Eric Schles

We just released a blog on whether the use of multimodal feature transfer can detect deception (phenomena like sarcasm, irony, and condescension) in conversation. The short answer... Yes. This particular write-up profiles a subset of our work at Symbl.ai on multimodal large language models, and is research led by one of our summer interns, Benjamin R. (from Prof. Larry Heck's AVA Lab at Georgia Institute of Technology). We have a paper currently under submission on this work, and we can't wait to share the full details with all of you soon. Blog Link: https://lnkd.in/g6m_wFDi #multimodal #featuretransfer #LLMs #deception #conversation #sarcasm #irony #language #AI #ML #NLP #LargeLanguageModels #SummerInternship #ResearchPapers #ACL #Llama2 #SymblDotAI

16

Hence the need to move AI onchain with projects like Morpheus AI

5

One library was missing on the list. It leverages constrained grammar sampling from llama.cpp to produce structured outputs with any LLMs... https://lnkd.in/etv3QWH4

"OpenAI basically set back progress towards AGI by probably 5 to 10 years. They caused this complete closing down of frontier research publishing. And now LLMs have sucked the oxygen out of the room" Provocative soundbite from François Chollet on Dwarkesh Patel's podcast. The increasing secrecy of frontier research has been a big disappointment - it is the opposite of the wonderful open source code & data momentum that has been steadily increasing over the years. On the other hand, the increased AI hype should bring more clever minds to the field, who spread their knowledge as they move employers or found their own businesses, and I would expect at least some of the capital fuelling the genAI boom will trickle into more AGI-like endeavours.

28

2 Comments

When assessing the performance of Large Language models (LLMs) we see the primary focus being technical benchmarks, such as graduate-level reasoning, coding proficiency, and multilingual math capabilities. These metrics are undeniably important, but only paint part of the picture. The recent performance comparison of models like Claude 3.5 Sonnet, Claude 3 Opus, GPT-4o, and Gemini 1.5 Pro underscores this point. While Claude 3.5 Sonnet excelled in graduate-level reasoning, coding, and multilingual math, it begs the question: How long until we start incorporating soft skills and behavioural traits into our model assessments? Soft Skills: The Human Touch in AI [1] Problem-Solving: Beyond solving complex equations, AI should demonstrate strong analytical and problem-solving skills, much like what we expect from top-tier talent. This involves tackling intricate financial problems and developing innovative solutions. [2] Communication: The ability to communicate technical concepts clearly and effectively to both technical and non-technical stakeholders is crucial. AI models need to bridge the gap between data science and business operations, ensuring everyone is on the same page. [3] Adaptability: As technologies evolve, so should our models. Adaptability in learning new technologies, adjusting to changing environments, and handling multiple tasks simultaneously is a trait we value in human employees and should equally prioritise in AI. [4] Teamwork: Collaboration is key. AI models should facilitate and enhance teamwork, contributing to a positive and productive work environment. They should be designed to integrate seamlessly with human teams, complementing their efforts. [5] Attention to Detail: Accuracy and precision in financial data and software development are non-negotiable. AI must exhibit meticulousness in its operations, ensuring high standards of quality and reliability. Behavioural Traits: Ethical and Proactive AI [1] Motivation and Initiative: Just as we look for candidates who demonstrate a proactive attitude towards learning and self-improvement, our models should show a capacity for continuous improvement and an 'interest' in advancing the fintech landscape. [2] Ethical Judgement: Understanding ethical considerations, such as data privacy, fairness, and regulatory compliance, is paramount. AI models must be designed with these principles in mind, ensuring they operate within ethical boundaries. Integrating these soft skills and behavioural traits into our assessment criteria will provide a more holistic view of an AI model's capabilities. It's not just about excelling in isolated technical tasks but also about how well these models can adapt, communicate, and integrate into broader business contexts. #AI #SoftSkills

13

🔝🚨Anomalous emergent behaviors in AI: Hermes 3, has "existential crises" when you use right inputs, a black prompt or ask "Who are you?" or "What do you see?". Here you are some examples of unexpected emergent behaviors at scala. Emergent behaviors are really relevant in AI because thosw appears without especific training and many times only a few times even in the LLM has been trained with the same datasets and procedures. #AI #AGI #LLM #research #artificialIntelligence https://lnkd.in/dyDUEFs4

3

1 Comment

This story can be used also as example for technical debt in software projects. I can’t say anything about the motivation of the author but, for sure I can state that any work around has a price! I will use Maxime Labonne post as example when talking about companies that try so sell you services without any sort of detail

