Janik Jaskolski

🤚🏼🛑  Here you have my slides from the API Conference New York As AI becomes increasingly API-driven, we face a critical question: 👉 If APIs are the arteries of AI, what’s flowing through them—just data, or also values? 👈 ⚖️ Why Ethics and Responsibility Can’t Be Optional When flawed decisions travel through widely used APIs, the harm multiplies. From biased credit scoring to opaque hiring filters, the consequences can be costly legally, reputationally, and ethically. With regulations like the EU AI Act on the rise, building responsible APIs is becoming a competitive advantage. 🛠️ The IGT-AI Framework To help structure this challenge, I recommend the IGT (Ikenna Governance Toolkit)-AI framework. It provides a blueprint for embedding ethical guardrails through: 1️⃣ Unified API & Access – Secure, centralized authentication and credentials. 2️⃣ Guardrails – Input/output filtering to stop harmful or non-compliant content. 3️⃣ Observability – Monitoring fairness, drift, and usage patterns. 4️⃣ GitOps for Governance – Making every configuration auditable and version-controlled. 🔑 Key Insights from My Talk ▶️ Transparency APIs should expose model version, confidence scores, and limitations. ▶️ Fairness Deployments should pass bias tests, with fairness metrics exposed in real time. ▶️ Accountability Clear ownership and tamper-evident logs ensure responsibility. ▶️ Privacy & Security Apply data minimization, encryption, and consent tracking at the API layer. ▶️ AI Gateways The central control point for implementing and enforcing all of the above. 💬 I’d love to hear your thoughts: How are you approaching responsibility in AI-powered APIs? What challenges do you see in balancing innovation with governance? I'm reading your comments! ✍ ✍ ✍

94

9 Kommentare

I believe you have already read "The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity," a paper released in June 2025 by Apple researchers under the spot because it enforced the idea that advanced AI systems known as Large Reasoning Models have a "complete accuracy collapse" when faced with high-complexity tasks and even reduce their reasoning effort past a certain difficulty threshold, challenging the wonderful tale that these models truly "think." 🛜 Source: https://shorturl.at/bTzwt So... where is the news? It seems we already forget about papers such as: - Stochastic Parrots and Pattern Matching Back in 2021, Emily Bender and her colleagues introduced the term "stochastic parrots" to describe LLMs. Why? These models excel at identifying statistical patterns within their extensive training datasets but they lack of a comprehension and understanding of the content they generate. The study from Apple corroborates this, demonstrating that LLMs perform optimally within their established pattern-matching capabilities but fail when tasks require a real understanding beyond rote memorization. - AI Skepticism by Gary Marcus Gary Marcus has been skeptic of deep learning's potential for years, thinking that it lacks the symbolic, systematic scaffolding needed for a true general intelligence: human reasoning relies on structured frameworks (think logic, rules, and abstract concept) while deep learning leans heavily on statistical correlations. The Apple study reinforces Marcus's point: when LLMs face complex problems requiring systematic planning, their lack of a deeper framework leaves them floundering. - Synthetic Reasoning Datasets In 2022, Abulhair Saparov and He He launched PrOntoQA, a synthetic dataset designed to test logical reasoning in LLMs. They found that while these models can handle individual deduction steps (e.g., “If A, then B”), they fail totally at proof planning figuring out which path to take among multiple valid options. Apple's research builds on this, and shows that LLMs not only struggle with planning but also give up when the going gets tough, reducing their effort instead of rising to the challenge. This might feel like déjà vu because, in many ways, it is but Apple's paper gets to this basis to add a new perspective: the problem of complexity. - Accuracy collapse: LRMs in the high-complexity scenarios have performance dropping to near-zero accuracy. - Effort reduction: Unlike humans, who might double down when faced with a tough problem, LRMs slack off. - The illusion exposed: The paper's title says it all. What we perceive as “thinking” is really a façade of fluency, built on pattern recognition rather than genuine cognitive processes. Ethically, it is misleading to say "reasoning" or "thinking" as it creates a cascade of problems discussed in the 1st comments. #ArtificialIntelligence #Illusion #AI #Reasoning #LLMs #Innovation #Apple

40

11 Kommentare

Attention Sinks and Compression Valleys in LLMs Are Two Sides of the Same Coin ... Why Do Language Models Waste Attention on Meaningless Tokens? Two puzzling behaviors have mystified AI researchers: language models inexplicably focus attention on semantically empty tokens, and their representations mysteriously compress in middle layers before expanding again. These seemed like separate quirks—until now. 👉 The Hidden Connection New research reveals these aren't separate phenomena but two sides of the same coin. Both stem from "massive activations"—features that spike to extreme magnitudes (10³-10⁶ times normal) on specific tokens like the beginning-of-sequence marker. The math is elegant: when one token's representation becomes massive, it mathematically forces the entire representation matrix to compress into a low-dimensional space. This compression creates the attention sinks as a byproduct. 👉 Three Phases of Processing This discovery unveils how transformers organize computation across their depth: Phase 1 (0-20% depth): Mix Early layers perform broad information mixing through diffuse attention patterns, building contextual understanding. Phase 2 (20-85% depth): Compress Massive activations emerge, forcing representations into compressed states while attention collapses onto sink tokens. This halts mixing but preserves essential information in fewer dimensions. Phase 3 (85-100% depth): Refine Token norms equalize, decompressing representations. Attention shifts to sharp positional patterns for final refinements. 👉 Why This Matters Different tasks peak at different phases. Embedding tasks (classification, retrieval) perform best during compression when features concentrate in low dimensions. Generation tasks need the full pipeline, requiring Phase 3's refinements for accurate next-token prediction. This explains why studies reach conflicting conclusions about "optimal" layers—they're measuring fundamentally different computational objectives. The work provides the first unified theory connecting attention patterns, representation geometry, and task performance in language models. Paper: "Attention Sinks and Compression Valleys in LLMs Are Two Sides of the Same Coin" by Queipo-de-Llano et al.

40

1 Kommentar