Souk's Notes - Unofficial Summaries
GCastRunning Tests in an Azure Playwright Workspace [GCast 209]
- An Azure Playwright workspace runs existing Playwright tests in the cloud. It is created in the Azure portal by searching resources for "Playwright workspaces" and giving a resource group, subscription, name, region (only a few regions offer it) and optional tags; creation takes about 30 seconds, and the workspace's Get Started tab lists the steps for wiring up a test project. The project is the one from the previous episode.
npm init @azure/playwright@latestadds the cloud config to the project. Run from bash, it created aplaywright.service.configfile (running on Linux with default Azure credentials) and added authentication settings to existing files; the recommended identity is Entra ID, so Giard signs in withaz loginand picks the tenant holding the workspace.- Tests reach the workspace through an environment variable and the service config.
PLAYWRIGHT_SERVICE_URLis set to the workspace endpoint, thennpx playwright testruns all tests or a named file such asexample.spec.js, with--configpointing at the service config — renamed from.tsto.jsbecause the project uses JavaScript. A maximum worker count is optional and lets many tests run simultaneously; his three tests passed in the cloud on one worker. - Persist the endpoint in a
.envfile instead of in memory. Installingdotenvwithnpm i --save-devand addingrequire('dotenv').config()to the config file makes it read thePLAYWRIGHT_SERVICE_URL=name-value pair (withoutexport) from.env; Giard notes a store such as Key Vault might be more secure. Rerun this way, the three tests passed again.
Polis PanditHow to Compete With AI Writing
- His central analogy: AI will do to writing what photography did to painting — not kill it, but force it to abandon what the machine already does well (fast, cheap, accurate depiction) and push into what it can't reproduce, the way photography pushed painters toward Impressionism and beyond.
- He argues LLMs cannot replicate genuine human experience — sensory detail, real suffering, a first kiss, the birth of a child — so writing grounded in real lived experience, especially personal essay, is terrain AI can't compete on; he cites his own essay on leaving his corporate job as an example only he could write.
- Specific craft advice: inject theatrical, surprising style (he cites Justice Scalia's opinion-writing as a model), use unexpected analogies and metaphors, and deliberately keep imperfections — fragments, atypical structure, even a stray typo — as visible signals of humanity against AI's "too pristine" prose.
- He credits embracing "weirdness" — an initially uncertain comparison of Tokyo to Tim Duncan and Hong Kong to Allen Iverson — with driving one of his best-performing pieces (366k+ views), arguing writers should risk the "weird zone" rather than compete with AI on safe, formulaic ground.
Dr. Jon Krohn939: Mixture-of-Experts and State-Space Models on Edge Devices — with Tyler Cox and Shirish Gupta
synopsis
Dell's Shirish Gupta and distinguished engineer Tyler Cox argue that mixture-of-experts and state-space models, normally awkward to run locally, can be deployed across an enterprise PC fleet as easily as in the cloud. Their case rests on IBM's open Granite 4.0 hybrids and on Dell's AI PC toolkit, which hides silicon diversity. Their position is that AI will be hybrid, with each inference routed to the cheapest engine that does the job.
Why the toolkit exists. Gupta says what was Dell Pro AI Studio, being renamed at recording as part of the Dell AI Factory, grew from customer research starting early the previous year. It addressed three problems: diverse silicon and runtimes slowing adoption by customers and software vendors, no enterprise-grade lifecycle management for on-device AI, and deployment far harder than the cloud. It has three parts: a PC model catalog in the Dell Enterprise Hub on Hugging Face; an on-device framework that lets existing cloud workloads point to the PC as localhost; and lifecycle management through the Dell Management Portal in Intune.
Why run locally. Gupta argues that inferences are not equal: scaling in the cloud sends costs up, and on-prem runs out of compute. Citing a prediction that 90% of AI workloads will be inference by decade's end, he says code autocomplete belongs on a PC while a full refactor may need a frontier model. Local inference saves about 200 milliseconds of latency, keeps queries private, and, as a recent AWS outage showed, stays available.
How it works. Cox calls the software largely glue. It integrates eight back-end runtimes "and counting" behind one API, routes models between NPU and GPU, and pushes the right model versions and dependencies to each machine in a 5,000-user rollout.
The Granite models. Gupta presents the IBM partnership as bringing Apache 2.0 models with open weights and training data that sit near the top of HELM. Cox lays out four variants: a 3-billion-parameter dense transformer Micro; a hybrid 4H Micro; and mixture-of-experts hybrids 4H Tiny (7 billion total, 1 billion active) and 4H Small (32 billion, 9 billion active). Sparse models, he explains, keep the knowledge of their total size and the speed of their active size.
State space models and hybrids. Cox traces state space models from 1960s spaceflight control to S4, S6 and Mamba. Granite uses a 9:1 ratio of Mamba to attention layers, since attention keeps local sensitivity that Mamba loses. Training data reaches 512k context and validation 128k. Eight 128k sessions on Micro use about 15 gigabytes against about 80 for a pure transformer. He cites a Meta FAIR paper that favors hybrids, and expects more of them.
Hardware and failure. Gupta's Lunar Lake figures include 88% more Teams battery life and up to 10x AI CPU performance. On the MIT claim that 95% of enterprise AI projects fail, he prescribes clear goals, humans in the loop through production, and traditional machine learning where it suffices.