6

🎉 KAN: Kolmogorov-Arnold Networks ⚡️TL;DR: 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. 👉 Paper Discussion by First Author, Ziming Liu : https://lnkd.in/gYwKMHfw pip install pykan 👉 Abstract: Inspired by the Kolmogorov-Arnold representation theorem, we propose Kolmogorov-Arnold Networks (KANs) as promising alternatives to Multi-Layer Perceptrons (MLPs). 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. We show that this seemingly simple change makes KANs outperform MLPs in terms of accuracy and interpretability. For accuracy, much smaller KANs can achieve comparable or better accuracy than much larger MLPs in data fitting and PDE solving. Theoretically and empirically, KANs possess faster neural scaling laws than MLPs. For interpretability, KANs can be intuitively visualized and can easily interact with human users. Through two examples in mathematics and physics, KANs are shown to be useful collaborators helping scientists (re)discover mathematical and physical laws. In summary, KANs are promising alternatives for MLPs, opening opportunities for further improving today's deep learning models which rely heavily on MLPs. ⚙️ Code: https://lnkd.in/ghpeJjyU 🦋 Docs: https://lnkd.in/gTBdBfYh 📄 Paper: https://lnkd.in/gcwniZuS 💡 Note: Checkout the “Author’s Note” section on GitHub ReadMe. https://lnkd.in/g7F3HPr7 Also see: Althernate implementation of KAN on GitHub 👉 efficient-kan: https://lnkd.in/gjGEEkpX 👉 fourier-kan: https://lnkd.in/gbjh5x5i #KAN #paper #code #docs #python #research #physics #pinn #ml

1

Anyone who has used long-context LLMs will quickly realize that performance at near the full claimed context size is not nearly as good as what you can get at 4-8k context. There are a lot of tricks that allow LLMs to accept longer inputs (positional interpolation, etc.) but they don't magically convert a 4k model to 32k model keeping the performance. The simple effective context length benchmark proposed here is very useful. RULER: What's the Real Context Size of Your Long-Context Language Models? (arXiv:2404.06654) https://lnkd.in/gAq_iuvF

25

1 Comment

https://lnkd.in/gWUpZECK Mixture of Agents (MoA) is an approach to harness the collective strengths of multiple LLMs to improve state-of-the-art quality. According to Together AI, it achieves a score of 65.1% on AlpacaEval 2.0, surpassing prior leader GPT-4o (57.5%).

5

"..Development will transition back to what it was back in the days before the waterfall process was even formalized in the sense that most time will be spent not iteratively debugging code but in rigorously specifying and documenting software systems up front, while LLMs compile those documents into the final program." Absolutely! Software development has long been an inefficient endeavor, often creating end products that hold users captive rather than enhancing their productivity. Conversational interfaces are the final piece of the puzzle for addressing this costly problem. We are entering a period where software applications will work for users rather than the other way around. #AI #Software #Disruption #Industry #TheInevitable

9

6 Comments

Yesterday, it was announced that NBC will be using generative AI to create personalized Olympic highlights reels and summaries using the voice of the iconic Al Michaels. voice. This will be available on Peacock. This was honestly inevitable when LLMs went mainstream. Here, you no longer need to sit through a half hour of Olympics summaries at night or hunt for videos on what sports you follow to find out what happened. No, now it can all come straight to you on an app. But wait, why stop there? What about customized educational content? Not every student learns the same way; why can't professors create lectures that are more problem solving-focused, more theoretical, or whatever appeals to and motivates a student? What about not just sports news, but will we eventually get all our news delivered to us in the voice of Walter Cronkite...or Donald Trump or Joe Biden? And why not movies? One could easily redo The Shawshank Redemption to have Morgan Freeman's "Redd" have the voice of Danny DeVito or Harry Shearer (the voice of Mr. Burns on "The Simpsons"). This sounds interesting and fun and cutting edge, but let's think downstream. Is this right and ethical, or more dangerous? A mix? I guess if the voice/image providers approve the use and are compensated, and it's clearly articulated that content is AI generated, AND we can verify accuracy, yeah, probably. But this is a lot to get right. In this scenario, Al Michaels clearly gave permission to NBC to use his voice, NBC is disclosing the AI use case, and the Olympic results available on other sources if you choose to look elsewhere for verification. But let's not delude ourselves into thinking that only responsible actors (pun slightly intended) will start mass delivering content. Check out the video below to see how real it sounds. And a bit more food for thought: 1. I've said this before: this is as bad as deepfake technology will be. Every day advancements are made, bringing us closer to AI being indistinguishable from reality. 2. Should we be comfortable with AI-generated news content being made normal? Separately, I'll post in the next day or so on Peacock, Apple TV+, Prime Video and other pay apps now hosting exclusive sports programming. I utterly detest this trend, but it is totally okay to screw consumers amirite? ..

