Souk's Notes - Unofficial Summaries

2024-07-14 · open the original ↗

• Inflation falling to 3% means the soft landing has arrived. The speaker notes most economists did not expect this level at this stage, and reads it as evidence that the Fed is doing its job, the US economy is being navigated competently, and the official numbers can be trusted.
• If inflation holds or falls, a September rate cut becomes likely. The speaker predicts a cut would bring a large stock market rally and rallies in other industries. The benefit lands first on big money and on banks holding hundreds of billions of dollars in held-to-maturity assets at a loss, not on ordinary people; the captions cut off there.

2024-07-12 · open the original ↗

• The UniFi Cable Internet modem replaces an ISP modem inside UniFi. The UCI costs $279 and has one DOCSIS 3.1 coax port and one 2.5 GbE port. A full 2.5 Gbps link needs one of the five UniFi gateways with a 2.5 GbE WAN port. It works behind a non-UniFi gateway, but then you get no statistics, and even in UniFi it shows only CPU, memory and bandwidth.
• ISP compatibility is narrow, so research before buying. It is US-only: Comcast Business only on the business side, Xfinity, Charter Spectrum or Cox on the residential side. Comcast caps third-party modems at 1.2 Gbps down and 40 Mbps up, Spectrum at 1 Gbps and 35 Mbps. It does not work on high-split DOCSIS, and some ISPs are rolling out DOCSIS 4.
• Static IPs on Comcast Business rule it out. Comcast Business allows static IPs only through its own gateway, so the host, who has a block of five static IPs, cannot use his unit. He is giving it away to a US subscriber.
• The main reason to buy is lease savings. The host pays Comcast Business a $32.95 monthly equipment fee, so a UCI would pay for itself in about 8.5 months and save over $300 a year after that. Activation is mostly reported online as under 10 minutes without calling support, with some calls running past an hour and one case unresolved even after a truck roll.
• Users report better performance more often than not. From Reddit and forum posts, the host found about 75% of users reporting faster speeds than with their ISP modem. Of the rest, half saw no change and half saw worse performance. He stresses this is anecdotal and reads it as a better-than-even chance of improvement.

2024-07-12 · open the original ↗

• Model data around your queries, and choose the partition key early. Cocchi, a PM on Microsoft's Azure Cosmos DB team working on the .NET and Go SDKs for the NoSQL API, says to denormalize, for example by embedding a person's addresses in the person document. Cosmos scales horizontally by adding machines, and the partition key must be chosen when a container is created. The logical partition is the property value, such as a city, which is hashed to a physical partition, the machine that stores it. A good key reduces cross-partition queries but cannot eliminate them.
• Size throughput in request units before you build. A point read, one item fetched by ID and partition key, costs about one RU for a 1 KB document. Cocchi points to the capacity calculator, which takes document size and requests per second, and warns that under-provisioning shows up as 429 throttling errors.
• Direct mode is fastest, but gateway mode limits open connections. In direct mode the .NET and Java SDKs connect straight to the backend nodes, opening four connections per physical partition, one per replica. With hundreds or, for some customers, thousands of partitions, that adds up, so the gateway's extra hop can be worth it. She says the client is heavy, carrying routing, caches and service logic, so create it as a singleton: one client per account for the application's lifetime.
• Consistency choices trade freshness against latency, especially across regions. Cosmos offers five modes. Strong consistency commits every write to all replicas, in every region, before acknowledging it. Cocchi recalls a limit of "like 5,000 miles or something" for strong consistency with multiple write regions. Bounded staleness lets reads lag by K versions or T time, and throttles writes if the lag exceeds those bounds. Session consistency lets a user read their own writes. She recommends at least two regions for any production app, and says the point-read SLA is under 10 ms.
• Use diagnostics and OpenTelemetry, and write your own retry policy. Every SDK request returns a diagnostic string listing retries, the regions contacted and the handler pipeline; support will ask for it when the P99 is affected. The .NET and Java SDKs now emit OpenTelemetry traces, with latency thresholds that capture diagnostics automatically and send them to tools such as Application Insights. The SDK retries 429s only a limited number of times, so she says the application should still handle cases such as 408 timeouts from a faulty pod or high client CPU. She was due to cover these topics at VSLive Redmond in August.

