Dan Bikel

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

My new standard attack is evolving to: "Why are you anthropomorphizing LLMs? Would you anthropomorphize 7zip or JPEG?" I'm reposting this article to help reframe the discussion. LLMs are, in fact, super lossy compressors. That's it. That's all they are. They don't care about your feelings because super lossy compressors don't have feelings and they don't have care. Stochastic parrot, or my new improved knockoff, stochastic pigeon, do not get to the heart of the problem. A compression / decompression algorithm is not a sentient being. End of story. Argument over. #WrittenByMe https://lnkd.in/gGaqyRMg

5

12 Comments

An interesting hybrid approach for conditional generation that leans into the strengths of both diffusion and autoregressive models - to produce text with a set of desirable properties: https://lnkd.in/eeiXR_MC It would be interesting to see this explored/evaluated further, across more use cases (and domains).

15

Sharing my experiment on creating a Streamlit based chatbot, that combines inputs from Agentic RAG using llamaindex (with openai and MistralAI), and langchain google genai (with Gemini-1.5-Flash), to extract information from images embedded within the PDF files. Was a nice experience to build this chatbot. Still many open questions, ways to improve the solution, in a production scenario..but nice to combine features from multiple libraries and come up with a solution. Please check it out! Thanks. #chatbot #gemini #openai #langchain #llamaindex https://lnkd.in/gPdtxhQF

19

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

16

1 Comment

There's more going on at Cloudian Inc than the mere signing of a new $23M round of funding: ‘What we have seen in the past year or so is that AI is making everybody question their data strategy and their cloud versus on-prem footprint. People are much more sensitive in terms of where they store their data. So essentially, what we're seeing is that enterprises are realizing that data is very fundamental to their AI journey. And at Cloudian, the data lake always sits at the foundation of everything else,’ says Cloudian CEO and Co-founder Michael Tso.

3

2 Comments

We have six interesting papers at #ACL2024 on #CodeLLM, #context usage of LLMs, #Low-resource language support with LLMs, #LVMs for Charts, #Multmodal RAG. Check those out: 1. XCodeEval: An Execution-based Large Scale Multilingual Multitask Benchmark for Code Understanding, Generation, Translation and Retrieval. Mohammad Khan, M Bari, Do Long, Weishi Wang, Md Parvez, and Shafiq Joty. In ACL'24 Main. https://lnkd.in/ghbrAX7p 2. On Context Utilization in Summarization with Large Language Models Mathieu Ravaut, Aixin Sun, Nancy Chen, and Shafiq Joty. In ACL'24 Main. https://lnkd.in/gEubck9H 3. Democratizing LLMs for Low-Resource Languages by Leveraging their English Dominant Abilities with Linguistically-Diverse Prompts Xuan Phi Nguyen, Mahani Aljunied, Shafiq Joty, and Lidong Bing. In ACL'24 Main. https://lnkd.in/gP_jGNhm 4. ChartInstruct: Instruction Tuning for Chart Comprehension and Reasoning. Ahmed Masry, Mehrad Shahmohammadi, Md Parvez, Enamul Hoque, and Shafiq Joty. In ACL'24 Findings. https://lnkd.in/g4nMXTR2 5. Data Augmentation using LLMs: Methods, Learning Paradigms and Challenges. Bosheng Ding, Chengwei Qin, Ruochen Zhao, Tianze Luo, Xinze Li, Guizhen Chen, Wenhan Xia, Junjie Hu, Anh-Tuan Luu, Shafiq Joty, and Shafiq Joty. In ACL'24 Findings. https://lnkd.in/gB_G76tN 6. XL-HeadTags: Leveraging Multimodal Retrieval Augmentation for the Multilingual Generation of News Headlines and Tags. Faisal Shohan, Mir Nayeem, Samsul Islam, Abu Akash, and Shafiq Joty. In ACL'24 Findings.

210

6 Comments

Since January, I've been up to my eyes in retrieval-augmented generation and large language models. It's the wild west of technology 🤖🤠 With so little guidance publicly available, the learning curve is steep! Our team frequently crawls ✨ Arxiv ✨ for inspiration because academia is leading the way 📚 My mistakes don't have to be yours! Here are 5 lessons I've learned the hard way: in production 😅🚀

3

Integrity. Ethics. Commercial safety. Move slow and build things.

8

While everyone seems to be focused on large language models (LLMs) and advancements in frontier models, with their vast capabilities and knowledge transference, it’s crucial not to overlook the unique value of lightweight models. Microsoft’s introduction of the Phi-3 family of small language models (SLMs) marks an exciting shift towards AI solutions that are not only powerful but also more accessible and efficient. The versatility of SLMs, which can outperform larger models, opens new possibilities for genAI applications in settings with limited resources. The idea of using a swarm of lightweight models working in a distributed fashion may be feasible. Such configurations could leverage the strengths of multiple smaller models to perform tasks collaboratively in remote or resource-constrained environments like tactical edge scenarios. For a deeper understanding of these small yet mighty models, take a look at this article on Microsoft’s latest advancements. #AI #MachineLearning #TechnologyInnovation

3

Watch this video to understand how LLMs store facts. The entire series is excellent to get an intuitive understanding. It all boils down to basic linear algebra. About 18 minutes into the video, it touches on sparse auto encoders and why LLMs generalize so well as the dimensionality increases

4

Will the future in #AI be prompt and go or agent. Andrew Ng (AI/ML guru) says that we should look at agents models. Just listen to Andrew Ng’s agentic AI talk. https://lnkd.in/gyjqN7gD He strongly believes that are patterns emerging and shows evidence from several technical papers published in the past year. In particular, get ready to unleash more with these 4 patterns: 🤔 Reflection (think Khanmigo) 🛠️ Awesome tools (magic image generators, etc.) 🎯 Strategic planning (define / delegate those tasks!) 🤝 Multi-agent teamwork so are you ready for #AI #agents?