- On-device AI needed to be as easy to deploy as the cloud. Gupta says Dell's toolkit, formerly Dell Pro AI Studio and being renamed within the Dell AI Factory, answered three problems: silicon and runtime diversity slowing adoption, no enterprise-grade fleet lifecycle management, and the gap with cloud ease. Its parts are a PC model catalog in the Dell Enterprise Hub on Hugging Face, an on-device framework that lets cloud workloads point at the PC as localhost via existing API standards, and lifecycle management through the Dell Management Portal in Intune.
- Inference should be routed by the token quality each task needs. Gupta cites a prediction that 90% of AI workloads will be inference by the end of the decade and forecasts hybrid AI across cloud, edge and PC. In his example, code autocomplete belongs on a small local model and a full refactor on a frontier model. Local inference also removes about 200 milliseconds of cloud latency and keeps queries private; he says a recent AWS outage showed its availability value.
- Hybrid Mamba-attention models keep long context cheap. Cox says Granite 4.0 H models interleave Mamba and attention layers at a 9:1 ratio because Mamba handles global context well but loses attention's local sensitivity. He cites IBM's figure of about 15 gigabytes for eight 128k-context sessions on the 3-billion-parameter Micro, against about 80 for a pure transformer. He relays a Meta FAIR paper recommending more hybrids for training and inference efficiency, and expects more of them.
- Sparse Granite variants pair large-model knowledge with small-model throughput. Cox describes Granite 4H Tiny (7 billion total, 1 billion active parameters) and 4H Small (32 billion total, 9 billion active) as mixture-of-experts hybrids, alongside 3-billion-parameter dense Micro models. He says Dell tunes inference parameters such as context length per device, from a 16-gigabyte AI PC to an RTX Pro 6000 Blackwell with 96 gigs of VRAM, and had day-zero support.
- AI PCs should be judged on CPU, GPU and NPU together. Gupta compares a Dell Pro Plus with Intel Core Ultra 200V (Lunar Lake) against an older non-AI-PC chipset: 88% more battery runtime in Teams meetings, 4.8x graphics, almost 10x on-device AI CPU and 4.2x AI GPU performance, over 6x on Mistral 7B and Llama 3.1 text generation, and up to 5.6x on Stable Diffusion 1.5 image generation.
All Things FinancialG-SIBs buy new treasury issuance and to raise cash to settle the SOFR exposes system stress 11-10-2025
- New Treasury issuance forces GSIBs to pre-fund cash, not signal a liquidity crisis. All new Treasuries go first to the nine GSIBs, who are obligated to buy at auction; ahead of a big offering they raise cash by lending securities overnight in the repo market (SOFR), pushing SOFR above the Fed's IORB rate.
- Last week's SOFR spike to ~29bps over IORB tracked a $125B Treasury offering, not a funding crunch. The Nov 4-6 3/10/30-year auctions totaled $41B, $41B and $43B; social-media "doom" posts framed the SOFR jump as a great-financial-crisis-style liquidity shortage.
- SOFR fell back below the Fed funds rate the day after settlement (Nov 6), confirming the cash-accumulation explanation. The author argues this pattern — banks stockpiling cash ahead of mandatory GSIB purchases — will recur with future large issuance and urges watching the "plumbing" rather than trusting alarmist commentary.
Loss DogThe $3.6M You're Leaving on the Table Over Your Career - (Debut Ep.) 11.10.25
synopsis
Launching the network's debut episode after a two-week hiatus, the hosts set the show's founding thesis in its first hour: narrowing the wealth gap isn't a policy fix, it's an individual-education problem, and the two things people are least taught -- how to use leverage and how to know their own worth -- are exactly what the coming Lost Dog platform is built to supply.
Format and mission. The hosts introduce One Lucky Dog as a three-day-a-week live show (Monday through Wednesday initially) built around recurring segments -- News You Can't Use, Unleashed, Bad Dog, Wow, and a closing market discussion -- distinct from their prior trading-focused shows. They stress the new show is not about trading itself but about entrepreneurship, self-worth, and building equity, with the forthcoming Lost Dog software (targeted for launch by late November, aimed at the $50,000-$300,000 salary range) positioned as the mechanism: an AI-driven tool combining math, data, and specialist input to give users a "two-sided market" on their own value.
Fiscal responsibility versus leverage. The day's central Unleashed segment, framed around the question of teaching fiscal responsibility amid national debt, pivots quickly into an argument that responsible use of leverage and risk-taking, not conservatism, is what separates people who reach the top of a "life distribution curve" from everyone else. Both hosts describe themselves as major risk-takers rather than fiscally cautious, arguing college debt, career risk, and financial leverage are tools worth using deliberately even though nobody formally teaches how; one host credits his own risk tolerance to watching his father go "all in" at every life stage despite claiming to bet cautiously, while the other notes his more risk-averse upbringing by contrast. They agree that learning to take calculated risk matters "a thousand times" more than learning fiscal restraint, since a mistake can be recovered from but an opportunity never taken cannot be, and connect the idea directly to building Lost Dog itself as a bet neither host had to take at this stage of their careers.