2024-07-11 · open the original ↗

• Feed algorithms reward frequency, which makes quality stand out. Rose, who blogs at samwho.dev, says curation algorithms reward daily posting and LLMs make low-quality volume easier to produce; he posts rarely and avoids naming specific technologies (load-balancing algorithms, not a particular implementation) so posts age well.
• Raw web standards are his bet on longevity. He uses web components without helper libraries and no framework, with TypeScript the only dependency he worries about over 10–20 years; PixiJS (about 600 KB, of which he uses a sliver) may give way to his own 2D renderer, which he says can recreate simpler visualizations in under 10 KB.
• Readers don't realise the visualizations are clickable. He takes at least half the blame: through several rounds of watching friends scroll past, a yellow button and then a static cursor both failed, and only an animated cursor labelled "click" worked, teaching him that people skim rather than read top to bottom.
• His posts are simulations, not drawn animations. Each widget is a configurable web component over domain-modelled classes (load balancers, servers, requests); randomness in the load-balancing post cost him evenings waiting for the right behaviour, so the queuing post he was paid to write for Encore's blog uses uniform timings, a realism trade-off he dislikes.
• The web should move beyond text for teaching. He calls himself a visual learner despite the research Hanselman cites against learning styles, says video has never worked for him, and, echoing Andy Matuschak, argues interactive media could make educational material 10x better than text.

2024-07-09 · open the original ↗

• Outsourced identity is now the norm, so integration is the hard part. Sempf, an application security architect, says he long told developers not to build their own identity system: code against corporate AD or use Google or Facebook OAuth so they never store passwords. He says providers such as Okta, Ping and Microsoft Entra now do this at scale. The risk has moved to integration, such as moving a 2004 Spring MVC app with its own password store onto Ping. Vendors' sample code also takes shortcuts, like repeatedly sending basic-auth credentials in a header, which he says is unsafe even over TLS.
• Misconfigured cloud systems are a frequent finding in his assessments. The open S3 bucket protected only by an unguessable URL is his classic example. Sempf says the expert who knew how to tune the on-premises servers, Cisco gear and WAF no longer applies once everything moves to Azure. Configuration has to be relearned, mistakes are "rampant", and providers keep changing their back ends.
• SQL injection never went away; he now finds it in GET parameters. ORMs parameterize queries, but teams hand-concatenate SQL when the ORM falls short, for example searching two companies' products after an acquisition. Because the examples always show form posts, Sempf says developers assume query-string parameters can't be injected, when they are easier to inject and less validated. He used to find one or two injections a month and lately one or two a week, all in GET parameters. The only real fix is a parameterized query at the database layer.
• Open source is a trust chain that few people examine. Beyond API keys and connection strings checked into public repos, Sempf points to dependency chains four or five deep, kept up by volunteers without oversight. He calls the XZ Utils backdoor, inserted by a community helper after the maintainer fell behind, a nation-state attack three years in planning that was caught by "blind luck", when someone noticed a timing change in performance testing. Giard counters that closed source also needs trust, but has more eyes on it.
• Use actionable security guidance and fix one thing at a time. Sempf says the OWASP Top 10 is mostly for persuading a CIO to fund security, while the Proactive Controls are its how-to side. He advises picking one control and improving it rather than being overwhelmed. For developers curious about the security world, he recommends a local BSides, of which he says there are 190 worldwide, usually with 300–400 attendees, over the fire hose of DEF CON.

2024-07-08 · open the original ↗

• Windows 11 Notepad now spellchecks and autocorrects by default. Microsoft had offered the feature to beta users since March 2024 and has now pushed it to all Windows 11 users. The host expects pushback from users who want Notepad unbloated and from coders who don't want code autocorrected. Each feature can be turned off separately in settings.

2024-07-08 · open the original ↗

• Graph-grounded chat queries your own meetings, emails and documents. Copilot for Microsoft 365 uses Microsoft Graph to answer questions about items in your organization. Giard says it respects their permissions and won't return anything you don't have access to. It lives at the Copilot link on Microsoft365.com, which goes to Microsoft365.com/chat.
• Built-in prompts handle common tasks such as recapping a meeting. View prompts brings up suggestions. The recap-meeting prompt lists recent meetings, then summarizes key takeaways and actions in separate sections, including who is responsible for each. In the demo it also linked a PowerPoint deck and documents mentioned in the meeting, which Giard had not attended.
• Free-form prompts keep context across follow-ups until New Chat. Giard asked for a summary of emails and Teams chats and channels that mentioned him in the past 14 days, then asked it to go back a week, and again a week, and it understood each follow-up. Clicking New Chat clears the history so a different topic starts fresh.
• Typing a slash lets you pick a specific meeting. Giard used it to choose an upcoming meeting and ask how to prepare for it. He got advice on what the meeting would cover and what to do beforehand.
• Graph supplies facts the prompt leaves out, such as reporting lines. Asked to write an email to "my boss's boss", declining a Tuesday 1 p.m. meeting, offering to watch the recording and asking how to help, Copilot identified the person through Graph without being told. It drafted a professional email drawing on that person's information. Giard advises being as specific as possible about the tone you want.

