Épisodes

  • ELO Ratings Questions

    ELO Ratings Questions
    Sep 18 2025
    Key Argument
    • Thesis: Using ELO for AI agent evaluation = measuring noise
    • Problem: Wrong evaluators, wrong metrics, wrong assumptions
    • Solution: Quantitative assessment frameworks
    The Comparison (00:00-02:00)

    Chess ELO

    • FIDE arbiters: 120hr training
    • Binary outcome: win/loss
    • Test-retest: r=0.95
    • Cohen's κ=0.92

    AI Agent ELO

    • Random users: Google engineer? CS student? 10-year-old?
    • Undefined dimensions: accuracy? style? speed?
    • Test-retest: r=0.31 (coin flip)
    • Cohen's κ=0.42
    Cognitive Bias Cascade (02:00-03:30)
    • Anchoring: 34% rating variance in first 3 seconds
    • Confirmation: 78% selective attention to preferred features
    • Dunning-Kruger: d=1.24 effect size
    • Result: Circular preferences (A>B>C>A)
    The Quantitative Alternative (03:30-05:00)

    Objective Metrics

    • McCabe complexity ≤20
    • Test coverage ≥80%
    • Big O notation comparison
    • Self-admitted technical debt
    • Reliability: r=0.91 vs r=0.42
    • Effect size: d=2.18
    Dream Scenario vs Reality (05:00-06:00)

    Dream

    • World's best engineers
    • Annotated metrics
    • Standardized criteria

    Reality

    • Random internet users
    • No expertise verification
    • Subjective preferences
    Key StatisticsMetricChessAI AgentsInter-rater reliabilityκ=0.92κ=0.42Test-retestr=0.95r=0.31Temporal drift±10 pts±150 ptsHurst exponent0.890.31Takeaways
    1. Stop: Using preference votes as quality metrics
    2. Start: Automated complexity analysis
    3. ROI: 4.7 months to break even
    Citations Mentioned
    • Kapoor et al. (2025): "AI agents that matter" - κ=0.42 finding
    • Santos et al. (2022): Technical Debt Grading validation
    • Regan & Haworth (2011): Chess arbiter reliability κ=0.92
    • Chapman & Johnson (2002): 34% anchoring effect
    Quotable Moments

    "You can't rate chess with basketball fans"

    "0.31 reliability? That's a coin flip with extra steps"

    "Every preference vote is a data crime"

    "The psychometrics are screaming"

    Resources
    • Technical Debt Grading (TDG) Framework
    • PMAT (Pragmatic AI Labs MCP Agent Toolkit)
    • McCabe Complexity Calculator
    • Cohen's Kappa Calculator

    🔥 Hot Course Offers:
    • 🤖 Master GenAI Engineering - Build Production AI Systems
    • 🦀 Learn Professional Rust - Industry-Grade Development
    • 📊 AWS AI & Analytics - Scale Your ML in Cloud
    • ⚡ Production GenAI on AWS - Deploy at Enterprise Scale
    • 🛠️ Rust DevOps Mastery - Automate Everything
    🚀 Level Up Your Career:
    • 💼 Production ML Program - Complete MLOps & Cloud Mastery
    • 🎯 Start Learning Now - Fast-Track Your ML Career
    • 🏢 Trusted by Fortune 500 Teams

    Learn end-to-end ML engineering from industry veterans at PAIML.COM
    Voir plus Voir moins
    4 min
  • The 2X Ceiling: Why 100 AI Agents Can't Outcode Amdahl's Law"