A pardon, a pay package, and a segment on differentiation. In the first "Bad Dog" rant, one host argues Trump's pardon of Binance founder CZ -- who, like his exchange, pleaded guilty to enabling money laundering -- reads as monetized clemency rather than mercy, pointing to a reported $2 billion deal between Binance and Trump's family crypto platform to host its stablecoin, and to a described market where lobbyists reportedly charge millions just to secure a meeting about a pardon. Later, discussing Tesla's board-approved trillion-dollar pay package for Elon Musk, both hosts say plainly they would have voted against it, contrasting the package with a reported $54,000 average Tesla salary with no equity attached; they debate, without fully resolving it, whether the real problem is one outsized number or executive compensation broadly, noting that even Bill Gates would have crossed $1 trillion in net worth had he not sold shares the prior summer -- though as a company founder rather than a hired executive being separately voted a package. The "Wow" segment on self-differentiation argues that normal, unforced conversational ability and likability matter as much as technical skill in hiring, since turnover is one of the most expensive costs a business can absorb, and that speed of decision-making is the clearest tell of future success: the hosts recount trading equity in their own company for a marketing deal in roughly 15 minutes, closing a quarter-billion-dollar exchange sale in a seven-minute phone call sealed on a verbal promise, and a hiring test in which only one candidate among several with perfect math SAT scores was willing to guess an answer on the spot rather than admit uncertainty.
The software demo and two caller exchanges. The show's first public preview of the Lost Dog platform uses a sample profile to walk through its methodology: an industry-average salary baseline ($185,599 in the demo), a computed personal value ($216,000) based on education, experience, and location, a negotiating target ($235,000), and a projected lifetime "opportunity cost" of not knowing one's worth ($3.64 million) -- all framed explicitly as the product's own methodology and marketing demonstration rather than an independently verified calculation. A first-time caller pursuing an MBA and worried about taking career risk without a financial safety net is told to reframe the question away from "hedging," since real risk-takers don't hedge; instead, he's advised to act whenever the probabilistic odds of success feel favorable (roughly 65-75%), on the reasoning that almost no setback is actually final and that both hosts were rejected by "a who's who" of major financial institutions before eventually succeeding. A second caller, a Lake Forest resident, questions the fairness of using a $185,000 CEO salary as an example given that local teachers and firefighters earn comparable pay with overtime; the hosts clarify neither of them takes a salary from Lost Dog, note they've historically capped their own pay at four to five times their lowest-paid employee, and cite the average S&P 500 CEO's $21 million annual pay and roughly seven-year tenure as the real benchmark the software is built to expose, alongside the separate point that 17% of employees at one friend's company reportedly decline free, fully matched 401(k) contributions -- a gap in market participation the hosts see as central to the wealth divide they're trying to address. The show closes on a sharply reversing market -- the NASDAQ swinging roughly 900 points and the S&P about 3% from Friday's lows to Monday's highs -- with both hosts calling the rally overextended given AI stocks "priced for perfection," while stopping short of predicting exactly when the move reverses.
- Lost Dog's software claims to quantify a lifetime "opportunity cost." Using a demo profile, the unreleased platform contrasts a $185,599 industry-average salary with a $216,000 computed value and a $235,000 negotiating target, then extrapolates the gap over a career into a $3.64 million lifetime figure -- a marketing demonstration of the product, not an independently verified result.
- A presidential pardon is framed as a direct conflict of interest. The rant notes Binance pleaded guilty to money laundering and founder CZ personally pleaded guilty to related charges, then points to Trump's family crypto platform closing a reported $2 billion stable-coin hosting deal with Binance around the same time, arguing the pardon reads as monetized clemency rather than mercy.
- Both hosts say they'd have voted against Musk's pay package. They call the trillion-dollar Tesla compensation deal emblematic of executive-pay excess -- contrasting it with a reported $54,000 average Tesla employee salary with no equity -- while still debating whether extreme CEO pay is a symptom worth fixing directly or an inevitable feature of scale, noting even Bill Gates would have crossed $1 trillion in net worth had he not sold shares the prior summer.
- Fast, binding decisiveness is treated as the clearest predictor of successful dealmaking. The hosts recount trading equity in their own company for a marketing deal in about 15 minutes, and separately closing a quarter-billion-dollar exchange sale in a seven-minute call sealed only by a verbal "yes"; in one hiring test recounted on air, only one candidate among several with top math SAT scores was willing to guess an answer on the spot rather than say "I don't know."
Dr. Jon KrohnUsing AI to Detect Early Signs of Alzheimer’s (with Marc Dupuis)
- Image processing could catch brain shrinkage too subtle for doctors. Dupuis says his master's research at Imperial College London, about 10 years before recording, used image processing on brain scans to detect early Alzheimer's. Back then brain shrinkage was one of the early physiological signs, but too subtle for a doctor to see until months later, when it was too late; Alzheimer's is hard to reverse, but early detection might slow it. He notes other markers have advanced a lot since.
- Brain-signal decoding is mostly a hard signal-versus-noise problem. Dupuis says much of his broader biotech master's covered signal processing of brain waves, for example from implants in monkeys controlling robotic arms. The difficulty is separating intent from everything else in billions of connections, such as whether the brain is reacting to what the eyes see or to the attempt to move the arm.