2024-07-05 · open the original ↗

• Bring the gateway up first, then add devices one at a time. The host says patience makes setup smoother. A UniFi OS gateway comes online either through the UniFi mobile app over Bluetooth or from a wired computer on a LAN port browsing to the IP on its screen, typically 192.168.1.1. The UI account needs multi-factor authentication, which Ubiquiti mandates as of July 2024.
• Add a local-only admin as an emergency backdoor. The host recommends this in case the UniFi cloud login fails. He also advises the Official release channel, scheduled auto-updates for UniFi OS and apps, and leaving the weekly cloud config backup on, or taking manual backups if you disable it.
• Change the default 192.168.1.x subnet if you might VPN sites together. Turn off auto-scale and change the third octet (his example is 192.168.4.1), but leave the /24 mask alone. The DHCP range defaults to .6–.254; widen or narrow it according to how many devices need static IPs.
• The guest checkbox isolates a network from your other subnets automatically. Ticking it creates firewall rules that allow internet access but block your other subnets. On the guest Wi-Fi he points to client device isolation (right for hotel-style networks, wrong where you cast or print), speed-limit profiles (his example is 10 Mbps each way) and a blackout schedule for business hours.
• DNS Shield will likely break the gateway's built-in ad blocking. The ad blocking filters DNS, but gives no visibility into what it blocks. Suspicious activity detection set to Auto is IDS, which only notifies; notify-and-block is IPS. It costs gateway CPU, but the host says home users would probably never notice.

2024-07-05 · open the original ↗

• Client demand for AI moved from strategy, to pilots, to production revenue. Pendse, ThoughtWorks' global director for AI strategy, says clients first asked whether generative AI was a threat or an opportunity. Next they wanted a handful of POCs, which at many companies became hundreds. Then they wanted platforms to standardize data versioning and security, and now they ask how to put these into production and make money from them.
• Easy RAG demos inflate expectations that production can't meet quickly. Dragging PDFs into a web tool yields a chatbot, so executives expect enterprise-wide scale "in another two weeks or two months". Pendse says security staff have to learn AI, and AI staff have to flag risks, to catch AI-specific issues. Those include whether a cloud agreement protects the data, staff uploading data classes that aren't allowed, and tools that by default use prompts to improve the model. A shared platform lets the enterprise make such decisions once instead of in every POC.
• AI makes money by freeing highly paid experts from mundane work. He calls RAG for knowledge access the first, perhaps only, enterprise generative-AI use case with traction. He cites a client team's "search, ask, do" progression: find the source document, then a chatbot, then agents that act. His examples are drug-discovery researchers reading papers and cleaning data, and financial analysts reading thousands of 10-Ks. For a question spanning proteomics, genomics, immunology and radiology, he says a multi-agent system fine-tuned per field could answer faster than five or six human experts.
• Cost, missing baselines and compliance slow adoption. Per-token spending scales unpredictably with demand, though Pendse notes prices are falling as OpenAI cuts costs and competition grows. Many firms can't quantify value because they never measured a baseline, so "the data house" has to be in order first. In financial services, he says the Federal Reserve's SR 11-7 requires traceability of the data used to train a model. LLM vendors won't provide that, so firms must either persuade the auditor or regulator or train their own model at great expense.
• Weak CI/CD practice gets worse once AI is layered on top. Pendse says continuous integration for AI is "one order of magnitude more complex". Teams that keep long-lived branches and rarely merge will see the same chaos spread to model and data versions. He argues fixing those processes is the route to showing traceable value, such as cost savings or new markets.

2024-07-04 · open the original ↗

• .NET Aspire is a bundle, not a framework. Edwards, a principal architect at Microsoft, describes it as templates, NuGet packages, a development-time orchestrator and full adoption of OpenTelemetry, together aimed at building distributed apps; he treats "cloud native" as a term that means different things to different people.
• Good abstractions let you punch through them. He says Web Forms' real sin was making it too hard to escape the abstraction for the 20% case; Aspire's high-level calls sit on ordinary lower layers (adding Postgres just adds a container and injects a connection string as an environment variable), and F12 shows the source.
• Components encode opinions developers would otherwise rediscover. Using Redis as the example, he says an Aspire component picks the client (StackExchange.Redis), registers it in DI with a set lifetime, binds options to configuration, references a community health-check package and wires up OpenTelemetry, all on by default; before Aspire, that client's logging took a raw TextWriter.
• The AppHost describes the whole app in C#. It extends the host-builder pattern to containers, executables, projects and parameters with references between them, and running it starts everything and tears it down; he concedes the top early feedback was that this lifetime is too simple, and that keeping containers running between sessions is a hard problem being worked on.
• Deployment changes are optional. He stresses the orchestration is mainly for development and an existing deployment can stay untouched; alternatively a deployment manifest feeds the Azure Developer CLI, tools Amazon is building, or the community "Aspirate" tool for Kubernetes.

