Sam Liang

Hugging Face demonstrated test-time scaling with Llama 1B, turning it into a "reasoner" that outperformed Llama 8B (8x larger!) in math. Test-time scaling has opened up a whole new dimension in model performance and AI research. I wrote about this when OpenAI released its o-1 model (the first 'reasoner' model). You can read it here: https://lnkd.in/gi737nsA At its core, a reasoner model uses Chain of Thought (CoT) reasoning—breaking problems into logical steps, solving them sequentially, and backtracking when needed. Scaling test-time compute enables this reasoning through two techniques: 1/ Self-Refinement: Models iteratively improve outputs, correcting errors over multiple passes. 2/ Search & Verifiers: Generate multiple outputs and evaluate them with verifier models (e.g., Process Reward Models for steps, Outcome Reward Models for final answers). Hugging Face’s contribution stands out because they’ve shown test-time scaling can be applied to any model, even smaller, open-source ones. They’ve also made it easier for others to replicate and advance the technique—helping the OS community catch up. That said, inference scaling is still far from efficient. Llama 1B required ~20 reasoning steps to outperform Llama 8B, a model only 8x larger. Nevertheless, inference scaling already achieves two critical things: – Pushing the frontier: Scaling inference pushes the boundaries of accuracy and delivers better results. – Smaller hardware requirement: Smaller models can match the performance of larger ones, reducing reliance on expensive infrastructure. The next frontier isn’t just about bigger models—it's about smarter inference. Excited to see where this leads!

62

4 Comments

Thrilled that mine and Qian Li's paper was just published in the Communications of the ACM! The big challenge we've been working on since grad school is how to make serverless computing work well for stateful applications. This is tricky because it's hard to make long-lived stateful workflows reliable in ephemeral serverless execution environments. The solution, which we talk about in this paper and are building at DBOS, Inc., is using database transactions to make programs durable. The key idea is to store the execution state of a serverless workflow in a database. That way, if the workflow is interrupted, it can automatically recover from where it left off by looking up its execution state in the database. If you know that stateful programs are durable, it becomes a lot easier to make them serverless. You can deploy them to ephemeral executors and leverage their durability to autoscale them and move them between executors as needed. That way, you can build incredibly reliable stateful programs that also benefit from the scalability and convenience of serverless.

113

9 Comments

Excited to share my first article about creating an application that uses an LLM to censor specific words/phrases from an audio file. Check out the GitHub link for more details: https://lnkd.in/dT32vAqc. #Google #OpenAI #Streamlit #LLM #AI #Python

12

GPT-4o says "9.11 is bigger than 9.9". It is extremely fascinating how AI models which can understand quantum mechanics, have read books from Tolstoy to Principia, and are now achieving super-human scores on knowledge benchmarks can fail at something as trivial as interpreting the decimal point system. These issues are at the heart of why AI models are considered unreliable. To build a world where AI models are ubiquitous, we need to figure this out ASAP. #AI #Reliability #LLMs

1

3 Comments

Finally official! I am thrilled to share the first fruit of my research, "Predictive Performance of Bayesian Stacking in Multilevel Education Data," co-authored with Dr. David Kaplan. This paper employed Bayesian ensemble methods to predict students' academic outcomes. We showed that Bayesian hierarchical stacking can exceed other stacking methods in terms of predictive performance by incorporating input-dependent weights in M-complete and M-open settings. While this marks a milestone in my academic journey, I am sure there is still a long way to go and this work is just the beginning! The paper can be found in JEBS through the following link: https://lnkd.in/gPmxU684

