Demis Hassabis

🔍 The Future of AI with OpenRLHF 🔍 As AI technology advances, aligning large language models (LLMs) with human values and intentions becomes increasingly critical. Reinforcement Learning from Human Feedback (RLHF) is a powerful technique addressing this challenge. However, traditional RLHF frameworks struggle with the complexity and resource demands of training models exceeding 70 billion parameters. Launching this week is **OpenRLHF** is an open-source framework designed to overcome these limitations by leveraging cutting-edge technologies such as Ray, vLLM, and DeepSpeed. Here's why OpenRLHF is a game-changer for AI development and why I am so excited! 1. Scalability - OpenRLHF efficiently scales RLHF training for LLMs beyond 70 billion parameters by distributing models across multiple GPUs, optimizing memory use, and minimizing computational overhead. 2. Performance Optimization - By integrating Ray for model scheduling, vLLM for accelerated sample generation, and DeepSpeed for enhanced training, OpenRLHF ensures superior performance and reduced training time. 3. Versatility and Usability - Fully compatible with the Hugging Face library, OpenRLHF supports various alignment techniques like Direct Preference Optimization (DPO), Kahneman-Tversky Optimization (KTO), and rejection sampling. It offers user-friendly, one-click training scripts for diverse models and algorithms. 4. Resource Efficiency - OpenRLHF’s scheduling and memory management reduce GPU memory fragmentation and communication overhead, enabling larger batch sizes and more efficient training processes. OpenRLHF sets a new standard for AI development and ResponsibleAI. Explore the full potential of OpenRLHF and its strategic benefits in the detailed report. Read more https://lnkd.in/eCF7Q9Pb #Innovation #AI #Leadership #RLHF #OpenSource #TechAdvancement #AILeadership #FutureOfAI

5

Working with some Large Language Models (LLMs) at Trionic Technologies LLP, Vatsal Patel asks a very interesting question: "Can we really exercise some influence or exploit LLMs to get focused answer from multiple choices?" On digging into it, we found a very interesting research paper that highlights the same! This paper, "Manipulating large language models to increase product visibility," shows how "strategic text sequences" could increase the visibility of the product in LLM-driven searching aids. Think about dropping a purposely designed message into a product description and seeing that one rise in the recommendation listing to the very top. Not only is it possible, but it's also now proven! Their research in optimizing these sequences employs techniques such as the Greedy Coordinate Gradient (GCG) algorithm to ensure that the very products, even if in real spaces, which were previously shrouded on digital catalogs, shine bright on the virtual shelf. This is almost revolutionary in SEO for AI-run search services in the sense of really changing the perception and priority of products by algorithms. The potential of businesses to harness this ethically is not less impactful and innovative in generating competitive edges. Let's see how strategic text sequences can help in innovations of digital marketing and e-commerce. Link: https://lnkd.in/dPE8zFtm #AI #MachineLearning #LLMs #DigitalMarketing #ecommerce

15

🚨 I've made a Google Colab Notebook for anyone interested in experimenting with the new Stable Diffusion SD3 model, via their API ✨ Google Colab Notebook 👉 https://lnkd.in/ei2Fx3aF Before using you will need to create a Stability.ai account (new accounts get 25 free credits) and create an API key. The notebook contains instructions for how to generate an API key and where to paste it in Google Colab. If you use up your free credits you will need to buy some more within your Stability account. Google Colab Notebook 👉 https://lnkd.in/ei2Fx3aF #stablediffusion #sd3 #api #stabilityai #googlecolab #python

15

5 Comments

I had a great time talking about Synthetic Data for Computer Vision at the #Bristol  MLOps Community #Meetup. If you couldn't attend in person, check out the video below.. #AI #ML #DeepLearning #SyntheticData