2024-07-02 · open the original ↗

• Enterprise AI demand centres on conversational agents where security is simple. Kotvis, a principal architect in Insight's AI-enabled apps and data group, says organizations want knowledge bases and help desks where role-level security is not in play. He sees enrolling enterprise security into AI, so people don't get information they shouldn't, as the biggest opportunity.
• Every legacy app gets rehosted, refactored or re-envisioned. Kotvis says scarce mainframe talent and the pull of cheaper, independently scalable cloud infrastructure drive modernization, especially in finance and insurance. After sorting an estate into commercial and custom apps, a custom app takes one of three paths. Rehosting is lift-and-shift with adjusted connection strings. Refactoring makes a working app able to use PaaS SQL or a distributed cache. Re-envisioning rebuilds an app that no longer delivers value, starting with UX and visual design. Of 100 apps, he estimates 5–10% are re-envisioned and 10–20% reach end of life.
• Finding people who understand the legacy system is the biggest delay. Old schemas hide "sins", such as columns tripled as first_name_1, first_name_2 and first_name_3, and cryptic mainframe abbreviations. Kotvis says that with full information and available stakeholders, projects would take half or a third of the time.
• Pragmatic 80/20 design with escape hatches beats dogma. Kotvis uses Entity Framework for most work but plans for the 10–20% that needs hand-written SQL, such as common table expressions or recursion, saying "dogma is kind of going to kill any momentum". He applies the same thinking to choosing tools like Apigee, MuleSoft or Logic Apps, and builds modularly so multi-cloud moves or acquisitions don't force a rewrite. He predicts brownfield skill will be highly valuable for the next five to ten years. Giard adds his own experience porting a .NET 1.1 app with help from Michael Feathers' book on legacy code.
• Dev containers and 12-factor principles make environments predictable and portable. Kotvis runs about 50 dev containers on a Linux server in his basement through VS Code, which he compares to GitHub Codespaces. They keep multiple Go or .NET versions apart and end "works on my machine". He uses the 12-factor app principles as guidance and notes that containers move easily between Azure Functions, Container Apps, Container Instances and App Service. Even so, shared files, caches or a socket backplane still need deliberate design.

2024-07-01 · open the original ↗

• A pump can be built for a hundredth of commercial cost. De Bock, an associate professor at the University of Otago and a paediatric endocrinologist, says the low-cost pump programme (LEAP) began in 2018 when a University of Canterbury mechanical engineer refused to proceed unless the device cost 1/100th the price; a national science challenge grant funded both the hardware and qualitative work on how people would view a non-commercial device.
• Pumps are essentially precise syringe drivers. He says the hard part is accurate, robust plunger delivery tested against international accuracy standards; he argues a cheap pump that lasts less long is more forgivable, like a cheap drill, unlike a several-thousand-dollar unit a parent fears a child will break.
• Automation ends the trade between quality and quantity of life. Injections can reach healthy glucose only with a very restrictive life, he says, citing data that people with type 1 make over 100 health decisions a day and describing episodic "diabetes burnout"; he endorses Hanselman's lane-keeping car analogy for the pump-sensor-algorithm loop.
• Technology acts as a health-equity equaliser. His team's cross-sectional study of everyone in New Zealand found CGM use removed the outcome gap predicted by socioeconomic status or ethnicity, including for Māori; he expects Chinese and South Korean entrants (10–14-day sensors) to push CGM prices down to compete with test strips, while much of the global south still gets only syringes and finger pricks.
• Don't build one and wear it yet. He says a study putting the pumps on people was about to be published, open access, but they are untested against everyday wear; his advice is to build one out of interest and pump into something other than yourself, since safety still has to be ironed out.

