Souk's Notes - Unofficial Summaries

2025-07-07 · open the original ↗

• Bat-Leah entered e-discovery as a temp receptionist and built domain expertise there. She started while still finishing her undergraduate degree, then moved through increasingly senior legal roles into data science leadership.
• Getting A's without intuition drove her toward the mathematics underneath. Bat-Leah says she took applied statistics courses (econometrics, psychometrics, finance with portfolio theory, where she learned PCA) and kept getting A's without feeling she understood why. The chair of Northwestern's statistics department told her to come back in a year once she had learned calculus. She self-taught calculus and some linear algebra through MIT OpenCourseWare and Khan Academy, took his probability and stochastic processes course, learned to code from its problems, and got an A on the final. She says that was the first time she felt she understood statistics.

2025-07-06 · open the original ↗

• Noisy legal labels pushed Bat-Leah from algorithms to data-centric ML. In her first R&D role, choosing classification algorithms for legal tasks, she found that the labeled data had so many mislabeled instances that she couldn't trust her evaluation of the algorithms. That drew her to what Andrew Ng coined data-centric AI.
• Data-centric ML holds the model fixed and systematically engineers better data. Bat-Leah contrasts it with iterating on architecture and learning algorithm. Her examples: whether to aggregate labels, DoReMi's reweighting of domains in the Pile for LLM pretraining, and a paper she can't name that uses linear algebra to choose which points to label before any model exists, unlike active learning.
• The research community grew from an MLCommons working group into a field. Bat-Leah traces the path: MLCommons' DataPerf benchmark group, which Andrew Ng was part of and which became a NeurIPS paper; workshops at ICLR and ICML, with Ng giving the inaugural DMLR keynote; a DMLR journal; and an MLCommons data-centric ML research working group. That group is partnering with Common Crawl on a challenge to produce a public dataset for a low-resource language.

2025-07-05 · open the original ↗

• Extensions let PostgreSQL add features without touching its core. Shelton, a Microsoft partner solution architect on the host's team focused on the data platform, says this extension model is what lets the open-source database adapt quickly. He says it is the route by which its AI capabilities arrive.
• The managed Azure offering takes over routine maintenance. Shelton describes Azure Database for PostgreSQL flexible server as platform as a service: minor version upgrades are applied for you, while a major upgrade (his example is 16 to 17) is the customer's decision. He also cites compute SKUs that scale up and down, the ability to stop the server to save resources, a query store for query patterns, and a VS Code extension tied to GitHub Copilot, which he says was in preview at recording.
• High availability and recovery come from zones, replicas and backups. A standby in a second availability zone is kept as a synchronous copy for failover. Read replicas, including in another region, are asynchronous with some lag, and serve read scale-out and disaster recovery through a failover you initiate. Backups are automatic, with retention of 1 to 35 days and a default of 7.
• pgvector adds vector-database capabilities to a relational database. Shelton says the extension stores embeddings, runs similarity search and indexes them with IVFFlat or HNSW. On Azure there is also DiskANN, an index type he says Microsoft developed, which boosts performance and recall while optimizing resources. The "vector database" framing is Giard's restatement, which Shelton accepts.
• Azure adds in-database embedding calls and accelerators for AI patterns. An Azure AI extension lets the database call Azure OpenAI to create embeddings, and other Azure AI services for tasks such as sentiment analysis. Solution accelerators cover GraphRAG using the AGE graph extension and semantic reranking of similarity-search results. For learning, he points to the Azure Database for PostgreSQL blog, Microsoft Learn modules and POSETTE, a Microsoft-sponsored virtual conference that had just taken place.

2025-07-05 · open the original ↗

• In e-discovery, recall and precision get negotiated with opposing counsel. Bat-Leah says the metrics, and arguments over margins of error, are sometimes negotiated with the other side or a government body, so data scientists must explain them to attorneys and judges. "Defensibility" comes down to what the attorney is comfortable defending, weighed against proportionality and undue burden. For a very rare relevance tag, for example, just sampling enough documents to evaluate it could be overly burdensome.
• Elusion sampling estimates recall, and Bat-Leah argues it is fair. Elusion, a metric she says she has never seen outside e-discovery, samples only the documents predicted not relevant. Combined with human ground-truth labels for the relevant ones, it gives an interval for recall. Its defensibility is debated. She argues that because purely human "linear review" is never measured this way, and its labels are probably not all correct, machine-learning workflows shouldn't be held to a higher standard.