Dr. Jon KrohnSemantic Layers 101: Keeping AI Analytics Consistent (with Marc Dupuis)
- AI analytics risks redefining a metric every time someone asks. Dupuis calls it a "justified fear" that AI given to everyone will reinvent metrics with each question, even when humans agree on definitions. He points to semantic layers and guardrails, with data teams supervising and working with business stakeholders, as the answer, and says the data team's job of keeping metrics consistent does not change.
- Dashboards will be fewer but matter more. Dupuis argues the dashboards that survive will be curated by data teams and track the core metrics such as churn, ARR and retention, with the team spending more time making sure they are correct. AI's role is the questions that have not yet reached the north-star metrics or OKRs.
- Explore a metric with AI before formalizing it. Dupuis's process is for a data scientist and a product manager or CSM to explore a hypothesis metric with AI on raw or curated data, often finding the question was wrong, rather than modeling it into the warehouse first. If the metric holds steady for a month or a quarter, it then goes through the "proper channels" into the warehouse, data models and a dashboard.
Mr. GlobalShut Up #education #college #truthmatters
- The reply rejects calling a degree a participation trophy. The short opens with one speaker claiming a degree only certifies four paid years of bare-minimum effort; the respondent counters by asking whether people from Kentucky upset that they could not vote in New York's election — prompting an official announcement that they could not — were college educated.
- One bad case does not justify denying everyone. The respondent concedes some college-educated people "aren't terribly bright" but calls the opposing view absolutist, comparing it to arguing that because some people commit SNAP fraud, nobody should get SNAP.
Dr. Jon KrohnAI Agents vs. Code Assistants: The Next Evolution in Data Work (with Marc Dupuis)
- AI plays two distinct roles in data work: coder and embedded collaborator. Dupuis says the first is code generation, writing, debugging or optimizing SQL and Python; in Fabi, he says, 90 to 95% of the code run is AI-generated, with manual correction possible. The second goes beyond automating a dashboard or a chart being sent: an agent embedded in the workflow interprets the table or chart and suggests the next best action. Fabi's AI summarization cells, for instance, turn any number of data frames into an executive summary using the user's choice of LLM.
Mr. GlobalSeems Legit #trump #news
- The clip is a two-line exchange with no market content. One unidentified speaker says people in "this room are going to prison, myself included", and another shouts back "Stop breaking the law". The transcript names no speaker and gives no context beyond the title's "Seems Legit" and #trump tag.
Dr. Jon Krohn938: Frontier AI Agents for Data Science — with Sphinx’s Rohan Kodialam
- Coding agents fail at data science because they never interpret the data. Kodialam, co-founder and CEO of Sphinx, which the host says has raised $9.5 million, argues data work differs from software engineering the way a poem differs from a technical paper. His example: a frontier coding agent handles a clean linear regression, but with outliers, invalid rows or non-linear shape its answers become "very variable". It treats the task as code that runs and outputs an R-squared, "mission accomplished", without looking at the data.
- Sphinx treats data as its own modality, built for AI to read. Kodialam says Sphinx builds only a representation-learning layer that turns data into context, and leaves text and code to frontier models. His illustration is five years of Walmart's stock price printed as about 80 pages of numbers, which is how an LLM sees data today, against a candlestick chart a human grasps at a glance. He adds that models are also poor at reading charts, giving vague answers on the correlation in a scatter plot.
- The agent works from the Python kernel and executes on the user's side. Kodialam says most coding agents treat the bash terminal as home base, while Sphinx uses the IPython kernel so it can transform in-memory objects, which makes Jupyter a natural interface. It works as a side chat operating at the level of actions, writes ordinary SQL, Python and markdown cells for auditability, and runs on the user's compute so Sphinx never takes their data. Users typically start with small tasks such as joining two tables before moving to fully agentic flows.
- Better tools mean more data scientists, not fewer. Kodialam says existing data-heavy clients mainly get "five times faster", while the more transformative case is a consumer packaged goods customer now hiring more data scientists because each produces more. He says some early customers' data teams have doubled, and cites a Harvard Business School figure that 80% of CEOs call data a top priority but 20% actually invest in it.
GCastCreating UI Tests with Playwright [GCast 208]
- Playwright is installed by an interactive npm initializer. With Node.js and an IDE (his is Visual Studio Code) as prerequisites,
npm init playwright@latestasks for TypeScript or JavaScript (he picks JavaScript), a folder for end-to-end tests (defaulttests), whether to add a GitHub Actions workflow (yes), and whether to install the Playwright browsers now. - By default every test runs once in each of three browser engines. The config installs Chromium (the engine under Chrome and Edge), Firefox and WebKit (the engine under desktop Safari), and engines can be added or removed later; so the two generated sample tests ran as six under
npx playwright test, andnpx playwright show-reportserves the per-browser results from a local web server. - A test is an awaited sequence of page actions followed by assertions. The samples
await page.gotothe playwright.dev URL and expect the title "Playwright", or click the link with role and name "Get started" and expect a heading named "Installation" to be visible; the test passes only if the assertion holds. - Codegen records clicks and keystrokes in a browser as test code.
npx playwright codegen <URL>opens an emulated browser beside a code window that writes a line per action; toolbar buttons add assertions (assert visibility, assert text), and the output can be switched from Node to Python, .NET with MSTest or NUnit, or Java with JUnit. Giard pasted a recorded run of a simple form into the spec file as a third test. - The recorded test failed only in WebKit, which he left unresolved. Three tests across three browsers made nine runs; the new one passed in Chromium and Firefox but in WebKit hit the 30-second timeout because it could not find and check a radio button. Giard guesses it may relate to Safari not being supported on Windows and calls it something to explore.