2024-06-30 · open the original ↗

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

2024-06-28 · open the original ↗

• Multithreaded Income is personal finance aimed only at technologists. Griffin, independent for 13 years, started it in September 2023 to talk about using tech skills to build several streams of income, prompted by a job market he says has been "deflating" for the past year or two. He aims for weekly episodes but often manages one every two weeks, and had about 40 out at recording.
• A frugal audience is very hard to sell to. His earlier podcast, Two Frugal Dudes, grew somewhat popular but earned perhaps a couple of hundred dollars, enough to pay the accountant, so it was shut down. He took from that the decision to narrow the new show to technologists and software developers.
• His guests fall into a few kinds of income thread. Moonlighters freelancing beside a day job, people who go fully independent after quitting or losing a job, and builders of SaaS or installed products, which usually start by solving their own problem. Some of those businesses have passed the million-dollar mark; one guest's conference stage timer earns about five figures a month with one developer. Others run conferences as businesses, write books or make courses.
• The show covers the business of software, not the tech. Griffin says nobody pays because a product is written in TypeScript or C# or built on Next.js or Laravel; they pay for the problem solved. He wants to cover what developers avoid: pricing, finding first customers, getting feedback and marketing.
• The plan is audience first, paid offerings later. The podcast earns nothing yet. Griffin intends to follow it with coaching, from group "masterminds" with built-in accountability up to one-on-one sessions, and with courses, starting with one on going independent as a consultant. He says he has not built those yet because he does not have an audience to sell to.

2024-06-25 · open the original ↗

• RAG lets an LLM answer with what your organization uniquely knows. Bogue defines retrieval augmented generation as bringing internal content to the model at query time, so answers use the organization's knowledge and language. He calls Copilot a wrapper around the model, which today is ChatGPT's, and says Microsoft could swap or add models without users having to care.
• M365 Copilot already does RAG by sitting on Microsoft Search. Bogue says it turns your prompt into a search query, retrieves results and generates a response from them. Because it goes through search it keeps existing permissions, so content you cannot open in search, such as an executive's salary, never reaches the answer. He found the Microsoft365.com portal chat the best way to use it at the time of recording.
• Better RAG answers come from tuning search. When answers go wrong, Bogue's remedy is search work rather than Copilot work: marking content as authoritative, building information architecture, or tuning search directly where no architecture exists. Connectors such as one for ServiceNow put line-of-business data in the index too. He expects questions like "what did we last sell this customer?" to become possible as small technical gaps close.
• Bogue credits Microsoft's responsible-AI handling, but calls it unfinished. His suicide-prevention research work gets shut down whenever the system does not read the word "suicide" as prevention, which he calls appropriate. He calls the commitment not to train models on a tenant's content or queries "huge", and says bias keeps creeping in however models are trained. His view is that AI will change jobs more than eliminate them, taking work such as Excel cell manipulation and slide design.
• Readiness for Copilot means change, content access and information hygiene. He points to a paper he wrote for AIIM, "Organizational Readiness for Generative Artificial Intelligence", still a draft then, which the host describes as about 30 pages. It covers change management so staff are not afraid, getting priority content into search so Copilot can reach it, and information architecture as hygiene.

2024-06-21 · open the original ↗

• Observability means getting ahead of problems rather than reacting. Crist, a solution architect in Elastic's partner organization, says teams traditionally waited for a "website's down" call, then pulled OS, database and firewall logs and CPU, memory and disk metrics separately and matched them by timestamp, sometimes in a spreadsheet. He defines observability broadly: uptime and performance, the customer's experience on a site, and successful rollouts.
• Logs, metrics and traces belong in one place under a common schema. Crist says Elastic began as search over petabytes of unstructured data; users started ingesting log files into it, and Elastic built tooling around that. Its Elastic Common Schema, which he says Elastic has donated to open source, keys on timestamp plus common fields such as host and IP address, so engineers need not switch tools. He says it also covers Kubernetes pods and containers.
• Traces and machine learning surface anomalies before outages. Traces show latency between services, normally 7–25 ms in his example, jumping to 100 ms or two seconds, and service maps show which service eats the most of a five-second response. Crist says Elastic ships "three or 400" machine learning jobs for anomaly detection, and correlation analysis finds which fields differ when a log rate goes from 100 to 10,000 entries a minute. Security examples include logins at 2 AM from users who log in 8 AM–6 PM, and data leaving for an unusual country.
• Elastic's AI assistant applies RAG to logs, knowledge base and business data. Crist describes asking in plain language whether an error matters or to summarize today's alerts, with the answer built from the customer's logs, metrics and traces through a model such as Azure OpenAI. Resolved problems go into a knowledge base for the next engineer. Customers adding revenue data and service-level objectives can ask the revenue impact of an outage, in his example $5,000 a minute; automated responses to alerts are what he sees as the future.
• Synthetic monitoring adds the user's view to the service's view. Crist says the roadmap includes scripted jobs that simulate a user, such as a Chrome browser in Japan searching a product and checking price and inventory, run every five minutes or hourly. He says this matters most for multinationals running thousands of pods worldwide. He also notes Elastic is API-driven, has a basic open-source version, and integrates with case tools such as ServiceNow.

