Axel Mendoza

🔬 Exciting Breakthrough in AI Interpretability by Anthropic ! 🔬 Eight months ago, our team embarked on a challenging journey to scale sparse autoencoders for recovering monosemantic features from transformers. Today, we're thrilled to share a significant milestone: we've successfully extracted high-quality features from Claude 3 Sonnet, Anthropic's medium-sized production model! Our research has unveiled a fascinating diversity of highly abstract features that both respond to and influence behaviors. Some key findings include: - Features representing famous people, countries, cities, and type signatures in code. - Multilingual and multimodal features, bridging text and images across languages. - Abstract and concrete instantiations of ideas, such as code vulnerabilities and discussions on security. 📊 Key Results: 1. Interpretable Features: Sparse autoencoders are producing clear, interpretable features even for large models. 2. Scaling Laws: We've utilized scaling laws to guide our training, enhancing efficiency and effectiveness. 3. Abstract Generalization: The features generalize across multilingual, multimodal, concrete, and abstract references. 4. Systematic Relationships: A systematic relationship between concept frequency and dictionary size needed for feature resolution. 5. Influence on Behavior: These features can be used to steer large models, influencing behavior significantly. Importantly, the findings also touch on crucial safety concerns, revealing features related to deception, sycophancy, bias, and dangerous content. This breakthrough marks a significant step forward in AI safety and interpretability, paving the way for safer, more reliable AI systems. Kudos to the Anthropic interpretability team for this remarkable achievement! 🎉 Link to Paper: https://lnkd.in/ehH-eCdr #AISafety #MachineLearning #AIResearch #SparseAutoencoders #Transformers #Anthropic #Claude3Sonnet #TechInnovation #ArtificialIntelligence #AIInterpretability

5

Kolmogorov–Arnold Networks seem like a paradigm shift in ML architectures. Curious to see how they evolve. "While MLPs have fixed activation functions on nodes ("neurons"), KANs have learnable activation functions on edges ("weights"). KANs have no linear weights at all -- every weight parameter is replaced by a univariate function parametrized as a spline. " https://lnkd.in/gdPk_qCQ

29

I am proud to be working at Martian, where our AI research team is pushing the boundaries of scaling Mechanistic Interpretability, arguably one of the most important areas of AI research. In recent weeks, we have seen a surge of progress in AI interpretability, driven by both Anthropic and OpenAI. While their recent techniques (linked in our blog post) involve deep and painstaking human inspection of individual neurons, our approach aims to automate this understanding at scale. The Martian research team, led by Shriyash U. and Chaithanya B., outlines a groundbreaking insight in this blog post: to truly automate the understanding of neural networks, we must evolve not only our techniques but also our theoretical foundations. They propose category theory as the key to this evolution. As our founder Etan Ginsberg would say, for more on the technical details, dive into the blog post. For those who are “normies” like me, you might wonder why this is so important! The goal of Mechanistic Interpretability is to make AI models understandable, so we can "read" them like a book. Currently, AI models are largely black boxes—even their developers are often surprised by what they can do. Understanding how to prompt them, what they excel at, and their inherent biases is a process of trial and error. Imagine the Possibilities: -Having detailed insights into an AI model's knowledge and capabilities. -Understanding its biases, strengths, and weaknesses. Knowing how to effectively prompt a model to get the desired answers. This level of understanding would revolutionize our interaction with AI, making it more transparent and predictable. At Martian, this research powers our LLM Router, which predicts the best large language model (LLM) for a given prompt based on a customer's cost, latency, and output quality preferences, a highly practical real-world application of Mechanistic Interpretability.  Follow Martian to be tuned into more exciting developments from our team! P.S. - As I write this, our founders just came out of a meeting with the research team of a major global tech company. They were delving into our technology, and the customer said, “You guys are better than OpenAI.” While I’m not sure of the exact context, it’s still incredibly flattering! #AI #MachineLearning #ArtificialIntelligence #TechInnovation #MechanisticInterpretability #Research #MartianAI #AITransparency

