Épisodes

  • Claude's Token Costs, Moonshot AI's $2B Raise & Agentic Design Patterns

    Claude's Token Costs, Moonshot AI's $2B Raise & Agentic Design Patterns
    May 8 2026
    (00:00:00) Claude's Token Costs, Moonshot AI's $2B Raise & Agentic Design Patterns
    (00:00:48) Token Budget Discipline for Developers
    (00:01:39) Moonshot AI's $2B Signal
    (00:02:44) Agentic Design Patterns in Production
    (00:03:39) What to Watch Next

    Your Claude Pro bill isn't growing because you're doing something wrong — it's growing because large context windows reward heavy use, and most teams haven't built the cost discipline to match. In this episode, we break down exactly why token budgets spiral inside 200K-context workflows, and what engineering-level fixes actually keep costs flat without sacrificing capability.

    We also unpack Moonshot AI's $2 billion raise at a $20 billion valuation. Their Kimi K2.6 model is now the second-most used LLM on OpenRouter, with annualised revenue topping $200M as of April. The signal isn't that Kimi is definitively better than Anthropic or OpenAI — it's that it's close enough, and cheap enough, that the tradeoff calculus has genuinely shifted for inference-cost-conscious builders.

    Finally, we look at what's emerging at the architecture layer. The agency-agents framework is trending on GitHub, and the design pattern it surfaces — structured specialist personas, explicit handoffs, validation checkpoints — reflects how serious production agent systems are actually being built. Not more capable chatbots. Choreographed teams.

    The through-line: larger models, larger contexts, and more capable agentic systems all create more surface area for cost and complexity to grow invisibly. The teams winning right now are treating token budgets, infrastructure choices, and agent architecture as first-class engineering decisions.

    For working developers and engineering leaders who want signal, not noise.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    5 min
  • Cursor's Context Breakdown, Claude's Limits & Agent Governance

    Cursor's Context Breakdown, Claude's Limits & Agent Governance
    May 7 2026
    (00:00:00) Cursor's Context Breakdown, Claude's Limits & Agent Governance
    (00:00:33) Cursor's Context Breakdown Is the Real Lever
    (00:01:28) Netskope's Embedded Agents Solve a Different Problem
    (00:02:10) Collibra and the Governance Gap
    (00:02:48) What to Watch Next

    This episode cuts through the noise on three developments that matter for developers running agents in production right now.

    Anthropic doubled usage limits for Claude Pro, Max, and Enterprise — ten hours of agent sessions instead of five, no peak-hour throttling, expanded Opus API capacity. It's a real capacity unlock, but capacity was rarely the binding constraint. Context is. Which is exactly why Cursor 3.3's new context usage breakdown deserves more attention than the Claude announcement. For the first time, developers can see line-by-line how much context their rules, skills, MCP integrations, and subagents are consuming. Bloated context is where agent performance collapses quietly — slower responses, higher costs, harder-to-trace errors. Doubling session time without addressing context hygiene just means running the same inefficiency longer.

    On the enterprise security side, Netskope's AgentSkope embeds six AI agents directly into their SASE platform — SOC triage, DLP, insider threat, access audits — with no data leaving the platform for external inference. The architectural constraint is the product. Forty percent of security alerts go uninvestigated not from lack of intelligence but lack of hands, and embedded agents answer that without adding latency or compliance exposure.

    Collibra's AI Command Center enters the governance layer with lifecycle tracking, ownership records, behaviour monitoring, and testing templates for regulated deployments. The catch: governance tooling assumes operational discipline already exists. For teams still in pilot mode, it's aspirational.

    The through-line: we're past the era of raw model availability as the gating factor. Instrumentation, ownership, and context discipline are the new constraints.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    4 min
  • Hackers Can't Use AI Tools — And What That Means for Your Team

    Hackers Can't Use AI Tools — And What That Means for Your Team
    May 6 2026
    (00:00:00) Hackers Can't Use AI Tools — And What That Means for Your Team
    (00:00:30) The Skill Floor Problem
    (00:01:13) Guardrails Holding on Mainstream Platforms
    (00:02:08) Pentagon AI Vendor Consolidation
    (00:02:58) What Developers Should Take From This

    A landmark study from the University of Edinburgh analysed over 100 million posts from underground cybercrime forums and returned a finding that cuts against the loudest fears in security: criminals can't get AI coding tools to work for them. Not because of ethics guardrails alone — but because AI is a capability multiplier, not a capability equaliser. Without a skill floor, the output is noise attackers can't evaluate or debug. This episode unpacks what that means for developers and engineering leaders thinking about productivity, competency gaps, and how their teams actually benefit from AI co-pilots.

