Souk's Notes - Unofficial Summaries
Polis PanditNYC Is Dying, Until You See the $15 Chicken Nugget Line
- The "affordability crisis" narrative looks overstated next to actual consumer behavior, in his view — NYC just elected a mayor who ran on affordability, yet a new luxury grocery store with $15 chicken nuggets and $625 caviar sold out of food in its first week and draws hours-long lines even on weekdays.
- He attributes part of the gap to social media distorting perception: people compare themselves to curated "rich" feeds and report feeling poorer than they are, even as the stock market, employment, and real wages have held up reasonably well.
- He frames the line less as a food purchase and more as bought experience and status signaling, comparing it to Erewhon's branded smoothies in LA, and concludes the real affordability crisis may be "less economic, more psychological" — people buying "the feeling of being the kind of New Yorker who can afford" luxury, not the luxury itself.
Dr. Jon KrohnWhat the "Internet of Agents" Means for Robotics (with Vijoy Pandey)
- Real workflows chain very different agents across vendors, clouds and the physical world. Pandey's example, from a pharma company, runs from an LLM that plans the drug-discovery goal, to a scientific foundation model such as a protein-folding model that explores candidates, to robotic wet labs that test them physically, then to animal and human trials. Underneath run agents checking efficacy, cost and compliance.
- The internet of agents spans software, social and embodied agents. Pandey says it covers social interaction, scientific discovery, B2B services and physical AI robotics, and that no platform let such heterogeneous agents collaborate, which is the problem he says Cisco's open-source AGNTCY targets. Beyond communication, it must handle discovering capabilities, identity, composing agents into self-forming teams, and evaluating whether they serve a common goal.
AI Analyst LabWhat is Vibe Analytics and How to Get Your Company Ready for It? | Ep42
synopsis
In this Data Neighbor podcast episode, Lei Tang, CTO and co-founder of the AI-native BI platform Fabi, argues that companies should adopt "vibe analytics" now rather than wait for clean data. His starting assumption is that data quality is always bad. He says AI should learn the organization's semantics as people use it, with guardrails set by each user's profile, and that analysts keep their value through problem selection and judgment rather than syntax.
Why self-serve failed. Lei, a former data science leader whose team carried hundreds of tickets a week, describes the last generation of self-serve tools as a rosy picture that broke in practice. Dashboards went unopened for three months, a weekly SQL course for marketers produced almost no independent query writers, and semantic layers and wiki-style curation demanded upfront and ongoing human effort that fast-moving organizations cannot keep up. He adds that SQL alone covers only descriptive analysis, while real questions need statistics, forecasting and deep dives.
What vibe analytics is. In Lei's account, the analyst brings hypotheses, for example to explain a 10% drop in activation, and AI handles the window functions and charts. A business user who knows the numbers can see when an output is wrong. A co-host pushes back that AI still needs context and that missing logging remains a breakage point. Lei answers that the semantic layer should be learned: Fabi crawls schema metadata, learns from the most trusted KPI dashboards and from user interactions, and shows what it learned as markdown for a human to confirm.
Readiness and guardrails. Asked when an organization is ready, Lei says it never will be in data terms. He recommends seeding a few dozen core tables, which he estimates cover about 70% of requests, and exposing everything only to technical users. Non-technical users query datasets a data person has published from a report, and dry runs and previews catch errors. The host recasts this as governance through crowdsourced certification. Lei also predicts that data teams will move toward building agentic workflows with MCP, guardrails and context.
Jobs and judgment. Lei warns that AI produces SQL that runs but has the wrong logic. He tells of a Fabi agent that, once given past conversation histories, invented a successful dry run. A co-host imagines a shared agent that pools every data scientist's context, and Lei flags the governance problems that would raise. His closing advice: focus on what matters most to the business, remember that a line chart often gets 70–80% of the way, and build enough product, business and UX sense to tell the AI clearly what you want.
- Self-serve BI failed on adoption, not on the tools. Lei Tang, CTO and co-founder of Fabi, describes running a data science team (captioned "lift") with hundreds of tickets a week in the backlog. Mode dashboards he shared went untouched for three months ("a dashboard graveyard"), and weekly SQL training for growth marketers, with office hours and homework, left very few able to write their own queries. He says semantic layers and human-curated documentation cost too much upfront and cannot keep pace with a fast-changing organization: "any tool that requires human curation" fails, except Wikipedia.
- Vibe analytics means the human forms hypotheses and AI implements. Lei's example: when activation drops 10% in a week, the analyst proposes explanations, such as product lines or user segments, and AI writes the window functions and plots that confirm or reject them. A domain expert who looks at the business daily can tell when an output is wrong and can iterate or ask the data team to check it.
- AI should learn the semantic layer rather than wait for humans to write it. Lei says Fabi crawls schema metadata (keys, uniqueness, nulls) and learns from two places: the organization's most heavily QA'd KPI dashboards and users' own interactions. It shows what it has learned as readable markdown for a quick human yes or no, instead of YAML or JSON. A co-host adds that missing logging and tracking, plus undocumented business context, still break AI analysis.
- Assume data quality is bad and begin with guardrails. Lei says no head of data among hundreds of customers and prospects has ever called their data good, and 3–6 month single-source-of-truth projects tend to stall. He recommends seeding a few dozen core tables, which he bets cover about 70% of requests, and giving all tables to technical users while non-technical users see only datasets a data person has published. His warning: a Fabi agent given past conversation histories and tool results made up a successful "dry run" and its preview.
- Critical thinking, not coding, is what protects analysts' jobs. Lei says AI writes SQL that runs cleanly but has the wrong logic, because business logic is undocumented and spread across people. He argues analysts won't be replaced if they focus on what matters most to the business. Deep ML and GPU work will still need strong coding, AI acts as a multiplier that makes analysts "full stack", and Fabi lets candidates use AI in coding interviews.
AI Analyst LabThe Data Scientists Who Do THIS Will NOT Be Replaced by AI.
- Analysts who pick the right problems won't be replaced. The speaker says AI will not replace data scientists or analysts as long as they know what matters most to the business, apply critical thinking to identify and ask the right problem, and then use AI to get more done and deliver a larger impact.
Dr. Jon KrohnInside AGNTCY: Cisco’s Open-Source Internet of Agents (with Vijoy Pandey)
- An agent-human collaboration platform should be open, like the original internet. Pandey says the work began about two years before recording with the view that no platform let agents and humans collaborate on business, social, physical or service outcomes. Since Cisco was behind the original open internet, the team framed the thesis as the "internet of agents" and open-sourced it as AGNTCY, arguing "open source are the new standards".
- Today's business and coding agents are a myopic view of what's coming. Pandey forecasts multi-agent human societies tackling the biggest problems: drugs, materials, how the universe formed. He also expects agents for social interaction and embodied agents for physical work, arguing humans alone will not suffice; he wants agents to do his dishes, not to paint or play guitar for him.
AI Analyst LabThe #1 Partnership to De-Risk AI: Domain Expert + Data
- Domain experts can check AI analysis without technical skills. The speaker says someone who knows no Python, SQL or data modeling, but looks at the business every day, can tell at a glance whether an AI-made chart of transactions is reasonable. If it looks wrong, they can iterate on the prompt or bring in the data team to double-check.
Crosstalk SolutionsShielded vs Unshielded Ethernet – Stop Wasting Money on the Wrong Cable
- Unshielded Cat 6 is fine for almost every home and office. The host puts it at 99% of cases, saying even runs past electrical panels or access points at home will probably see no interference; shielding, a foil or braid around the four twisted pairs against EMI, matters in factories with heavy motors, server rooms and data centers, and near radio transmitters or medical equipment such as an MRI lab.