Master InvestorAndrew Ross Sorkin: Why it pays to be an optimist
- Churchill, on a 1929 US lecture tour, watched the crash unfold from the floor of the New York Stock Exchange and later praised America's "unshakable faith in a golden future." Sorkin recounts Churchill losing his own fortune in the crash, witnessing what he believed was someone jumping from a window from his Plaza Hotel room, and still writing that the American "speculative machine" is "not built to prevent crises, but to survive them."
- Sorkin argues it has paid, historically, to be a "professional optimist" in US markets rather than a skeptic. He says that even through the Depression and the 2008 crisis, staying invested — and even speculating — has beaten holding cash over the past 100 years.
Scott HanselmanThe Past Still Boots with the Interim Computing Museum's Stephen Jones
synopsis
Stephen Jones, formerly senior engineering manager at Living Computers: Museum + Labs and now at the Interim Computer Museum near Seattle, ends up arguing that vintage computing is best preserved as working, touchable machines sustained by a membership community, and that the Living Computers collection survives rather than having been scattered.
Running machines, not artifacts. Jones traces remote access to vintage systems back to a 1997 call for accounts on a PDP-10 clone, well before the museum opened in 2012, and says the passion predates Paul Allen's money. Computers behind glass celebrate only industrial design; Living Computers' seven full-time engineers were what made it interactive. A community model. ICM is a new nonprofit, chartered in April, focused on covering lease and utilities, with 356 members and sponsors and a first surplus quarter. Jones borrows from SDF.org, a 1987 BBS that became a nonprofit Unix ISP and self-sustaining on memberships. A $36-a-year membership gives access to 26 remote systems, and members can visit Saturdays to Mondays, 10 to 4, by appointment. Hanselman, a member, promotes Benevity donation matching.
Why the past matters. Hanselman worries younger developers are losing low-level understanding; Jones points to the range of ages at events like the Portland Retro Gaming Expo and to sidelined ideas from Lisp machines to the Connection Machine. Both praise tools that connect old hardware to modern systems, and ICM's lab of about 3,000 magnetic media artifacts.
What happened to Living Computers. Jones calls GeekWire's June 2024 coverage cruel, saying only about 30 LCM items were auctioned. Most of the collection is at the Mims museum in Roswell, Georgia, or at ICM's 3,000 sq ft space in Tukwila. The name "interim" conveys urgency: everything is ephemeral.
- Computers belong running, not behind glass. Jones, formerly senior engineering manager at Living Computers: Museum + Labs and now at the Interim Computer Museum, argues a machine behind glass only celebrates industrial design, while a running one conveys what it did. He says Living Computers had seven full-time engineers, "unheard of" for a technical museum, since most outsource exhibit building and interpretation.
- The new museum survives on community, not a benefactor. ICM, a nonprofit that got charitable status in April and joined Benevity for matched donations in May, focuses on covering lease and utilities; Jones reports 356 members and sponsors and its first surplus quarter. He draws the model from SDF.org, a 1987 BBS turned nonprofit Unix ISP that he says survived the dot-com bust on membership; a $36-a-year membership gives access to 26 remote vintage systems, real, hybrid and simulated.
- Living Computers' collection was not gutted, contrary to coverage. Jones calls GeekWire's June 2024 articles cruel and a missed opportunity, saying only about 30 LCM items went to the Christie's auction alongside space art and other memorabilia. He says most of the rest is at the Mims museum in Roswell, Georgia, which took the Living Computers name and is preparing an Apple 50th exhibit, or at ICM's roughly 3,000 sq ft space in Tukwila, near SeaTac, open to members by appointment.
- Legacy media is degrading, which is why the museum is "interim." ICM's preservation lab holds about 3,000 magnetic media artifacts, from disk packs to DECtapes, and uses techniques such as baking tapes to restore binders. Jones says the name expresses urgency: data remains trapped on esoteric carriers while the media slowly degrades.
- Retro tools bridge old machines and the present. Hanselman argues "we're forgetting how to drive stick shift" and says running multitasking Unix on a simulator taught him how the cloud descends from the Altair. Both cite adapters that put old machines online and read old disks; Hanselman says he recovered a never-released DOS 4 beta that Ray Ozzie sent him on floppies. Jones adds that history's sidelined ideas, such as Lisp machines and the Connection Machine, show the best concepts do not always win.
Dr. Jon KrohnWill AI Tools Like Fabi Replace Data Scientists? (with Marc Dupuis)
- Self-serve AI analytics changes data scientists' work rather than replacing them. Dupuis says that when business users explore data themselves, even imperfectly, they bring data scientists much better initial questions instead of filing a Jira or Linear ticket. He says this is the change he sees most with customers: a more collaborative, data-literate working relationship.
- Business understanding and statistics will outvalue memorized coding. Dupuis argues that as AI gets good at coding, recalling syntax such as a Plotly chart matters less. He says data scientists should learn how the business makes money and its P&L, and lean on fundamentals in statistics and ML to design experiments and supervise the AI. He calls it the most exciting time for data scientists in five or ten years.
