IBM Technology Video Summaries

We explore how code reviews have evolved from structured Fagan inspections to AI-driven outcome reviews, highlighting the shift from implementation-focused practices to validating business intent and impact.

- The fundamental building blocks: code, docs, and architecture
- Early review eras: Fagan inspections and pair programming
- The pull request and automation eras: consensus reviews and CI/CD checks
- The AI era: LLM-assisted analysis, human judgment, and evidence-based outcome reviews

This is ideal for developers, engineering leaders, and technology professionals seeking to understand how AI is transforming development workflows and what it means for the future of code review.

Pre-trained LLMs freeze their knowledge at release, but the world keeps changing, leading to confident hallucinations and costly mistakes in AI agents. This video explains how connecting an LLM to a live knowledge layer via a web data infrastructure layer solves that problem.

■ The Problem with Pre-trained Models
- Knowledge stops growing after release; real-world events, prices, and products are unknown
- Hallucinated answers can mislead humans and cause AI agents to fail at scale

■ The Solution: Knowledge Layer and Web Data Infrastructure
- Fresh, reliable, detailed web data is delivered at inference time
- Five key elements: grounded, deep, fresh, formatted, and timely data

■ Real-World Challenges and Best Practices
- Direct web scraping is blocked by CAPTCHAs and anti-bot technologies; raw HTML is inefficient for LLMs
- E-commerce data decays quickly; prices and stock should be refreshed at inference time

■ Summary
- Reliable AI systems depend on a trustworthy data layer, not just a larger model

This video is for developers, AI engineers, and anyone building AI agents that need real-time, accurate information. Viewers will learn why garbage-in-garbage-out happens and how to design a web data layer that keeps AI answers grounded and current.

In this episode, we cover three major stories shaping the AI landscape: Nvidia's acquisitions of Hugging Face and Poolside, IBM's new dual-processor mainframe chip developed with Arm, and the reveal of the mystery model Ox Alpha as GLM-5.3. The discussion explores what these developments mean for open-source AI, enterprise infrastructure, and model competition.

■ Nvidia's Open-Source AI Play
- Acquisition of Hugging Face for $12.9 billion and a $6 billion deal with Poolside
- Analysis of Nvidia's bet on an open model ecosystem and its hardware-centric incentives

■ IBM's Dual-Processor Mainframe
- Combining IBM Z and Arm instruction sets on a single chip
- Bringing AI inference closer to mainframe data and the implications for reliability

■ Mystery Model Ox Alpha
- Revealed as GLM-5.3, an open-source Chinese model with high throughput claims
- The stealth launch strategy and its effect on model visibility

This episode is for AI practitioners, enterprise architects, and technology strategists seeking a clearer understanding of the evolving AI infrastructure stack and the forces driving consolidation and innovation.

Leaderboard scores alone don't guarantee real-world AI performance. This breakdown explains the gap between benchmark results and production behavior, covering the trade-offs between accuracy, performance, and cost, and introducing the two main types of benchmarking: model evaluation and system evaluation.

■ Model Evaluation
- Standardized benchmarks like MMLU and execution-based coding tests, plus LLM-as-a-judge for open-ended responses

■ System Evaluation
- Key metrics such as time to first token, latency, and throughput, and why workload shape matters

■ Agent Evaluation
- Evaluating each step in the agentic chain, from intent understanding to tool calls and safety checks

■ Key Takeaway
- Real benchmarking requires testing with your own data, traffic patterns, and success criteria

Viewers building chatbots, RAG systems, coding assistants, or agents will learn how to set up meaningful benchmarks and avoid common evaluation mistakes.

This episode of Security Intelligence examines the rapid advancement of open-weight AI models in cybersecurity, a novel defensive use of prompt injection, and a targeted social engineering campaign against security conference attendees. The panel discusses the implications of GLM-5.3's accelerated cyber capabilities, the potential of context bombing as a defense mechanism, and the lessons from a ClickFix campaign aimed at Black Hat and DEF CON attendees.

