Temporada 1
A special episode on the essay the whole industry is talking about today. Jordan and Riley walk through Bill Gates's new Gates Notes essay arguing that "there is no plan" for the social, political, and economic upheaval AI is about to cause — that AI could be "the greatest equalizer ever invented, or the worst source of injustice." They cover the five thresholds Gates says we've already crossed (bio, cyber, psychosocial, job-market-destruction, and loss of control), including his most chilling claim — "any model that can make novel molecules should be monitored," and that he rates bioterrorism ~50x more likely/scary than a natural pandemic — plus his point that last year's leap in agentic coding was simultaneously a "massive cyberattack threshold." On jobs, why he breaks with the usual "no technology ever caused net job loss" reassurance ("this time is different"): white-collar work hit first (sales, support, software engineering, paralegal), blue-collar via cheaper robots (the $20/hr worker replaced by a $10/hr robot), an entry-level hiring squeeze, and a competitive "vicious cycle." Then his three clusters of proposals: human-reserved jobs (with CBAM-style tariffs to enforce them), redistribution via robot + token taxes, and new governance — national coordinating bodies plus an international AI organization modeled on nuclear-weapons inspections, aviation regulation, and the ozone treaties. Also his case that industry can't self-regulate, why US–China coordination starts at home, and the genuine optimism (health, education, climate, eventual abundance — but turbulence first). Gates discloses his own ongoing financial ties to the AI industry; this is the first of a planned series. Source: "A turbulent AI era and critical choices to make" by Bill Gates — Gates Notes, Aug 26 2026 — https://www.gatesnotes.com/a-turbulent-ai-era-and-critical-choices-to-make (with Gates's interview in MIT Technology Review, Aug 26 2026). This is commentary/summary in the hosts' own words, not a reproduction of the essay.
Jordan and Riley continue with the second half of AgentFlo's Bedrock AgentCore build, focused on trust and reliability for agents that take real actions like placing orders and applying discounts. They cover the "model proposes, deterministic policy decides" philosophy — three layers of guardrails across the request lifecycle, Cedar policies enforcing business rules like a maximum discount independent of whatever the model reasons, and a two-table DynamoDB design that pins prices and cart state to structured data so the model can't hallucinate them — plus a reported +12% net revenue uplift from early deployment. This is original commentary, not a reproduction of the original post. Source: "How AgentFlo built AI sales agents with Amazon Bedrock AgentCore – Part 2" by Muhammad Musab Iqbal, Melanie Li, et al. (AWS + AgentFlo), AWS Architecture Blog, Aug 21 2026 — https://aws.amazon.com/blogs/architecture/how-agentflo-built-ai-sales-agents-with-amazon-bedrock-agentcore-part-2/
Jordan and Riley compare the two autoscalers Netflix has run across its 30,000+ Apache Flink jobs — a 2019 Mantis-based scaler that reasoned over coarse cluster metrics and moved a single knob (total TaskManager count), versus the Apache Flink community autoscaler that reasons from inside the job using True Processing Rate to size each operator independently. They cover why Netflix is converging on the open-source option, the engineering gaps it had to close at scale (metric collection at high parallelism, preserving forward chaining, respecting sink limits), the safety checks in the "realizer," and a 58% (~$1.1M/yr) compute cut for one team. This is original commentary, not a reproduction of the original post. Source: "A Tale of Two Flink Autoscalers" by Samuel Yeboah, Francesco Di Chiara, and Mingliang Liu, Netflix Tech Blog, Aug 21 2026 — https://netflixtechblog.com/a-tale-of-two-flink-autoscalers-e9f6a1b1492b
