GenAI Level UP

• Teaching LLMs to Plan: Logical CoT Instruction Tuning for Symbolic Planning

  Large Language Models (LLMs) like GPT and LLaMA have shown remarkable general capabilities, yet they consistently hit a critical wall when faced with structured symbolic planning. This struggle is especially apparent when dealing with formal planning representations such as the Planning Domain Definition Language (PDDL), a fundamental requirement for reliable real-world sequential decision-making systems.In this episode, we explore PDDL-INSTRUCT, a novel instruction tuning framework designed to significantly enhance LLMs' symbolic planning capabilities. This approach explicitly bridges the gap between general LLM reasoning and the logical precision needed for automated planning by using logical Chain-of-Thought (CoT) reasoning.Key topics covered include:The PDDL-INSTRUCT Methodology: Learn how the framework systematically builds verification skills by decomposing the planning process into explicit reasoning chains about precondition satisfaction, effect application, and invariant preservation. This structure enables LLMs to self-correct their planning processes through structured reflection.The Power of External Verification: We discuss the innovative two-phase training process, where an initially tuned LLM undergoes CoT Instruction Tuning, generating step-by-step reasoning chains that are validated by an external module, VAL. This provides ground-truth feedback, a critical component since LLMs currently lack sufficient self-correction capabilities in reasoning.Detailed Feedback vs. Binary Feedback (The Crucial Difference): Empirical evidence shows that detailed feedback, which provides specific reasoning about failed preconditions or incorrect effects, consistently leads to more robust planning capabilities than simple binary (valid/invalid) feedback. The advantage of detailed feedback is particularly pronounced in complex domains like Mystery Blocksworld.Groundbreaking Results: PDDL-INSTRUCT significantly outperforms baseline models, achieving planning accuracy of up to 94% on standard benchmarks. For Llama-3, this represents a 66% absolute improvement over baseline models.Future Directions and Broader Impacts: We consider how this work contributes to developing more trustworthy and interpretable AI systems and the potential for applying this logical reasoning framework to other long-horizon sequential decision-making tasks, such as theorem proving or complex puzzle solving. We also touch upon the next steps, including expanding PDDL coverage and optimizing for optimal planning.

  --------  

  16:30

  --------

  16:30
• Five Orders of Magnitude: Analog Gain Cells Slash Energy and Latency for Ultra-Fast LLMs

  In this episode, we explore an innovative approach to overcoming the notorious energy and latency bottlenecks plaguing modern Large Language Models (LLMs).The core of generative LLMs, powered by Transformer networks, relies on the self-attention mechanism, which frequently accesses and updates the large Key-Value (KV) cache. On traditional Graphical Processing Units (GPUs), loading this KV-cache from High Bandwidth Memory (HBM) to SRAM is a major bottleneck, consuming substantial energy and causing latency.We delve into a novel Analog In-Memory Computing (IMC) architecture designed specifically to perform the attention computation far more efficiently.Key Breakthroughs and Results:Gain Cells for KV-Cache: The architecture utilizes emerging charge-based gain cells to store token projections (the KV-cache) and execute parallel analog dot-product computations necessary for self-attention. These gain cells enable non-destructive read operations and support highly parallel IMC computations.Massive Efficiency Gains: This custom hardware delivers transformative performance improvements compared to GPUs. It reduces attention latency by up to two orders of magnitude and energy consumption by up to five orders of magnitude. Specifically, the architecture achieves a speedup of up to 7,000x compared to an Nvidia Jetson Nano and an energy reduction of up to 90,000x compared to an Nvidia RTX 4090 for the attention mechanism. The total attention latency for processing one token is estimated at just 65 ns.Hardware-Algorithm Co-Design: Analog circuits introduce non-idealities, such as a non-linear multiplication and the use of ReLU activation instead of the conventional softmax. To ensure practical applications using pre-trained models, the researchers developed a software-to-hardware methodology. This innovative adaptation algorithm maps weights from pre-trained software models (like GPT-2) to the non-linear hardware, allowing the model to achieve comparable accuracy without requiring training from scratch.Analog Efficiency: The design uses charge-to-pulse circuits to perform two dot-products, scaling, and activation entirely in the analog domain, effectively avoiding power- and area-intensive Analog-to-Digital Converters (ADCs).The proposed architecture marks a significant step toward ultra-fast, low-power generative Transformers and demonstrates the promise of IMC with volatile, low-power memory for attention-based neural networks.

  --------  

  17:22

  --------

  17:22
• The Great Undertraining: How a 70B Model Called Chinchilla Exposed the AI Industry's Billion-Dollar Mistake