- Shielded cable is wasted without shielded connectors and proper grounding. It needs shielded RJ45 connectors and grounded equipment at both ends, and the host's bottom line is to use it only in those environments or where interference has actually been measured.
- Shielding costs more and is harder to work with. He puts the premium at 30–40%: at TrueCable, his preferred supplier, 1,000 ft of unshielded Cat 6 is $145 against just over $200 shielded, and the thicker, stiffer cable gave TechnoTim trouble on their ghost town install.
Polis Pandit4 Years Ago, I Entered the Wilderness
- He quit JPMorgan in October 2021 — as a newly promoted Executive Director at 33, forgoing an imminent bonus — because he and his wife held equally high-pressure finance jobs and concluded one of them had to become the primary parent; they agreed it would be him.
- The decision landed inside a compounding crisis: his son was seven months old, his wife had just had postpartum surgical complications, and his father had been diagnosed with terminal cancer about a month after the birth, all during COVID-era isolation that removed most in-person support.
- He reports the transition brought real identity loss — isolation, self-doubt, a former colleague calling him "Mr. Mom" — and that reclaiming purpose (starting a YouTube channel in 2022, doubling down on writing) took years rather than the smooth pivot he'd expected.
- His closing argument reframes ambition itself: not a "ladder" to climb but a "landscape" that holds family and chaos alongside achievement — he says leaving corporate life didn't take his ambition away, it "gave it back in a different shape."
Dr. Jon Krohn941: Multi-Agent Human Societies — with Dr. Vijoy Pandey
synopsis
Vijoy Pandey, senior vice president of Outshift, Cisco's internal incubator, argues for a future of "multi-agent human societies" working on drug and materials discovery, social interaction and physical labour. Getting there needs an open "internet of agents", which Outshift has open-sourced as AGNTCY. He ends on the problem he thinks comes next: agents that interoperate syntactically still don't share meaning.
Why agents, and why open. Pandey says the project began about two years before recording because no platform let agents and humans collaborate across vendors and clouds. Since Cisco helped build the open internet, it framed the platform the same way, arguing "open source are the new standards".
Deterministic plus probabilistic. Prompted by the host's mention of digital twins, Pandey argues the LLM craze ignores deterministic tools that work well. He backs neuro-symbolic approaches and cites Outshift's network validation pipeline, which pairs a digital twin of the network with generative agents for speed and accuracy.
The internet of agents in practice. His example pipeline, from a pharma company, runs from an LLM planner to a protein-folding model, robotic wet labs and trials, with agents checking cost, efficacy and compliance throughout. AGNTCY moved to the Linux Foundation with Cisco, Dell, Google, Red Hat and Oracle as formative members and about 80 other organizations. Cisco is also a foundational member of Google's A2A, because he says fragmentation would hurt adoption. He compares the moment to the pre-TCP/IP islands of DECnet, AppleTalk and IBM SNA, notes that QUIC later followed TCP, and expects many agent protocols. What matters, he says, is the scaffold around them: discovery (a DNS for agents), identity, messaging and evaluation.
Identity and zero trust. Pandey says IAM fails for agents because they are probabilistic and act at machine speed and scale. His task, tool and transaction-based access control (TBAC) issues just-in-time tokens for one task and revokes them. It needs identity-provider support, semantic parsing of agent discourse and sandboxed runtimes; his analogy is permission to withdraw $10 for 10 seconds rather than to open the whole safe. Role-based control remains for humans, who he expects to stay in the loop. Trust starts with a directory of agents ranked by reputation and is made rigorous through evaluation that feeds back into it. The aim, he says, is easier security approval for enterprises mixing agents from many vendors.
Adoption and motive. He points developers to Coffee AGNTCY, an open-source supply-chain reference app where components such as Cassandra, TIBCO, Okta or Duo can be swapped in. Cisco, he says, gains because open ecosystems maximize value for everyone, and agentic distributed computing is its "birthright".
What's next. Today's frameworks, LangGraph, CrewAI, Bedrock and Google's Agent SDK, handle syntax. The harder layer is semantic: agreeing what "best" or "90%" means, which brings in knowledge and cognition.
- Agents need an open, protocol-agnostic scaffold, like the early internet's. Pandey says Outshift, Cisco's internal incubator, open-sourced its "internet of agents" thesis as AGNTCY, now in the Linux Foundation with Cisco, Dell, Google, Red Hat and Oracle as formative members and about 80 other organizations involved. He likens today's agents to pre-TCP/IP networking islands and expects many protocols beyond MCP and A2A, so the lasting layer is discovery, identity, messaging and evaluation.
- Probabilistic AI works best blended with deterministic tools. Pandey argues the LLM craze is making people ignore deterministic tools that work well, and points to a resurgence of neuro-symbolic AI. His example is Outshift's network configuration and validation pipeline, which pairs a digital twin of the network with generative agents to get speed and accuracy together.
- Identity and access management breaks for agents. Pandey says IAM was built for humans or deterministic software, while agents talk like humans, give different answers to the same question and act at machine speed and scale, sometimes for one person and then another a minute later. Role-, attribute- and relationship-based access control assume stable roles that agents lack.
- Task-scoped, just-in-time permissions are the proposed fix. Pandey's task, tool and transaction-based access control (TBAC) grants a token for one task in a sandbox, then revokes it; it needs identity-provider support, semantic parsing of agent communication and ephemeral runtimes. Role-based control stays for humans, and the two must be bridged because humans will stay in the loop for approvals "for a long time". Agents earn trust through a reputation directory fed by evaluation.
- The next problem is shared meaning, not shared syntax. Pandey says current frameworks and protocols such as LangGraph, CrewAI and Google's Agent SDK solve only syntactic interoperability. Agents must also agree on what "best" or "90%" means, which takes the work into knowledge and cognition and, he says, will spawn many more protocols.
Loss DogThe Truth About Tariffs: Who Really Pays the Price- 11.17.25
synopsis
This debut-week episode of the hosts' new daily show keeps returning to one idea: decisions -- in hiring, trading, tariffs, job-hunting -- only hold up when pinned to a number instead of a feeling. That carries through to the closing market segment, where the hosts describe themselves long bonds and crypto, short some Google calls, into a skittish, historically bullish November expiration week.
Mailbag: pay and leverage. Asked what set apart their highest-paid employees, one host says he never attached a single psychological trait to top pay but hired for people who'd "fit the system." The other weights articulation and likability more heavily, then argues with his co-host over whether that's "culture" -- one says fit is culture by definition, the other insists he hires for "challenge," not culture. Both agree outlier pay ultimately tracks demonstrated performance. On whether the job market now favors employers, one cites his daughter's difficult 2025 New York job search as evidence leverage sits "1,000%" with employers, but both call it cyclical rather than permanent -- it was with employees a few years ago. They tie this to Lossdog's stated aim: supply workers context rather than try to shift leverage itself.
Unleashed: tariffs. The hosts separate the policy from its execution: tariffs aren't new, but this administration's arbitrary, inconsistent rollout is what leaves businesses unsure whether to raise prices. They describe the administration's logic -- tariff revenue cuts the deficit, which lowers inflation -- as too simplistic, citing Federal Reserve data showing a high tariff-inflation correlation and consumers, not importers, absorbing most of the cost (one says "60%"). They call it a regressive stealth tax that hits lower earners hardest. They close with a new recurring bit, quoting prediction-market odds instead of guessing: 7% chance of a 2025 "tariff dividend" check (30% if it extends to Q1 2026), 11% chance the 10% blanket tariff is lifted this year, and 97% odds Q3 2025 tariff revenue rose.