    On the guardrails front, Claude, Codex, and similar mainstream platforms are proving more resistant to jailbreak attempts than many predicted. Attackers falling back on WormGPT and jailbroken alternatives are finding them resource-intensive and noticeably worse. Model-level restrictions are functioning — for now. AI-assisted crime is gaining ground only in low-skill, high-volume vectors: bots, romance scams, SEO fraud. Complex attack chains remain largely unaffected.

    The structural story: the Pentagon has awarded AI contracts for classified military networks to seven vendors — Google, Microsoft, AWS, Nvidia, OpenAI, Reflection, and SpaceX. Anthropic is not on the list, following a public dispute over AI ethics positioning. Vendor positioning on defence contracts is now an active policy decision, not a procurement formality. For developers building enterprise AI systems, understanding where the major platforms sit on government contracts matters more than ever.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    5 min
  • Pentagon's AI Vendor List: What Anthropic's Exclusion Signals for Enterprise

    Pentagon's AI Vendor List: What Anthropic's Exclusion Signals for Enterprise
    May 5 2026
    (00:00:00) Pentagon's AI Vendor List: What Anthropic's Exclusion Signals for Enterprise
    (00:00:22) Anthropic Excluded After Contract Dispute
    (00:00:59) GenAI.mil Now Operational
    (00:01:46) Automation Bias as Operational Risk
    (00:02:29) Vendor Lock-In and Enterprise Parallels
    (00:02:54) What Developers Should Watch Next

    The Pentagon just made its classified AI contractor list public, and the seven companies on it — Google, Microsoft, AWS, Nvidia, OpenAI, Reflection, and SpaceX — tell a governance story that matters well beyond national security contexts. Anthropic's absence is the headline: the company walked away after the Pentagon declined contractual protections against autonomous weapons and surveillance of US citizens. OpenAI now fills the classified role Claude would have occupied.

    This isn't a capability or benchmark story. It's a procurement and governance story. For developers and engineering leaders, that distinction is critical. Safety boundaries don't live only in model cards and responsible-use policies — in high-stakes deployments, they become contract terms. And contract terms can remove you from the table entirely.

    The episode also covers GenAI.mil, now operational and compressing months-long military workflows into days — a productivity pattern that should feel familiar to any team that has shipped an internal AI tool. What's different is the operational stakes. The contracts include human-in-the-loop language, but the practical detail of override mechanisms and decision thresholds remains thin.

    The deeper risk flagged here is automation bias: the well-documented tendency for human operators to defer to AI recommendations under time pressure, regardless of what the contract says. Georgetown's CSIS has flagged this specifically for battlefield contexts. The lesson transfers directly to enterprise: human oversight clauses are a governance floor, not a solution.

    Finally, with Anthropic out, OpenAI holds the dominant position in classified military AI. That vendor concentration dynamic is one every team building on a single model provider should be watching closely.

    A YesWee production.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    4 min
  • Big Tech Cuts Junior AI Roles — Startups Move the Other Way

    Big Tech Cuts Junior AI Roles — Startups Move the Other Way
    May 4 2026
    (00:00:00) Big Tech Cuts Junior AI Roles — Startups Move the Other Way
    (00:00:13) Story One — Big Tech Cuts Entry-Level AI Roles
    (00:01:06) Story Two — Small Teams Moving the Other Direction
    (00:01:46) Story Three — Model Selection Gets Practical
    (00:02:33) Close

    The entry-level AI engineering market just split in two, and if you're hiring or job-hunting, the implications are immediate. Large tech companies have quietly stopped backfilling junior AI roles — agentic tooling now handles the code review, boilerplate generation, and debugging passes that early-career engineers used to own. The on-ramp into big tech is shrinking fast.