Polis PanditThe Trouble With Populism
- Populism is a rhetorical strategy, not an ideology — a "people vs. elites" morality play that appears on both left and right — and he traces the pattern historically through Juan Perón's Argentina and Mussolini's Italy before applying it to Trump and Zohran Mamdani.
- He argues Trump's populism runs on grievance and nostalgia (targeting immigrants, "liberal elites," the "deep state") while Mamdani's runs on moral purity (landlords, corporations, and "the rich" as villains against virtuous average New Yorkers) — different content, the same structural mechanism.
- The core danger he identifies is shared by both: populism delegitimizes good-faith opposition by recasting disagreement as either corruption or moral defect, which is why he predicts Mamdani's administration will underdeliver — populism wins campaigns but "almost always results in horrible governance" because solving problems removes the grievance that fuels the movement.
Crosstalk SolutionsThis One Box Powers Starlink and Peplink — Anywhere!
- One enclosure houses a Starlink Mini and a mobile Peplink router. Shown at the Peplink Tech Summit in Orlando, the Antenna Max S includes a PoE splitter powering both from a single cable and built-in cellular antennas for the Peplink, and can be mounted on an RV or mobile command center or deployed in the field; the host, whose store sells Peplink, pitches it as "the product for you".
Loss DogTom Sosnoff Wants You to Join the Pack. Here's Why.
- The new show pivots from trading talk toward general debate and commentary. One Lucky Dog is pitched as unfiltered, uncensored and uncontrollable — a shift from the hosts' trading-focused catalog (named examples: iron condors and covered calls) toward politics, listener calls, and "wisdom" from 40 years of entrepreneurship.
AI Analyst LabAvoid AI Failure: Why Scale Breaks Systems | Ep40
synopsis
The hosts of the Data Neighbor podcast interview Jigyasa Grover, an ML engineer at Uber, author and course instructor. She argues that agents fail at scale through weaknesses in the parts around the model: data, evaluation, memory and security. Her conclusion is that scalable evaluation, adversarial evaluation included, is the real barrier before agents can run autonomously, and that until then high-stakes decisions need a human in the loop. Critical reasoning is the human skill she would keep sharp.
The lifecycle over the model. She tells listeners not to chase the newest model and to think end to end instead. Well-governed data separates a working model from a production system. Evaluation is roughly "100x" harder than in classical ML because it has to be qualitative as well as quantitative, and every user behaves differently. The host adds that the same user's tolerance for imprecision changes with context.
Agent types and guardrails. Using Google's Agent Development Kit as her reference, she separates LLM agents from workflow agents (sequential, parallel or loop). Her travel-assistant example orders a weather check before the packing list, with events and an itinerary built in parallel. In her view workflows are themselves a guardrail. Constraints still need dynamic calibration through feedback, and a co-host adds personalization as a further guardrail.
Memory. She breaks agent memory into long-term, mid-term and short-term memory, and the work into retaining, retrieving and reasoning. Context-window limits and latency make it hard to retrieve exactly enough. She also admits she can't map human mid-term memory onto anything in current systems.
Where traditional ML stays. Asked where LLMs are being forced in, she says structured prediction such as recommendations belongs to traditional ML, which is faster, cheaper and easier to measure. LLMs augment it through explanations and feature engineering. The host describes using LLMs to create variables for causal models.
Evaluation and attack surfaces. Looking ahead to consumer agents, she says the battle is evaluation for agents that negotiate with other companies' agents. She lists the attacks: LLM jacking, prompt injection to pull another user's card number, memory poisoning, compromised API keys and social engineering. Her countermeasures are specialist adversarial testing, continuous monitoring, and role-based memory access with no cross-contamination. The host observes that most evals he sees check answer quality, not security.
Wider effects. The hosts move on to children who trust voice assistants over their parents. Jigyasa warns of over-reliance and misinformation, and a co-host argues for foundational education on how such systems work. She describes her own split of tools at the time of recording: GPT-5 for productivity, Claude Sonnet in Cursor for code, Glean for onboarding and Perplexity for research. She also raises the dilution of the human signal in training data as AI content spreads, content-provenance labels, and a shift from SEO to "GEO", optimizing to appear in LLM answers.
- Production quality comes from the whole lifecycle, not the newest model. Jigyasa Grover, an ML engineer at Uber who previously worked at Twitter, says it is fine not to run the latest model version. Curated, representative, well-governed data is what separates a working model from a production system. Evaluating LLM and agentic systems is "100x" harder than classical ML, because it needs qualitative as well as quantitative measures and no one can simulate every user's interaction or tool trajectory.
- Workflow agents add the reliability that LLM agents lack. Following Google's Agent Development Kit, she separates LLM-based agents from workflow agents, which can be sequential, parallel or loop-based. In her travel-assistant example, the weather check has to run before the packing list, while a parallel agent builds an itinerary from local events. Too many guardrails stifle creativity and too few allow unsafe results, so she calibrates constraints with monitoring and feedback, for instance when weather data is unavailable.
- Agent memory comes down to retain, retrieve and reason. She describes long-term memory (preferences), mid-term memory (boarding passes, upcoming tickets) and short-term memory (the current context). Because context windows are limited and latency rises with what goes into them, the hard part is retrieving "nothing more, nothing less" than what the task needs. A co-host summarized this as the three Rs.