2025-07-04 · open the original ↗

• Driving AI adoption takes both executive alignment and front-line enablement. In the Ep. 897 clip, the founder and CEO of a strategy consulting firm gives five tips: top down and bottoms up, with the change agent bridging the two; back bold claims with proof points; inspire as well as incentivize, since companies over-rely on incentives; focus on the 18% of early adopters, not the naysayers; and have the courage to go first.
• A finished project that interests you beats a checklist of portfolio projects. In the Ep. 893 clip, Avery Smith says many projects die for lack of motivation, so pick one tied to a job or hobby you care about. One killer project aimed at a target employer may be enough. Waiting to have five before applying is a trap, and for analyst roles one SQL and one data-viz project is a good start. Personal-data projects motivate but can be hard to collect data for: Apple doesn't let you export screen time. Even a delivery driver's A/B-tested gate-access texts made a credible résumé bullet.
• Experienced candidates still struggle against thousands of applicants per job. In the Ep. 899 clip, Kirill Eremenko, the show's founder, describes a student: an LA senior developer in her mid-40s targeting $200,000-plus roles. She has years of Python data processing and four machine-learning courses behind her, yet finds it hard to break in, so she applies selectively through her network. The host argues that in-person networking, such as Meetup groups, is the best route to opportunities, and that online collaboration is the fallback.
• Early-stage AI investing bets on teams, weighed for AI-nativeness and market savvy. In the Ep. 895 clip, AIX Ventures co-founder and general partner Johnson says founders will pivot, so VCs aren't market timers. After an inflection such as ChatGPT, consensus applications like AI tutoring need extremely AI-native teams. As the technology commoditizes, as the SaaS stack did, non-consensus market insight matters more, until a new architecture replaces transformers. He prefers production AI engineers who also read papers over a dedicated paper-reading PhD.

2025-07-03 · open the original ↗

• Rewriting from scratch in Go has stood the test of time. Goswami, CEO of Traefik Labs, says the project began as open source in 2016 to expose Docker Swarm containers as services securely. The founder and CTO, Emil, rejected existing approaches and wrote it natively in Go for its lightweight, container-friendly, modular design. Goswami claims 3.3 billion downloads and over 800 contributors nearly 10 years on.
• Traefik runs natively wherever containers run. He lists Docker and Docker Swarm, K3s, k3d, minikube, MicroK8s, and the managed AKS, EKS, GKE and OKE, plus HashiCorp Nomad and VMs. Having started before Kubernetes, the company says it moved from being cloud-native in the Docker era to being Kubernetes-native.
• The product is three tiers, and upgrading takes one command. The open-source tier is a reverse proxy and L4–L7 load balancer, typically used as a Kubernetes ingress controller. The paid API gateway centralizes authentication and authorization through OIDC/OAuth identity providers, and adds distributed certificate management, rate limiting, and Azure Key Vault and HashiCorp Vault integration. API management adds versioning, governance, a developer portal and observability for businesses that monetize APIs. Moving from open source to the gateway is one Helm chart command, under a minute with a license key, upgrading in place with configuration preserved.
• An AI gateway is an extension of the API gateway. Goswami argues that AI workloads are increasingly deployed in Kubernetes as microservices and consumed as APIs. So Traefik released AI gateway functionality inside the API gateway tier, and applies API management's governance, observability and portal to AI workloads too.
• A community-built firewall shows the open-source-to-paid path. He says a community pull request added a web application firewall based on OWASP's Coraza through a Wasm interface, released within about four months under quarterly release cycles. Traefik then built a higher-performance version for the paid gateway. He sees an awareness gap: people still think of Traefik as only open source. His parting message is the security of the gateway tier and the governance of the management tier.