23

"Why did Santa's AI-driven sleigh keep running into trees?" So much 🔥 from Julie Wang and John Allard in today’s OpenAI live stream! They covered Reinforcement Fine-Tuning, medical research, and shared some truly punny jokes. Moments like these remind me why I love my team: brilliant people solving fascinating problems—and having fun along the way! https://lnkd.in/gJnNiTXc If you'd like to participate in the Research Program by sharing data, check out https://lnkd.in/gZYehx9T or ping me. Reinforcement Fine-Tuning is a new technique for teaching the new models how to reason. It's kinda like how AlphaGo taught a model how to play by "grading" games, then learning what "good" moves are by comparing moves in winning vs losing games. This opens up a whole new paradigm for models, where you're not just teaching them the answer, but teaching them how to arrive at the right answer. What is Reinforcement Fine-Tuning (RFT)? It’s a groundbreaking technique for teaching models how to reason. Think of it like AlphaGo: the model learns by being “graded” on games and understanding what makes moves in winning games better than losing ones. With RFT, prompts are like the “games” and chain-of-thought reasoning is like the “moves.” By evaluating the "winning" reasoning steps, the model learns which ones lead to better answers. This means we’re not just teaching models the answers—we’re teaching them how to think through problems and consistently arrive at the right solutions. My team’s mission is simple: give you the world’s best model for your domain. If you’re finding it hard to get API responses to work for your business—whether it’s a performance issue, cost concern, or accuracy challenge—let me know. We have tools to improve outcomes, and I love working with customers to see how we can do even more. Oh, and Santa’s sleigh? Julie gave us the answer: “Because he hadn’t ‘pine’-tuned yet!” 🥁

26

1 Comment

Exciting news in the world of AI and biology! 🧬🤖 EvolutionaryScale unveiled ESM3, a groundbreaking generative AI model that's set to revolutionize protein design. Here's why it's a game-changer: • ESM3 can simultaneously reason over protein sequence, structure, and function - a first of its kind[1][2]. • Trained on 2.78 billion proteins from Earth's natural diversity, it's the most compute-intensive biological model ever created[3]. • In a stunning demonstration, ESM3 generated a new fluorescent protein, simulating 500 million years of evolution[3][4]. This breakthrough could accelerate innovations in drug discovery, materials science, and even carbon capture[3]. By making biology programmable, ESM3 opens up a new frontier for scientific research and technological advancements. #AI #Biology #Innovation [1] https://lnkd.in/dewKawN5 [2] https://lnkd.in/der6SFtj [3] https://lnkd.in/dZAXRQ6D [4] https://lnkd.in/dX6gtGms [5] https://lnkd.in/dCJ-F979

8

The field of LLMs is getting more and more interesting. So many real world use cases are emerging, making a different shift in digital transformation. Stay tuned for more details. #LLMs #digitaltransformation

9

This Demo won us “Most technically impressive” at the Supabase hackathon at Y Combinator Finn Metz, Sricharan Guddanti, and I worked through the night last weekend to build something we’ve been missing: an interactive AI tutor. As students at University of California, Berkeley, we rely on LLMs daily, but there’s currently no natural way to interact with them. We built a tutor that actually works with you—walking through homework at your pace and answering questions in a way that’s easy to follow. Huge thanks to judges Jared Friedman and Eli B. from YC, as well as Supabase founders Paul and Ant. We’re launching soon—sign up for the waitlist at reeviewai.com and follow us at reeview.ai. If a tool like this sounds helpful, we’d love it if you shared it with your friends!

65

4 Comments

This blog post by FAR.AI considers the utility of "pacing", taking your time, to plan. This reminds me of Andrew Ng's complaint that we now demand LLMs to write without ever using the backspace - AI systems should be able to go back and revise earlier output. This idea of postponing final decisions is also present when we consider how to label datasets with a crowd of noisy AI algorithms. Check out the comments for some tables comparing the performance of NTQR's Jury Evaluation Theorem versus majority voting. It pays to evaluate the crowd first, then use their decisions to arrive at final conclusions. NTQR's documentation page (https://lnkd.in/gypDcd7H)

