Paul Golding

LLMs and Optimization: An Unbeatable Combo!

19

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

I took Meta’s brilliant paper, and asked Anthropic’s Claude 3.5 Sonnet to write a provocative article summarising it from the perspective of what this many mean for closed foundation models. Read below: Although, safety and alignment wise … it’s pretty easy to finetune these things if you know what you’re doing to turn them into subversive comedic monsters 😉🤔😂 Not that I would do such a thing. No sir! (full paper @ https://lnkd.in/d2qs6DPn ) —— Your Move, Sam Altman: Meta's Llama 3.1 Throws Down the Gauntlet In a bold move that has sent shockwaves through the AI community, Meta has unveiled Llama 3.1, a family of LLMs that not only rival the performance of OpenAI's vaunted GPT-4 but do so with a twist: they're giving it away for free. That's right Sam, the game has changed. While OpenAI has been guarding its golden goose behind an API paywall, Meta has decided to release not 1, not 2, but 3 powerhouse models into the wild. We're talking 8B, 70B, and a whopping 405B parameters of open-source AI goodness. It's like they're handing out Lamborghinis at a go-kart race. But it's not just about size (though let's face it, 405 billion parameters is nothing to sneeze at). Llama 3.1 is flexing on multiple fronts: Performance: The 405B model is going toe-to-toe with GPT-4 and Claude 3.5 Sonnet across a smorgasbord of tasks. It's like watching David take on Goliath, except David is also a giant and he's livestreaming the whole thing. Efficiency: While OpenAI's models are busy sipping mai tais in their cushy cloud servers, Llama 3.1’s 8B model can run on a toaster. Okay, maybe not a toaster, but certainly on your average laptop. Democratizing AI? More like starting an AI revolution. Transparency: Meta's not just showing their cards; they're explaining the whole trick. Training process? Here you go. Data curation? Have at it. Evaluation methods? It's all yours. It's so open, it makes OpenAI look like Fort Knox. Safety: They've got Llama Guard watching the gates. It's like they've hired a bouncer for their AI, but instead of keeping people out, it's keeping the AI in check. Multimodal Potential: Vision and speech capabilities are waiting in the wings. It's like Meta's saying, "Oh, you thought language was impressive? Hold my beer." So, what's the catch? There isn't one. Meta's basically dropped a nuke of openness on the AI landscape, and the fallout is going to reshape everything. To Sam Altman and the rest of the closed-source crew, the message is clear: the era of AI gatekeeping is over. Meta's not just inviting everyone to the party; they're giving away the DJ booth, the dance floor, and the whole damn club. In conclusion, Llama 3.1 isn't just a model; it's a manifesto. It's Meta saying, "We're not just in the game; we're changing it." So, Sam, the ball's in your court. Are you going to open up, or are you going to be left behind in a world where state-of-the-art AI is as open as a 24-hour diner? Your move, Sam. Your move.

7

1 Comment

Yet another indicator about why LLMs are not databases. They are non-deterministic (in other words, they are stochastic), meaning that even with the same prompts and contexts, different host GPUs can change the output of the LLM in different ways. Even with RAG, replicability is not guaranteed. The underlying message here? LLMs value as generators is unquestionable, and it is a good tool for natural language prompts. However, LLMs likely will never be reliable enough to serve as databases, precisely because they are chaos machines: they are extraordinarily dependent upon starting conditions. https://lnkd.in/gBDei_ye

277

35 Comments

RAG ... a necessary but *not* sufficient element of reliable NIR ... "Neural Information Retrieval". As revolutionary as LLMs are, NLU is not solved. The old NLU benchmarks are obsolete, but the challenge of genuinely *understanding* and fully responding to specific user intent remains before us. And many of the old tools will find new applications in a new synthesis. Call it Neural Information Retrieval. https://lnkd.in/eD72UBQ4

18

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

How is Mark Zuckerberg going to make money from open-sourced AI models? Meta has just released its Llama-3 70B model, an open-sourced LLM that is tied with GPT-4 turbo as the best AI model on the market. What's super interesting about Meta's strategy is that, unlike other giants like Google DeepMind and OpenAI who have kept their models closed-source, Zuck has decided to open-source Llama, which means making the source code and/or the trained model publicly available for anyone to use, tinker with, and distribute freely. It's like what Google did with Android, keeping the mobile operating system open so people can build on it for lots of different use cases and without restriction to try and fight Apple's hegemony and drive advertising sales – which is Google's core revenue generator. And from the sounds of it, Meta has taken a similar approach and poured billions into Llama as a way to put the middle finger up at the closed-source competition by making it free to the entire world. This does a couple of things: 1. It accelerates AI adoption (good) 2. Allowing others to tinker helps drive innovation and is potentially another route to superintelligence (very good) 3. It undermines competition (also very good) But I wouldn't mistake this as some newly developed sense of altruism. The key question is: where will the return on investment come from if Zuckerberg isn't directly monetizing his AI models? Evan Armstrong writes that the crazy thing here is that Zuck might be scared that AI will drive consumer attention away from social media platforms. In other words, users spend more time talking to AI instead of scrolling through Instagram or Facebook. To counter this threat, Meta is integrating AI chatbots into all of its platforms as a way to collect valuable data and increase its ad targeting capabilities. The strategy might pay off, or it might not. But it's a bold, multi-billion dollar move that could shape the next decade and maybe the rest of our lives.