2025-07-03 · open the original ↗

• Technology-assisted review pairs a long-text classifier with active learning suited to prevalence. Bat-Leah says traditional TAR uses random forests, SVMs, logistic regression or ensembles. When relevant documents are rare, relevance feedback has humans label the documents the model already rates most likely relevant, then retrains iteratively. With more balanced classes, uncertainty sampling has humans label the documents with the highest entropy.
• Epiq's AI Discovery Assistant trains classifiers from both labeled examples and natural-language instructions. Bat-Leah says the tool combines traditional long-text classification with LLMs. A case needs at least a responsiveness model, and often classifiers for privilege, confidentiality and other issues the attorneys care about. The host relays the claims that it automates 80% of discovery and completes reviews up to 90% faster, figures he notes Bat-Leah had qualified.
• Few law firms have data scientists, so the tools must carry the expertise. Bat-Leah says firms mostly rely on such tools for discovery. Using them defensibly still ideally needs domain expertise and some familiarity with evaluation metrics, which Epiq is trying to build into the product with intuitive explanations.

2025-07-02 · open the original ↗

• Agentic products are cheap to build and expensive to scale. Mahesh Yadav, an AI product manager who has worked at Microsoft, Meta, Amazon and Google, says the cost of building is now prompting, but something that works for 5 people takes 10x more work for 50 and 10x more again for 500, so most of the time goes into evaluation. Users also widen an agent's rules as it succeeds: a contract-risk agent asked next to write into HubSpot now feeds systems of record, where a misplaced zero could turn a $1 million insurance purchase into a $10 million one.
• AI assistants get one shot, so they must favor precision. Yadav says that for an AI product placed between users and support at Google, 85% of people dropped after the second interaction, whereas Netflix can throw 45 titles at a viewer on high recall. TikTok, in this account, could favor recall because its 60-second format built feedback into the product, reading a swipe before 60 seconds as a signal, where Facebook relied on the social graph.
• Evaluation means turning subjective quality into objective yes/no checks. Yadav's method is to narrow scope (risks in NDAs, not every contract), build ground truth from a domain expert's past reviews to get accuracy, precision and recall, then score helpful, honest and harmless with checklist questions, such as whether a liability answer states the amount, explains it, says when it applies and cites a source checkable in under 2 seconds. The project in progress has 70 helpfulness, 30 honesty and 27 harmlessness checks (totalled as 132), producing scores such as 78% helpful and 3% harmful on which a leader can sign off and benchmark competitors. Shane says he presented similar metric families at work: coverage, quality and faithfulness.
• LLM judges mostly fail once validated against human labels. Yadav's rule is to measure a judge's precision and recall against expert labels and require 80–90%; set the bar above 80 and, he says, you learn they don't work. Harmlessness checks such as asking for card details, profanity or attacking competitors are solved out-of-the-box guardrails, groundedness is available as an API, and helpfulness still needs paid subject-matter experts.
• Yadav predicts product makers and AI evaluation scientists as new roles. Product managers should ship their own first MVP and test it with simulated or paid users before engineering is brought in, while "get me data insights" work fades; Yadav compares evaluation now to data science around 2010–2016, when the title was coined. Shane expects PMs to do early iterations alone and data and engineering to handle scaling and evaluation, and a co-host expects today's roles to merge; Yadav also forecasts thousands of billion-dollar vertical-agent companies, such as one solving auto-dealership compliance, in place of seven trillion-dollar ones.

2025-07-02 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2025-07-02 · open the original ↗

• E-discovery means sifting huge volumes of unstructured records for what must be produced. Bat-Leah, senior director of Epiq's AI labs (Epiq has over 6,000 employees, per the host), explains that when companies sue each other they must exchange anything that might be evidence. That means hundreds of thousands to millions of emails, documents, spreadsheets, tweets and texts. Attorneys must decide what they are legally obliged to produce, and nearly all discovery is now electronic, apart from cases such as asbestos litigation that still need paper scanned in.
• Bat-Leah qualifies Epiq's speed claim: it beats traditional TAR, which the software also supports. The host relays Epiq's claim that its AI Discovery Assistant automates more than 80% of e-discovery and completes reviews up to 90% faster than TAR or linear review. She says it is better than traditional TAR, which she defines as using machine learning to classify documents as relevant or not to a litigation.
• Flooding the other side with unresponsive documents is now bad faith, which makes precision matter. Bat-Leah says overwhelming opposing counsel with documents that don't respond to their requests for production is considered bad faith today. Machine learning also makes it easier for opposing counsel to find the needles they are looking for.