LLMs are terrible at basic math. Why? Here's why: Unlike people, LLMs don’t conceptualize mathematics. They simply recognize patterns in the training data (text) they’ve learned from. When asked what 2 + 2 is, an LLM recalls instances of seeing ‘2’, ‘+’, ‘2’ and so on, allowing it to approximate solutions but not match the precision of a calculator. That’s how we end up with LLMs asserting that 2 + 2 = 5. But for analytical teams, like finance, that makes LLMs really hard to adopt in day-to-day workflows. The productivity gains of having AI do calculations are negated by needing to fact-check and audit every calculation and number. LLMs are also not deterministic, so no amount of prompting is going to guarantee you’ll get the correct answer every time. And that’s just for basic math – the outputs I’ve seen from LLMs on more complex calculations and analyses can get very wonky. Add thousands of rows of data and you’re squarely in hallucination land. So what’s the solution? Calculate things manually? No. The good news is that this is changing quickly and there are options: 1 - AI is getting better at math. Researchers continue to train LLMs with more specialized datasets that include a much wider array of mathematical problems and solutions. The newest models, like OpenAI’s latest family of models, are incorporating things like symbolic reasoning and “chain of thought” methods to break complex problems into smaller, logical steps the LLM can more easily process. Results have been really promising and we’ve seen each new iteration of models continue to beat previous benchmarks. 2 - Leverage AI agents For intensive workflows, and to really make sure your outputs are correct, working with companies that are building high-performing AI agents is likely your best bet. Agentic workflows allow LLMs to be used for what they are great at (inferring user intent, making decisions, etc.) while outsourcing the things they struggle with (complex calculations) to other parts of a broader system. Text-to-SQL, “tool-use”, few-shot prompting, and other strategies can be combined with real-time data integrations to dramatically improve performance on analytical operations. At Concourse we’ve spent the better half of 2024 getting our AI Agents to be really good at math, and I’m so excited to continue growing our capabilities to tackle even more complex mathematical problems for our customers in 2025. Reach out if you’re interested in learning more about how we’ve built Concourse and the impact we’re already delivering for finance teams.

42

5 Comments

Hello fellow engineers & Pythonistas! 👋 I’m excited to share that I’ll be speaking about Pydantic and FastAPI at the upcoming Python Meetup on October 21 in Singapore! 🚀 Thanks to Martin Brochhaus for this opportunity! 🙏 https://lnkd.in/gYPbASS9 Let’s connect, learn, and innovate together! If you’re attending, I would love to hear your thoughts or any suggestion you might have. Looking forward to seeing you all there! #Python #FastAPI #Pydantic #PythonMeetup

50

3 Comments

Week 3 at Headstarter Fellowship: Diving Deep into AI Chatbot Development 🤖 https://lnkd.in/gQs-w42f Wrapped up an intense week building a sophisticated AI-powered chatbot. We started with a Next.js foundation, leveraging its API routes for our backend needs. The frontend came to life with Shadcn UI components styled with Tailwind CSS for that sleek, modern look. But the real magic happened when we integrated the AI. We implemented Retrieval Augmented Generation (RAG) using Langchain and Pinecone, which allows our chatbot to pull relevant information from a specified knowledge base (Right now it just has some old information about me). Then came the fun part: we integrated both Groq's Mixtral-8x7B and OpenAI's GPT-4o, building a custom routing mechanism to choose the best model based on query complexity. The chatbot isn't just smart; it's multilingual too! We added support for English, French, and German responses. And to keep things dynamic, we implemented streaming responses - it's pretty cool to see the AI "think" in real-time. While each of these components might seem straightforward on its own, the real challenge was orchestrating everything into a seamless, cohesive application. It's one thing to implement RAG or integrate multiple AI models, but to make them work together smoothly? That's where the real fun begins! There's still room for improvement, of course. We're looking at refining the query complexity determination and optimizing RAG performance. But I'm proud of what we've accomplished in just a week. It's a testament to how quickly we can build complex AI applications with modern tools and frameworks. This project has been a fantastic learning experience, thanks to Bill Zhang for the great help, pushing me to think holistically about AI application architecture. Can't wait to see how we'll apply these skills in future projects! #HeadstarterFellowship #AIEngineering #ChatbotDevelopment #HeadstarterAI Faizan Ahmed

5

🚀 New Medium Article Alert! 🚀 I’ve just published an article where I explore Overview of LLMs, GPT Models, RAG, and LlamaIndex Read it now 👉 https://lnkd.in/gARHY54f #AI #AZURE #LLM #RAG #GPT #Tech #Programming #Medium