6

1 Comment

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

New research tackles the fight against misinformation with AI-powered counter-narratives! A fascinating paper explores using Large Language Models (LLMs) to automatically evaluate the effectiveness of Counter Narratives (CNs) in combating online misinformation. The study reveals that LLMs, particularly chat-aligned models, excel at generating and evaluating CNs when fine-tuned with domain-specific data. This promising approach could revolutionize how we address misinformation and promote healthier online discourse. #misinformation #disinformation #cognitivesecurity #generativeAI #LLM #AI #techforgood https://lnkd.in/gHrVk9bs

11

I was left a bit disappointed with the lack of AV support at the IJCAI conference for remote* speakers/participants (more on that later), but I am happy to be able to share the full recording of my invited talk at the AI Governance Workshop with you! My talk - titled "Can LLMs Really Understand?" - goes into the details of how LLMs attend to human conversation data, as compared to other kinds of data. It is based on a recent paper that we released on this topic. My thanks again to all the organizers of the AIGov Workshop - Asim Munawar, Djallel Bouneffouf, Baihan Lin, PhD, Irina Rish, Lauri Goldkind - for their gracious invite, and I look forward to comments on the talk! #InvitedTalk #AIGovernance #LLMs #AI #ML #NLP #LargeLanguageModels #ConversationalAI #ConversationAI #SymblDotAI Symbl.ai #NebulaLLM * Although I understand the motivations behind trying to foster in-person conference events, it is unconscionable to actively discourage remote speaking and participation in this day and age. There are any number of reasons that can prevent physical attendance at an event - the biggest one being that conferences are often held in countries that gate entry with cumbersome and time-consuming visa requirements. I was left very disappointed that the IJCAI organization would ignore the fact that not all passports are created equal in this world. I will probably write something longer about this experience in due course. Chime in if you have had similar experiences in the past.

27

1 Comment

The latest advances in artificial intelligence (particularly large language models) continue to reverberate. Even for an “old school” AI person like myself (who cut his teeth with Prolog) it is clear that there has been a step change in our ability to create computer systems that can interact with humans using natural language. GPT-4 et al are exhibiting early signs of “common sense” and have encoded useful conceptual representations of the world. The debate rages on as to whether this is “intelligence”, but to an engineer like me, it sure seems useful! #LLM #AI #Intelligence #Programming https://lnkd.in/ebAgNapK

14

4 Comments

I recently took a deep-dive into an article written by Laurin Brechter on AI Quantum Intelligence's Blog about A/B Testing. Folks, this isn't just ordinary A/B testing. 💡 Brechter is discussing Bayesian A/B Testing. With traditional A/B testing, businesses can experiment to see how different versions of a webpage or marketing asset perform in terms of click-through rates and user engagement. Conversion rates- the percentage of visitors who turn into customers takes the helm in this strategy. So how is Bayesian A/B Testing different? Applying Bayes' Theorem in these tests provides a flexible and complete picture of the full range of possible outcomes and their corresponding probabilities. ✔ Gone are the days of just aiming for an increase in each conversion. ✔ Instead, it presents a wealth of statistical data that gives robust insights into user engagement and click-through rates. My prediction on this? 🔮 The use of Bayesian A/B Testing marks a new dawn in business and marketing optimization strategies. Harnessing the power of detailed Bayesian A/B testing would give businesses the agility to adapt their marketing strategy to meet more precise, evidence-based objectives. What's particularly insightful here is: - The ability to capture a full range of possible outcomes, not just the best/worst case scenarios. - It allows businesses to better pivot with their marketing strategies based on these robust insights. The inspiring thing about Bayesian A/B Testing is how it mirrors the horizon of possibilities: expansive, unknown, but reachable and fathomable through technology, analytics, and a dash of boldness! Remember: Business optimization is not just about playing, but also about winning the right moves. Now that’s an invigorating thought! Let’s redefine boundaries and challenge norms by embracing dynamic tools like Bayesian A/B testing geared towards sharp, prescient decision-making. #ABTESTING #BAYESIANTHEORY #BUSINESSOPTIMIZATION #MARKETINGSTRATEGIES #AIQUANTUMINTELLIGENCE.

4

AI is going to make data-driven decisions 100% democratized! Jenica Zhong and Jessica F. have built an extraordinary product Tile (YC W24) to achieve this and enable all the Dwights out there to explore and investigate data in a modularized manner 😎 One of the most exciting features of Tile is the ability to have explainability on how to came to a certain result and how you can fine-tune these steps to get to where you need to be. Let me remind you, when you use SQL or Python to do it, things can be super complex, and backtracking can be extremely challenging. With Tile (YC W24), not only data scientists will be empowered to run no-code queries, but it will also give the ability to make data-driven decisions on real-time data to non-technical people! Check out this super interesting demo of how Dwight (from The Office) would use Tile to make data-driven sales decisions at Dunder Mifflin 👇 I am an investor in Tile (YC W24), and am excited to see the product grow and empower non-tech professionals across various industry verticals to make informed decisions, and get handy analytics! Interested to try it out yourself? Sign up for a free trial here: https://lnkd.in/dmk6NBUf ♻ Share this with your network to let them know of this amazing product coming out in the market

155

8 Comments

Dive into the world of Generative AI and explore its impact on code generation! 🤖 Dr. Bob Sutor recently tested five browser-based code generators for HTML files, revealing a fascinating range in quality and completeness. Discover the results and unlock the potential of AI in software development. #AI #CodeGeneration #TheFuturumGroup

7