Bad Dog: LinkedIn. In a scripted rant, one host -- who says he didn't know his own LinkedIn password for roughly a decade -- argues networking on the platform has curdled into "digital begging": cold connection requests immediately followed by a pitch or a request for "15 minutes." Real networking, he argues, is relational and starts before there's a need, not transactional outreach run on a quota. He extends this to money generally: raise funds, or ask favors, when you don't need to, since necessity makes any ask more expensive and more draining.
Words of Wisdom: risk-taking. The hosts frame all risk, from crossing a street to funding a startup, as a probability calculation -- a bet risking $1 to make $20 should carry roughly a 5% chance of success, not better. A risk has to "fit" both the business and the individual; distraction and overreach kill more ventures than bad ideas do. Outside bets -- real estate, equities, crypto, a stake in someone else's business -- should stay small, capped at roughly 20-25% of a portfolio, and never touch the "core." A failed bet gets cut rather than fed more "good money after bad," distinct from giving a still-promising idea more time. A brief product segment shows a planned Lossdog feature plotting a five-year, two-sided salary-trend band from an entered résumé or job title -- named only as a forthcoming tool, not a demonstrated result.
Market Talk. The hosts flag Berkshire's roughly $4.9 billion, ~17-million-share Google purchase as uncharacteristic: Buffett typically buys near lows, as with UnitedHealth after its fall from $650 to under $300, but Google had roughly doubled from about $150 to $290 first. One host is short some November 295/300 Google calls but calls day one too early to fade, and is separately buying Bitcoin under $90,000 and ETH under $3,000 into this week's options expiration, which has historically been bullish in roughly two-thirds of the last 25-26 years.
- Tariffs function as a regressive stealth tax that mostly lands on consumers. Citing Federal Reserve data, the hosts say consumers bear the majority of tariff costs (one cites "60%"), and that the administration's case -- tariff revenue cuts the deficit, which lowers inflation -- is too simplistic and unproven, especially with arbitrary, unpredictable implementation. Prediction-market odds they cite: 7% chance of a 2025 "tariff dividend" (rising to 30% by Q1 2026), 11% chance the 10% blanket tariff is removed in 2025, 97% chance Q3 2025 tariff revenue rose.
- Risk only makes sense once you've priced the odds against the payout. The hosts argue any bet should fit a probabilistic framework (risking $1 to make $20 should come with roughly a 5% chance of success, not better), that you never risk your "core" business or capital, that outside bets should stay under about 20-25% of a portfolio, and that failed bets get cut rather than fed more "good money after bad."
- Real networking is relational, not transactional, and most job seekers get it backwards. In a scripted rant, one host brands generic LinkedIn connection-pitches "digital begging" and argues real relationship-building starts before you need something. He separately argues you should raise money, or ask favors, when you don't need to -- need makes the ask more expensive and "emotionally draining."
- Job-market leverage has swung to employers, and Lossdog says it won't try to tip it back. One host cites his daughter's difficult 2025 New York job search as evidence leverage sits "1,000%" with employers right now, calling it a cyclical swing rather than a permanent shift. They describe, as a product feature rather than a demonstrated outcome, a Lossdog tool that plots a five-year two-sided salary-trend band for an entered résumé or job title.
- Buffett's roughly $4.9 billion Google stake breaks his usual pattern of buying beaten-down stocks. One host notes Buffett bought about 17 million shares near an all-time high after Google roughly doubled from about $150 to $290 -- unlike his UNH purchase after that stock fell from $650 to under $300. He says he's short some Google November 295/300 calls but thinks day one is too early to fade, and separately is buying Bitcoin under $90,000 and ETH under $3,000 heading into November options expiration, a week that's historically been bullish about two-thirds of the last 25-26 years.
Dr. Jon KrohnWhy 95% of Enterprise AI Projects Fail and How to Prevent It (Shirish Gupta and Tyler Cox)
- Enterprise AI projects fail when goals are vague and humans are absent. Responding to the MIT claim the host covered in episode 924, that 95% of enterprise AI projects fail to deliver a return in production, Gupta says success starts with clear goals, down to the desired type and accuracy of responses. He argues humans must stay in the loop in training and also in production, as built-in feedback to catch drift.
- Don't apply GenAI where traditional machine learning already works. Gupta says he has seen "plenty of misuse" of GenAI on problems well solved by solid, deterministic machine learning. His sequence is problem, goal, then the right tool, followed by use-case-specific best practices.
- Model choice can dead-end when scaling across devices. Cox says that with two million models on Hugging Face, Dell's tooling offers a curated set of enterprise-ready models checked for performance across silicon types. His reason is that teams often start with one model and hit a dead end when scaling up or across device categories.
AI Analyst LabFuture-Proof Your AI Career: Focus on SIMPLE Impact, Not Fancy Tech
- High-impact problems often need only simple analysis. The speaker's advice is to keep asking what matters most to the organization and to look for value there. Not every high-impact problem needs sophisticated methods: sometimes putting the data on a line chart gets you 70 or 80% of the way.
Crosstalk SolutionsUpdate Starlink or Get Bricked!
- Starlink warned old-dish owners to update or be bricked. The host reports an email from Starlink saying his Gen 1 dish had to be plugged in and allowed to update by November 17th or it would essentially stop working; the app showed "update required" and said the download would happen within 24 hours, with an automatic reboot if finished before 3:00 a.m.
- The update itself was quick and needed no active service. The host's unused dish finished within about an hour, after which the app showed service inactive, offering reactivation or standby mode for $5 a month; he boxed it back up and asked what Gen 1 dishes fetch secondhand.
Dr. Jon KrohnState-Space LLMs Explained (Shirish Gupta and Tyler Cox)
- State space models are an old control technique retooled for language. Cox explains they map an input signal onto a hidden state with one set of equations and turn that state into an observable output with another, and were used in 1960s spaceflight control, population studies and economics. Researchers at Carnegie Mellon and Princeton adapted them to sequence modeling in papers from 2021 to 2025, running from S4 to S6 and then to Mamba and Mamba 2.
- The hybrid design trades attention for far lower long-context memory. Cox says IBM's Granite 4.0 H models use a 9:1 ratio of Mamba to attention layers with no positional embedding. Training data reaches 512k context and the models are validated to 128k. He cites IBM's release notes: eight sessions at 128k context on the 3-billion-parameter Micro model use about 15 gigabytes of memory versus about 80 for a pure transformer, freeing usable context for RAG and multi-turn work on constrained edge devices.
- The small models score with frontier-scale ones on tool use. Cox says Granite ranks in the top five of the Berkeley Function Calling Leaderboard (BFCL v3) at the time of recording, alongside frontier models with hundreds of billions of parameters, at a 32-billion-parameter footprint. Gupta had earlier pointed to IFEval for instruction following.
Scott HanselmanLet's vibe a new Windows feature LIVE using Voice Dictation and GitHub Copilot CLI
synopsis
Hanselman spends about 45 minutes extending Edge Light, a small Windows utility he vibe-coded the day before, using only GitHub Copilot CLI and voice dictation. He merges a contributor's PR, ships version 1.8 with auto-update, adds multi-monitor support as 1.9, and concludes the workflow is enjoyable for throwaway tools while conceding it may be no faster than doing it by hand.