19

🍎 Building an AI-Powered Calorie Counter with Hugging Face Spaces 🚀 Deploying Large Language Models (LLMs) for real-world applications can be tricky, but Hugging Face Spaces makes it effortless! I’m excited to share my latest Medium blog, where I detail how I created an AI-powered calorie counter app 🥗 using Hugging Face Spaces. 💡 Why Hugging Face Spaces? Traditional LLM deployment involves challenges like: ⚙️ Managing hardware and infrastructure. 📈 Scaling apps for high traffic. 🖥️ Building intuitive user interfaces. Hugging Face Spaces tackles these head-on with: ⚡ Quick Deployment: Get started with frameworks like Gradio in minutes. 🛠️ Managed Infrastructure: Handles hardware, scaling, and backend setup. 🌐 Instant Sharing: Generates a public link to showcase your app effortlessly. ✨ How My App Works Using image-to-text models, my app identifies food items from photos 📸 and delivers detailed nutritional insights like calorie, fat, and carb counts. Hugging Face Spaces enabled me to focus on building the app’s core features while taking care of the backend complexities. 🔗 Explore the Full Journey In my Medium blog, I cover: 1️⃣ The step-by-step guide to deploying the calorie counter app. 2️⃣ Challenges in hosting LLMs and how Spaces resolves them. 3️⃣ Tips for creating scalable, user-friendly AI apps. 📖 Read the blog here: https://lnkd.in/dxb9uR-S   💻 Try the calorie counter app and see my other projects here: https://lnkd.in/durbADd2 Hugging Face Spaces is a game-changer for developers and researchers, democratizing access to powerful AI tools. Whether you're building a calorie tracker, chatbot, or any innovative solution, this platform has you covered! 💡 💬 Let me know what you think, and feel free to share ideas for future projects! #AI #MachineLearning #LLM #HuggingFace #CalorieCounter #Innovation #Tech

17

Keeping up with research doesn't have to be so painful: LLMs to the rescue 🧠 If you’ve ever tried staying up to date with the latest research on platforms like arXiv, you know how time-consuming (and overwhelming) it can be. That’s why we created a demo notebook to show how LLMs can simplify this process for you. Here’s what the demo does: 🔍 Monitors arXiv RSS feeds for new papers. 🤖 Uses an LLM to determine if a paper is relevant to your field of study. 📬 Sends you an email with a summary, key takeaways, and a link to any relevant papers. How it works: 1️⃣ Parses arXiv RSS feeds to collect titles, abstracts, and URLs. 2️⃣ Asks an LLM to assess relevance, why the paper is relevant, and extract the main takeaway. 3️⃣ Emails you only the most relevant results—no more wasting time! 📩 Want to see the demo in action? Check out the link in the first comment. 🛠️ This notebook is the first demo in our new AI & LLMs Demo Repository designed to help you learn practical applications of AI in real-world workflows. At Limai, we’re passionate about helping businesses and individuals leverage AI to solve problems and unlock new opportunities. If you want to learn more about how you can leverage AI to grow your business just get in touch! #ArtificialIntelligence #LLMs #ResearchTools #Arxiv #GenerativeAI

17

1 Comment

I have to write outreach emails/messages regularly to potential clients. I wanted to automate the writing process. I could use LLMs to write it for me, but I hate it when it sounds more like a robot and it's even worse when it hallucinates 😞 So I created a RAG-based LLM application (running locally) using LlamaIndex that takes in all my previous emails, proposals, and portfolio projects and generates content that sounds more like me 😎 . To top it off, it often identifies and highlights connections between our previous work and the client's industry (See the demo below for a closer look.) Note: This is just a demo, I haven't provided it with much data yet. I think it can do even better with more personalized content. #LLM #RAGtechnique #AI #outreach #codeclever

26

7 Comments

Huge congratulations to the LanceDB team, Chang She and Lei Xu, on securing an $8M funding round from CRV! Their mission to build The Database for Multi-modal AI is an exciting one. It's amazing to see how far they've come since Chang's days as a star VP Eng at Tubi. Check out the TechCrunch article for more details on their journey. #AI #technews #fundinground

103

1 Comment

Super interesting talk by Thomas Mullaney on the engineering challenge of inputting Chinese characters on a keyboard and how it was influenced by the intricate relationship between culture, technology, and geopolitics. I've been fascinated by this issue for a long time ever since I learned how to type Chinese with pinyin. One interesting difference between typing English and Chinese is that when you type English, every letter on screen is essentially WYSIWYG. But in Chinese, if you use pinyin for example the computer has to guess what characters you mean. It essentially becomes a search/predictive text problem, which is super interesting as we enter the age of LLMs which are just next token predictors. Can’t wait to read “The Chinese Computer”! #computerhistorymuseum

63

4 Comments