2024-06-20 · open the original ↗

• Older developers should mentor with "yes, and", not stop energy. Hanselman, now Microsoft's VP of developer community, says his identity is teacher, not company, and his goal is making early-career people succeed, including 40-something career switchers. Elders offer kind, patient mentorship and historical context, and should answer a junior's idea with "here's what we learned when we tried" rather than "we did that before".
• Information is easier to get, but getting noticed is harder. Hanselman stands by his "dark matter developers" post: for every social-media star there are a thousand people doing good work, and being noticed may not matter. Social media is "algorithmically designed to make us feel bad", so the only comparison worth making is against yourself yesterday.
• Don't make juniors earn the high level by learning everything below. Hanselman compares insisting on assembler and dump debugging to making his 16-year-old learn stick shift when 3–4% of American cars have one. His standing advice is to go one level below your comfort zone (V8 for JavaScript, the CLR for .NET), and he tells a story of a Black Girls Code student who debugged his "broken toaster" by asking whether the neighbours had power.
• Hire for curiosity, not for the language someone already knows. Hanselman relays Scott Hunter's line about hiring a basketball team for height, meaning intellectual height, and cites a Microsoft ASP.NET engineer who was a Python person before joining. The hosts agree programmers are different "species", and good leaders fit each to the work they love.
• Feedback thrown as a rock goes unread. Revisiting a 2015 panel where he called a comment "either elitist or mean", Hanselman says the critic was 75% right but the delivery buried it. His Build message: Microsoft is 300,000 people and "not organized enough to be as evil" as accused, its engineers are findable, so vote with your feet and engage them with empathy.

2024-06-18 · open the original ↗

• AI services are pulling a late-moving public sector into the cloud. Martensson, an Azure consultant who runs his own company, says government lags because of decades of legacy IT and some of the most sensitive data there is, such as citizens' patient records, governed by codes and EU rules. Because the large AI services run on Azure, using them on that data means moving it to the cloud, possibly petabytes of it.
• His region is already migrating, ahead of its peers. He works for Region Skåne, which he calls Sweden's third-largest region, with a large centralized IT department; Azure's data centers in Sweden "make the conversation smoother". Its Azure subscriptions went from about 40, few in use and not set up uniformly, to over 100 and rising, and he says local Microsoft representatives tell them other regions and communes are mostly behind. He knows of no separate government cloud in Europe like Azure's in the US, so the region uses public cloud as is, starting with things like its public website, classed as critical infrastructure because it informs citizens in a crisis and so must stay available.
• The public body must own cloud governance, not its suppliers. Martensson says the region needs its own core team, which he sits on, deciding "what goes" and telling its long-standing server and operations partners. That competence does not exist in-house in the public sector, so it must be trained or hired; he calls the work honourable because "this is my tax money at work".
• Meet the regulation, not the legacy system built for it. When told everything must go through an old central control system, he asks for the actual requirement, since no law mandates that system, only governance. He refuses to "weigh the cloud down" by reproducing it in Azure and uses the platform's built-in auditing and other services instead.
• AI may replace one of two doctors reading pathology images. In a current project, Martensson says, pathology photographs that traditionally need two doctors to check for tumours are reviewed by AI, which flags regions of interest, plus one doctor. The two-doctor and AI-plus-doctor results are being compared to see if they are "within reason"; his case is that the AI does not tire and scales, saving doctors' time.

