OpenAI’s Sora 2, Anthropic’s Claude Sonnet 4.5, Google Research’s ReasoningBank, oLLM Snyk body { margin: 0; padding: 0; -webkit-text-size-adjust: 100% !important; -ms-text-size-adjust: 100% !important; -webkit-font-smoothing: antialiased !important; } img { border: 0 !important; outline: none !important; } p { Margin: 0px !important; Padding: 0px !important; } table { border-collapse: collapse; mso-table-lspace: 0px; mso-table-rspace: 0px; } td, a, span { border-collapse: collapse; mso-line-height-rule: exactly; } .buttontext { text-transform: inherit } .ExternalClass * { line-height: 100%; } .em_defaultlink a { color: inherit; text-decoration: none; } .em_footer a { color: #979797; text-decoration: underline; } .em_purple a { color: #8a2ac2 !important; text-decoration: underline !important; } .em_g_img+div { display: none; } a[x-apple-data-detectors], u+.em_body a, #MessageViewBody a { color: inherit; text-decoration: none; font-size: inherit !important; font-family: inherit !important; font-weight: inherit !important; line-height: inherit; } @media only screen and (max-width: 100%; } .em_wrapper { width: 100%; } .em_hide { display: none !important; } .em_full_img img { width: 100%; height: auto !important; max-width: 100%; } .em_center { text-align: center !important; } .em_side15 { width: 100%; } .em_ptop { padding-top: 20px !important; } .em_pbottom { padding-bottom: 20px !important; } .em_h20 { height: 20px !important; font-size: 1px !important; line-height: 1px !important; } .em_hauto { height: auto !important; } u+.em_body .em_full_wrap { width: 100%; width: 100%; } .em_pad { padding: 20px 15px !important; } .em_ptb { padding: 20px 0px 20px !important; } .em_pad1 { padding: 20px 15px 10px !important; } .em_pad2 { padding: 10px 15px 20px !important; } .em_ptb1 { padding: 30px 0px 20px !important; } .em_plrb { padding: 0px 15px 20px !important; } .em_h10 { height: 10px !important; line-height: 0px !important; font-size: 0px !important; } .em_wrap_50 { width: 100%; } } @media screen and (max-width: 100%; height: auto !important; } .em_img_1 img { width: 100%; height: auto !important; } .em_img_2 { width: 100%; height: auto !important; } .em_img_2 img { width: 100%; height: auto !important; } .em_img_3 { width: 100%; height: auto !important; } .em_img_3 img { width: 100%; height: auto !important; } .em_img_4 { width: 100%; height: auto !important; } .em_img_4 img { width: 100%; height: auto !important; } } The future of secure AI-driven development is here, and DevSecCon25 is leading the conversation! Join us on October 22, 2025 for this one-day event to hear from leading experts in AI and security from Qodo, Ragie.ai, Casco, Arcade.dev, and more! The full agenda includes: Mainstage - Hear inspiring keynotes from leaders in AI and cybersecurity. Expect forward-looking insights, industry thought leadership, and a vision of what’s next in the world of secure AI. AI demos track - Bring your laptop and join us for interactive, hands-on demos under the theme “Build and Secure with AI.” You'll leave with skills you can immediately apply. AI security track - Cutting-edge talks exploring the evolving security challenges of the AI era. Discover how to safeguard AI-driven applications, gain visibility into models, and secure agents across the SDLC. Snyk innovation track - Experience the latest advancements from Snyk in this dynamic track featuring live product demos, major announcements, and customer success stories. Don't miss this opportunity to gain the knowledge and strategies needed to embrace the AI revolution securely. Save your Spot Manage Preferences | Book a Demo| Contact Us| Community SponsoredSubscribe|Submit a tip|Advertise with UsWelcome to DataPro Expert Insights #152We’reexcited to bring you another packed edition full of deep dives, practical tutorials, andcutting-edgeupdates in Data & AI. This week,we’rethrilled to welcomeNishant Arora, Solutions Architect at AWS, to our newsletter portfolio, who will be sharing deep-dive insights on how AI is reshaping industries. In this issue, Nishant unpacks how trustworthy machine learning is being built into automotive and manufacturing systems, where safety, explainability, and regulatory readiness are mission critical.Butthat’sjust the beginning. This issue also rounds up some of the most important developments across the AI ecosystem:🔹OpenAI’s Sora 2brings the next generation of video and audio generation, blending realism, controllability, and creativity.