AI EngineerReviewed by the Liquid Gold team

Highly useful for learning, with unusually specific, actionable detail from practitioner talks that mostly distinguishes demonstrated results from claims.

• Signal: Concrete architectures and numbers with caveats — Credit Karma's governance stack with a disclosed ~90% token reduction 'without paired answer-quality study,' In The Pocket's retrieval hit rate of 97% on only 63 questions framed as retrieval-only, OutRival's 'Tokenmaxx' explicitly noting HarborBite is a synthetic proxy rather than realized savings.
• Signal: Transferable engineering lessons and exercises, e.g., Coder's 'locate persistent tokens and rehearse an incident' and Sonar's 'measure the whole path from proposed change to accepted change.'
• Limit: Several talks report company-internal figures (Monzo's 10% of merged PRs, Coinbase's hundreds of millions of transactions) that are not independently verifiable; a few talks (swyx's keynote, the spatial-memory talk) are conceptual or propose applications rather than demonstrating deployments.

Content is largely evidenced and carefully hedged, with curation notes themselves flagging claim boundaries.

• Signal: Multiple descriptions explicitly mark claims as reported, relayed, or unvalidated — the secondhand banking incident is 'presented as a relayed account, not an independently documented incident,' and the solo-builder talk notes it is 'not a controlled study.'
• Signal: Viewer comments add corroboration and pushback, e.g., a first-hand ~50% serving-stack improvement matching the SemiAnalysis claim, and critiques of weaker talks (Kormos's presentation called 'mostly charts and jargon').
• Limit: Heavy reliance on vendor/company self-reporting (Coinbase, Monzo, Coder, Spranic) with no independent verification; some figures like the x402 transaction volume and 90%-off-the-laptop prediction are speculative.

Some talks connect to speakers' employers' products, but the sampled content reads as conference talks sharing practice rather than funnels to paid offers.

• Signal: Speakers from Coder, Coinbase, Sonar, and Spranic tie lessons to their companies' approaches (e.g., Coder's workspace governance), a mild soft-promo pattern.
• Signal: No curated record mentions paid courses, coaching, affiliate links, or upsells; technical detail dominates descriptions.
• Limit: Metadata alone cannot fully rule out embedded sponsorships or product pitches in the talks themselves.

Each curated video is a distinct, substantive talk from a different speaker and organization with no templated or near-duplicate content.

• Signal: Seventeen records cover non-overlapping subjects — merge queues for agent-scale commits, agentic payments, spatial memory, M&A-driven AI transformation, harness debates — with distinct metrics, architectures, and case stories per talk.
• Limit: These are recorded conference talks, so originality reflects the source event's curation; a few shorter entries (swyx's keynote) carry less material than others.

The sampled comments are unusually substantive: technical first-hand results, constructive critiques, and genuine debate about the talks' claims.

• Signal: First-hand corroboration and experience-sharing, e.g., a viewer achieving ~50% gains via better batching, and a code-search practitioner reporting Serena's 60% token cut with worse accuracy.
• Signal: Critical engagement with methods, e.g., caution that self-disagreement analysis can mask shared failures, concerns about MCP gateway blast radius, and multiple commenters calling the Kormos talk manager-style without a built-system demo.
• Signal: Extended threaded debate on the harness talk (13 comments) covering model-vs-harness boundaries, hobbyist data flywheels, and tool-generation.
• Limit: This is a newest-first selected sample with no like counts, author identities, or totals; it measures discussion quality in the sample, not overall audience engagement, and four videos have no supplied comments at all.

Newest video 4 days ago · 872 uploads in the last 12 months