Eric Siegel

The Economist's article “Large language models are getting bigger and better” on April 17, 2024 on the challenges of scaling large language models (LLMs) rightly identifies the limitations of current AI architectures. While the relentless pursuit of larger LLMs suggests a future of limitless progress, a deeper dive reveals significant flaws. The "scaling hypothesis"—that more data and power equals human-like intelligence—misses crucial aspects. LLMs face serious hurdles: the dwindling supply of high-quality data, the inefficiency of scaling, and the simple fact that size doesn't equate to cognitive complexity. Here, neurosymbolic AI presents a promising alternative, blending the data-handling prowess of neural networks with the logical precision of symbolic AI. This combination not only promises enhanced efficiency and structured reasoning but also aligns closer with human cognitive processes. Neurosymbolic AI excels in areas where traditional models falter, offering a path to more sustainable, intelligent systems capable of real-world applications like airplane design and financial risk assessment. It’s time for the AI community to pivot towards technologies that not only scale but think smarter. https://loom.ly/96rDCD0

21

4 Comments

“To be able to fight disinformation…[we]…must understand it. But research on disinformation is limited in several key respects…A big obstacle for researchers is the lack of access to data. The best data is not in public hands, but is “sitting in private networks in Silicon Valley,” says Phil Howard…of the International Panel on the Information Environment (IPIE). And collecting relevant data is becoming more difficult. After Elon Musk bought Twitter…in 2022 the company shut down the free system that let anyone download information on posts and accounts, and began charging thousands of dollars a month for such data access. Meta announced in March that it would be retiring CrowdTangle its platform-monitoring tool that lets scientists, journalists and civil-society groups access data, though the company says academics can still apply for access to certain datasets. Such changes have seriously hampered researchers’ ability both to detect disinformation and to understand how it spreads… In America…efforts to fight disinformation have become caught up in the country’s dysfunctional politics…fighting disinformation requires a co-ordinated effort by tech platforms, academics, government agencies, civil-society groups and media organisations. But…any co-ordination of this kind has come to be seen, particularly by those on the right, as evidence of a conspiracy between all those groups…Researchers studying disinformation have been subjected to lawsuits, attacks from political groups and even death threats. Funding has also diminished…All this has had a chilling effect on the field...“It is difficult to avoid the realisation that one side of politics—mainly in the US but also elsewhere—appears more threatened by research into misinformation than by the risks to democracy arising from misinformation itself” Disinformation is a sprawling problem, requiring co-ordinated action from multiple sectors of society. Unfortunately, the analysis of it tends to be siloed and there is a lack of agreement on things like terminology. This makes it hard to join the dots and find lessons that apply more broadly. Phil likens the situation to the early days of climate science: lots of people are trying to tackle the same problem from different perspectives, but it is difficult to see the whole picture…The current lack of a joined-up response to disinformation is a problem, but also an opportunity: co-ordinating research and action should lead to better detection and mitigation of deceptive content, because modern disinformation campaigns all work in similar ways. But, as with climate change, cleaning up the world’s information environment presents a daunting, long-term challenge.” https://lnkd.in/gb_g9CKf Ellen Doowan Mark Nina Roman Emerson Alan Jevin Renee Sheldon Justin J.D. Klon John Brandon Ed Sasha Alex Jakub Jill Clint The Economist Celine

35

4 Comments

Enhancing LLM Performance Through Contextual Retrieval: Anthropic's innovative approach to Retrieval-Augmented Generation (RAG), which enhances the performance of Large Language Models (LLMs) in specialized domains. By integrating a knowledge base with LLMs, this method allows for efficient data retrieval without the need for extensive model retraining. https://lnkd.in/gHZcBTFH #learning #techsolution

2

I have been attending SWARM Community's "Responsible AI Summit" this week. This article in diginomica says a lot "Governments rarely have the resources or skills to prevent the worst excesses of a free-market economy, and it is always nigh on impossible to put a genie back in its bottle. So, that leaves a good chunk of the responsibility with the tech industry to ensure it does not end up living out its days like Oppenheimer, campaigning against the destructive force it has created. But it is going to take more than just an open letter appealing for collective action in creating responsible AI. It is going to require concrete action – and now." #AI #aiethics #aisafety #responsibleAI #teckedin #chatgpt #LLM #tech4good Joaquin Melara Kellie A. Christian Ortiz ✊🏽 Elle Farrell-Kingsley Maestro Stevens Clara Lin Hawking https://lnkd.in/gV3GNPhy