🔹Anthropic’s Claude Sonnet 4.5sets new benchmarks in coding, reasoning, and agentic capabilities.🔹Google Research’s ReasoningBankshows how LLM agents can self-evolve by learning from both successes and failures.🔹oLLMdemonstrateshow 100K-context LLMs can now run on consumer GPUs with SSD offload.🔹Tutorials and explainers,from building advancedAgentic RAG pipelinesand experimenting withLangGraph workflows, to demystifying theGini Coefficient, preparing video data withvid-prepper, and writing your firstTriton GPU kernel.🔹And finally,a timelylook at howensemble-based fact-checking systemscan catch and neutralize repeating false claims before they spread.This edition is designed to give you bothstrategic insightsinto AI’s role in industry transformation andhands-on knowledgeto stay ahead in your technical work.Let’sdive in.Cheers,Merlyn ShelleyGrowth Lead, PacktTrustworthy Machine Learning in Automotive: Safety, Explainability, and Regulation Readiness– Written by Nishant Arora, Solutions Architect at AWSArtificial intelligence (AI) and machine learning (ML) are redefining the automotive industry. Cars are no longer just mechanical systems; they are intelligent, adaptive, and connected machines. Advanced driver-assistance systems (ADAS), predictive maintenance tools, and self-driving algorithms promise safer and more efficient transportation. Yet, the integration of ML also raises pressing concerns:can we guarantee these systems behave safely, explain their choices, andcomply withstrict automotive standards?Unlike recommendation systems or digital assistants, automotive MLoperatesinlife-critical environments. A single wrong decisionidentifyinga pedestrian, miscalculating braking distance, orfailing to detectsensor faults could have irreversible consequences. This is whytrustworthinessis not just a desirable property, but apreconditionfor adoption at scale.Safety as the Core of TrustIn safety-critical applications, evaluating ML performance goes beyond accuracy. What matters is whether the system preserves safe operation under all circumstances. A useful framing is:P (Safe|Model Decision)This probability expresses the likelihood that, given a model’s action, the outcome is safe. Accuracy alone does not guarantee that the rare but dangerous cases are adequately addressed.Equally important is theability to measure uncertainty. For example, an object recognition system in an autonomous car must know when it is unsure if a shadow is a pedestrian or just road texture. This can be modeled as predictive variance:Var(y∣x,θ)whereyis the outcome for inputxunder model parameters θ. Systems that quantify uncertainty allow safer fallback strategies such as driver takeover or conservative control.Safety can also be built directly into model training. A combinedobjectivefunction might looklike: L=Laccuracy+λ⋅LsafetywhereLaccuracyreflects predictive performance andLsafetypenalizes unsafe decisions, weighted by factorλ. In this way, the model learns not only to be correct, but also to respect predefined safety boundaries.Finally,confidence calibrationis vital. Regulators often require that predicted probabilities align with actual outcomes, ensuring that an ML model’s confidence is trustworthy:E[∣y^−y∣]≤εwhereεrepresentsthe maximum allowable deviation. Poor calibration can create dangerous overconfidence even when classification accuracy is high.Explainability: Building Human TrustEven a safe system will not be widely adopted if engineers, regulators, and customers cannot understand how it works. This is whereexplainable ML (XAI)becomes indispensable.Some prominent methods include:>> Feature attribution tools(e.g., SHAP, LIME) that show which sensor inputs or environmental factors most influenced a model’s decision.