    The 2X Ceiling: Why 100 AI Agents Can't Outcode Amdahl's Law"
    Sep 17 2025
    AI coding agents face the same fundamental limitation as parallel computing: Amdahl's Law. Just as 10 cooks can't make soup 10x faster, 10 AI agents can't code 10x faster due to inherent sequential bottlenecks.📚 Key ConceptsThe Soup AnalogyMultiple cooks can divide tasks (prep, boiling water, etc.)But certain steps MUST be sequential (can't stir before ingredients are in)Adding more cooks hits diminishing returns quicklyPerfect metaphor for parallel processing limitsAmdahl's Law ExplainedMathematical principle: Speedup = 1 / (Sequential% + Parallel%/N)Logarithmic relationship = rapid plateauSequential work becomes the hard ceilingEven infinite workers can't overcome sequential bottlenecks💻 Traditional Computing BottlenecksI/O Operations - disk reads/writesNetwork calls - API requests, database queries Database locks - transaction serializationCPU waiting - can't parallelize waitingResult: 16 cores ≠ 16x speedup in real world🤖 Agentic Coding Reality: The New Bottlenecks1. Human Review (The New I/O)Code must be understood by humansSecurity validation requiredBusiness logic verificationCan't parallelize human cognition2. Production DeploymentSequential by natureOne deployment at a timeRollback requirementsCompliance checks3. Trust BuildingCan't parallelize reputationBad code = deleted customer dataRevenue impact risksTrust accumulates sequentially4. Context LimitsHuman cognitive bandwidthUnderstanding 100k+ lines of codeMental model limitationsCommunication overhead📊 The Numbers (Theoretical Speedups)1 agent: 1.0x (baseline)2 agents: ~1.3x speedup10 agents: ~1.8x speedup 100 agents: ~1.96x speedup∞ agents: ~2.0x speedup (theoretical maximum)🔑 Key TakeawaysAI Won't Fully Automate Coding JobsMore like enhanced assistants than replacementsHuman oversight remains criticalTrust and context are irreplaceableEfficiency Gains Are LimitedReal-world ceiling around 2x improvementNot the exponential gains often promisedSimilar to other parallelization effortsSuccess Factors for Agentic CodingWell-organized human-in-the-loop processesClear review and approval workflowsIncremental trust buildingRealistic expectations🔬 Research ReferencesPrinceton AI research on agent limitations"AI Agents That Matter" paper findingsEmpirical evidence of diminishing returnsReal-world case studies💡 Practical ImplicationsFor Developers:Focus on optimizing the human review processBuild better UI/UX for code reviewImplement incremental deployment strategiesFor Organizations:Set realistic productivity expectationsInvest in human-agent collaboration toolsDon't expect 10x improvements from more agentsFor the Industry:Paradigm shift from "replacement" to "augmentation"Need for new metrics beyond raw speedFocus on quality over quantity of agents🎬 Episode StructureHook: The soup cooking analogyTheory: Amdahl's Law explanationTraditional: Computing bottlenecksModern: Agentic coding bottlenecksReality Check: The 2x ceilingFuture: Optimizing within constraints🗣️ Quotable Moments"10 agents don't code 10 times faster, just like 10 cooks don't make soup 10 times faster""Humans are the new I/O bottleneck""You can't parallelize trust""The theoretical max is 2x faster - that's the reality check"🤔 Discussion QuestionsIs the 2x ceiling permanent or can we innovate around it?What's more valuable: speed or code quality?How do we optimize the human bottleneck?Will future AI models change these limitations?📝 Episode Tagline"When infinite AI agents hit the wall of human review, Amdahl's Law reminds us that some things just can't be parallelized - including trust, context, and the courage to deploy to production." 🔥 Hot Course Offers:🤖 Master GenAI Engineering - Build Production AI Systems🦀 Learn Professional Rust - Industry-Grade Development📊 AWS AI & Analytics - Scale Your ML in Cloud⚡ Production GenAI on AWS - Deploy at Enterprise Scale🛠️ Rust DevOps Mastery - Automate Everything🚀 Level Up Your Career:💼 Production ML Program - Complete MLOps & Cloud Mastery🎯 Start Learning Now - Fast-Track Your ML Career🏢 Trusted by Fortune 500 TeamsLearn end-to-end ML engineering from industry veterans at PAIML.COM
    Voir plus Voir moins
    4 min
  • Plastic Shamans of AGI