2024-06-14 · open the original ↗

• Open-source big data descends from Google's published papers. Brust, a journalist and analyst who was once a developer, says Google published papers on MapReduce and the Google File System, and a Yahoo team built an open-source implementation that became Hadoop and then the company Hortonworks. Hortonworks later merged with Cloudera, which kept the name though, he says, mostly Hortonworks' technology survived; Brust discloses Cloudera is his client. He says Hadoop has since given way to Spark.
• Microsoft arrived late because big data meant Linux and Java. Brust says Microsoft entered by partnering with Hortonworks to port Hadoop to Windows, then ported its own distribution back to Linux, followed by its HDInsight cloud service; around 2010 the world was not Microsoft-friendly. The cloud lowered the entry cost of a rack of commodity servers and abstracted away the Linux setup that Windows and Visual Studio developers would not have done. He calls Amazon's Redshift arguably the first cloud data warehouse, and notes Google's BigQuery.
• Fabric is real integration, not a rebranding. The industry's problem now, Brust argues, is fragmentation: even Microsoft customers had to stitch together Azure SQL Data Warehouse, Data Factory, Event Hubs and Blob or Data Lake Storage. Fabric unifies them technologically and economically, as SaaS paid for as one pool of compute; it is Spark-based for data engineering and data science, SQL Server-based elsewhere, and includes KQL databases from Kusto. He traces the lineage from SQL Server's OLAP Services, acquired with Israel's Panorama, to Amir Netz, now a technical fellow and "mastermind" behind Fabric. Brust is a 20-year Microsoft MVP and was in Fabric's early adopter program.
• Rivals are extending products rather than unifying platforms. Brust says Google is adding capabilities to BigQuery, and Snowflake is pitching its warehouse as a data platform, but he does not think that equals putting an abstraction over "seven or eight specialized platforms". He grants that the Cloudera Data Platform is also a platform. In his view the 2010–2012 rush of invention was the "scatter" phase and the industry is now in the "gather".
• AI's role in data will be assistive, starting with governance. Brust says the data world does not yet know what to do with AI for its mainline work. He hopes for easier generative AI and machine learning, sees AI populating data catalogs and drafting governance policy as a starting point for data stewards to correct, and says self-healing ETL pipelines that adapt to schema changes are "science fiction-y" but starting to happen.

2024-06-13 · open the original ↗

• XPF runs unmodified WPF apps cross-platform by swapping WPF's renderer. James says Avalonia forked WPF in December 2021 and replaced milcore, the unmanaged DirectX 9 compositor, with Avalonia's renderer, leaving PresentationCore and PresentationFramework largely untouched. Apps and third-party control suites like Telerik, Syncfusion and DevExpress still believe they run on WPF; Hanselman calls it "a wildly successful lie".
• A pure-WPF port takes seconds; native dependencies are the real work. Per James, porting means changing the csproj's project SDK to XPF and adding a license property. Hanselman's Baby Smash, with 106 lines of XAML but calls into winmm, user32 and speech synthesis, ran on James's Mac after about five minutes of commenting out the speech synthesizer.
• Avalonia draws every pixel itself and needs little from the OS. The only native control is the window. Rendering sits on Skia, with DirectX pluggable, and embedded devices like Schneider Electric's HMIs get pixels pushed straight to the Linux framebuffer. James says this makes supporting many Linux flavours easier, including a sale to a Chinese conglomerate on a Chinese distro on Arm hardware, though text rendering is the hardest part.
• Total code reuse across devices means "you must hate your users." James separates cross-platform from cross-paradigm: modal dialogs, async browser dialogs and single-view apps all change by device. WebAssembly and iOS/Android are internal previews. Hanselman suggests 70–80% reuse is good; James thinks good architecture gets higher, and calls Apple's Catalyst ports on the Mac subpar.
• Paid XPF exists to fund open-source Avalonia. James says Avalonia is about nine full-time people, bootstrapped rather than VC-backed, and earned its revenue from support and porting services. XPF targets a single-digit share of WPF apps as a stepping stone, so a free XPF is not possible; he offers a Hanselminutes landing page with a 30-day trial.

2024-06-13 · open the original ↗

• Every UniFi network needs four components, often combined in one box. They are a gateway (router/firewall), wireless access points, switches and the UniFi Network controller; the Dream Machine Pro combines gateway, switch and controller without Wi-Fi, while UniFi Express combines gateway, Wi-Fi and controller with a single LAN port. The host recommends beginners run the controller on the gateway, so it is always on and new devices find it easily.
• Beginners should pick a gateway with a switch, not built-in Wi-Fi. The host prefers the Cloud Gateway Ultra to UniFi Express: the Ultra costs $129 and has a four-port switch; Express costs $149 with a Wi-Fi 6 radio, but supports fewer devices (up to 30 UniFi devices) and its small antenna covers about an apartment or small home. The Ultra lets you place stronger access points where coverage is best. Most gateways handle up to 1 Gbps; above that, check each product page's throughput with IDS/IPS enabled.
• A small PoE switch covers most home wiring. His pick is the 8-port, AC-powered Switch Ultra 60 W costs $159 and can power two or three PoE devices; the $199 Lite 16 PoE (8 of 16 ports PoE) is the step up, and he runs one in his garage. He also likes the $29 Flex Mini, powered by USB-C or PoE, for wiring devices behind a TV.
• Beginners should skip Wi-Fi 7 and buy Wi-Fi 6 access points. Wi-Fi 7 costs more and few clients can use it, though he expects that to change in four or five years. Of the PoE models under $150 he rates the $129 U6 Plus best, about double the U6 Lite's throughput for $30 more; the U6 Pro at $159 is fine too, and he suggests a U6 Mesh outdoors.
• With access points, less is more. He advises starting with one or two centrally placed access points in a typical wood-and-drywall home and adding more only where coverage proves lacking; without PoE switches, each needs a PoE injector. He built the UI Product Selector site he uses here to filter and share builds, and the video opens and closes with a pitch for Rogue Support's network design service.