>> Surrogate models, such as simple decision trees approximating a deep neural network, whichmake the decision boundary more interpretable.>> Rule-based explanations, translating complex outputs into understandable logic:“if road is slippery and braking distance exceeds threshold, reduce speed.”Such techniques allow developers to debug failures, give regulators evidence for certification, and help buildpublic confidencein ML-driven cars.Regulation and Safety StandardsTraditional automotive safety is governed by standards likeISO 26262, which defines processes and Automotive Safety Integrity Levels (ASILs). These were designed for deterministic, rule-based software. ML, by contrast, is probabilistic and data-driven, creating new challenges for compliance.To bridge this gap, companies are adoptingverification and validation (V&V) frameworkstailored for ML. These include large-scale simulation testing, corner-case scenario generation, and monitoring model drift once systems are deployed. The aim is not just to test for accuracy, but to produceaudittrailsand evidence of robustness that regulators can certify.Looking ahead, standardswilllikely evolveto explicitly account for ML, requiring documentation of uncertainty estimates, explainability reports, and continuous monitoring logs.Emerging Pathways to Safer MLSeveral technological approaches show promise in making automotive ML more trustworthy:Cloud-NativeMLOpsCloud platforms now allow continuous retraining and redeployment of ML models asconditions shift (e.g., new road layouts or changing weather patterns). With automated testing pipelines, everynew versioncan be checked against safety and compliance metrics before deployment.Digital Twins and Safety-Constrained Reinforcement LearningDigital replicas of cars and environments enable billions of simulated test miles without real-world risk. Reinforcement learning agents can be trained with explicit safety constraints, ensuring that unsafe behaviors are never reinforced.Self-Monitoring Agentic AIFuture systems may integrate agentic AI that audits its own behavior in real-time. Such systems could flag potential regulatory violations, halt unsafe actions, or escalate control to human operators. Thisrepresentsa step toward vehicles thatself-enforce compliancerather than relying solely on external oversight.Conclusion: Toward a Trustworthy FutureAI in automotive promises safer roads, lower maintenance costs, and smarter mobility. But none of this progress matters unless these systems areprovablysafe, transparent, and regulationready.Automakers must embed safetyobjectivesdirectly into training and evaluation. Regulators must expand standards like ISO 26262 to incorporate probabilistic models. Cloud providers and technology partners must deliver the infrastructure for continuous monitoring and compliance assurance.The next era of mobility will not be defined merely by how advanced ML models become, but by how muchtrustsociety places in them. Only when AI systems are demonstrably safe, explainable, and aligned with regulatory frameworks will we see widespread adoption of truly autonomous and intelligent vehicles.References➖ ISO 26262:2018. Road Vehicles – Functional Safety.International Organization for Standardization.➖ Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete Problems in AI Safety.arXivpreprint arXiv:1606.06565.➖ Doshi-Velez, F., & Kim, B. (2017). Towards a rigorous science of interpretable machine learning.arXivpreprint arXiv:1702.08608.➖ Kendall, A., & Gal, Y. (2017). What uncertainties do we need in Bayesian deep learning for computer vision?Advances in Neural Information Processing Systems (NeurIPS).➖ Shapley, L. S. (1953). A value forn-person games.Contributions to the Theory of Games, 2(28), 307–317. (Basis for SHAP explainability methods).➖ National Highway Traffic Safety Administration (NHTSA). (2020). Automated Vehicles 4.0: Preparing for the Future of Transportation.U.S. Department of Transportation.Highlights in Data & AI⭕Sora 2 is here -OpenAI:OpenAI has launchedSora 2, a next-generation video and audio generation modelthat’smore physicallyaccurate, realistic, and controllable than its predecessor. It supports synchronized dialogue, sound effects, and cameo features that let users insert themselves into generated scenes. Released with a new iOS app, Sora 2 emphasizes creativity, social connection, and responsible usage.