    Plastic Shamans of AGI
    May 21 2025
    The plastic shamans of OpenAI 🔥 Hot Course Offers: - 🤖 Master GenAI Engineering - Build Production AI Systems - 🦀 Learn Professional Rust - Industry-Grade Development - 📊 AWS AI & Analytics - Scale Your ML in Cloud - ⚡ Production GenAI on AWS - Deploy at Enterprise Scale - 🛠️ Rust DevOps Mastery - Automate Everything 🚀 Level Up Your Career: - 💼 Production ML Program - Complete MLOps & Cloud Mastery - 🎯 Start Learning Now - Fast-Track Your ML Career - 🏢 Trusted by Fortune 500 Teams Learn end-to-end ML engineering from industry veterans at PAIML.COM
    Voir plus Voir moins
    11 min
  • The Toyota Way: Engineering Discipline in the Era of Dangerous Dilettantes

    The Toyota Way: Engineering Discipline in the Era of Dangerous Dilettantes
    May 21 2025
    Dangerous Dilettantes vs. Toyota Way EngineeringCore ThesisThe influx of AI-powered automation tools creates dangerous dilettantes - practitioners who know just enough to be harmful. The Toyota Production System (TPS) principles provide a battle-tested framework for integrating automation while maintaining engineering discipline.Historical ContextToyota Way formalized ~2001DevOps principles derive from TPSCoincided with post-dotcom crash startupsDecades of manufacturing automation parallels modern AI-based automationDangerous Dilettante IndicatorsPromises magical automation without understanding systemsFocuses on short-term productivity gains over long-term stabilityCreates interfaces that hide defects rather than surfacing themLacks understanding of production engineering fundamentalsPrioritizes feature velocity over deterministic behaviorToyota Way Implementation for AI-Enhanced Development1. Long-Term Philosophy Over Short-Term Gains// Anti-pattern: Brittle automation scriptlet quick_fix = agent.generate_solution(problem, { optimize_for: "immediate_completion", validation: false});// TPS approach: Sustainable system designlet sustainable_solution = engineering_system .with_agent_augmentation(agent) .design_solution(problem, { time_horizon_years: 2, observability: true, test_coverage_threshold: 0.85, validate_against_principles: true });Build systems that remain maintainable across yearsEstablish deterministic validation criteria before implementationOptimize for total cost of ownership, not just initial development2. Create Continuous Process Flow to Surface ProblemsImplement CI pipelines that surface defects immediately:Static analysis validationType checking (prefer strong type systems)Property-based testingIntegration testsPerformance regression detectionBuild flow:make lint → make typecheck → make test → make integration → make benchmarkFail fast at each stageForce errors to surface early rather than be hidden by automationAgent-assisted development must enhance visibility, not obscure it3. Pull Systems to Prevent OverproductionMinimize code surface area - only implement what's neededPrefer refactoring to adding new abstractionsUse agents to eliminate boilerplate, not to generate speculative features// Prefer minimal implementationsfunction processData(data: T[]): Result { // Use an agent to generate only the exact transformation needed // Not to create a general-purpose framework}4. Level Workload (Heijunka)Establish consistent development velocityAvoid burst patterns that hide technical debtUse agents consistently for small tasks rather than large sporadic generations5. Build Quality In (Jidoka)Automate failure detection, not just productionAny failed test/lint/check = full system haltEvery team member empowered to "pull the andon cord" (stop integration)AI-assisted code must pass same quality gates as human codeQuality gates should be more rigorous with automation, not less6. Standardized Tasks and ProcessesUniform build system interfaces across projectsConsistent command patterns:make formatmake lintmake testmake deployStandardized ways to integrate AI assistanceDocumented patterns for human verification of generated code7. Visual Controls to Expose ProblemsDashboards for code coverageComplexity metricsDependency trackingPerformance telemetryUse agents to improve these visualizations, not bypass them8. Reliable, Thoroughly-Tested TechnologyPrefer languages with strong safety guarantees (Rust, OCaml, TypeScript over JS)Use static analysis tools (clippy, eslint)Property-based testing over example-based#[test]fn property_based_validation() { proptest!(|(input: Vec)| { let result = process(&input); // Must hold for all inputs assert!(result.is_valid_state()); });}9. Grow Leaders Who Understand the WorkEngineers must understand what agents produceNo black-box implementationsLeaders establish a culture of comprehension, not just completion10. Develop Exceptional TeamsUse AI to amplify team capabilities, not replace expertiseAgents as team members with defined responsibilitiesCross-training to understand all parts of the system11. Respect Extended Network (Suppliers)Consistent interfaces between systemsWell-documented APIsVersion guaranteesExplicit dependencies12. Go and See (Genchi Genbutsu)Debug the actual system, not the abstractionTrace problematic code pathsVerify agent-generated code in contextSet up comprehensive observability// Instrument code to make the invisible visiblefunc ProcessRequest(ctx context.Context, req *Request) (*Response, error) { start := time.Now() defer metrics.RecordLatency("request_processing", time.Since(start)) // Log entry point logger.WithField("request_id", req.ID).Info("Starting request processing") // Processing with tracing points // ... // Verify exit conditions if err != nil { metrics.IncrementCounter("processing_errors", 1) logger.WithError(err).Error("Request processing failed") } return resp, err}13. Make Decisions Slowly by ConsensusMulti-stage ...
    Voir plus Voir moins
    15 min
  • DevOps Narrow AI Debunking Flowchart