24

Hallucinations in LLMs are inevitable, no seriously! Josh Fritzsche and I have worked hard at Voze to find a way to tune, prompt and even plead 🙏 with LLMs to prevent them from the occasional 🍄 hallucination. It seemed like the better we have done the harder the edge cases have been to solve and eventually we realized that they tend to frequently happen when LLMs predict or make anticipations on "nothing". I remember at one point nearly going insane with frustration because when LLMs had a hallucination, they were EPIC! 😱 Then we pushed through it and benefited for the effort! How did we solve this common problem? We didn't (Not completely), we got better at identifying the issues as they happened. Once we identified in the data we were able to start looking for patterns. The same catalysts that cause it to happen once if left unchanged caused it over and over again. Our solution is astoundingly simple. Once you know that it's likely to happen..... let it! Accept it will 🍄, but also be preemptive in not focusing on it and use alternative data sources, generators or providers. Change the conditions that cause it but continue to monitor it as it happens. As Dan Caffee is fond of saying to me on a regular bases. "Dusty we need the system to be adaptive and provide mechanisms to learn from it's past." Well he is right! All we needed was data and the people that contribute to it's quality. Huge props 👏 to strong support from Kathy and Janelles audit teams. We were armed with a knowledge and a desire to learn from past mistakes. Armed with confidence that we could correct these problems we could focus on the data to prevent their impact going forward. When doing millions of LLM (Generative AI) calls regularly Josh and I have learned that hallucinations will happen. We also know exactly how we will handle and prevent it going forward! Failure is not doing anything and accepting defeat.

4

2 Comments

Open AI has released the Model Spec, a detailed description of how GPT models have been instructed to handle different types of prompts. This includes concepts such as respecting the chain of command (yes, developers win over end users!), asking clarifying questions and avoiding sharing personal information. If you're looking for details on what's under the hood of GPT4, you'll be disappointed, but it is still a good resource to bookmark if you want to understand at a high level how GPT models are taught to behave in a certain way. #openai #llm #gpt4 https://lnkd.in/egJWNxhj

9

Very insightful work by Anthropic’s interpretability team. And an amazing paper, with outstanding writing and figures. The idea is very simple: interpret LLMs by leveraging sparse autoencoders as surrogate models of the MLP of transformer blocks, which allow one to disambiguate the superposition of features captured by a single neuron. A simple idea, but a very careful and complex execution, as it is often the case in our line of work. The paper goes into many details and provide a large array of insights, although the gist of the implementation remains obfuscated due to the closed source nature of Claude. Too bad, because this is the kind of work that we need to better understand and eventually trust LLMs. This is demonstrated by the authors in the section ‘Influence on Behavior’, where they show that clamping some features to either high or low value during inference is “remarkably effective at modifying model outputs in specific, interpretable ways”. Hopefully this kind of work is going to be replicated and generalized to open-weights models, such that we have new ways to steer their behavior. https://lnkd.in/eVym7f_f #interpretability #xai #explainableai #steerableai #anthropic #claude #anthropic

3

🔥🚀 Exciting News from MIT: Introducing FlowMap for Deep Learning-based Structure from Motion! 🔥🚀 MIT unveils FlowMap, a groundbreaking self-supervised, end-to-end differentiable SfM method delivering COLMAP-level accuracy for 360° scenes. FlowMap revolutionizes the process by precisely solving camera poses, intrinsics, and dense depth of video sequences without relying on unique image features or manual matching. This innovation outperforms prior gradient-descent based methods and rivals COLMAP, marking a significant advancement in SfM technology. 👉 Explore the project page, paper, and code: https://lnkd.in/gyyinuTu #Reconstruction #SfM #AI #DeepLearning #Innovation

7

Stole a few points from Prof Christopher Manning, who talked about LLMs and language modeling in general in the latest TWIML AI Podcast: 📌 Humans acquire language skills in a way very different from LLMs. We need millions of words, compared to billions or even trillions of tokens for LLMs. LLM researchers may want to investigate and learn from how humans acquire language skills. 📌 LLMs cannot reason. However, there are other deep learning models, such as AlphaGo, that can. LLM researchers may want to look into how to integrate that type of reasoning/searching/planning capability into LLMs. 📌 LLMs' world models should enable search and discovery. Although Prof. Manning didn’t call this out explicitly, my understanding is more similar to a knowledge graph type of structure. 📌 Next-gen LLM idea: a soft form of locality and hierarchy. Transformers attend every token to every other token, which is very inefficient, while human language can be modeled by n-grams most of the time. #artificialintelligence #machinelearning #deeplearning https://lnkd.in/euzwMQ6p