⭕Introducing Claude Sonnet 4.5 \ Anthropic:Anthropic has launched Claude Sonnet 4.5, its most powerful and aligned coding model yet. Excelling at reasoning, math, and real-world computer use, it powers complex agents with long focus spans. Alongside, Anthropic released upgrades to Claude Code, new app features, and the Claude Agent SDK. Available today via API, apps, and extensions, pricing matches Claude Sonnet 4.⭕Google AI Proposes ReasoningBank: A Strategy-Level I Agent Memory Framework that Makes LLM Agents Self-Evolve at Test Time.Google Research introducesReasoningBank, a memory framework that lets LLM agents learn from both successes and failures without retraining. By distilling interaction traces into reusable reasoning strategies, agents self-evolve across tasks. Paired with Memory-aware Test-time Scaling (MaTTS), the approach boosts effectiveness by up to 34% and reduces interaction steps by 16% on web and software-engineering benchmarks.⭕Meet oLLM: A Lightweight Python Library that brings 100K-Context LLM Inference to 8 GB Consumer GPUs via SSD Offload—No Quantization Required.oLLMis a lightweight Python library for running large-context Transformers on a single NVIDIA GPU by offloading weights and KV-cache to SSDs. Supporting models like Llama-3, GPT-OSS, and Qwen3-Next-80B, it avoids quantization, uses FP16/BF16 with FlashAttention-2, and enables up to 100K tokens on 8–10 GB VRAM. Designed for offline workloads, it trades throughput for practicality.⭕How to Build an Advanced Agentic Retrieval-Augmented Generation (RAG) System with Dynamic Strategy and Smart Retrieval?This tutorialdemonstratesan Agentic Retrieval-Augmented Generation (RAG) system that goes beyond basic document lookup. Using embeddings, FAISS, and a mock LLM, the agent decides when retrieval is needed, selects strategies (semantic, multi-query, temporal, hybrid), and synthesizes context-aware responses. The result is a more adaptive, transparent, and intelligent RAG pipeline for practical use.⭕Beyond ROC-AUC and KS: The Gini Coefficient, Explained Simply.This tutorial explains the Gini Coefficient as a classification metric alongside ROC-AUC and KS. Using the German Credit dataset, it walks through sorting predictions, plotting Lorenz curves, calculating areas, and deriving Gini. The result shows how Gini measures a model’s ability to rank positives over negatives, with higher valuesindicatingstronger separation and near-perfect classification.⭕How to Build Effective Agentic Systems with LangGraph?Amid rising agentic frameworks for powerful models (GPT-5, Gemini 2.5 Pro), this article introducesLangGraph, an agentic framework that abstracts state, tool-calling, and routing to build workflows. Using a document CRUD example, it shows graph-based intent routing (nodes, edges, state). Pros: easy setup, open-source, cleaner code. Cons: some boilerplate and framework-specific errors. Compared withLangChain/LlamaIndex/CrewAI,LangGraphbalances control and productivity.⭕Preparing Video Data for Deep Learning: Introducing Vid Prepper:This new product,vid-prepper, is an open-source Python package designed to make video preprocessing for machine learning and deep learning faster and more efficient. It provides tools for analyzing metadata, filtering out problematic files, standardizing formats/codecs/frame rates, detecting shots and objects, and converting videos into tensors. The goal is to reduce costs, avoid training bottlenecks, and simplify large-scale video data preparation.⭕Learning Triton One Kernel At a Time: VectorAddition.This tutorial introduces GPU programming with Triton by walking through vector addition as your first kernel. It explains GPU architecture basics,optimisationprinciples like reducing memory bandwidth costs and operatorfusion, andshows how Triton abstracts CUDA complexity. By writing a simple vector addition kernel, readers learn how to map work across threads, manage memory, and build efficient GPU code.