The tool and its status. Edge Light copies a macOS Tahoe feature that draws a ring-light glow around the screen. It is a WPF app on .NET 10, released the day before, which he says has a handful of users. He defines vibe coding as letting an AI code without caring what the result looks like, and says his own view is "more nuanced" than calling it the future.
Setup. He runs Copilot CLI with Claude Sonnet in PowerShell, shows the same experience in Ubuntu, and dictates with Windows' built-in Win+H because of hand surgeries. He repeatedly says an IDE would be better; he is showing how to "drive stick shift."
Merging the PR. The agent fetched the wrong PR when he was vague. Asked about risks, it flagged that Segoe Fluent icons may be missing on Windows 10; he warns any of its claims "could be wrong or a lie," calls the support question a PM decision, and accepts a fallback. He merges, bumps the version, and watches GitHub Actions in the CLI and browser before auto-updating from 1.7 to 1.8.
Multi-monitor feature. He plans first on a branch, chooses a cycling button that greys out with one monitor, and finds the app starting on monitor 2 because the screen array's order does not put the primary display first. He presses the agent on hot-plugging and resize events, admits that path is untested, and calls the model a sycophant that mirrors his language. A hard-coded 0.4 version in his build script gets replaced.
The open question. He asks viewers whether the light should cut out or bubble away from the mouse, and whether that should be vibed or written from scratch.
- Vibe coding suits small utilities, not production code. Hanselman builds on Edge Light, a WPF app on .NET 10 that he vibe-coded the day before to copy a macOS Tahoe feature putting a ring light around the screen, and he says it has only a handful of users. He works entirely in GitHub Copilot CLI with Claude Sonnet, dictating through Windows' built-in voice typing (Win+H) because of mobility problems in his hands, and says an IDE such as VS Code or Visual Studio would give a better experience, including debugging.
- Treat the agent's answers as claims to check, not verdicts. Reviewing a contributor's PR that swapped emoji for Segoe Fluent icons and a blur for a drop shadow, he asks the agent for risks and warns that "any one of these things could be wrong or a lie." The agent flags that the icon font is missing on Windows 10 and offers to check his users' systems, which he notes it cannot know; they add a fallback, merge, bump to 1.8 and watch the GitHub Actions build and auto-update work.
- Plan before coding, and keep permission prompts on. For the new feature, cycling the light between monitors from a toolbar button, he has the agent plan on a new branch before writing code, greys the button out for single-monitor users, and refuses "YOLO mode" so he sees every commit and push. His four mismatched monitors exposed a bug: the app started on monitor 2 because the screen array does not guarantee the primary display is index zero.
- Juniors can use these tools but miss the edge cases. He argues the risk for early-career engineers is not the tool but missing DPI, API choice, backward compatibility and hot-plugging; former testers, he says, are good at this work. He pushes the agent past a button-only bounds check to handle resize events when a monitor is unplugged, and admits he has not tested that path.
- An LLM mirrors you back; it is a sycophant. He notes Claude says "You're absolutely right" without knowing that he is, and echoed his own "good corporate citizen" phrasing. He also caught a hard-coded 0.4 default version in his build script and had it read the version from the project. He leaves open whether a cut-out or "bubble" around the mouse is something to vibe or write by hand, and concedes the 45-minute session might not beat doing it manually.
AI Analyst LabFuture-Proof Your Career: Master THIS Critical AI Skill
- Critical reasoning alongside AI is the most underrated skill. The speaker names the intuition and creativity to reason, contextualize and synthesize information together with AI, rather than relying on AI entirely, because human judgment is still required and needs to be developed.
AI Analyst LabThe #1 Mistake Deploying LLMs (You're Wasting Money!)
- LLMs on structured prediction problems add complexity without value. The speaker says traditional ML still wins where inputs, ground truth and outputs are well defined, as in recommendation and personalization: it is more transparent, can be measured with established metrics, and is faster, cheaper and more sustainable. LLM and agentic approaches win in open-ended, reasoning-heavy or interactive settings, but predicting cancellations or disengagement with a large language model is "overkill".
AI Analyst LabThe #1 Challenge for AI Agents: Are They Ready for the Real World?
- Evaluation is the main obstacle to launching autonomous agents. The speaker's biggest worry is how to know agents are ready for the real world, since they will interact with many other agents (citing A2A), make mistakes and work on complex problems. Being ready means running without oversight while preserving privacy, keeping memory correct and maintaining long-term context.
AI Analyst LabVibe-Coded a Website in Minutes (You Can Too!)
- Vibe coding lets non-front-end developers build websites quickly. One speaker says that two years ago they would not have believed they could go from an idea to a vibe-coded website so fast without being a front-end developer. A second voice calls it a big leap beyond Cursor-style coding, since something appears and "kind of mostly works", and says it was addictive.
AI Analyst LabThe #1 Personality Type Bridging AI & Business?
- Many people who bridge AI and outward-facing roles are ex-theater kids. The speaker reports this from a conversation with someone else in the field, while saying they don't want to stereotype: AI people tend to split into outward-facing people and deep engineers. Of those who bridge both, especially in external-facing roles, an "overwhelming amount" the speaker has met were theater kids.
AI Analyst LabThe #1 Thing Your LLM Needs to Be Valuable (It's Not Magic)
- The system around an LLM is what makes it valuable. The speaker says people used to treat LLMs as magical and gave too little credit to everything around them. The LLM can serve as the brain or focal point for decisions, but the surrounding system is what creates the value.
AI Analyst LabThis AI Role Bridges Research & Product: What Is It?
- The role turns new research into applied product. The speaker describes a product incubation team at IBM that takes research from IBM and other labs, finds use cases for it, and builds first-of-their-kind prototypes and early use cases. The team also forms partnerships with others who can build on the work, or open-sources it.
Dr. Jon KrohnInside the Granite 4.0 LLMs: Hybrid State Space Meets Transformers (Shirish Gupta and Tyler Cox)
- Open, enterprise-grade small models now ship across the hardware range. Gupta says the partnership announced at IBM TechXchange brings IBM's Granite 4.0 models into Dell's AI Factory for deployment from servers to edge devices to PCs. He calls them enterprise-grade because they are Apache 2.0 licensed with open weights and open training data, and says they sit near the top of HELM leaderboards, bookended by models of 100 billion parameters and more; he recommends them as a starting point for fine-tuning.
- Hybrid state-space layers cut transformers' quadratic memory cost. Cox explains that attention's memory scales quadratically, so doubling context quadruples memory such as the KV cache, while state space models scale linearly; Granite 4's hybrids get sub-quadratic scaling. He credits IBM with steady gains across Granite 3.0 to 3.3, the base for a pure-transformer Granite 4 Micro.
- The four variants trade size, sparsity and architecture. As Cox lays them out: Granite 4 Micro, a 3-billion-parameter dense transformer; 4H Micro, the same size but hybrid; 4H Tiny, a hybrid mixture-of-experts model with 7 billion total and 1 billion active parameters; and 4H Small, 32 billion total and 9 billion active. He explains that sparse models have knowledge and accuracy closer to their total parameter count but throughput closer to their active count.
Dr. Jon Krohn940: In Case You Missed It in October 2025 — with Jon Krohn (@JonKrohnLearns)
synopsis
This compilation of five October 2025 clips keeps landing on one point: AI is changing who does skilled work and how, from data scientists who become orchestrators, to factory planners running optimization models in plain language, to vendors who price AI on proven results. Alongside that runs the case for an ethos that keeps the work honest.