- Traditional ML still wins on well-defined prediction. In her view, recommendation and personalization with clear inputs, ground truth and outputs belong to traditional ML, and using an LLM there is overkill. LLMs instead augment it: her current work uses them for post-hoc explanations of rankings, and she suggests LLM-extracted features for traditional models. The host describes using LLMs to build causal-model variables that couldn't be captured before.
- Scalable evaluation and security block truly autonomous agents. Jigyasa expects agents to deal with other companies' agents, using an Expedia agent contacting Hyatt's and Air Canada's agents as her example. She lists threats including LLM jacking, prompt injection to extract the previous user's credit card, and memory poisoning through repetition. Her defenses are adversarial simulation, continuous anomaly monitoring, memory hygiene with role-based access control, and a human in the loop for high-stakes decisions. The host notes that most evals he sees measure answer quality rather than security.
Scott HanselmanHalo 4 - Services in Azure with Caitie McCaffrey
synopsis
This episode, uploaded in 2025, was recorded on 7 February 2013 as Hanselminutes show 358. Caitie McCaffrey, a services engineer at 343 Industries, ends up arguing that Halo 4's backend worked because the team rebuilt it from the ground up for Azure, assuming that any machine can be pulled away at any time, rather than porting what ran in physical data centers.
What the services are. McCaffrey lists the services behind the Halo games, Halo Waypoint and mobile clients. Stats records kills, headshots and weapons, aggregating lifetime totals on the server but not in real time. Presence is real time and more robust than the standard Xbox Live service. Title files deliver playlists, updated weekly. There are also user-generated content and cheat-and-ban services; Xbox Live handles matchmaking and achievements.
Queues and event processing. Stats go into Azure Service Bus queues, are acknowledged once stored durably, and are processed later, with the goal of tolerating two to three days of downtime without data loss. McCaffrey rebuilt cheat-and-ban detection with complex event processing on StreamInsight, running the event stream through rules from community management; bans now arrive within 10 to 15 minutes instead of the next day. A separate Hadoop BI store keeps heavy queries off production.
Deployment and launch. They deploy by VIP swap between staging and production during working hours, which forced idempotent, fault-tolerant design. Launch week, the load peak for a game, meant at least five people in the office around the clock and fixes once or twice a day. They worked with the Azure team on core capacity, and a perf test took down production Service Bus queues before launch.
Frameworks and tooling. The team used Orleans, a Microsoft Research actor-model framework not yet public at the time, ETW-based logging, and portal-driven deployments, with scripted one-click deploys still to come. Her closing advice is to design services for the cloud's pricing and failure model rather than lifting them over unchanged.
- Halo 4's backend was rebuilt from scratch for Azure. McCaffrey, a services engineer at 343 Industries, speaking in a show recorded on 7 February 2013 (show 358), says it was the first Halo whose services ran entirely on Azure, and that the team rebuilt and re-architected every service rather than porting it, because a game's load peaks at launch and then tapers. Elastic capacity let them avoid expensive hardware that goes out of warranty while still supporting the long tail; older Halo titles then still ran in physical data centers.
- Queues let the game hand off stats without waiting. The stats service writes game data durably to Azure Service Bus queues and acknowledges at once, so players keep playing while processing happens later in a single transaction. She says the goal was to survive the service being down for two to three days without losing data; presence, by contrast, is fully real time. During a performance test they took down production Service Bus queues, and the Service Bus team fixed the problem before launch.
- Cheat-and-ban detection moved from nightly batches to near real time. McCaffrey wrote it using complex event processing on StreamInsight in Azure: rules from the community management team are defined first and the event stream from games runs through them. Signals include players muting someone or leaving when they host, and bans now land within 10 to 15 minutes of detection instead of the next day, with durations that escalate for repeat offenders.
- Build for an environment where anything can fail. Because deployments are VIP swaps between full staging and production deployments, done during the day without draining traffic, services had to be idempotent and fault tolerant. She warns that lifting server code into Azure unchanged "will probably work" but will not be cost-effective; the team also gave each developer their own account after shared-account cores were left running overnight.
- Orleans simplified scaled-out code, then internal-only. The team used Orleans, a Microsoft Research actor-model framework whose single-threaded "grains" let her write code synchronously while it handled message addressing; at recording it was not publicly available. Logging ran on ETW with an internal viewer called MDS, because Azure's diagnostic tables did not scale to their log volume, and scripted one-click deployments were still unfinished.
Dr. Jon KrohnThe Origin of Fabi: Fast BI for the AI Era (with Marc Dupuis)
- The company grew from a product manager's dependence on a data scientist. Dupuis says he moved from data science into a decade-plus of product management, while his co-founder, who has a PhD and led data science teams at companies including Lyft and Clari, handled his data questions by pulling data in a notebook and sending back spreadsheets or screenshots. Fabi, short for "fast BI", aims to make that collaboration far more efficient with AI.
- "Vibe analytics" means building data apps and analyses mainly through AI. Dupuis says fully or semi-autonomous AI lets semi-technical users such as product managers, CSMs and founders build dashboards and workflows by talking to it, while freeing data scientists for higher-value work than coding.
- AI shouldn't query the warehouse live on every question. Dupuis warns that handing a marketer an MCP server or direct connection would run up data warehouse costs, so a vibe-analytics platform has to control how AI reaches the data.