7

Fantastic News for NLP Enthusiasts: LangChain and Hugging Face Announce New Partner Package! 🤗🤗 A powerful collaboration between LangChain and Hugging Face! 🤗 They've released a new Python package, `langchain_huggingface`, designed to seamlessly integrate the latest advancements from Hugging Face into the LangChain ecosystem. This is fantastic news for anyone working on Natural Language Processing (NLP) tasks. Here's what you can expect: Effortless Access to Cutting-Edge Models: Easily leverage the extensive collection of pre-trained models available on Hugging Face directly within your LangChain workflows. Enhanced Functionality: LangChain's robust linguistic toolkit, including tokenization and syntactic analysis, empowers you with a deeper understanding of language structures, leading to more nuanced NLP applications. Streamlined Pipelines: Construct optimized NLP pipelines that efficiently handle diverse tasks, from text classification to machine translation. Unified Framework: Deploy your combined models effortlessly, achieving greater flexibility in tackling various NLP challenges. This partnership signifies a significant step forward for the NLP community. By combining the strengths of LangChain and Hugging Face, `langchain_huggingface` paves the way for more efficient and effective NLP development. Ready to get started?Simply install the package using `!pip install langchain_huggingface` in your notebook . #NLP #LangChain #HuggingFace #MachineLearning #ArtificialIntelligence Share your thoughts 💭💭 . https://lnkd.in/gEaXzWgB

2

1 commentaire

If you want to understand why LLMs are not the solution for AGI (Artificial General Intelligence), I highly recommend listening to Lex Fridman podcast with Yann LeCun : Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI T͟L͟D͟R͟: 1. LLMs do not understand the physical world 2. LLMs have not the ability to remember 3. LLMs have no persistent memory 4. LLMs are not able to reason and to plan

27

5 commentaires

In this episode, we discuss Why do small language models underperform? Studying Language Model Saturation via the Softmax Bottleneck by Nathan Godey, Éric de la Clergerie, Benoît Sagot. This paper investigates the phenomenon of performance saturation in small language models, attributing the issue to a mismatch between the model's hidden layer size and the complexity of the targeted probability distribution. The softmax bottleneck, a known limitation in neural networks, is identified as a contributing factor to this mismatch, leading to suboptimal performance due to the emergence of degenerate latent representations during late pretraining stages. The study demonstrates that models with fewer than 1000 hidden dimensions are particularly susceptible to this effect, resulting in decreased effectiveness upon evaluation.

8

Llama 3 by Meta was just released last week .. and it is already on the Chatbot Arena leaderboard 🏆. The battle for the best (open ?) LLM continues stronger than ever ! The 𝐂𝐡𝐚𝐭𝐛𝐨𝐭 𝐀𝐫𝐞𝐧𝐚 is a crowdsourced open platform for blind comparisons of LLMs, developed by LMSYS. It is one of the most popular ways to evaluate LLMs, and companies releasing new models thrive to reach the top of the leaderboard https://lnkd.in/g_sWj5pF. 𝐇𝐨𝐰 𝐝𝐨𝐞𝐬 𝐢𝐭 𝐰𝐨𝐫𝐤 ? ▶ The approach is very simple: users enter prompts, and the platform generates responses from two different LLMs, then users chose the best answer, without knowing which model generated which answer ▶ Scores are calculated with the Elo ranking systems ▶ They have collected hundreds of thousands of votes so far 𝐖𝐡𝐲 𝐢𝐬 𝐭𝐡𝐢𝐬 𝐩𝐨𝐰𝐞𝐫𝐟𝐮𝐥 ? ▶ Realistic Evaluations: traditional benchmarks are not as helpful as they might seem because they can be manipulated and tailored for optimization.. with the Chatbot Arena, LLMs are evaluated through real-world user interactions for more realistic evaluations ▶ The data is always fresh = no risk of data contamination, as compared to static benchmarks ▶ Continuous Feedback: Rankings are updated based on ongoing user votes, to ensure up-to-date performance metrics ▶ Transparent Process: the Elo rating system is a clear and explainable method 𝐒𝐨𝐦𝐞 𝐥𝐢𝐦𝐢𝐭𝐚𝐭𝐢𝐨𝐧𝐬 𝐨𝐟 𝐭𝐡𝐞 𝐂𝐡𝐚𝐭𝐛𝐨𝐭 𝐀𝐫𝐞𝐧𝐚 ▶ Short user interactions might not reveal more advanced model capabilities ▶ Lack of diversity and complexity in the prompts ▶ Crowdsourcing is not representative: it is likely that most of the platform users are in a similar group/race/gender, so their preferences might not fairly reflect the entire population ▶ Users tend to be biased towards more verbose responses 👌 Overall, I love this platform, and it is very user-friendly so if you haven't yet tried it, give it a go! -- #AI #GenAI #Llama3 #LLMs #ChatbotArena