8

Today, Google Cloud and Anthropic made some exciting announcements: 📣 Claude 3 Opus, Anthropic’s most intelligent model, is now generally available on Vertex AI ⚙️ Tool use is available for the Claude 3 model family on Vertex AI, enabling Claude to act as an autonomous agent 📈Provisioned throughput is available for the Claude 3 model family on Vertex AI

1

These findings motivate our design of a novel framework to investigate hidden representations in LLMs with LLMs, which we call Patchscopes. The key idea behind this framework is to use LLMs to provide natural language explanations of their own internal hidden representations. Patchscopes unifies and extends a broad range of existing interpretability techniques, and it enables answering questions that were difficult or impossible before. For example, it offers insights into how an LLM's hidden representations capture nuances of meaning in the model's input, making it easier to fix certain types of reasoning errors. While we initially focus the application of Patchscopes to the natural language domain and the autoregressive Transformer model family, its potential applications are broader. For example, we are excited about its applications to detection and correction of model hallucinations, the exploration of multimodal (image and text) representations, and the investigation of how models build their predictions in more complex scenarios. Google DeepMind

9

The LLM will see you now: Google and DeepMind released a paper on Monday describing Med-Gemini, a group of advanced AI models targeting healthcare applications. The authors claim that Med-Gemini is outperforming competing models such as OpenAI’s GPT-4. During the pandemic we saw the rise of Telehealth which has largely come back to earth but could doctors be consulting more often with AI? (We know they already Google stuff) #Healthcare #AI

12

9 Comments

Are newer smaller models better then GPT-4? (read till the end) The chart shared by Peter Gostev shows his analysis results, demonstrating that smaller foundational models (6B, 7B, and 8B parameters) are becoming increasingly efficient and effective. Over time, these smaller models have surpassed the original GPT-4 in performance (represented by the green line). This is significant because smaller models require less computational power and resources, making them more cost-effective while still delivering high-quality results. What's the Elo Rating? Elon Rating is a scoring system that ranks models based on their performance in head-to-head evaluations. A higher Elo rating means the model performs better in comparison to others. For example, if Model A has a higher Elo rating than Model B, it means Model A is more likely to outperform Model B in similar tasks. The Elo rating system, while useful for comparing language models (LLMs), has several limitations and potential downsides when applied in this context — such as sensitivity to comparison order, lack of transitivity, and reliance on relative scoring. Therefore the need for complementary evaluation methods (e.g., MMLU scores) for a more comprehensive assessment of model performance (holistic view). For example, a model with a high Elo score might excel at generating responses that are stylistically appealing or persuasive but could lack depth in reasoning or factual accuracy. Human evaluators may unconsciously favor models with more fluent, coherent, or stylistically pleasing outputs, even if those outputs lack substance. This can inflate the scores of models with better "style" rather than better "intelligence." Does it mean the results are irrelevant? Absolutely not! We have learned that for specific situations where factual accuracy isn't critical, we can employ smaller models that require MUCH less computational resources (energy consumption) and obtain stylistic results comparative to LLMs. Note: This post has been written with the help of AI.

4

Bit by bit, we're slowly reinventing the semantic web. The article here explores an idea that I've been thinking about as well. There are two key advantages to integrating knowledge graphs via RAG vs. trying to work JUST from LLMs - the fist is that you have the ability to name entities and taxonomies, ensuring topical synchronization via taxonomies. However, a second aspect is that you can teach an LLM to utilize a specific data structure by passing it a schema. However, there's a corollary to that. If you can teach an LLM how to build schemas, then the LLM can use this to design a schema for its own purposes to optimize reasoning. Now, I'm going to make a contrafactual argument here to others (people I respect) to say that I think an LLM is perfectly capable of reasoning, because we don't understand reasoning. An LLM is incapable (and even here I'd quibble) to utilize intent, but if it is specifically informed to create a "castle in the sky" graph that allows it to perform reasoning tasks, it should be able to traverse that graph with appropriate constraints, because I've used that technique myself when using LLMs to design schemas. Put another way, in order for an LLM to reason, it needs to become an ontologist. https://lnkd.in/gW6D9TeW

44

6 Comments

Today, Google Cloud and Anthropic made some exciting announcements: 📣 Claude 3 Opus, Anthropic’s most intelligent model, is now generally available on Vertex AI ⚙️ Tool use is available for the Claude 3 model family on Vertex AI, enabling Claude to act as an autonomous agent 📈Provisioned throughput is available for the Claude 3 model family on Vertex AI

23

The rise of video applications has spurred interest in video-language pre-training, where large-scale video and weak captions are used for representation learning. This approach leverages models like transformers for tasks such as video captioning, action recognition, and retrieval. Key datasets like Howto100M and Kinetics aid in training robust models. Pre-training methods include single-stream (e.g., VideoBERT, HERO) and two-stream (e.g., CBT, UniVL) techniques. These advancements enable better understanding of video content, facilitating improvements in multimodal tasks and the development of more efficient models for diverse applications. https://lnkd.in/gby2wEjX

2

Doing #GENAI? Often after creating a #LLM , it feels like the "Talking Parrot", garbles at times . We need the responses to be in context and TEXT to SPARQL to be accurate. #ontology can help. Recommend article by Juan Sequeda: Problems : 1️⃣ To what extent can the accuracy increase by leveraging the ontology of a knowledge graph to detect errors of a #SPARQL query and an LLM to repair the mistakes? 2️⃣ What errors are most commonly presented in SPARQL queries generated by an LLM? 🧪 hypothesis: An ontology can increase the accuracy of an LLM-powered question-answering system that answers a natural language question over a knowledge graph.

8