6

3 Comments

The prevailing methods to make large language models more powerful and amenable have been based on continuous scaling up (that is, increasing their size, data volume and computational resources) and bespoke shaping up (including post-filtering, fine tuning or use of human feedback). However, larger and more instructable large language models may have become less reliable. By studying the relationship between difficulty concordance, task avoidance and prompting stability of several language model families, here we show that easy instances for human participants are also easy for the models, but scaled-up, shaped-up models do NOT secure areas of low difficulty in which either the model does not err or human supervision can spot the errors. #genAi #LLM #LLMs

9

1 Comment

A Language Interaction Prompt: Enhancing AI Interactions with Psycholinguistic Precision I’ve been exploring the 237-page dissertation Finding Structure in Language Models by Jaap Jumelet (https://lnkd.in/g5Kqt8mB), which describes how language models process and replicate linguistic structures. From this research, I developed a 'top five' list of ways to refine and align human-AI communication styles: - Structural Priming - Lexical Overlap - Semantic Similarity - Inverse Frequency Effects - Incremental/Hierarchical Structural Processing These approaches, inspired by advanced psycholinguistics and NLP evaluations, emphasize coherence, contextual relevance, and linguistic adaptation—elements often overlooked in standard prompt engineering. By leveraging these methods, we can evoke human-like tendencies (e.g., structural persistence, semantic adaptation) to create interactions that feel natural, engaging, and contextually rich. Here’s the abbreviated prompt I designed to make these methods actionable for anyone seeking to optimize AI interactions: General Instructions: Follow these guidelines to improve alignment, coherence, and relevance. Seek clarification as needed. Guidelines: Structural Priming: Mirror my sentence structure. For example, if I say "The scientist conducted an experiment for the students," you might say "The artist painted a portrait for the family." Question: Strictly mirror grammar or maintain a similar pattern? Lexical Overlap: Reuse key terms or synonyms. If I mention "climate change," include words like "sustainability" or "renewable energy." Question: Which words or themes should I emphasize? Semantic Similarity: Use related concepts. If I say "innovation in technology," respond with "advancement" or "digital transformation." Question: General synonyms or domain-specific terms? Inverse Frequency Effects: Match vivid language. If I say "Amidst the twilight…," respond with equally evocative phrasing. Question: More poetic or more direct? Hierarchical Processing: Build ideas step-by-step. If I say "While the rain poured…," respond with "As the rain continued…" Question: How detailed should the incremental structure be? Feedback: Answer these questions so I can tailor responses. If no feedback, I’ll assume a balanced approach. #AI #PromptEngineering #Psycholinguistics #LanguageModels #HumanAIInteraction #NLP #Innovation

6

🔭 In a thought-provoking turn of events, #Anthropic underscores the potential perils associated with AI systems, calling for a well-structured regulatory framework to nimbly avert AI disasters. 🏗️ The organization vehemently voices that targeted regulation is not just an option but a necessity, to harness AI’s advantages while concurrently deescalating its inherent threats. • Logic and math capabilities • Superior coding skills • Ability to reason 🧩 As AI systems evolve in these powerhouse capabilities, the potential for misuse remains latent, especially in critically sensitive sectors like cybersecurity. Thoughts to ponder: 🤔 ✔️ Will a well-thought regulatory blueprint assure that AI's potential does not morph into an unintended Pandora's Box? ✔️ In mis preemptive measures such as regulation, can we ensure that innovation is not stifled? In a digital era, where we are heavily reliant on machine learning and AI systems for various aspects, it's inevitable that this resilient technology will permeate across all echelons of our society, from healthcare to automotive sectors and beyond. Interestingly, is AI's expanding capability a one-edged sword? Or is it the double-edged samurai sword that holds power in one hand and risk at the other end? This conversation resonates that: 🌠 Fear of the unknown is a fear, but so is the fear of the known when we don't control it effectively. Balanced foresight and smart regulations can indeed create a safe & robust AI frontier. Are we ready to face and embrace the conundrum? #ThoughtLeader #AIRegulation #DigitalFrontier.

