Juan David Gil López’s Post

Chinese AI research lab DeepSeek has unveiled its latest reasoning models, DeepSeek-R1 and DeepSeek-R1-Zero. DeepSeek-R1 is fully open-source and distributed under the MIT license. The lab has released six distilled models, ranging from 32 billion to 70 billion parameters. They are designed to address tasks in math, code generation, and reasoning. https://lnkd.in/gkUgpXGG

Do today's LLMs, by design, face fundamental, intractable problems with formal and symbolic reasoning? Can these limitations be overcome by integrating different formal languages or programming systems into existing AI systems? Until solved, this seems like a pretty significant obstacle to achieving robust, consistent, human-level AI performance on general tasks. https://lnkd.in/gd46Yqyq

It's much easier to get a working AI application prototype using large language models than traditional machine learning techniques. But that's less than half the battle. Over time, if you don't have ways to monitor and debug outputs of LLMs, then you will fall into the same post-production quagmire of many ML apps. That's why it's important for AI Engineers to invest in evals to ensure their system meets a baseline level of performance. I'll be frank: I don't know much about evals. But I know it's a blind spot for me. That's why it's one of the areas I want to learn about the most over the next few days at the AI Engineer World's Fair Conference. I'd also be grateful if anyone has some good resources to share around how to build good evals for LLM-based applications.

Imagine a world where software engineering becomes more efficient, reliable and less prone to pesky bugs? This dream might be closer than you think, thanks to advancements in Artificial Intelligence (AI), particularly Large Language Models (LLMs). Our latest article unpacks Meta's TestGen-LLM, the game-changing tool that's revolutionising unit testing. Discover how AI is reshaping the landscape of software development and ensuring bug-free code. https://lnkd.in/drba2vxM

Enjoying the AI DC Guild journey. Just completed this fascinating course: Carbon Aware Computing for GenAI developers from DeepLearning.AI Exploring how LLM can have a significant carbon footprint to trying out one strategy for carbon aware ML development.

Phi-4 shows what’s next in AI. Despite its compact size, Phi-4 outperforms significantly larger models such as GPT-4o (its teacher model!) and Gemini Pro 1.5 in areas like mathematical reasoning. What’s fascinating about Phi-4 is its remarkable training strategy. Even after several epochs on synthetic data, the team observed no overfitting. In fact, the 12-epoch version surpassed models trained on more unique web tokens. It’s not about the size of the model but the intelligence behind its architecture and training methods!

No Cost : 5-Day Generative AI Intensive Course with Google November 11 - Friday, November 15 Day 1: Foundational Models & Prompt Engineering - Explore the evolution of LLMs, from transformers to techniques like fine-tuning and inference acceleration. Get trained with the art of prompt engineering for optimal LLM interaction. Day 2: Embeddings and Vector Stores/Databases - Learn about the conceptual underpinning of embeddings and vector databases, including embedding methods, vector search algorithms, and real-world applications with LLMs, as well as their tradeoffs. Day 3: Generative AI Agents - Learn to build sophisticated AI agents by understanding their core components and the iterative development process. Day 4: Domain-Specific LLMs - Delve into the creation and application of specialized LLMs like SecLM and Med-PaLM, with insights from the researchers who built them. Day 5: MLOps for Generative AI - Discover how to adapt MLOps practices for Generative AI and leverage Vertex AI's tools for foundation models and generative AI applications. All you need is basic knowledge of Python, a Kaggle account, and AI Studio.

Research on GenAI is moving forward at an astonishing pace: as you can read in the article below, a new and optimised way, to get an LLM to solve complex problem has been discovered: it outperfors well known frameworks while requiring 10x-40x less computation power (so :dollarcoin: :dollarcoin: ). What’s really cool is that they built this prompt-ing framework based on how humans reason when they need to solve complex problem!! https://lnkd.in/dM9uzeQV

DeepSeek, a Chinese AI research company, has released DeepSeek-R1, an open-source AI model that rivals OpenAI's o1 in reasoning tasks, including mathematics and coding. DeepSeek-R1 is available under the MIT license, allowing for free use and modification. The model's largest version contains 671 billion parameters, which may require substantial hardware resources to run locally. However, DeepSeek has also released distilled versions, such as DeepSeek-R1-Distill with 1.5B, 7, 8B and their 32B which outperform OpenAI's o1-mini across various benchmarks and are more feasible for local deployment on consumer-grade hardware. I tried the tiny model with only 8B parameters running local and it is quite impressive...

Open-source AI continues to advance, with models such as DeepSeek-R1 now competing with proprietary systems like OpenAI's o1 in reasoning, math, and code. https://lnkd.in/dw4vCYU8