Master InvestorAndrew Ross Sorkin: What the crash of 1929 can teach us
- Sorkin says writing his new book was an emotional challenge, partly because the success of "Too Big to Fail" made him fear he couldn't replicate it. He describes wondering whether that earlier book's impact was "lightning in a bottle" — skill, or luck and timing.
- He frames the new book not as a warning of an imminent repeat, but as a history lesson meant to instill humility about predicting what comes next. He points to 1929 and 2008 as the parallels he draws on, saying the goal is to help prevent repeating history rather than to forecast its return.
Dr. Jon Krohn937: How to Design AI-First Products — with Marc Dupuis
synopsis
Marc Dupuis, co-founder and CEO of Fabi.ai, an AI-native business intelligence platform, argues that AI analytics does not replace data scientists. It moves them away from routine coding toward business understanding, experiment design and supervising the AI. It works only when the AI is bounded to curated data, gives users a choice of autonomy, makes every action reversible, and delivers insights into workflows rather than dashboards.
Origin. Dupuis started as a data scientist after a neurotechnology master's at Imperial College London and spent more than ten years as a product manager on data-heavy SaaS. His co-founder, who has a PhD and led data science at companies including Lyft and Clari, used to answer his data questions by pulling data in a notebook and sending back spreadsheets or screenshots. Fabi, "fast BI", is meant to replace that loop. Its "vibe analytics" lets semi-technical users build data apps and workflows by talking to an AI that can run fully or semi-autonomously.
What happens to data scientists. Dupuis says business users who explore data themselves bring better questions instead of tickets. As AI codes well, recall of syntax loses value; what gains is understanding how the business makes money and its P&L, plus strong statistics and ML for setting up experiments and checking the AI. He calls it the most exciting time for data scientists in five or ten years.
Guardrails. Taking up the host's point about cost and data quality, Dupuis describes a caching layer that stores each warehouse pull as a data frame, and advises giving AI only a few curated tables. Human collaboration matters as much: Fabi's "smartbooks", reactive AI-enabled notebooks, can be shared so a data scientist checks a product manager's pull or seeds a query for them to explore. He thinks business users mostly know their limits.
Workflows over dashboards. Data teams' most common complaint, Dupuis says, is dashboards nobody opens or that get exported at once. Dashboards still matter for top-level metrics. Otherwise he tells teams to always choose an automated workflow that pushes insights into Slack, Teams, email or spreadsheets, or acts directly, because dashboard ROI depends on behaviour change.
Designing AI-first. Citing Andrej Karpathy's autonomy scale, Dupuis says users should pick anything from hand-coding to full AI control, which requires an AI able to act on every interface element. Every action must be reversible; Fabi snapshots accepted suggestions with diffs and reverts. AI-generated reports answer only from data frames the data scientist has granted, and say so when the data is missing. In Fabi, he says, 90 to 95% of the code run is AI-generated.
Data quality. Dupuis reconciles two of his own positions, spend most of the time cleaning, but don't wait for perfect data, by distinguishing the warehouse model from the working dataset. A perfect data model is unnecessary, but messy inputs such as a revenue spreadsheet deserve about 90% of the effort before AI analysis.
Metrics. The data team still owns consistent north-star metrics, and AI risks recalculating them on every question. New metrics should be explored with AI for a month or a quarter before being built into the warehouse. On Goodhart's law he has no fix, and argues metrics such as ARR shift with pricing changes and board narratives, so they cannot be set in stone.
- Self-serve AI analytics is safe only on curated, bounded data. Dupuis says Fabi caches each AI-written SQL pull as a Python data frame so follow-up questions don't hit the warehouse, and urges customers to expose only a few curated core tables rather than thousands. In published smart reports, the AI answers follow-ups only from the data frames a data scientist granted and replies "I don't have the data" otherwise, which he presents as the guard against hallucination or the AI "going rogue".
- Data teams should automate workflows instead of building dashboards. Dupuis says the most common complaint from data teams is dashboards that few people view, or that get exported straight away; one data scientist joked they were lucky if someone clicked one by accident. Since ROI on dashboards is hard to prove and depends on behaviour change, his rule is to always pick the workflow: push AI-generated insights into Slack, Teams, email or spreadsheets, or take the action directly.
- AI-first products need adjustable autonomy and fully reversible actions. Citing Andrej Karpathy's autonomy scale and his Iron Man suit example, Dupuis says users should choose anything from hand-coding to the AI doing 100%. Product managers mostly use the fully autonomous mode and data scientists the semi-autonomous one, so the AI must be able to act on any element of the interface. Because AI work takes wrong turns, Fabi snapshots each accepted suggestion with a diff and a revert.
- AI needs well-framed data, not a perfect data model. Dupuis says warehouses rarely hold only clean, wide, current tables; there is a "constellation" of messy spreadsheets around them. His advice is not to wait for a perfect model but, before using AI on messy data such as a revenue spreadsheet, to spend 90% of the time cleaning it: removing merged cells and charts and fixing names.
- Metrics should stay flexible because the business keeps redefining them. Asked about Goodhart's law, Dupuis says he has no magic fix and that quota-setting shows smart people will game incentives. He argues setting metrics in stone invites heartache: ARR can mean something different the next quarter if the product team adds credit- or consumption-based pricing, or if executives change the story they tell the board.