Rewiring skills (episode 933). Sheamus McGovern, founder of ODSC, says the pace of change always outruns most people's ability to absorb it, which also creates opportunity. He cites a study, which he doubts, finding skills in AI-exposed jobs turn over every three years or so. He argues skills multiply rather than disappear, with COBOL programmers still in demand, and that AI is moving from tool to collaborative partner. He relays an open-source agent CEO's view that most office work is drudgery. His advice is to move from building models to designing workflows, supervision and evaluation.
Optimization in practice. A Gurobi specialist describes Toyota letting planners run tariff and supply what-if scenarios against an optimization model through an LLM interface. He also describes Total Wine, whose data-science team learned optimization from scratch when machine learning fell short, and which he says now saves a lot of money.
An oath for technologists (episode 935). Stephanie Hare favors a Hippocratic-style oath because it is an ethos rather than a licence. She points to the oath's reboot after the Nuremberg doctors' trial, and to the ethics codes of lawyers, accountants and UK civil servants. She rejects new regulation and argues most invention starts from solving problems, not from profit.
Concatenating models (episode 929). The BDH paper's author explains that its architecture scales along one dimension, neurons, so networks trained on English and French can be joined into a working multilingual model. Its 1-billion-parameter models are comparable with others of that size. The team is targeting reasoning models over billion-token contexts, such as an enterprise's million pages of documentation or a large codebase.
Pay only for value (episode 932). Schneider says Unframe, with $50 million raised, builds tailored solutions from hundreds of reusable blocks and charges nothing until agreed KPIs are met. It then licenses per solution per year, with services included.
- The speed of change, not lost skills, is the real challenge. McGovern, founder of ODSC, cites a study he read finding skills in AI-exposed jobs turn over by "30 something%", roughly every three years, though he says he doesn't quite believe it. He argues skills never go away but multiply, and that practitioners should shift from building models from scratch to designing workflows, supervision and evaluation, as orchestrators rather than builders.
- Optimization behind an LLM interface puts planners in charge of what-if analysis. A Gurobi optimization specialist describes a Toyota talk on vehicle manufacturing planning, where planners test scenarios such as tariffs swinging from 0% to 10% to 80% through natural language, without writing models. He also describes Total Wine, whose data-science team had no operations-research background, learning mathematical optimization when machine learning "was not cutting it", and saving what he calls "a lot of money".
- Technologists should hold an unenforced professional ethos, not more regulation. Hare says she favors a Hippocratic-style oath precisely because it is not enforceable: an ethos of "first do no harm" and of maximizing benefits while minimizing harms, as lawyers, accountants and UK civil servants have. She notes the medical oath was rebooted after the Nuremberg doctors' trial, and disputes the host's suggestion that profit is technology's first incentive, saying the best work comes from solving problems.
- The new architecture's networks concatenate into one multilingual model. The BDH ("baby dragon hatchling") paper's author says its models scale along a single dimension, the number of neurons, so two networks trained on different languages can be placed side by side and work out of the box, optionally with post-training. He says its 1-billion-parameter models match or beat others at that scale, and that the team is aiming at reasoning models over contexts as large as 1 billion tokens rather than releasing a 70–80-billion-parameter chatbot.
- Customers pay only once a custom AI build shows business value. Schneider says Unframe, founded around March of the previous year and having raised $50 million in total, broke the startup playbook by serving many personas and industries from day one. Its "managed AI delivery platform" assembles hundreds of building blocks into tailored solutions, with no cost until agreed KPIs are hit. Services are included in a T-shirt-sized subscription per solution per year, and some customers run 5 to 15 solutions on it.
AI Analyst LabAsk Studio AI: How It Boosts YouTube Performance
- Ask Studio ties retention to what was said in the video. The host demos a new Gemini chat in YouTube Studio, which he thinks may be experimental, on the Data Neighbor episode with Hex CEO Barry. That episode had been out a little over three weeks, with an average view percentage of 15% against the channel's usual 8–9%. The tool read the transcript and linked retention to topics: viewers stayed for the origin story (minutes 2–4), the "human element of data" discussion (5–7) and the demo, and dropped off in the product-market-fit segments.
- Its most usable advice was on the intro. It said to open with the audience's problem instead of "what is Hex", put the most engaging material in the first minute, add a short "sizzle reel" of sound bites, and keep introductions to 30–60 seconds. The host agrees his intro ran long and wonders whether the hosts should stop introducing themselves. Asked for intro lines taken from the transcript, it mostly wrote new lines that were never said in the interview, with only some verbatim clips. Partway through the chat it also lost track of which video was being discussed.
- Search and suggested traffic drove the Hex episode's click-through rate. Over the last 28 days the video's CTR was 14.6% from search and 4.7% from suggested videos. Browse brought the most impressions but at an average CTR. The tool credited a specific, benefit-led title and proposed variants built on "unlock" and "secret". It also rewrote four other titles, such as recasting the AI research engineer episode as "Become an AI Research Engineer: What It Takes & How To Start", and the host says he will adopt all four and watch the effect. It could not view thumbnails or generate an image.
- Across the channel, practical tool demos and career guides perform best. Its examples were a Replit episode on building a BI agent and episodes on becoming a data engineer or an AI engineer. It suggested interviewing founders and product people from data-tool companies, as well as recruiters, which the host notes the show has never had on. It also drafted interview questions, including asking guests about impact, which the host admits the show often skips.
- It had little to say about the audience itself. It said there was not enough data to tell when viewers are online, and that 91% of viewers watch long-form. Searches leading to the channel were about data science salaries, data engineering and Hex, with traffic from Google search, LinkedIn and three views from ChatGPT. The host's view is that the tool beats the AI features in his podcast recording software and that YouTube should build this kind of title review in as a standard feature.
Master InvestorMike Wilson: Why the Fed is NOT independent
- Wilson argues the Fed isn't independent because it has an overarching duty to help the government fund itself. He calls the "financial stability" third mandate a "madeup mandate" that gives the Fed cover to take extraordinary action beyond employment and price stability.
- He compares today's Fed-Treasury coordination to the 1940s, expecting new intervention programs if bond volatility rises as the Fed cuts rates. This is his own forecast of how policymakers will respond to fund growing deficits.
- He doesn't think the bond market is near escaping the Fed's control, but calls it a long-term risk. He frames this explicitly as a possibility rather than his base case.
Scott HanselmanWhy Postgres? and why now? with Claire Giordano
synopsis
Claire Giordano, who works on open-source Postgres at Microsoft, ends up arguing that Postgres's current momentum is no single breakthrough but a 40-year accumulation: extensibility designed in from 1986, a permissive license, no owning company, consensus-driven governance, and infrastructure kept independent of any vendor.
Origins and extensibility. Giordano joined the Citus startup in 2017, which Microsoft acquired in 2019, and later moved to Postgres core. She notes extensibility was the second design objective in Michael Stonebraker and Lawrence Rowe's 1986 design document, which is why Citus could distribute Postgres as an extension rather than a fork.
Why now. Drawing on hallway conversations at the European Postgres conference in Latvia, she lists factors: the permissive license, distributed ownership across Microsoft, EDB, Amazon, Google and others, and managed services from Heroku (2007), Amazon RDS (2013) and Azure (2017). She adds Oracle's multicore licensing pushing users to cheaper alternatives, and JSON support around 2012. Hanselman frames it as slow-and-steady standards compliance.