10

The Human-or-Machine Issue In the June 2024 issue of Communications of the ACM, David Harel reflects on Alan Turing's "imitation game" and introduces the broader "human-or-machine" (H-or-M) issue. As machines increasingly mimic humans, Harel explores new social, psychological, and technical questions. Will humans care whether their interacting agent is human or machine? How will behaviors and emotional responses differ in human-machine interactions compared to human-human ones? #HumanComputerInteraction #ImitationGame #Interactions #HumanBehavior

3

2 Comments

How do we ensure evaluations of LLMs (aka “evals”) are rigorous and trustworthy? 🤔 A recent paper by Anthropic tackles this question head-on, offering practical recommendations for measuring model performance with precision. TL;DR ▪️Use the Central Limit Theorem: Treat eval scores as part of a “question universe” to measure underlying skill—not just lucky guesses. ▪️Account for Clusters: Some evals have related questions (e.g., same text passages). Ignoring this can lead to overconfidence in results. ▪️Reduce Randomness: Techniques like re-sampling answers or using token probabilities can cut noise, sharpening results. ▪️Focus on Paired Differences: Compare models using shared questions to reduce variance and amplify insights. ▪️Power Matters: Fewer questions = less statistical power. Power analysis ensures we ask enough to detect meaningful differences. Evals are more than benchmarks—they’re our window into what LLMs can (and can’t) do. By embracing this approach, we can make AI adoption clearer and more robust. 🔗 Link to the paper: https://lnkd.in/d5uDbkqA 🔗 Link to the blog: https://lnkd.in/dj9kJCWE #ai #genai #anthropic #evals

48

AI Capabilities Plateauing? Is "Bigger is Better" Losing Its Edge? A recent article by Reuters highlights these challenges and the industry's pivot toward new paradigms, such as reasoning and inference-time computation. The focus is shifting from brute force scaling to models that innovate and think independently. For years, scaling up pre-training—feeding vast amounts of data into ever-larger models—has been the cornerstone of AI advancements. But according to Ilya Sutskever, co-founder of OpenAI and Safe Superintelligence (SSI), we may be hitting a plateau. Recent findings suggest that simply increasing model size is yielding diminishing returns. Yann LeCun puts it succinctly: "Unless the learning system can 'invent' solutions and answers to new inputs by itself (through an ability to reason) instead of merely receiving and adapting previously-learned answers, the plateau will be with us." Implications for the Industry - Are we ready for AI systems that think and reason rather than memorize and adapt? - How do we redesign learning paradigms to enable true reasoning and problem-solving? This isn’t just a technical shift—it’s a philosophical one. AI’s evolution depends on its ability to move from scaling to reasoning. Reuters Article: https://lnkd.in/e2uyZVVW #AI #MachineLearning #ScalingLaws #Innovation

8

3 Comments

The interfaces we have today will become useless: Anthropic’s new “computer use” feature in Claude 3.5 represents a major shift in AI capabilities, marking the start of a future where humans might no longer understand software interfaces as we do today. The model can now use computers as people do—navigating screens, clicking buttons, and typing. This could render current interfaces obsolete as AI automates complex tasks across industries. However, the feature remains experimental and error-prone, though expected to improve rapidly with developer feedback. This innovation opens new possibilities for the future of human-computer interaction. For more details, check the full article here. https://lnkd.in/eqfbmhdQ

38

2 Comments

The recent growth of LLMs underscores the incredible complexity of the human brain. While LLMs can suffer from issues like hallucinations and an innate inability to select data intelligently, advancements continue to push the boundaries of what's possible. For instance, a recent paper on Ferret-UI demonstrates innovative approaches by splitting images into halves for LLM processing. However, it's important to remember that the human brain's ability to focus on relevant regions based on context is a product of evolution and extensive knowledge. LLMs are evolving, and though it might sometimes seem like solving a jigsaw puzzle, each piece contributes to a greater understanding. The race for more powerful GPUs is part of this journey, and while it may not immediately solve the challenge of achieving general AI, it is driving significant progress. The hype around LLMs may seem overwhelming, but these developments are essential steps towards converging AI with human-like understanding and capabilities.

6