84

18 Comments

🚀 Unlocking the Power of Text Embedding with Cohere and Hugging Face 📚 Text embedding lies at the heart of the remarkable capabilities of Large Language Models (LLMs). It transforms textual data into numerical vectors, enabling machines to understand and process language effectively. In my latest notebook, we delve into the world of text embedding using Cohere's Embed endpoint and the rich email intent classification dataset from HuggingFace. Join me on this exciting journey as we harness the potential of text embedding to unlock deeper insights and drive innovation in natural language processing. Notebook link: https://lnkd.in/dXbFcFdJ _Love and AI_ 🤍🤖 #TextEmbedding #Cohere #HuggingFace #NLP #AI

1

1 Comment

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

Following our release of posteriors, Normal Computing’s library for Uncertainty-Aware LLMs, we're excited to share how this new library makes large-scale Bayesian learning in PyTorch more accessible than ever! Our newest paper on "Scalable Bayesian Learning with Posteriors" by Sam Duffield, Kaelan Donatella, Johnathan Chiu, Phoebe Klett, and Dan Simpson is now available on arXiv. In this paper, we show the key benefits of using posteriors for Bayesian machine learning: • Better generalization and out-of-distribution predictions by averaging over multiple plausible models • Coherent online learning of new data without catastrophic forgetting • Decomposing uncertainty into aleatoric (data) and epistemic (model) components The paper also unifies SGD and SGMCMC (Stochastic Gradient Markov Chain Monte Carlo) methods through a temperature parameter. SGMCMC extends SGD by adding noise to the updates, allowing it to sample from the posterior distribution of parameters instead of just finding a point estimate. This unification leads to a modification of deep ensembles, which we call "parallel SGMCMC". By adding noise to the parameter updates during training, parallel SGMCMC allows the ensemble members to approximate the posterior distribution more accurately compared to standard deep ensembles. Read the full paper here: https://lnkd.in/evwdDnbX

42

2 Comments

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

#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

I was chatting about translation with Andrew Ng at an event several weeks ago. It's fun to see him post a workflow with the seed of the AI Review concept that's live in LILT today. === This is along the path to what many in the industry are doing. Two comments: 1. Replace step 2 with a formalized reflection approach such as AutoMQM. 2. BLEU isn't discriminative at the accuracy levels modern MT systems achieve. Use a trained metrics like BLEURT or COMET to see real differences between systems. 20 years ago we called this idea automatic post-editing, where you'd train a second string transducer to rewrite initial MT output. However, those systems were very difficult to train and cascade. Using LLMs in cascades is significantly more effective.

42

Explainable AI - What actually is it? Explainable AI is reentering the discourse...after a rather quiet period (because let's face it - it is really hard for LLMs). I wanted to bring up several important questions to address when deciding to discuss or work on this topic. There is room for all of the answers, but DO have a clear answer before you jump into this space. 1. What does "Explainable" actually mean in your context? Not super obvious. Are you trying to explain the exact mathematical/programmatic set of steps that went on between the input and the output? This works well for a Logistic Regression or a Rule-base model, but not so much for a multi-billion parameter model. Are you trying to approximate which features of the input had largest effect on the output? Can be applied to large models, but many explanation methods will have a degree of error. Are you trying to explain the underlying "human" process (captured in the training data) that connected the inputs and the outputs? Arguably more useful in certain contexts, but may not connect to the actual model internals. 2. Why are you creating explanations? Many different worthwhile goals here. They will help answer question 1, and the general path you take. A few possibilities: - Building Trust: Proving to a human that a model is doing something meaningful and accurate. - Evaluating for Bias and Unfairness: Proving to a human that a model is not relying to "biases and stereotypes" to make decisions. - Progressing Research: Fundamentally how a model works can help build better models. 3. Who are you building explanations for? Explanations that are helpful for an ML Engineers are not the same ones that are helpful for a doctor. An ML Engineer may want a higher-fidelity approach that "highlights" the weights in the model; a doctor may want to see several reference MRI scans from the training data that led to a diagnosis but not actually care about the underlying neural network. ⁉ What other fundamental questions did I miss? PS: I chose not to address the notions of "interpretability" vs "explainability" here because that would get us way past the character limit.

7