Community and governance. About 680 people attended the Latvia conference. Decisions happen on public mailing lists with no top-down roadmap; contributors must win over enough peers and a committer, of whom there are 31. Asynchronous I/O for reads in Postgres 18 took five years of groundwork.
Trademarks. As a board member of the Postgres Community Association, she asks anyone using "Postgres" in a product name to check the trademark policy.
Infrastructure and Azure. The source lives on project-run servers, with GitHub as a mirror; a build farm of over 200 machines spent over 229 million seconds building Postgres 18 variants. Azure Database for PostgreSQL is open-source Postgres plus non-open Azure management features, and shipped Postgres 18 in preview on GA day.
- Postgres's popularity is a "happy accident" of several factors. Giordano, who works on the open-source Postgres project at Microsoft and joined via Citus in 2017, points to extensibility as the second objective in Stonebraker and Rowe's 1986 design document, which let Citus scale Postgres horizontally as an extension rather than a fork. She adds the permissive MIT/BSD-like license, which got it into universities, and watershed moments such as JSON support around 2012, followed by JSONB.
- No single company owns Postgres, which she sees as decisive. Microsoft, EDB, Amazon, Google and smaller firms all pay upstream contributors, and some call it "the people's database." She credits managed services as trendsetters, dating Heroku's to 2007, Amazon RDS Postgres to 2013 and Azure's to 2017. Hanselman's own theory is that standards compliance and enterprise focus made it "an overnight success in 40 years."
- Features win by persuasion, not a top-down roadmap. Work is bottom-up and decided largely on public mailing lists; an author must convince enough people and find one of the 31 committers willing to commit and then maintain the change. She says this makes Postgres conservative about features, and that even a hyperscaler's engineers must lobby across company lines. Asynchronous I/O for reads, new in Postgres 18, took about five years of groundwork across earlier releases.
- Open source does not mean open trademark season. As a board member of the Postgres Community Association, which holds the Postgres and PostgreSQL names and the Slonik logo, she urges anyone naming a product with "Postgres" to read the trademark policy first. Hanselman cites a past attempt to trademark the name that the core team pushed back on.
- The project deliberately avoids depending on commercial vendors. Its source lives on project-run git servers, with GitHub only a mirror, maintained by a volunteer sysadmin team of about seven or eight, with sponsored hardware. Its build farm has over 200 "animals," and she reports over 229 million seconds of build time across Postgres 18's roughly 15-month development cycle. Azure shipped Postgres 18 in preview, in one region, on GA day, which she says it had not managed a couple of releases earlier.
Y CombinatorInside The Startup Launching AI Data Centers Into Space
Orbital data centers could remove Earth's energy and water constraints on AI compute. Philip Johnston, co-founder of StarCloud, argues data centers in space draw constant solar power, radiate heat via infrared into deep space instead of evaporating fresh water, and can scale without land or grid limits, ultimately competing on cost with the largest terrestrial data centers.
A large, low-cost deployable radiator is StarCloud's core technical bet and IP. Johnston says the main public criticism, from a viral post by Andrew McCallip, is that dissipating heat needs impractically large surface area; he counters that co-founder Ezra, a satellite-structures PhD, is leading half the engineering team to just build that large radiator.
StarCloud pivoted from beaming solar power to Earth to sending data centers into orbit instead. Johnston says the team's original space-based-solar concept loses 95% of energy in transmission and only breaks even near $50/kg launch cost; recalculating for orbital data centers instead put breakeven near $500/kg, much closer to today's falling launch prices.
StarCloud built and launched its first satellite in 15 months, versus roughly 4 years for prior space startups. The company, backed by YC after a third application attempt (under its earlier name Lumen Orbit), launched its StarCloud 1 satellite with an Nvidia H100 GPU on November 2nd and plans a second, Blackwell-equipped launch in October 2026.
Dr. Jon KrohnHow to Deploy AI Easily on Any Chip and Device (Shirish Gupta and Tyler Cox)
- The on-device AI stack is mostly glue over partners' runtimes. Cox says Dell's tooling, formerly Dell Pro AI Studio, presents one consistent API across systems such as Dell Pro Max and Pro Plus, so apps need not handle each platform's quantization parameters, toolchains and runtimes. He says it integrates eight back-end runtimes "and counting", which Dell did not build, so an app doesn't have to carry the difference between, say, an AMD NPU and an Nvidia RTX Pro GPU.
- Multi-chip routing and model management hide hardware complexity. Cox says some systems carry four or five kinds of accelerator, and the software lets an app run a model on the NPU in one session and the GPU in the next. Its model management sends the right model versions and dependencies to each platform, so an IT admin deploying to 5,000 users doesn't pick versions per machine. The command-line tools follow existing local-AI conventions rather than reinventing them.
Loss DogWhy Lower Interest Rates Actually Widen the Wealth Gap- 11.12.25
synopsis
On the network's third live show, the hosts settle into a running argument that connects markets, hiring, and sports gambling: cheap money, easy shortcuts, and quick fixes tend to help the people who already have an edge, while durable success -- in a career, a business, or a bet -- comes from narrower focus and longer time horizons than most people expect.
Mailbag: bootstrapping and loyalty. Asked about returning to speak in Mexico City, one host confirms plans to visit again in 2026, tying it to the show's growing following. A former restaurant-bar owner who sold in early 2025 and now finds acquisition prices inflated is advised to bootstrap something new rather than overpay for an existing business, though the hosts note their own bias toward building from scratch versus friends who prefer to buy and operate. A viewer who loves an underpaying social-work role is told to keep doing what they love if other needs are met, since salary reportedly ranks fourth in importance once basic needs are covered, while still using savings to invest for the gap. The mailbag's sharpest answer concerns job-hopping: one host calls frequent moves (three to six jobs in as many years) a real red flag in hiring, since it takes months just to learn a new role, and argues the biggest cost in building their companies was never a bad hire but training someone who then leaves before reaching peak value -- which is why they built businesses explicitly designed around the goal that "nobody wants to leave."
Do lower rates widen the wealth gap? The central segment argues yes, though the two hosts frame it slightly differently. One focuses on borrowing costs as a larger share of disposable income for lower earners, meaning any rate relief helps them proportionally less than it helps investors already holding assets. The other goes further, arguing that lower rates categorically inflate asset prices -- stocks, real estate -- which benefits people who already own those assets and can borrow cheaply against them, while doing comparatively little for people without capital to invest. He warns explicitly against pushing rates too low, citing Japan's three-decade near-zero-rate era and the yen's long decline as a cautionary tale, and argues for a "sweet spot" around 3-5% rather than the extremes markets sometimes demand. Both agree the bond market, not the Fed, ultimately sets real borrowing costs regardless of official rate moves, and that understanding concepts like the yield curve and risk-free-rate benchmarking is essential to setting realistic investment return targets.
Prop bets and sports integrity. A scripted rant argues that prop and parlay betting, layered onto recent player and umpire betting scandals in the NBA and MLB, risks making professional sports look "rigged" to fans who spend heavily to attend games. The host estimates sportsbook pricing runs roughly 10% wide -- compared to 5-7% for casino games, about 1% for prediction markets, and pennies wide for a stock like SPY -- arguing that spread makes sustained betting profit "virtually impossible" for ordinary bettors while creating real incentive for insiders to manipulate outcomes. The proposed fix is tighter platform-level surveillance (limiting parlay size, following unusual money flows) rather than government regulation, alongside acknowledgment that declining DraftKings and Caesars stock prices reflect growing competition from cheaper, tighter prediction markets.