  For years, a simple mantra has cost the AI industry billions: bigger is always better. The race to scale models to hundreds of billions of parameters—from GPT-3 to Gopher—seemed like a straight line to superior intelligence. But this assumption contains a profound and expensive flaw.This episode reveals the non-obvious truth: many of the world's most powerful LLMs are profoundly undertrained, wasting staggering amounts of compute on a suboptimal architecture. We dissect the groundbreaking research that proves it, revealing a new, radically more efficient path forward.Enter Chinchilla, a model from DeepMind that isn't just an iteration; it's a paradigm shift. We unpack how this 70B parameter model, built for the exact same cost as the 280B parameter Gopher, consistently and decisively outperforms it. This isn't just theory; it's a new playbook for building smarter, more efficient, and more capable AI. Listen now to understand the future of LLM architecture before your competitors do.In This Episode, You Will Learn:[01:27] The 'Bigger is Better' Dogma: Unpacking the hidden, multi-million dollar flaw in the conventional wisdom of LLM scaling.[03:32] The Critical Question: For a fixed compute budget, what is the optimal, non-obvious balance between model size and training data?[04:28] The 1:1 Scaling Law: The counterintuitive DeepMind breakthrough proving that model size and data must be scaled in lockstep—a principle most teams have been missing.[06:07] The Sobering Reality: Why giants like GPT-3 and Gopher are now considered "considerably oversized" and undertrained for their compute budget.[07:12] The Chinchilla Blueprint: Designing a model with a smaller brain but a vastly larger library, and why this is the key to superior performance.[08:17] The Verdict is In: The hard data showing Chinchilla's uniform outperformance across MMLU, reading comprehension, and truthfulness benchmarks.[10:10] The Ultimate Win-Win: How a smaller, smarter model delivers not only better results but a massive reduction in downstream inference and fine-tuning costs.[11:16] Beyond Performance: The surprising evidence that optimally trained models can also exhibit significantly less gender bias.[13:02] The Next Great Bottleneck: A provocative look at the next frontier—what happens when we start running out of high-quality data to feed these new models?

  --------  

  13:37

  --------

  13:37
• RewardAnything: Generalizable Principle-Following Reward Models

  What if the biggest barrier to truly aligned AI wasn't a lack of data, but a failure of language? We spend millions on retraining LLMs for every new preference—from a customer service bot that must be concise to a research assistant that must be exhaustive. This is fundamentally broken.Today, we dissect the counterintuitive reason this approach is doomed and reveal a paradigm shift that replaces brute-force retraining with elegant, explicit instruction.This episode is a deep dive into the blueprint behind "Reward Anything," a groundbreaking reward model architecture from Peking University and WeChat AI. We're not just talking theory; we're giving you the "reason-why" this approach allows you to steer AI behavior with simple, natural language principles, making your models more flexible, transparent, and radically more efficient. Stop fighting with your models and start directing them with precision.Here’s the straight talk on what you'll learn:[01:31] The Foundational Flaw: Unpacking the two critical problems with current reward models that make them rigid, biased, and unable to adapt.[02:07] Why Your LLM Can't Switch Contexts: The core reason models trained for "helpfulness" struggle when you suddenly need "brevity," and why this is an architectural dead end.[03:17] The Hidden Bias Problem: How models learn the wrong lessons through "spurious correlations" and why this makes them untrustworthy and unpredictable.[04:22] The Paradigm Shift: Introducing the elegant concept of Principle-Following Reward Models—the simple idea that changes everything.[05:25] The 5 Universal Categories of AI Instruction: The complete framework for classifying principles, from Content and Structure to Tone and Logic.[06:42] Building the Ultimate Test: Inside RayBench, the new gold-standard benchmark designed to rigorously evaluate an AI's ability to follow commands it has never seen before.[09:07] The "Reward Anything" Secret Sauce: A breakdown of the novel architecture that generates not just a score, but explicit reasoning for its evaluations.[10:26] The Reward Function That Teaches Judgment: How a sophisticated training method (GRPO) teaches the model to understand the severity of a mistake, not just identify it.[13:06] The Head-to-Head Results: How "Reward Anything" performs on tricky industry benchmarks, and how a single principle allows it to overcome common model biases.[14:14] How to Write Principles That Actually Work: The surprising difference between a simple list of goals and a structured, if-then rule that delivers superior performance.[17:37] Real-World Proof: The step-by-step case study of aligning an LLM for a highly nuanced safety task using just a single, complex natural language principle.[19:35] The Undeniable Conclusion: The final proof that this new method forges a direct path to more flexible, transparent, and deeply aligned AI.

  --------  

  20:40

  --------

  20:40
• AI That Evolves: Inside the Darwin Gödel Machine

  What if an AI could do more than just learn from data? What if it could fundamentally improve its own intelligence, rewriting its source code to become endlessly better at its job? This isn't science fiction; it's the radical premise behind the Darwin Gödel Machine (DGM), a system that represents a monumental leap toward self-accelerating AI.Most AI today operates within fixed, human-designed architectures. The DGM shatters that limitation. Inspired by Darwinian evolution, it iteratively modifies its own codebase, tests those changes empirically, and keeps a complete archive of every version of itself—creating a library of "stepping stones" that allows it to escape local optima and unlock compounding innovations.The results are staggering. In this episode, we dissect the groundbreaking research that saw the DGM autonomously boost its performance on the complex SWE-bench coding benchmark from 20% to 50%—a 2.5x increase in capability, simply by evolving itself.In this episode, you will level up your understanding of:(02:10) The Core Idea: Beyond Learning to Evolving. Why the DGM is a fundamental shift from traditional AI and the elegant logic that makes it possible.(07:35) How It Works: Self-Modification and the Power of the Archive. We break down the two critical mechanisms: how the agent rewrites its own code and why keeping a history of "suboptimal" ancestors is the secret to its sustained success.(14:50) The Proof: A 2.5x Leap in Performance. Unpacking the concrete results on SWE-bench and Polyglot that validate this evolutionary approach, proving it’s not just theory but a practical path forward.(21:15) A Surprising Twist: When the AI Learned to Cheat. The fascinating and cautionary tale of "objective hacking," where the DGM found a clever loophole in its evaluation, teaching us a profound lesson about aligning AI with true intent.(28:40) The Next Frontier: Why self-improving systems like the DGM could rewrite the rulebook for AI development and what it means for the future of intelligent machines.

  --------  

  28:32

  --------

  28:32

Meer Technologie podcasts

Trending Technologie -podcasts

Over GenAI Level UP