■ GLM-5.3 and Open-Weight AI Models
- Cyber capabilities developing faster than expected through post-training
- Panel discussion on offensive vs. defensive AI progress and the need for automated patching

■ Context Bombing as a Defense
- Tracebit's use of prompt injections to deter malicious AI agents
- Honeypot-like approach, its limitations, and the importance of defense in depth

■ Post-Black Hat Social Engineering
- ClickFix campaign targeting conference attendees with fake encryption keys
- Human vulnerability, the risk to cybersecurity professionals, and the need for strong security hygiene

This episode is for cybersecurity professionals and organizations seeking to understand emerging AI threats and defensive strategies. Viewers will gain insights into balancing AI development, adopting layered defenses, and maintaining vigilance against social engineering attacks.

We break down the difference between an AI model and the agentic harness, and show why the harness often determines how capable an AI agent really is. You'll learn the three core components that make up a harness: tools, memory, and the agentic loop.

■ Core concepts
- AI model: the neural network, a "brain in a jar"
- Agentic harness: the wrapper that gives the model tools, memory, and a loop

■ Inside the harness
- Tools: file access, code execution, web browsing, computer use, MCP
- Memory: context window, instruction files, compaction, search
- Agentic loop: plan, act, observe, and continuous verification

■ Why it matters
- Most recent capability gains come from better harnesses
- The same model can perform differently depending on the harness

This is for developers, AI enthusiasts, and anyone evaluating AI products. After watching, you'll be able to distinguish model performance from harness design and ask more precise questions about what AI can do.

This video explores why AI coding tools often produce changes that are technically correct but architecturally harmful, and what capabilities they need to truly respect an existing codebase. It argues that speed without understanding can lead to fast chaos in production systems.

■ Repo Awareness
- Finding the right context without overwhelming the model
- Recognizing existing utilities and patterns

■ Architectural Context
- Respecting service layers and system rules
- Avoiding duplicate helpers and unnecessary dependencies

■ Planning, Verification, and Boundaries
- Making reasoning visible before patching
- Running checks and asking approval for risky changes

Developers and engineering teams will learn what to look for in AI coding assistants and how to guide them toward safer, more maintainable contributions.

We explore the common 'what versus why' problem in data retrieval, illustrated by a pharmacy prescription denial where the SQL database says 'no' but the reason is buried in a PDF. We then outline a six-step workflow for an AI agent that acts as a digital librarian, knowing when to query SQL and when to perform semantic search, and finally stitching the results into a single grounded answer.

■ The What vs. Why Problem
- Pharmacy denial example: SQL holds the coverage status, PDFs hold the policy rationale
- A common issue across industries with fragmented data

■ The Digital Librarian Agent Workflow
- Six steps: read the question, identify what/why, build queries, execute, compile, answer
- LLMs handle reasoning; tools handle database queries and data reformatting

■ From Data Retrieval to Reasoning
- Connecting SQL tables to PDF policy documents
- Turning siloed data repositories into a reasoning engine

Developers, data professionals, and AI practitioners will gain a practical framework for designing agentic workflows that combine structured and unstructured data. By the end, you will understand how to approach similar integration challenges and move from simple data retrieval to delivering complete answers.

This episode of Mixture of Experts covers major AI industry developments, including IBM's new partnership with OpenAI, Stripe's acquisition of OpenRouter, and insights from Ramp's AI index on enterprise spending. The panel discusses how these shifts signal a move toward AI orchestration, routing, and cost governance rather than model ownership.

■ IBM-OpenAI Partnership
- Enterprise AI orchestration and multi-model strategy
- IBM as a neutral enterprise AI integrator

■ Stripe Acquires OpenRouter
- $7B deal and the rise of AI gateways
- Routing as the new economic layer

■ Ramp AI Index and AI Legislation
- Uneven AI adoption and the shift to FinOps
- AI-generated bills in Congress and the need for governance

This episode is valuable for technologists, business leaders, and AI practitioners seeking to understand market trends, cost management, and the evolving role of consulting and infrastructure in enterprise AI.