    DevOps Narrow AI Debunking Flowchart
    May 16 2025
    Extensive Notes: The Truth About AI and Your Coding JobTypes of AI

    • Narrow AI

      • Not truly intelligent
      • Pattern matching and full text search
      • Examples: voice assistants, coding autocomplete
      • Useful but contains bugs
      • Multiple narrow AI solutions compound bugs
      • Get in, use it, get out quickly

    • AGI (Artificial General Intelligence)

      • No evidence we're close to achieving this
      • May not even be possible
      • Would require human-level intelligence
      • Needs consciousness to exist
      • Consciousness: ability to recognize what's happening in environment
      • No concept of this in narrow AI approaches
      • Pure fantasy and magical thinking

    • ASI (Artificial Super Intelligence)

      • Even more fantasy than AGI
      • No evidence at all it's possible
      • More science fiction than reality
    The DevOps Flowchart Test

    1. Can you explain what DevOps is?

      • If no → You're incompetent on this topic
      • If yes → Continue to next question

    2. Does your company use DevOps?

      • If no → You're inexperienced and a magical thinker
      • If yes → Continue to next question

    3. Why would you think narrow AI has any form of intelligence?

      • Anyone claiming AI will automate coding jobs while understanding DevOps is likely:
        • A magical thinker
        • Unaware of scientific process
        • A grifter
    Why DevOps Matters
    • Proven methodology similar to Toyota Way
    • Based on continuous improvement (Kaizen)
    • Look-and-see approach to reducing defects
    • Constantly improving build systems, testing, linting
    • No AI component other than basic statistical analysis
    • Feedback loop that makes systems better
    The Reality of Job Automation
    • People who do nothing might be eliminated
      • Not AI automating a job if they did nothing
    • Workers who create negative value
      • People who create bugs at 2AM
      • Their elimination isn't AI automation
    Measuring Software Quality
    • High churn files correlate with defects
    • Constant changes to same file indicate not knowing what you're doing
    • DevOps patterns help identify issues through:
      • Tracking file changes
      • Measuring complexity
      • Code coverage metrics
      • Deployment frequency
    Conclusion
    • Very early stages of combining narrow AI with DevOps
    • Narrow AI tools are useful but limited
    • Need to look beyond magical thinking
    • Opinions don't matter if you:
      • Don't understand DevOps
      • Don't use DevOps
      • Claim to understand DevOps but believe narrow AI will replace developers
    Raw Assessment
    • If you don't understand DevOps → Your opinion doesn't matter
    • If you understand DevOps but don't use it → Your opinion doesn't matter
    • If you understand and use DevOps but think AI will automate coding jobs → You're likely a magical thinker or grifter