⭕Building Fact-Checking Systems: Catching Repeating False Claims Before They Spread.This article explores how automated fact-checking systems can catch repeating false claims before they spread. It introduces previously fact-checked claim retrieval (PFCR), where claims are matched against existing verified ones, saving time and improving accuracy. Using retrieval–rerankerpipelines and ensemble methods that combine lexical and semantic models, the approach makes fact-checking faster, scalable, multilingual, and more reliable in today’s digital information ecosystem.See you next time!*{box-sizing:border-box}body{margin:0;padding:0}a[x-apple-data-detectors]{color:inherit!important;text-decoration:inherit!important}#MessageViewBody a{color:inherit;text-decoration:none}p{line-height:inherit}.desktop_hide,.desktop_hide table{mso-hide:all;display:none;max-height:0;overflow:hidden}.image_block img+div{display:none}sub,sup{font-size:75%;line-height:0} @media (max-width: 100%;display:block}.mobile_hide{min-height:0;max-height:0;max-width: 100%;overflow:hidden;font-size:0}.desktop_hide,.desktop_hide table{display:table!important;max-height:none!important}} body { margin: 0; padding: 0; -webkit-text-size-adjust: 100% !important; -ms-text-size-adjust: 100% !important; -webkit-font-smoothing: antialiased !important; } img { border: 0 !important; outline: none !important; } p { Margin: 0px !important; Padding: 0px !important; } table { border-collapse: collapse; mso-table-lspace: 0px; mso-table-rspace: 0px; } td, a, span { border-collapse: collapse; mso-line-height-rule: exactly; } .buttontext { text-transform: inherit } .ExternalClass * { line-height: 100%; } .em_defaultlink a { color: inherit; text-decoration: none; } .em_footer a { color: #979797; text-decoration: underline; } .em_purple a { color: #8a2ac2 !important; text-decoration: underline !important; } .em_g_img+div { display: none; } a[x-apple-data-detectors], u+.em_body a, #MessageViewBody a { color: inherit; text-decoration: none; font-size: inherit !important; font-family: inherit !important; font-weight: inherit !important; line-height: inherit; } @media only screen and (max-width: 100%; } .em_wrapper { width: 100%; } .em_hide { display: none !important; } .em_full_img img { width: 100%; height: auto !important; max-width: 100%; } .em_center { text-align: center !important; } .em_side15 { width: 100%; } .em_ptop { padding-top: 20px !important; } .em_pbottom { padding-bottom: 20px !important; } .em_h20 { height: 20px !important; font-size: 1px !important; line-height: 1px !important; } .em_hauto { height: auto !important; } u+.em_body .em_full_wrap { width: 100%; width: 100%; } .em_pad { padding: 20px 15px !important; } .em_ptb { padding: 20px 0px 20px !important; } .em_pad1 { padding: 20px 15px 10px !important; } .em_pad2 { padding: 10px 15px 20px !important; } .em_ptb1 { padding: 30px 0px 20px !important; } .em_plrb { padding: 0px 15px 20px !important; } .em_h10 { height: 10px !important; line-height: 0px !important; font-size: 0px !important; } .em_wrap_50 { width: 100%; } } @media screen and (max-width: 100%; height: auto !important; } .em_img_1 img { width: 100%; height: auto !important; } .em_img_2 { width: 100%; height: auto !important; } .em_img_2 img { width: 100%; height: auto !important; } .em_img_3 { width: 100%; height: auto !important; } .em_img_3 img { width: 100%; height: auto !important; } .em_img_4 { width: 100%; height: auto !important; } .em_img_4 img { width: 100%; height: auto !important; } }
Read more
United States
Great Britain
India
Germany
France
Canada
Russia
Spain
Brazil
Australia
Singapore
Canary Islands
Hungary
Ukraine
Luxembourg
Estonia
Lithuania
South Korea
Turkey
Switzerland
Colombia
Taiwan
Chile
Norway
Ecuador
Indonesia
New Zealand
Cyprus
Denmark
Finland
Poland
Malta
Czechia
Austria
Sweden
Italy
Egypt
Belgium
Portugal
Slovenia
Ireland
Romania
Greece
Argentina
Netherlands
Bulgaria
Latvia
South Africa
Malaysia
Japan
Slovakia
Philippines
Mexico
Thailand