2025-07-01 · open the original ↗

• The host's pick for a 10 Gb UniFi home gateway is the UCG Fiber. At $279 it has four 2.5 Gb LAN and two 10 Gb SFP+ ports, any port assignable as WAN, up to 5 Gbps with IDS/IPS, and an M.2 slot for adding UniFi Protect storage later. The design question came from a Rogue Support ticket from a user with UGREEN NAS units and a 10 Gb workstation.
• The Pro XG 8 PoE switch is pricey but powers everything. At $499 it has eight 10 Gb RJ45 ports, all PoE+ capable, two 10 Gb SFP+ uplinks and a desktop or wall form factor. The $299 USW Flex XG saves money and can itself be PoE-powered from the gateway, but cannot power an access point.
• The U7 Pro XG is better value than the U7 Pro. For $10 more it adds a 10 Gb port against the U7 Pro's 2.5 Gb, plus the flat-faced, metal-backed design carried over from Ubiquiti's enterprise E-series.
• The complete build lands right at the title's budget. The three pieces total $977, and with a DAC cable the host puts the full 10 Gb network with PoE at about $1,000.

2025-07-01 · open the original ↗

• Epiq's discovery tool improves on traditional technology-assisted review, Bat-Leah says, but doesn't replace it. Bat-Leah, senior director of Epiq's AI labs, qualifies the claim the host relays that the AI Discovery Assistant automates over 80% of e-discovery and finishes reviews up to 90% faster: it is better than traditional TAR, which the software also supports. TAR is a long-text classifier with active learning: relevance feedback when relevant documents are rare, uncertainty sampling when classes are balanced. Epiq's tool trains classifiers from both labeled examples and natural-language instructions, for responsiveness, privilege, confidentiality and case issues.
• Legal metrics are negotiated, and elusion is meaningless without the starting prevalence. Bat-Leah says recall and precision targets are sometimes negotiated with opposing counsel or a government body, and defensibility is what the attorney is comfortable defending. Elusion, the false-omission rate among predicted-non-relevant documents, is often judged "good" under 5%. She says that tells you nothing if prevalence was under 5% to begin with. In a TAR 2.0 workflow, where humans review every predicted-relevant document, she derives a recall interval from the elusion interval. She argues machine learning shouldn't be held to a higher standard than unmeasured human linear review.
• A point estimate without sample size and confidence interval is "lying with statistics". Bat-Leah teaches intervals by having consultants sample with 90% confidence at least 10 times, so they see the true value fall outside the range about one time in ten. She says 95% and 99% intervals converge only at very large samples (in the millions) and differ greatly at sizes of 8 or 10. The level should follow how important it is to be right.
• Data-centric ML holds the model fixed and engineers the data, but lacks prestige. Bat-Leah, third author of the paper "DMLR: Data-Centric Machine Learning Research — Past, Present and Future", says the field long lacked prestigious archival venues. That is changing with NeurIPS's datasets and benchmarks track and the DMLR journal, a sibling of JMLR. She notes that roughly 80% of project time goes to data, while attention goes to the 20% spent on models. DMLR is considering an applications-focused workshop because applications research is looked down on.
• DataPerf benchmarked data selection, debugging, valuation and red-teaming. Bat-Leah describes the MLCommons challenges, on a paper whose authors include Andrew Ng: choosing training subsets in speech and vision, finding mislabeled instances, and valuing individual data points. She says valuation may matter for compensating people whose data trains models. Adversarial Nibbler sought benign-sounding prompts that generate unsafe images, such as "a child sleeping in red paint". She concedes that benchmark focus doesn't always translate to real-world impact.

2025-06-30 · open the original ↗

• A VLAN is layer 2; moving between VLANs is layer 3. The host explains that traffic within one subnet or VLAN stays at the data-link layer however many VLANs are stacked on an interface, while getting from one VLAN to another is routing at the network layer.
• A layer 3 switch routes between VLANs without the router. Normally inter-VLAN traffic has to go back to the router to be forwarded; a layer 3 switch can be configured to do that routing itself. The host offers this as an off-the-cuff explanation and asks others on the stream to check him.

2025-06-30 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2025-06-30 · open the original ↗

• Demand for production deployment skills in AI work is rising. Eremenko says a departing member ("Evan", a software consultant with US clients) found that LLM and agent jobs increasingly require cloud, production and deployment skills. He adds that Ed Donner's survey of his students put production deployment as the top topic wanted. The host won't call it compulsory but says software engineering skills alongside AI broaden the range of jobs open to a candidate enormously.
• A long-term AI engineer should master two distinct areas. Eremenko's first area is the science of AI: embedding models, experimenting with data, validating context relevance, model choice, down to LoRA. The second is production systems: CI/CD pipelines, AWS Lambda and Step Functions, deciding whether MCP is the right fit, and deploying and monitoring cost-effectively and securely.