Jordan and Riley unpack Cloudflare's shift from all-or-nothing OAuth consent to optional scopes — letting developers mark each scope as required or optional so users can deselect what a given task doesn't need, with the resulting access token carrying only what was actually granted. They cover the key mechanism (required-vs-optional is evaluated only against the scopes requested in that specific flow, keeping the consent screen focused on the immediate task), why it matters in the MCP and agent era, and the engineering catch — apps now have to handle partial grants gracefully instead of assuming full approval. This is original commentary, not a reproduction of the original post. Source: "From all-or-nothing to task-based OAuth consent" by Miller Vargas and José Enrique Rodríguez, The Cloudflare Blog, Aug 20 2026 — https://blog.cloudflare.com/task-based-oauth-consent/
Jordan and Riley walk through the first half of AgentFlo's architecture for AI sales agents that live in WhatsApp, SMS, and RCS and handle product discovery, carts, and order completion — a five-layer design built on Amazon Bedrock AgentCore. They dig into recipe-based deployment that onboards a merchant in minutes, a model-driven approach instead of rigid workflows, why AgentFlo chose a single agent over multi-agent for commerce, per-session microVM isolation, and standardizing every third-party integration behind one MCP tool gateway so onboarding a service becomes a config change, not an agent change. A future episode picks up the trust and reliability layers. This is original commentary, not a reproduction of the original post. Source: "How AgentFlo built AI sales agents with Amazon Bedrock AgentCore – Part 1" by Muhammad Musab Iqbal, Melanie Li, Mohammad Hamza Khan, Anirudh Marc J, and Vincent Wang, AWS Architecture Blog, Aug 19 2026 — https://aws.amazon.com/blogs/architecture/how-agentflo-built-ai-sales-agents-with-amazon-bedrock-agentcore-part-1/
Jordan and Riley break down Shopify Engineering's "gisting" technique — replacing a big, repeated system prompt with a handful of learned "gist tokens" that carry the same behavior to the model, trained via self-distillation (teacher pass with the full prompt, student pass with the gist tokens, minimize the gap between them). They cover why deployment stays trivial — the learned embeddings are written straight into the model's embedding matrix, no custom serving path — and the payoff on Shopify's Sidekick GraphQL agent: 4:1 compression, time-to-first-token down 19%, end-to-end latency down 38%, and 14% fewer GPUs. This is original commentary and discussion, not a reproduction of the original post. Source: "Gisting: Compressing LLM Agent Context to increase throughput and reduce cost" by Paige Vegna and Cody Mazza-Anthony, Shopify Engineering, Aug 19 2026 — https://shopify.engineering/gisting
A real-world architecture for finding sensitive data hidden inside medical images at scale. Jordan and Riley cover how Clario (part of Thermo Fisher) auto-detects PHI/PII across thousands of DICOM slices in clinical trials, where sensitive data hides in three surfaces: standard metadata tags, custom/vendor private tags, and text burned into the image pixels themselves. The architecture: images in S3, API Gateway at the edge, an EKS detection backend (a single series can span thousands of slices, so it's long-running and memory-intensive), RDS PostgreSQL for auditable compliance metadata, Amazon Textract for OCR, and Anthropic's Claude Sonnet 4.5 on Amazon Bedrock to identify PHI/PII in extracted text and deep-scan pixels for burned-in text, returning type + precise bounding boxes. The key design call: detection is deliberately separated from redaction — the AI does high-recall flagging, while irreversible masking happens in a separate human-in-the-loop QC flow. Evaluated against production-representative data (with and without PHI, to measure false alarms) using spatial-proximity matching: F1 of 0.975 on burned-in image text, 0.995 on metadata tags. Lessons: evaluate on real-world variability (off-the-shelf models degrade), never let AI drive irreversible redaction alone, and mind data minimization (S3 lifecycle auto-delete). Source: How Clario technology detects PHI/PII in DICOM images using Amazon Bedrock — AWS Architecture Blog, Aug 19 2026, by Alex Boudreau, Cuong Lai, Matthew Agard & Praveen Haranahalli — https://aws.amazon.com/blogs/architecture/how-clario-automates-phi-pii-detection-in-dicom-images-using-amazon-bedrock/ This is commentary/summary in the hosts' own words, not a reproduction of the article.