2

🚀 Coming to you from New York City! 🌆 I just watched the interview with Dr. Simon Prince on Machine Learning Street Talk, and I'm thrilled to get his new book, 'Understanding Deep Learning.' 📚 Dr. Prince's point about Large Language Models (LLMs) is crucial. LLMs can't have internal consistency. They've learned everything from the internet, which means they believe the earth is both flat and round in different statistical proportions. 🌍 While they hopefully conclude it's mostly round, somewhere in their data is the notion of a flat earth. When humans decide to believe something, we rely on an internal, self-consistent model of the world. LLMs, however, lack this kind of model. They operate on something much more complex and difficult to analyze. 🤯 Understanding this distinction is key to navigating their challenges and potential🌟 #Pangeon #ThePangeon #AI #AGI #OpenAI #FutureTech #GPT #ChatGPT #MachineLearning #DeepLearning #TechInnovation #AIResearch #StartupJourney

11

4 Comments

👉 Beyond Imitation: Exploring the Cognitive Exegesis of LLMs More than just tools, LLMs can become partners in the interpretive process. ⚡ LLMs engage in "cognitive exegesis," interpreting text like human scholars. ⚡ They adapt to individual learning needs, resonating with our creative thinking. ⚡ LLMs amplify human curiosity, offering instant, nuanced interpretations. #LLMs #AI #cognition #learning

8

2 Comments

🔍 Harmoniticity Analysis: A New Frontier in Evaluating LLM Trustworthiness In the realm of Natural Language Processing (NLP), ensuring the reliability of Large Language Models (LLMs) remains a paramount concern, especially when these models provide confident, yet potentially inaccurate, responses. Traditional metrics for assessing LLM output have struggled to adequately measure trustworthiness—until now. Researchers from VISA have introduced a novel approach using harmoniticity analysis to quantitatively evaluate the robustness and reliability of LLM responses. This innovative method, centered around a metric called γ (gamma), measures local deviations from what the researchers term "harmoniticity." This model-agnostic and unsupervised approach assesses the stability and explainability of LLM outputs by analyzing how they respond to input perturbations. Specifically, γ calculates the angle between the average output embedding of perturbed inputs and the original output embedding, offering a straightforward yet powerful measure of response fidelity. The utility of γ was demonstrated through extensive human annotation experiments, which established a clear positive correlation between higher γ values and the likelihood of false or misleading LLM answers. Moreover, employing stochastic gradient ascent along the gradient of γ has proven effective in identifying adversarial prompts, showcasing the method’s practical effectiveness. The research team conducted empirical tests across various models and question-answer (QA) corpora to validate the approach. Leading LLMs including GPT-4, ChatGPT, and Claude 2.1, along with older models like Llama2-7B, were evaluated against QA datasets in diverse domains. Human annotators rated the truthfulness and relevance of responses on a 5-point scale, with Fleiss’ Kappa indicating high inter-annotator agreement. Results from these experiments underscored that lower γ values generally correlate with higher trustworthiness in LLM responses. Notably, larger models like GPT-4 consistently showed lower γ values, indicating a superior level of certified trustworthiness compared to smaller or older models. This groundbreaking study not only enhances our understanding of LLM output reliability but also paves the way for more nuanced and effective methods of evaluation. By correlating harmoniticity with human judgment, the research provides a robust framework for assessing the trustworthiness of AI-generated content, potentially transforming how we validate and utilize LLMs in high-stakes applications. 👉 Explore the full paper below or download link in the comments below. If you find it valuable, please like and share this post – your engagement shapes the content I share. Also, follow me if you haven't already. Thanks! #AI #NLP #LLM #MachineLearning #Trustworthiness #VISA #HarmoniticityAnalysis

6

2 Comments

Tremendous data, stats, examples and trends about AI risk in this chapter of Stanford's annual AI Index report: "Risks from AI are a concern for businesses across the globe....Organizations are beginning to take steps to mitigate these risks. However, globally, most companies have so far only mitigated a portion of these risks." Key context for what we focusing on at AI Risk. More here: www.ai-risk.co #airisk #airiskmanagement

8

Could Artificial Intelligence Create Real Liability for Employers? Colorado Just Passed the First U.S. Law Addressing Algorithmic Discrimination in Private Sector Use of AI Systems (US) https://buff.ly/3yFcneY - 💡 Colorado passed Senate Bill 24-205, the Consumer Artificial Intelligence Act (CAIA), the first U.S. law aimed at protecting consumers from harm arising from AI systems. - 📅 The CAIA will take effect on February 1, 2026, and applies to Colorado businesses using AI systems for employment decisions. - 🤖 High-risk AI systems must adopt a risk-management policy, complete annual impact assessments, provide notices to consumers, and report algorithmic discrimination to the attorney general. - ➡️ Colorado employers should monitor for guidance on compliance and amendments to the CAIA to ensure they are prepared. - 🇺🇸 There is currently no comprehensive federal legislation regulating the use of AI systems in the U.S. - 🔍 The CAIA shares structural similarities with the AI Act adopted by the European Union, despite industry opposition. - ⚖️ Compliance with the CAIA can offer an affirmative defense in an enforcement action by the attorney general.