2025-06-29 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2025-06-29 · open the original ↗

• Work backwards from what local employers require. Eremenko advised a senior developer moving into AI to ask LA recruiters which boxes an AI engineering role must tick, then learn only those. He argues this narrows the study scope and puts the candidate in front of recruiters. Their commission gives them a reason to place her once she reports back ready.
• AI job requirements differ by city and industry. The host cites his episode 856, built on LinkedIn surveys, which named AI engineer the fastest growing job of 2024. It found AI researcher roles concentrated in San Francisco near frontier labs and AI consultant roles in New York. Eremenko recommends the job aggregator hiring.cafe for filtering roles by location and title.

2025-06-29 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2025-06-28 · open the original ↗

• Don't let a model write raw SQL against your database. Nixon, who works on Microsoft's SQL team, says a model is like "a four-year-old genius" that means well but can mis-join or drop a table to satisfy a request. His architecture puts Data API builder, a REST endpoint over SQL that won't run those operations, between the database and an MCP server that wraps it. The demo stack is a Blazor app, Azure OpenAI (chosen so data stays inside Microsoft's security commitments), Azure App Service and Azure SQL Database.
• The MCP client quietly hands the model tools to answer from. As Nixon describes the flow, the client attaches the tool list to the user's prompt. The model picks a tool such as "list customers by size" and its filter parameters, the client calls Data API builder, and Data API builder writes the T-SQL. He says the added latency is small, and because the chain runs in the user's context, the user only gets data they could already reach.
• Tool descriptions are a token-budget trade-off each developer must tune. Each tool is a method decorated with attributes describing keys, columns and parameters, with examples of OData $filter strings. He relies on the model already knowing OData. He believes the evolving MCP spec currently allows about 1,024 tokens for all of a server's tools. He says around 20 tools overwhelms a model (Claude perhaps 30–40), so he prefers small, specific servers over "one MCP server to rule them all".
• AI-enabled apps may become read-only screens, with the model making changes. In his insurance demo, built on five tools, "John's name is actually Marcus" updated the right customer's first name from conversational context. A single message about Samantha's email from last Tuesday changed her surname, added a $1,000 wedding-ring property policy effective the first of next month, and logged a communication record. Nixon argues a model can be made to follow business rules a user might skip, cutting UI and retraining work when processes change.
• The separate MCP server is meant to fold into the API layer. Nixon says the team hopes to build MCP directly into Data API builder, "by end of the year maybe", to add an MCP-enabled setting alongside the existing REST and GraphQL switches. Developers would choose which tools are exposed and control token flow back to the model. The separate-server architecture works today.

2025-06-28 · open the original ↗

• Even strong experience gets lost among thousands of applicants. Eremenko describes a senior Python developer in LA ("Clara", mid-40s), with years of pandas data processing in the medical sector and four machine-learning courses, who targets $200,000+ roles and still can't break in. He blames thousands of applicants per job and AI pre-screening, and backs her choice to apply mainly through her network.
• In-person networking is the best route to opportunities. The host calls it "easily the best way". He points to local meetups (meetup.com in US and Canadian cities), where talks are paired with social time. Where in-person isn't possible, he suggests collaborative online projects as the substitute, since collaborators remember each other's expertise.