    🔥 Hot Course Offers:
    • 🤖 Master GenAI Engineering - Build Production AI Systems
    • 🦀 Learn Professional Rust - Industry-Grade Development
    • 📊 AWS AI & Analytics - Scale Your ML in Cloud
    • ⚡ Production GenAI on AWS - Deploy at Enterprise Scale
    • 🛠️ Rust DevOps Mastery - Automate Everything
    🚀 Level Up Your Career:
    • 💼 Production ML Program - Complete MLOps & Cloud Mastery
    • 🎯 Start Learning Now - Fast-Track Your ML Career
    • 🏢 Trusted by Fortune 500 Teams

    Learn end-to-end ML engineering from industry veterans at PAIML.COM
    Voir plus Voir moins
    11 min
  • No Dummy, AI Isn't Replacing Developer Jobs

    No Dummy, AI Isn't Replacing Developer Jobs
    May 14 2025
    Extensive Notes: "No Dummy: AI Will Not Replace Coders"Introduction: The Critical Thinking ProblemAmerica faces a critical thinking deficit, especially evident in narratives about AI automating developers' jobsSpeaker advocates for examining the narrative with core critical thinking skillsSuggests substituting the dominant narrative with alternative explanationsAlternative Explanation 1: Non-Productive EmployeesOrganizations contain people who do "absolutely nothing"If you fire a person who does no work, there will be no impactThese non-productive roles exist in academics, management, and technical industriesReference to David Graeber's book "Bullshit Jobs" which categorizes meaningless jobs:Task mastersBox tickersGoonsWhen these jobs are eliminated, AI didn't replace them because "the job didn't need to exist"Alternative Explanation 2: Low-Skilled DevelopersSome developers have "very low or no skills, even negative skills"Firing someone who writes "buggy code" and replacing them with a more productive developer (even one using auto-completion tools) isn't AI replacing a jobThese developers have "negative value to an organization"Removing such developers would improve the company regardless of automationUsing better tools, CI/CD, or software engineering best practices to compensate for their removal isn't AI replacementAlternative Explanation 3: Basic Automation with Traditional ToolsSoftware engineers have been automating tasks for decades without AISpeaker's example: At Disney Future Animation (2003), replaced manual weekend maintenance with bash scripts"A bash script is not AI. It has no form of intelligence. It's a for loop with some conditions in it."Many companies have poor processes that can be easily automated with basic scriptsThis automation has "absolutely nothing to do with AI" and has "been happening for the history of software engineering"Alternative Explanation 4: Narrow vs. General IntelligenceUseful applications of machine learning exist:Linear regressionK-means clusteringAutocompletionTranscriptionThese are "narrow components" with "zero intelligence"Each component does a specific task, not general intelligence"When someone says you automated a job with a large language model, what are you talking about? It doesn't make sense."LLMs are not intelligent; they're task-based systemsAlternative Explanation 5: OutsourcingCompanies commonly outsource jobs to lower-cost regionsJobs claimed to be "taken by AI" may have been outsourced to India, Mexico, or ChinaThis practice is common in America despite questionable ethicsOrganizations may falsely claim AI automation when they've simply outsourced workAlternative Explanation 6: Routine Corporate LayoffsLarge companies routinely fire ~3% of their workforce (Apple, Amazon mentioned)Fear is used as a motivational tool in "toxic American corporations"The "AI is coming for your job" narrative creates fear and motivationMore likely explanations: non-productive employees, low-skilled workers, simple automation, etc.The Marketing and Sales DeceptionCEOs (specifically mentions Anthropic and OpenAI) make false claims about agent capabilities"The CEO of a company like Anthropic... is a liar who said that software engineering jobs will be automated with agents"Speaker claims to have used these tools and found "they have no concept of intelligence"Sam Altman (OpenAI) characterized as "a known liar" who "exaggerates about everything"Marketing people with no software engineering background make claims about coding automationCompanies like NVIDIA promote AI hype to sell GPUsConclusion: The Real Problem"AI" is a misnomer for large language modelsThese are "narrow intelligence" or "narrow machine learning" systemsThey "do one task like autocomplete" and chain these tasks togetherThere is "no concept of intelligence embedded inside"The speaker sees a bigger issue: lack of critical thinking in AmericaWarns that LLMs are "dumb as a bag of rocks" but powerful toolsLeft in inexperienced hands, these tools could create "catastrophic software"Rejects the narrative that "AI will replace software engineers" as having "absolutely zero evidence"Key Quotes"We have a real problem with critical thinking in America. And one of the places that is very evident is this false narrative that's been spread about AI automating developers jobs.""If you fire a person that does no work, there will be no impact.""I have been automating people's jobs my entire life... That's what I've been doing with basic scripts. A bash script is not AI.""Large language models are not intelligent. How could they possibly be this mystical thing that's automating things?""By saying that AI is going to come for your job soon, it's a great false narrative to spread fear where people worry about all the AI is coming.""Much more likely the story of AI is that it is a very powerful tool that is dumb as a bag of rocks and left into the hands of the inexperienced and the naive and the fools could create ...
    Voir plus Voir moins
    15 min
  • The Narrow Truth: Dismantling IntelligenceTheater in Agent Architecture