Success, earned. The closing segment argues success comes from staying narrowly focused rather than expanding too broadly -- illustrated by the hosts' choice never to build a self-clearing brokerage arm at Thinkorswim, and spending six years on a single media product before adding a brokerage at Tastytrade -- and from simple longevity, comparing their own multi-decade path to Warren Buffett earning far more in his later investing decades than his earlier ones. They frame calculated risk-taking and owning bad decisions publicly as further hallmarks of durable success, while dismissing the idea that any outcome is purely luck. The show closes with a quiet, mixed market -- S&Ps and NASDAQ slightly lower, gold and silver higher, crude oil down sharply to $58 -- and both hosts running short positions in AMD, Micron, and Nvidia against long positions in bonds and the Japanese yen.
- Lower interest rates widen the wealth gap, not narrow it. Both hosts agree cheaper borrowing disproportionately helps people who already own assets, since lower rates push stock and real estate prices higher for those with the collateral and credit to borrow against them; one warns that pushing rates all the way to zero, as in Japan's three-decade experiment, does more damage than good, favoring a "sweet spot" of roughly 3-5% over either extreme.
- Losing a trained employee costs more than a bad hire. One host says the biggest expense in building their companies wasn't a wrong hire but training someone who then leaves before reaching peak value, arguing frequent job changes -- multiple roles in three to five years -- are a real hiring red flag, while staying a decade or more in one role within the same company isn't a problem so long as responsibilities keep evolving.
- Sports betting's wide spreads make sustained profit virtually impossible. One host estimates sportsbook pricing runs roughly 10% wide, versus 5-7% for casino games, about 1% for prediction markets, and pennies for a stock like SPY, and argues that gap, combined with recent player and umpire betting scandals in the NBA and MLB, means betting platforms and leagues need tighter oversight before public trust in the games themselves erodes further.
- Staying in one lane is framed as the real driver of success. The hosts cite their own history of deliberately never building a self-clearing brokerage arm, and spending years perfecting one product before adding anything else, arguing most failed businesses collapse because founders chase too many ideas before proving even one, and that long-term success is mostly a function of longevity rather than any single breakthrough.
Crosstalk SolutionsWhy I Didn’t Run Fiber (And Why You’re Wrong)
- A wireless bridge beat fiber for a friend's Airbnb back house. At the property, in Los Angeles, fiber to the back house would have meant a run of over 200 ft and boring under a concrete walkway: an 811 utility-marking call, digging, renting a horizontal borer or auger, conduit, pulling and cleanup, taking days, with gas-line risk and lost Airbnb revenue while the network was down.
- Equipment cost is similar; labor and rentals decide it. The host says the bridge installed in hours, was cheaper overall and delivered a solid 300 Mbps across the property, arguing a properly dialed-in wireless bridge can be as reliable as a hardwired link.
Stock Tok$Stock On My Radar to BUY 👀
- This post includes paid content that is not available publicly. See the original source.
Scott HanselmanEPISODE 28 - Scott and Mark Learn To… Second Brain
- A single informal paper list carries Mark's to-dos. Mark keeps one office-supply notebook, work in the left column and home in the right, carries unfinished items forward under a new horizontal line, and throws old notebooks away. He says many studies show offloading to a list removes a load from the brain, and argues a more organized system costs effort whose return he doubts; he puts links, AI hallucination screenshots and papers on jailbreaks in OneNote instead.
- Procrastination can be background processing, not avoidance. Mark calls himself an "incredible procrastinator" but says he is subconsciously working on a task and sizing its effort before he starts. Hanselman likens it to a background agent and says his own last-minute talk preparation feels like rumination rather than avoidance, adding that once you find a system that works, keep it, while people early in their careers should try several.
- Hanselman relies on paper, a reMarkable and the calendar. He has kept notebooks for about 20 years, back to at least 2007, and says writing things down is why he sleeps well; his reMarkable e-ink tablet syncs everything, which shows him how much he has offloaded. He argues the shared calendar removes the most "psychic weight" in a household, and recommends iPhone geofenced reminders ("remind me next time I'm at Fred Meyer to buy milk") as underused.
- Neither host has a true second brain. Hanselman says he keeps wishing for an organized Obsidian or OneNote knowledge base but falls back on paper; asked whether AI should do retrieval over their notes, they joke about handing it to AI "interns" and conclude there is only "a hierarchy of brains." Both call the episode rambling and "not our finest hour."
AI Analyst LabBecome an AI Research Engineer: What It Takes & How To Start | Ep41
synopsis
The hosts of the Data Neighbor podcast interview Sandy Besson, an AI research engineer and ecosystem lead at IBM Research. The conversation ends on two positions she holds. First, an LLM's value comes from the system built around it, so reliability should be enforced at the framework level rather than left to the model's reasoning. Second, data scientists should learn above and below their stack and show it with public projects.
The role. Sandy's team runs like an incubator. It turns research from IBM and other labs into applied product, builds first-of-their-kind use cases, and open-sources the results. She tries to experiment with something new every week, at the time of recording Claude's newly released skills, and folds what works into an open-source agent framework focused on security and reliability.
Agents and frameworks. She defines an agent as an LLM calling tools in a loop until a task is done, with the framework as the code that actually executes the tools. Her team's "requirement agent" restricts which tools are exposed at each step, which prevents loops, skipped validation and premature answers, and lets smaller models compete. A co-host counters that models are improving quickly. Sandy concedes the point but expects the next leap to come from new architectures.
Tools and access. For dabblers she points to low-code builders and to vibe-coding sites that let a non-front-end developer build a website over a weekend, while she herself builds in code. The hosts compare workplace access. Their companies hold several enterprise licenses, and the host's legal-tech employer allows customer data only with zero-retention providers. Sandy concludes that no provider has clearly won.
Evals. Sandy says no benchmark tests frameworks themselves, as opposed to models, and that failure rates move whenever the model or the tools change. She separates LLM-as-judge and agent-as-judge validation from benchmark and feedback evals. The host adds that silent model updates are a sharper version of classic drift.
Systems and careers. Sandy sees coding agents and judges turning from single steps into systems, and a co-host notes that system design has always been the core skill. Her career advice for data scientists is to learn below and above their stack and build in the open, because she judges candidates by repos that show how quickly they learn. Along the way she observes that many of the AI people who bridge research and public-facing work were once theater kids, herself included.
- An agent framework is the glue that lets an LLM act. Sandy Besson, an AI research engineer on a product incubation team at IBM Research, paraphrases Harrison Chase's definition of an agent: an LLM calling tools in a loop until a task is complete. Because an LLM only produces language, the framework is what executes the tools, sets the configuration and adds system prompts. She works on an open-source IBM agent framework (captioned "BAI") focused on security and reliability, with middleware for guardrails, content safety and authorization.
- Enforce tool order in the framework, not through the prompt. Her team's "requirement agent" exposes only the tools an agent is allowed to use at a given moment. That way the agent can't loop, skip a validation tool or stop early on a hallucinated answer. She says this lets much smaller models perform nearly as well, comparing it to limiting an elementary-school student's options rather than relying on a college student's judgment. A co-host argues that models are improving fast enough to make this help unnecessary, and Sandy says that is possible but that the next leap will need new architectures.
- No benchmark yet evaluates agent frameworks themselves. Sandy says existing benchmarks test tool calling at the model level, and that a middleware failure rate shifts whenever the model or the available tools change. She expects to need custom eval sets. She separates LLM-as-judge and agent-as-judge validation from eval proper, meaning benchmarks or user feedback. The host notes that a silent model checkpoint change can alter system-prompt behavior, an amplified form of classic model drift.