2025-06-27 · open the original ↗

• Barry's Pulitzer photo shows his son celebrating a Nintendo, not the prize. He won in 1988 for distinguished commentary; a surprise announcement cancelled a Key West trip with his son Rob, so he promised a Nintendo game, and the front-page photo caught Rob's joy at that. Rob, now 44, runs a data-assisted investigative team at the Wall Street Journal that has itself won a Pulitzer.
• Book tours mean uninformed interviewers and no ethics. Barry relays David McCullough's story of a 1970s morning-TV host asking "Who is Normandy Beach?" about The Path Between the Seas, and tells how, booked for Oprah's "Things We Do in Secret" in the mid-'90s to return a stolen Hyatt towel-charge sign ($75), he stole a second Hyatt sign in St. Louis as a backup when FedEx could not promise delivery.
• Barry doesn't think computers make us dumber, and AI can't do humor. Despite his Substack post "Stupidification", he calls computers useful tools; of at least a hundred AI-written "Dave Barry columns" people emailed him, he says "all of them sucked", and an AI-written book with his title appeared on Amazon about the day his came out. He switched from Windows to Apple to get his phone and computer to talk, but says he uses Microsoft Word all the time.
• The readers who hated him supplied the most material. A throwaway jab at Neil Diamond's "I Am... I Said" ("not even the chair") drew perhaps a thousand angry letters, a follow-up column, weeks of reader mail and eventually a book on songs people hate; a "Hoosier" joke drew a couple of hundred letters from Indiana, his favourite pointing out there are no caribou there.
• You can, contrary to his teachers, joke your way through life. Elected class clown at Pleasantville High (class of 1965), he started on obituaries and municipal meetings at Pennsylvania's Daily Local News, wrote humor "ad libs" on the side, and only went full-time in his 30s. At 77 he calls the memoir a more serious, partly sad book that his editor and agent pushed him into, and says he has no plans for another.

2025-06-27 · open the original ↗

• Barry's Pulitzer photo shows his son celebrating a Nintendo, not the prize. He won in 1988 for distinguished commentary; a surprise announcement cancelled a Key West trip with his son Rob, so he promised a Nintendo game, and the front-page photo caught Rob's joy at that. Rob, now 44, runs a data-assisted investigative team at the Wall Street Journal that has itself won a Pulitzer.
• Book tours mean uninformed interviewers and no ethics. Barry relays David McCullough's story of a 1970s morning-TV host asking "Who is Normandy Beach?" about The Path Between the Seas, and tells how, booked for Oprah's "Things We Do in Secret" in the mid-'90s to return a stolen Hyatt towel-charge sign ($75), he stole a second Hyatt sign in St. Louis as a backup when FedEx could not promise delivery.
• Barry doesn't think computers make us dumber, and AI can't do humor. Despite his Substack post "Stupidification", he calls computers useful tools; of at least a hundred AI-written "Dave Barry columns" people emailed him, he says "all of them sucked", and an AI-written book with his title appeared on Amazon about the day his came out. He switched from Windows to Apple to get his phone and computer to talk, but says he uses Microsoft Word all the time.
• The readers who hated him supplied the most material. A throwaway jab at Neil Diamond's "I Am... I Said" ("not even the chair") drew perhaps a thousand angry letters, a follow-up column, weeks of reader mail and eventually a book on songs people hate; a "Hoosier" joke drew a couple of hundred letters from Indiana, his favourite pointing out there are no caribou there.
• You can, contrary to his teachers, joke your way through life. Elected class clown at Pleasantville High (class of 1965), he started on obituaries and municipal meetings at Pennsylvania's Daily Local News, wrote humor "ad libs" on the side, and only went full-time in his 30s. At 77 he calls the memoir a more serious, partly sad book that his editor and agent pushed him into, and says he has no plans for another.

2025-06-27 · open the original ↗

• Software VPNs can leak a remote worker's real location. The host says Crosstalk support gets many requests to mask a worker's location from an employer that locks access to a home IP; with a VPN, Tailscale or ZeroTier, forgetting to enable it or a dropped tunnel exposes the real location. GL.iNet sponsored the video.
• The Flint 3 is GL.iNet's new tri-band Wi-Fi 7 home router. It has four 2.5 Gbps LAN ports, one 2.5 Gbps WAN port, 2.4/5/6 GHz bands with MLO, preconfigured VPN client profiles (PIA, NordVPN, Mullvad and others), AdGuard Home, dynamic DNS and parental controls. As of recording it was on pre-order, with both early-bird price tiers already sold out.
• AstroWarp fixes a single home router as everyone's internet exit. GL.iNet's cloud service, which the host likens to SD-WAN, joins its routers and client apps (phone, Windows, Mac) into one central network at astrowarp.net; you draw links between devices and toggle "access the internet via" the chosen exit node.
• The host's test kept his home IP across three different WANs. With a Slate 7 travel router on T-Mobile Home Internet at a remote office, and later on hotel Wi-Fi in repeater mode, what-is-my-IP checks showed his home Comcast Business connection and the same IP once AstroWarp was enabled, with no re-configuration after moving.