    But the story doesn't end there. Smaller companies and startups are moving in the opposite direction, actively recruiting AI-native junior talent — developers already fluent in Cursor, comfortable building on Claude or Copilot, and thinking natively in agentic patterns. When your team is five people, that fluency is a genuine force multiplier.

    On the model side, the one-model-fits-all era is over. Production teams are now making model selection decisions based on workflow fit: cost versus context window, speed versus safety constraints. DeepSeek's low pricing and open weights have put visible pressure on premium vendors, and thin-wrapper businesses built on a single API are feeling the squeeze. Task-specific reliability is beating raw benchmark performance. And permissive open-source licensing has quietly become a competitive moat, not just a philosophical stance.

    This episode covers the structural hiring shift across big tech and startups, the practical framework engineering teams are using to choose models in 2024, and why open-source momentum is reshaping vendor purchasing decisions. No hype — just the signal that changes how you build and hire.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    3 min
  • Junior AI Hiring Cuts, Model Selection Shifts & Open-Source Moats

    Junior AI Hiring Cuts, Model Selection Shifts & Open-Source Moats
    May 3 2026
    The AI developer market is reorganising faster than most teams are tracking. In this opening episode, three structural shifts come into focus — and together they paint a clear picture of where leverage actually sits in 2024.

    First: the hiring bifurcation. Large tech companies are cutting junior AI positions as agentic frameworks absorb the scaffolding work those roles once owned. But smaller companies are doing the opposite — actively recruiting AI-native developers who built with these tools from day one and don't need to unlearn old habits. Two distinct labour markets are forming simultaneously, and knowing which one you're competing in changes every decision.

    Second: model selection is no longer about brand loyalty or raw benchmark performance. Teams are choosing models based on workflow fit — weighing cost, context window size, and task-specific reliability against each other. DeepSeek's open weights and aggressive pricing accelerated this shift. One-model-fits-all is no longer a defensible strategy.

    Third: open-source momentum has crossed from a community value into a real business moat. Permissive licensing gives companies control over cost structure, eliminates vendor lock-in, and enables fast pivots. Production deployments of multi-agent systems with interrupt-driven, human-in-the-loop checkpoints are now standard architecture — and the best tooling for that pattern is largely open-source.

    For engineering leaders and developers building with AI today, these are not slow-moving trends to monitor from a distance. They are decisions landing on teams right now.

    This episode includes AI-generated content. A YesOui.ai Production.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    7 min
  • The Two-Tier Developer Market: AI Hiring Splits, New Models & Agentic Frameworks

    The Two-Tier Developer Market: AI Hiring Splits, New Models & Agentic Frameworks
    May 2 2026
    The developer labor market is fracturing along a single fault line: AI. Large tech companies are pulling back from entry-level engineering hires, citing AI automation as justification for reduced headcount. At the same time, smaller, faster-moving companies are hiring aggressively — specifically for AI-native talent with hands-on experience in agentic workflows and model integration. This episode maps that split clearly and explains why where you apply now matters as much as what you know.

    On the model front, three distinct releases have reshaped the selection landscape in a short window. GPT-5.5 repositions OpenAI's flagship as an operating layer for agentic business tasks. DeepSeek V4's one-million-token context window and low token costs change the architecture decisions teams are willing to make. Anthropic's Opus 4.7 carves out space for regulated, instruction-sensitive environments where predictability outweighs raw capability. The era of one-model-fits-all purchasing is ending — task-specific selection is now the baseline expectation.

    The third story is Craft Agents from Lukilabs, an agentic framework released under Apache 2.0 that hit GitHub Trending shortly after launch. The permissive licensing is the signal: in a market where thin-wrapper businesses face real margin pressure from cheaper open models, frameworks that let teams own their stack are gaining genuine traction. Agentic UI patterns — interrupt-driven approval flows, generative UI state synchronisation — are also moving from research into production, though wire format standards remain unresolved.

    All three stories point to the same structural shift: control is moving toward builders who can make independent decisions about models, tooling, and workflows. This episode gives you the framework to act on that now.

    This episode includes AI-generated content. A YesOui.ai Production.

    This episode includes AI-generated content.
    Afficher plus Afficher moins
    7 min