10

2 commentaires

Baal, our Bayesian Active Learning library is working on a major version and we want to know more about you! If you use Baal for Active Learning, Uncertainty Estimation or Bayesian Deep Learning, we would **love** to talk to you! 😎 In more detail, we want to understand when our users use our library and how. You can take a spot in our Calendly:

14

Want to test Meta, Google, and Mistral models all on one platform ? **Groq.com** Why Groq? Because speed matters. GroqCloud supports a diverse array of models, enabling you to select the best for your needs: ➡ llama3-8b-8192 by Meta (Context Window: 8,192 tokens) ➡ llama3-70b-8192 by Meta (Context Window: 8,192 tokens) ➡ mixtral-8x7b-32768 by Mistral (Context Window: 32,768 tokens) ➡ gemma-7b-it by Google (Context Window: 8,192 tokens) Groq's commitment to GenAI inference speed is transforming real-time AI applications with its LPU. What is an LPU? ➡ Language Processing Unit, is a revolutionary processing system designed specifically for LANGUAGE applications. Unlike traditional CPUs (Central Processing Units) and GPUs (Graphics Processing Units), which handle a broad range of tasks. LPUs are optimized for the rapid processing demands of LLMs This specialized architecture allows for more efficient handling of sequential data, significantly reducing the time per word processed. ➡ That means you get faster outputs You can run Llama-2 70B at over 300 tokens per second per user!

24

📢 Hugging face is directly integrated with GenAI developers favorite tool "langchain"🦜 Get set go!🚀 💡 pip install langchain-huggingface https://lnkd.in/ddHreVrs #huggingface #GenerativeAI #langchain

22

MARS5 TTS: Open Source Text to Speech with insane prosodic control! 🔥 > Voice cloning with less than 5 seconds of audio > Two stage Auto-Regressive (750M) + Non-Auto Regressive (450M) model architecture > Used BPE tokenizer to enable control over punctuations, pauses, stops etc. > AR model predicts L0 coarse tokens, refined further by the NAR DDPM model followed by the vocoder Great job Camb AI team! Kudos for open sourcing the artifacts - looking forward to what comes next ;)

2 035

70 commentaires

💥💥💥 Anthropic successfully identified how millions of concepts are represented in a large model 🤯 👉 Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet | Anthropic From the blog post: Eight months ago, we demonstrated that sparse autoencoders could recover monosemantic features from a small one-layer transformer. At the time, a major concern was that this method might not scale feasibly to state-of-the-art transformers and, as a result, be unable to practically contribute to AI safety. Since then, scaling sparse autoencoders has been a major priority of the Anthropic interpretability team, and we're pleased to report extracting high-quality features from Claude 3 Sonnet, 1 Anthropic's medium-sized production model. We find a diversity of highly abstract features. They both respond to and behaviorally cause abstract behaviors. Examples of features we find include features for famous people, features for countries and cities, and features tracking type signatures in code. Many features are multilingual (responding to the same concept across languages) and multimodal (responding to the same concept in both text and images), as well as encompassing both abstract and concrete instantiations of the same idea (such as code with security vulnerabilities, and abstract discussion of security vulnerabilities). Some of the features we find are of particular interest because they may be safety-relevant – that is, they are plausibly connected to a range of ways in which modern AI systems may cause harm. In particular, we find features related to security vulnerabilities and backdoors in code; bias (including both overt slurs, and more subtle biases); lying, deception, and power-seeking (including treacherous turns); sycophancy; and dangerous / criminal content (e.g., producing bioweapons). However, we caution not to read too much into the mere existence of such features: there's a difference (for example) between knowing about lies, being capable of lying, and actually lying in the real world. This research is also very preliminary. Further work will be needed to understand the implications of these potentially safety-relevant features. KEY RESULTS Sparse autoencoders produce interpretable features for large models. Scaling laws can be used to guide the training of sparse autoencoders. The resulting features are highly abstract: multilingual, multimodal, and generalizing between concrete and abstract references. There appears to be a systematic relationship between the frequency of concepts and the dictionary size needed to resolve features for them. Features can be used to steer large models (see e.g. Influence on Behavior). We observe features related to a broad range of safety concerns, including deception, sycophancy, bias, and dangerous content. 👉 https://lnkd.in/dg5jxymr #machinelearning