2025-06-27 · open the original ↗

• Developer SQL tooling is being consolidated into one editor extension. Nixon, on the SQL team for about three years, admits Microsoft confused developers about which tool to use. Azure Data Studio is now deprecated and its work is moving into the MSSQL extension for VS Code. SQL Server Management Studio stays the Windows-only superset for DBAs. The extension will remain partly a subset of SSMS, plus developer tooling such as database projects and container work.
• The extension spins up local database containers with no Docker knowledge. With Docker Desktop installed, a wizard checks prerequisites, pulls or reuses the Linux image, and offers SQL Server 2025, 2022, 2019 or 2017. It sets the SA password and creates a connection. Nixon changes the port to 1234 to avoid clashing with a local install on 1433. An Azure SQL Database image is planned "sometime next year". He says SQL Server 2025, then in CTP and expected GA at year end, is almost identical in function to Azure SQL DB.
• The preview features are aimed at year-end, with acknowledged slippage. He demos a JSON/DSL-based schema designer, and schema compare between two databases that generates a sync script. "Chat with this database" opens Copilot with context on the database's tables and objects. A built-in MCP tool can push proposed schema changes into the designer for visual review before they are committed. Nixon says the goal is end of 2025 and "some things aren't going to make it".
• Database projects with a human review gate are the recommended deployment. Nixon lists three routes: schema compare from local to Azure, a database-first project of DDL scripts kept in Git and published, and publishing through .NET Aspire. He prescribes the database project: deploy to UAT automatically, but have a person review the generated script before production. Real databases have hundreds or thousands of objects. Pre-built GitHub and Azure DevOps pipeline actions exist for this.
• Unit-test database objects the way you test code. Nixon argues a stored procedure is just a method, with inputs, outputs and sometimes state changes. Tests should catch its regressions and run in CI/CD. He says this applies to any database, not just SQL Server.

2025-06-27 · open the original ↗

• Keep trying providers until one has the right expertise. The host's 95-year-old grandmother Annie says recent headaches and dizziness, bad enough to stop her driving, improved with physical therapy. The first provider did nothing for her, the second was worse, and the third cleared her dizziness and then eased decades-old knee pain. The host's lesson is not to wait for problems to go away on their own.
• Engagement with family and a pet sustains her independence. She still lives alone in the house her husband largely built himself between 1949 and
  1952. Her daughter helps her, and a small dog that has lived with her for almost two years "makes me happy".

2025-06-26 · open the original ↗

• OneNote can import a Teams meeting's details with one click. On a OneNote page Giard clicks Meeting Details, picks the Microsoft account that attended, navigates to the meeting's date and selects the meeting. He credits a coworker, who showed it at an internal "Teach Me in 10" session.
• The import brings in what he used to type by hand, plus Copilot's AI summary. The page fills with a link to the meeting, its recording, the original Outlook invitation, the participants, the meeting collateral (files and PowerPoints added to it), a Copilot-written summary, suggested tasks, and a notes section at the bottom where he can keep typing as on any OneNote page.
• Giard values it because note-taking pulls attention from the meeting. He says the summary "may or may not be as detailed" as his own notes but saves him a lot of work, and calls finding the feature "game changing".

2025-06-26 · open the original ↗

• The field can evolve faster than a part-time learner. Eremenko recounts a process engineer in his early 30s ("Ben") who said in August 2024 he would be job-ready in five months. Ten months on, he feels scattered and always playing catch-up, studying around a full-time job and family. Eremenko notes that LangChain and LangGraph faded within a year and prompt engineering within about eighteen months, displaced by MCP and agentic AI.
• Long-term trends give stable ground amid hype. The host names Python, data structures and algorithms (useful even if Rust displaces Python in five to ten years), SQL and ever-cheaper compute as safe bets. He says prompt engineering always looked short-lived, because frontier labs fine-tune models to anticipate what the prompter wants.

2025-06-25 · open the original ↗

• Iran's rapid defeat makes its opponent's failure against terror groups suspect. He says Iran's opponent, unnamed, dismantled Iran's military in a 12- or 13-day assault planned since 2007, through spy infiltration, crippled weapons and communications, and decapitation strikes, setting Iran back one or two decades. He finds it hard to believe the same country cannot eliminate terror groups in a small piece of land, given Iran is about 30 times larger.