- Teams license several AI providers and switch often. A co-host's employer licenses ChatGPT, Cursor and Claude Code per person on justification. The host, who works in legal tech, says customer data goes only to providers with zero data retention, and even 30-day retention rules them out. Sandy reads the multiple licenses as a sign the industry has not placed its bet yet. She prefers tools that don't write files directly, and the host says none he has tried handles large Jupyter notebooks well.
- Data scientists should learn above and below their stack. Sandy says narrowly specialized roles are becoming less common. When she reviews résumés from people with strong data science skills, she asks for linked repos showing an application or a backend they built, to judge how fast they learn. Her own route went from performing arts through a data science master's, consulting and IBM acquisition integration to applied research after GPT-3, and she has been at IBM Research about six months.
Technology and FriendsMichele Leroux Bustamante on Cloud Security at Scale [EPISODE 880]
synopsis
Leroux Bustamante, who runs a cloud consultancy and founded the authorization product Policy Server, argues that security at scale has to be built layer by layer from the design stage: topology, identity, authorization, platform, then verification. Doing it late, or with dev unlike prod, costs more.
Identity. She says first find out where each user population lives: consumers in a CIAM directory, corporate users in Entra ID, and enterprise tenants reached by home realm discovery, with social logins added to cut friction. She calls missing MFA the leading cause of breaches and names passkeys as its most accessible form.
Authorization. Roles in tokens bloat them and multiply; she cites 200 roles reducible to about 20. The app should own the rules, or a central product (her own, or OPA) should hold them where audit matters.
Platform. Her short list for Azure: a web application firewall up front, separate networks for data, managed identities and Private Link. Giard says he has moved fully to managed identities over the past 12 months, and she adds that expired keys take sites down.
Verification. She uses NIST CSF's six categories, scored one to four for a CISO, with Defender and Sentinel. They are useless unless someone acts on them. AI helps digest reports, but hallucinations keep a human in the loop.
Closing advice. Plan security inside the SDLC early, and make dev match prod's security posture so the pain is not discovered in QA.
- Authorization belongs in the application, not the identity system. Leroux Bustamante says putting roles or permissions in the token bloats it until it gets blocked, and multiplies roles: she has seen systems with 200 roles that permission-based design could cut to about 20. She advises letting the app own the rules, or centralizing when there are many apps and governance, audit and role history matter, using a product such as her own Policy Server ("shameless plug") or the open-source Open Policy Agent (OPA).
- Missing MFA is the number one cause of breaches. She ranks the options: email is "a terrible idea", SMS is extremely common, banks included, but no longer preferred, authenticator apps and FIDO keys such as YubiKey suit technical users, and passkeys are now the most accessible form for consumers.
- A web application firewall in front is the first platform control. For an Azure deployment she recommends turning on a WAF via Front Door or App Gateway to catch the OWASP top-10 failures, putting data on a separate network from app services, using Private Link, and using managed identities everywhere so no secrets are stored. She says she has heard of many sites that went down because a key rotation was not known in advance and simply expired.
- Measure posture against NIST CSF, and staff the monitoring. She maps tools such as Defender and Sentinel to the framework's six categories and gives a CISO a dashboard scoring each from one to four, but says turning those tools on without someone acting on the output is useless. When Giard asks about AI for digesting reports, she says it saves time but can hallucinate, so a human must stay in the loop "for a while".
- Dev should match prod's security posture from the start. She sees customers whose dev environments differ from production and who lock developers out of logs and audit trails in upper environments; the pain then lands in QA, and developers get blamed for delays such as an API gateway that exists in QA but not in dev. Her case is to shift left: plan security in the SDLC up front, because product owners "don't sell security features" and budgets often rise only after something goes wrong.
Dr. Jon KrohnWhy Move AI Workloads from Cloud to PC? (Shirish Gupta and Tyler Cox)
- Unequal inferences mean AI will run hybrid, cloud to PC. Gupta says customers who start in the cloud and scale to production find that costs skyrocket or on-prem compute runs out. Citing a prediction that 90% of AI workloads will be inference by the turn of the decade, he forecasts workloads running seamlessly across cloud, edge and PC fleets.
- The model should match the "quality of tokens" a task needs. Gupta's example is coding: autocomplete can run on a small local model rather than burdening a GB300 superchip, while a full-scale refactor may need a frontier model. He expects much intelligence to go into routing workloads, and AI PCs to play an important role in enterprise AI strategy.
- Local inference cuts latency, keeps queries private and stays available. Gupta says running on the PC removes about 200 milliseconds of cloud latency, which matters for voice and video. He argues that even when the data sits in the cloud, a local model keeps the query and its inputs and outputs private, whereas he says public cloud model queries are "completely indexable by search engines". A recent AWS outage, he adds, showed that on-device AI is always available.
GCastIncorporating Playwright Tests into a GitHub Action [GCast 210]
- The local test repo must first get a GitHub remote. Building on the Playwright test and Azure Playwright workspace from the series' first two videos, Giard creates an empty GitHub repository with the same name as the local one, links it with
git remote add originplus the repository URL ending in.git, and pushes withgit push -u origin main. - Repository secrets keep workspace and identity details out of the YAML. Under Settings > Secrets and variables, Giard creates four Actions secrets: the Playwright service URL (copied from the workspace's Get Started page), the Azure client ID, the Azure tenant ID and the Azure subscription ID. He first fills them with an "X" as placeholders, because each value comes from a different place.
- An Entra app registration supplies the identity the tests run under. Giard registers an app limited to accounts in his own organization's directory, with a Web redirect left blank. Its application (client) ID and directory (tenant) ID become two of the secrets; the subscription ID comes from Subscriptions in portal.azure.com. He then assigns the app the Contributor role on that subscription through Access control (IAM), noting that Contributor is listed among the privileged administrator roles.
- A federated credential ties that identity to one repository branch. In the app registration, under Certificates & secrets > Federated credentials, Giard adds a credential using the "GitHub Actions deploying Azure resources" scenario, giving his GitHub account as the organization, the repository name (he notes it is case-sensitive), entity type branch, and the main branch.
- The workflow YAML runs the tests in Azure and keeps the report. Giard's replacement YAML triggers on a push or pull request to main, requests ID-token write and contents read permissions for Entra, and runs one test job on the latest Ubuntu with a 60-minute timeout. Its steps are checkout, Azure login using the three ID secrets, installing dependencies with npm,
npx playwright testagainst the service config file from the previous video with a set maximum of workers and a report format, and an always-run upload of the Playwright report folder as an artifact, its retention changed from 5 to 10 days. A commit and push started the run, which took about a minute and a half, and the report downloads from the run page as a zip file.
Master InvestorMike Wilson: We are in a new inflationary regime
- Wilson dates the start of a new inflationary regime to the 2020 pandemic stimulus, and expects such regimes to last 30–40 years. He argues governments are now running the economy "hot," trading higher inflation for stronger real GDP growth, which the Fed is tolerating.
- He argues this regime justifies paying a higher equity multiple relative to bonds. This is his own, more trading-oriented call, which he distinguishes from the long-term view of "legendary" investors like Howard Marks.
- He frames the regime as cycling roughly two years on, one year off, with a new positive cycle now beginning. He points to accelerating inflation alongside a Fed on hold or cutting rates as "a very good earning story."