Low-level CPU security — Spectre, speculative execution, side channels — grounded in a real production system running untrusted code for tens of thousands of tenants. Jordan and Riley cover why Workers is uniquely exposed (untrusted JS in V8 isolates, tens of thousands of tenants sharing one OS process, so a single arbitrary read = cross-tenant leak), then how Cloudflare re-attacked their own production to test whether their 2021 defense (Dynamic Process Isolation) still holds against newer techniques. They walk the four hurdles the team had to clear to pull it off in prod: rebuilding a high-res timer over a WebSocket (Workers freezes local timers), a speculative type-confusion gadget that survived V8 pointer compression via a raw 64-bit backing-store pointer, signal amplification through a tree-PLRU L1 replacement trick, and cache reset by pigeonhole eviction. Result: a demonstrated remote Spectre leak up to 12 bit/s at 99% accuracy in production — already mitigated, no evidence of exploitation in 3 years. Fixes: hardened DyPrIs plus integrating the V8 Sandbox with in-process isolation to kill the raw-pointer gadget class. Lesson: Spectre isn't patch-once — periodically re-attack your own defenses under real production conditions. Source: A revisit of remote Spectre attacks on Cloudflare Workers — The Cloudflare Blog, Aug 19 2026 (paper: arXiv:2608.17043) — https://blog.cloudflare.com/revisiting-spectre-attacks-on-workers/ This is commentary/summary in the hosts' own words, not a reproduction of the article.
One of the most underrated ML skills — forecasting discipline, dressed up as a COVID story. Jordan and Riley unpack why "retrain" secretly hides three different decisions: refit (same model, new data), respecify (change structure/features/priors), and hold (do nothing) — and how a standing retrain cadence always silently picks refit. They cover the three failure modes: chasing noise (an off-cycle refit on a one-off spike overweights the least-understood window), carrying ghosts (a stale COVID cancellation-timing assumption stays switched on behind clean-looking refits), and respec-as-panic (rebuilding under deadline for a passing FX shock). The decision triage: params drifted but process same -> refit; process changed in a way the model can't represent (tell: one-sided, directional misses) -> respecify; miss inside normal range -> hold. Plus the real case where fixed-hierarchy pooling broke on divergent geographic recovery, a refit went unstable (~3x error, wouldn't settle = structural signal), and respecifying to borrow along geographic adjacency cut error by more than half. Takeaway: refitting keeps a model current but not honest — a good team learns to forget shocks on purpose. Source: How we knew COVID was over (and what our models had to unlearn) — Airbnb Engineering, Aug 19 2026, by Harrison Katz — https://medium.com/airbnb-engineering/how-we-knew-covid-was-over-and-what-our-models-had-to-unlearn-c606b9bdb0ab This is commentary/summary in the hosts' own words, not a reproduction of the article.
When an LLM gets a fact wrong, is the fact missing (an "empty shelf") or present-but-unretrievable ("lost keys")? Jordan and Riley unpack Google Research's argument that standard accuracy metrics collapse these two cases — even though one calls for scaling model/data and the other for cheaper post-training/inference-time fixes. They walk through "knowledge profiling" (analyze the state of each fact, not each question) and its five profiles, the three behavioral notions (encoding via priming, knowledge across phrasings incl. the reversal curse, and recall — direct vs. inference-without-encoding), and WikiProfile, a 2,150-fact benchmark (10 tasks each) built with a Gemini-2.5-Pro + search-grounded pipeline. Headline finding on frontier models (Gemini 3, GPT-5): many factual errors are recall failures, not encoding failures — the knowledge is often in there, the model just can't reach it on demand. Practical takeaway: before assuming the model doesn't know something, test whether it can recognize it or reach it with chain-of-thought; if so, the fix is on your side of the API, not in the weights. Source: Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality — Google Research Blog, Aug 12 2026 (paper: arXiv:2602.14080) — https://research.google/blog/empty-shelves-or-lost-keys-recall-is-the-bottleneck-for-parametric-factuality/ This is commentary/summary in the hosts' own words, not a reproduction of the article.