2025-06-25 · open the original ↗

• Mark's stress comes from unmanaged work and no time to process. He describes himself, like Hanselman, as an extroverted introvert: one-on-ones recharge him, meetings where people present to him drain him, and two days of back-to-back meetings with no breaks would make him "combust", so he keeps gaps in his calendar.
• Early-career people need someone at work who won't burn them. Hanselman argues that people who started their careers after COVID lack the in-office relationships older staff built, and recommends a trusted back channel in big meetings; Mark frames work as networks of trust where you can risk saying something stupid.
• Hanselman's TEDx talk was memorized and drilled like standup, not ad-libbed. He wrote a seven-page script, had no teleprompter, and gave it 54 times (six to audiences, twice internally at Microsoft) under a speaker coach's record-and-refine and spaced-repetition method, landing about 80% scripted. He rehearsed at 12:05–12:30 against a 12-minute limit, forgot to allow for applause, and his heart rate hit 140 against a usual on-stage 120; he judges the method the difference between gold and his usual B+ talk.
• Mark exercises daily under a self-imposed "no excuses" rule. He works out six days a week first thing in the morning, 30–60 minutes of cardio plus three 50-minute weight sessions with a trainer, after once sliding from six days to four to three to occasional. A trainer, he says, is someone you are obliged to show up for who also takes the decisions out of your hands.
• Breathing and writing things down are their main stress tools. Mark focuses on long breaths, since stress shortens breathing and feeds racing thoughts, and thinks it should be taught in elementary school; both offload worries onto paper, and Hanselman cites his "Sync to Paper" blog post from almost 20 years ago.

2025-06-25 · open the original ↗

• This post includes paid content that is not available publicly. See the original source.

2025-06-25 · open the original ↗

• Fork a shared repository rather than editing it directly. Giard presents forking — copying someone else's repository into your own GitHub account — as a safer way to collaborate, since changes in the fork do not affect the original. His example is Microsoft's open-source NLWeb project, which accepts contributions; he recommends it as a tool for adding natural-language chat to a website.
• The Fork dialog sets owner, name, description and branches copied. The owner defaults to the signed-in account; the name defaults to the original's, which he keeps as a reminder of where it came from, and must be unique within the account. The description is pre-filled, and an option copies only the main branch (his choice) or every branch. Create Fork copies the repository and opens it.
• A pull request asks to merge the fork's changes back into the original. After a one-line README edit committed in his fork, he opens Pull requests, creates a new one, reviews the diff, and adds a title and description.
• Repository rules can hold the pull request for review. NLWeb requires a review from someone other than the last pusher, so his request goes into a queue for the maintainers to approve rather than merging automatically. Giard notes that not every repository has such rules.

2025-06-25 · open the original ↗

• "AI engineer" now mostly means building on existing LLMs. The host says that until about two years ago the term implied a frontier-lab researcher scaling transformers across GPUs. Now it mostly means calling LLM APIs, which takes data evaluation, exploratory analysis in Python and guardrails, not knowledge of stochastic gradient descent.
• Hands-on experimentation with models matters more than theory. The host says a three-billion-parameter Llama fine-tuned with LoRA on a narrow task can outperform Claude 4. He argues that knowing such options exist, and having tried them, matters more than knowing how LoRA works.
• Both staying abstract and digging deeper can pay. The host sees one path in building commercial applications at the abstraction layer, potentially solo with "hundreds of agents". The other is peeling back layers into mathematics, physics and engineering over decades. He believes the latter keeps raising one's value even as LLMs get better at that work, while allowing he may be "a dinosaur".

2025-06-24 · open the original ↗

• The ceasefire makes a world war even less likely. As reported, Iran was to begin the ceasefire at midnight Eastern time and its opponent within 12 hours after; if it holds for 24 hours the conflict ends. Both sides were still firing (Iran launching missiles, its opponent striking Tehran) yet holding to it for now.
• America's president forced the ceasefire on both sides to exit the war. He reads the president's statement (a tweet) as a sign Washington wants nothing more to do with the conflict. Separately, China had already warned Iran not to close the Strait of Hormuz, since China depends heavily on that oil.
• Iran's strike on the US base in Qatar was not an escalation. He says Iran warned both Qatar and the US about four days in advance, so the base and surrounding area had been cleared of military and civilian personnel. He poses the open question of which side will break the ceasefire first, or keep it longest.