerging Technology Innovators, Inc. #eti Sam Evans

5

6 Comments

OpenAI's "structured outputs" are a bigger thing than most seem to think - because they change the nature of the product in the eyes of engineers. The number 1 thing I hear from engineers working with LLMs is, "I cannot trust the output", which instantly turns many engineers into sceptics. As OpenAI now guarantees 100% accuracy for their API responses (using "structured outputs"), it makes data integration much easier. At the same time, they again reduced the price of their new model by 50%... This means fewer errors, more reliable integrations, and, ultimately, more intelligent decision-making in the products everyone is building. For those of you not yet using LLMs as part of your product stack, does this change the way you approach the technology? #ArtificialIntelligence #Reliability

16

2 Comments