2024-06-11 · open the original ↗

• An administrator builds five pieces before anyone gets a workstation. In the Azure portal under Microsoft Dev Box, Giard creates a Dev Center (about 12 minutes this time, 10–15 before), then a project inside it (under five minutes), a Dev Box definition, a Dev Box pool, and finally access control on the project. The project's optional Dev Box limits cap how many developers can use it at once, for cost or practical reasons; he leaves it off, and says tags only add name-value pairs for sorting and filtering reports.
• The definition fixes the image and hardware, and hardware sets the price. Created under Dev Box Configuration in the Dev Center, it takes an image (he picks Visual Studio 2022 on Windows 11 Enterprise) pinned to a version or to "latest". More CPUs and RAM cost more per running hour; he takes the cheapest, 8 vCPUs and 32 GB RAM, with 256 GB of SSD storage, and turns on hibernation to save money after inactivity. Custom image templates are deferred to another video.
• The pool ties a definition to a network and a shutdown schedule. Created under the project's Manage section, it uses a Microsoft-hosted network by default, which Giard says usually makes sense, with your own network as an option for specific configuration. He enables auto-stop at 7:00 p.m. US Central as a cost-saving measure, leaves single sign-on off, and must confirm he holds the licences for the Windows and Visual Studio software; creation took about seven minutes.
• Two roles on the project's Access Control (IAM) tab grant access. DevCenter Dev Box User lets someone create dev boxes from the definition; DevCenter Project Admin suits someone taking over administration, for example while you are on vacation. Assigning a role to a user took about a second.
• Users create boxes at devbox.microsoft.com and connect by browser or RDP. The portal warns creation averages 25 minutes and can take up to 65, with an email when ready; his took about half an hour and arrived with Visual Studio and PowerShell installed. The browser needs nothing extra, but without single sign-on it asks for a password; Giard recommends the RDP desktop client where you have it, keeping the browser for a borrowed machine.

2024-06-11 · open the original ↗

• Responsible AI has moved from principles to operational practice. Sandford, Microsoft's Developer Engagement Lead in Asia, says this year's talks are about operationalizing responsible generative AI. She advises a pinch of salt with influencers' claims: by the time a trick is shown it is usually blocked, and one TikTok clip she checked presented as a Microsoft release a deepfake project that was only research, used generated rather than real faces and was not even in preview.
• Protection works as defence in depth across four levels. Sandford lists the foundation model's training, then system or "meta" prompts that set purpose and rules (no harm, no offensive content, no copying others' material), then content filters, then the user experience layer. She describes jailbreaking as getting a model to reveal or work around its hidden rules, often through a "fictional story" without constraints, and says it used to be possible before GPT-3.5, when there was no system prompt.
• Content filters are adjustable because some uses need sensitive topics. Sandford says jailbreak and safety filters are sliders that can all be turned on, which suits a front-end customer-service bot. A health application's patients may need to discuss sexual or violent content, so there the filters stay off and the conversation is controlled through more detailed system prompts.
• Grounding settings and red-teaming limit hallucination. She says OpenAI's scale runs from zero, fully grounded and answering "I don't know", to four, while Azure OpenAI stops at two, enough in Microsoft's view for an imaginative story. A company bot grounded in its data can be set to zero and told in the prompt not to make answers up; temperature controls how far repeat answers diverge. The final level is red teams trying to break the bot, with improvement that never ends.
• Humans stay accountable for what their AI agents say. Sandford cites an airline whose chatbot told a bereaved customer he could claim a funeral discount within 90 days after buying; the airline argued in court the bot was its own legal entity, and she says the judge held it responsible for all its agents, AI or human. She adds that transparency means telling users they are talking to a bot and handing off to a human where needed, and uses a bank manager openly lending only to straight white men aged 30–40 to show something can be transparent without being fair.