2

In Episode 13 of "So What ?", we discuss the following news articles: The AI Revolution Is Already Losing Steam The AGI Hype Is Damaging the AI Opportunity You can watch the episode here: https://lnkd.in/eCp2a8_9 Previous episodes can be watched here: https://lnkd.in/gNyuQt5E #data #analytics #ai #ml #datascience #designedanalytics

looks like I'm not the only one trying to mine arxiv I'm using pymupdf, but txtai has some examples scraping pdfs I like the layout detection this article has plus the ability to remove references rag over arxiv has become my new pet project (sorry stocks, chasing you was exhausting and not directly in line with the latest ml trends). eventually though, whatever I cook up for arxiv will be applied to 10ks https://lnkd.in/eJf8k8_z

1

Iteration Head: A Mechanistic Study of Chain-of-Thought Vivien Cabannes, Charles Arnal, Wassim Bouaziz, Alice Yang, Francois Charton, Julia Kempe Abstract Chain-of-Thought (CoT) reasoning is known to improve Large Language Models both empirically and in terms of theoretical approximation power. However, our understanding of the inner workings and conditions of apparition of CoT capabilities remains limited. This paper helps fill this gap by demonstrating how CoT reasoning emerges in transformers in a controlled and interpretable setting. In particular, we observe the appearance of a specialized attention mechanism dedicated to iterative reasoning, which we coined "iteration heads". We track both the emergence and the precise working of these iteration heads down to the attention level, and measure the transferability of the CoT skills to which they give rise between tasks. 👉 https://lnkd.in/dzWgCZAP #machinelearning

4

1 commentaire

Over the past year, I've realized that AI can be an incredibly valuable asset for software developers. The key lies in mastering its prompts. Here is the breakdown of Good and Bad AI prompts : Good Prompt ✔ Clear and Specific: Write a Python function that takes a list of numbers and returns the average, handling empty lists gracefully. Context Provided: I'm building a shopping cart application. Can you generate some unit tests for the product addition functionality? Desired Format: Create a pseudocode algorithm for a depth-first search on a binary tree. Bad prompts - Vague and Ambiguous: Make me some code for my Real estate project." (AI has no idea what kind of functionality you need) Missing Context: Write a sorting algorithm. (Without specifying the data type or desired sorting order, the AI might generate something unusable) Overly Complicated: Develop a secure, scalable, and user-friendly e-commerce platform with AI-powered product. If you learn how to leverage AI, you simply can be the best in your Job. #ai #solutionarchitect #leader #thought #openai #gemini #meta #Llhama #chatgpt

11

Armand Ruiz, if the goal is to advance #AI towards a more human-like understanding and creativity, Ilya's viewpoint may be seen as more aligned with these aspirations. The emphasis on deep understanding can drive innovations that make AI interactions more natural and contextually aware. For practical and technical applications where reliability, efficiency, and scalability are paramount, Yann's viewpoint may be more appropriate. Focusing on the statistical and structural capabilities of LLMs aligns with optimizing and deploying AI technologies in real-world scenarios where deep "understanding" is less critical than performance and accuracy.

1

Already 8 months when i presented autogen V.02for Gen Ai startup program with Microsoft incubator startup n STATION F and now we are here. ❤️ This library has demonstrate how to use #agents / multi LLM - SLM with decision pattern ( tooling RAG, script , code ,apis ) to adress usecases and pipelines of creation of business values. ( and not only for a chatbot...) Autogen rises ( a lot ! ), new capabilities, low-code studio appears... Framework orchestration in python, java, c#... demonstrated how to create application with #AI. 🔥 Gen AI is obviously mature for business and pratical usecases are already in production.🚀 🏃 And if you don't know about #Autogen ? >> Rush into youtube / website to understand why it must be a must known for your business 🔥 One of the next step ( low code for autogen Studio ): https://lnkd.in/ea64EF5b the library : https://lnkd.in/e4W7_gHT 📜 First demo live ( french) : https://lnkd.in/eTchhs5n 📜 First demo presential :https://lnkd.in/ehmjwAgZ #agents #microsoftfrance #llm #slm #genai #ai #autogen

28