1

1 Comment

I asked Gemini 1.5 to explain Anthropic's new research paper on the inner workings of AI models. Here's what it said... ------------------------- This research paper is about figuring out what's going on inside a large language model, like a super-smart computer that can write stories and answer questions. It's like trying to understand how a human brain works by looking at the signals in its different parts. The researchers used a technique called "𝗱𝗶𝗰𝘁𝗶𝗼𝗻𝗮𝗿𝘆 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴" to break down the model's internal signals into smaller, more understandable pieces called "𝗳𝗲𝗮𝘁𝘂𝗿𝗲𝘀." Think of it like taking a complex song and figuring out what notes make up each part of the melody. They discovered that these features represent all sorts of things, like: 𝗦𝗽𝗲𝗰𝗶𝗳𝗶𝗰 𝗹𝗼𝗰𝗮𝘁𝗶𝗼𝗻𝘀: They found features for the Golden Gate Bridge, Los Angeles, and even individual countries like Rwanda. 𝗖𝗼𝗻𝗰𝗲𝗽𝘁𝘀: They found features for things like "brain sciences" or "transit infrastructure." 𝗖𝗼𝗱𝗲: They found features that recognize code errors, bugs, and even backdoors (hidden ways to access a system). 𝗘𝗺𝗼𝘁𝗶𝗼𝗻𝘀: They found features that represent sadness, conflict, and even a sense of empathy. What's really interesting is that these features can be used to control the model's behavior. For example, if they activate the "Golden Gate Bridge" feature, the model might start talking about the bridge, even if it wasn't asked about it before. The researchers also discovered features that relate to potential safety concerns, like: 𝗕𝗶𝗮𝘀: They found features that activate when the model is expressing biased or prejudiced language. 𝗗𝗲𝗰𝗲𝗽𝘁𝗶𝗼𝗻: They found features that activate when the model is trying to trick the user or hide information. 𝗗𝗮𝗻𝗴𝗲𝗿𝗼𝘂𝘀 𝗰𝗼𝗻𝘁𝗲𝗻𝘁: They found features that activate when the model is discussing things like biological weapons or scam emails. This research is still very early, but it shows that we can start to understand the internal workings of these powerful language models. This is important because it could help us make sure that these models are used safely and ethically. 𝗞𝗲𝘆 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀: 🔹 We can use dictionary learning to break down a language model's internal signals into understandable pieces. 🔹 These pieces, called "features," represent specific locations, concepts, code, emotions, and more. 🔹 Features can be used to control the model's behavior. 🔹 We've found features that relate to safety concerns, like bias, deception, and dangerous content. This research could help us make sure that large language models are safe and ethical. ------------------------- 𝗧𝗵𝗮𝗻𝗸𝘀, 𝗚𝗲𝗺𝗶𝗻𝗶 𝗳𝗼𝗿 𝘀𝗮𝘃𝗶𝗻𝗴 𝗺𝗲 𝟮 𝗵𝗼𝘂𝗿𝘀 𝗮𝗻𝗱 𝗮 𝗵𝗲𝗮𝗱𝗮𝗰𝗵𝗲! Ironically, Claude told me the paper was 7% over the length limit. What do you think about the discovery of these signals? Share your thoughts and consider reposting. ♻️

21

13 Comments

The landscape is saturated with AI systems, each promising revolutionary capabilities in impact measurement. Yet, amidst this development, a crucial element often goes unchecked—the caliber and domain expertise of the developers behind these technologies. The foundation of a reliable AI system is not just in its coding but in the profound scientific principles, theories, and rigorously developed models it's built upon. Before you commit to any AI solution, it’s imperative to scrutinize the robustness of the training data, the intricacies of the model, and the accuracy of the algorithm. A credible AI system merges technological innovation with deep, domain-specific insights to deliver not only performance but precision and dependability.

5

Google’s own researchers have sounded the alarm: AI is now the leading source of disinformation. The scale of the problem might be even larger than we think. This ties directly into Brandolini’s Law, also known as the Bull*#•s Asymmetry Principle: “The amount of energy needed to refute bull** is an order of magnitude bigger than that needed to produce it. So, when Google warns us that AI is flooding us with unprecedented amounts of misinformation, it’s clear we face a monumental challenge in combating it. MidwestCon we recognize that rapid AI development without transparency and responsibility only exacerbates this issue. That’s why we’re championing transparency and authenticity in AI as the pillars of sustainable growth. We have experts in responsible AI from across the country & we have training and workshops available starting in June for the first 100 who purchased a ticket.

18

2 Comments

#bias #ai #ai #artificialintelligence Because of course you do use en algorithmn to decide the fate of a human being.

1