    The Narrow Truth: Dismantling IntelligenceTheater in Agent Architecture
    May 14 2025

    how Gen.AI companies combine narrow ML components behind conversational interfaces to simulate intelligence. Each agent component (text generation, context management, tool integration) has direct non-ML equivalents. API access bypasses the deceptive UI layer, providing better determinism and utility. Optimal usage requires abandoning open-ended interactions for narrow, targeted prompting focused on pattern recognition tasks where these systems actually deliver value.

    🔥 Hot Course Offers:
    • 🤖 Master GenAI Engineering - Build Production AI Systems
    • 🦀 Learn Professional Rust - Industry-Grade Development
    • 📊 AWS AI & Analytics - Scale Your ML in Cloud
    • ⚡ Production GenAI on AWS - Deploy at Enterprise Scale
    • 🛠️ Rust DevOps Mastery - Automate Everything
    🚀 Level Up Your Career:
    • 💼 Production ML Program - Complete MLOps & Cloud Mastery
    • 🎯 Start Learning Now - Fast-Track Your ML Career
    • 🏢 Trusted by Fortune 500 Teams

    Learn end-to-end ML engineering from industry veterans at PAIML.COM
    Voir plus Voir moins
    11 min
  • The Pirate Bay Hypothesis: Reframing AI's True Nature

    The Pirate Bay Hypothesis: Reframing AI's True Nature
    May 14 2025
    Episode Summary:

    A critical examination of generative AI through the lens of a null hypothesis, comparing it to a sophisticated search engine over all intellectual property ever created, challenging our assumptions about its transformative nature.

    Keywords:

    AI demystification, null hypothesis, intellectual property, search engines, large language models, code generation, machine learning operations, technical debt, AI ethics

    Why This Matters to Your Organization:

    Understanding AI's true capabilities—beyond the hype—is crucial for making strategic technology decisions. Is your team building solutions based on AI's actual strengths or its perceived magic?

    Ready to deepen your understanding of AI's practical applications? Subscribe to our newsletter for more insights that cut through the tech noise: https://ds500.paiml.com/subscribe.html

    #AIReality #TechDemystified #DataScience #PragmaticAI #NullHypothesis

    🔥 Hot Course Offers:
    • 🤖 Master GenAI Engineering - Build Production AI Systems
    • 🦀 Learn Professional Rust - Industry-Grade Development
    • 📊 AWS AI & Analytics - Scale Your ML in Cloud
    • ⚡ Production GenAI on AWS - Deploy at Enterprise Scale
    • 🛠️ Rust DevOps Mastery - Automate Everything
    🚀 Level Up Your Career:
    • 💼 Production ML Program - Complete MLOps & Cloud Mastery
    • 🎯 Start Learning Now - Fast-Track Your ML Career
    • 🏢 Trusted by Fortune 500 Teams

    Learn end-to-end ML engineering from industry veterans at PAIML.COM
    Voir plus Voir moins
    9 min