Souk's Notes - Unofficial Summaries

2023-10-14 · open the original ↗

• Generative AI still needs human curation, even at scale. Turner, VP of data at the consultancy Kin and Carta, says predictive and generative AI overlap but differ in their maths: generative AI becomes part of the product and can surprise with new combinations. One client is letting it generate output for a million products, and he says that cannot be done without human curation; he prefers to treat "hallucinations" as wrong best guesses from probabilities.
• Projects should start from a shared data foundation, not a model. He says the work begins with listening to the client, then building a data platform that many solutions can use. That means agreed definitions (what is a customer, when a sale counts, what the geographies are), a data dictionary, stewardship, governance and quality checks. About 90% of the time goes on cleansing and preparation, and the product on top is then a much lighter lift.
• Waiting for clean data is a mistake, because it never arrives. Turner says the bigger and more successful the company, the more embarrassed it is by its data, and the stock consultant answer is two years of preparation. His answer is that acquisitions and new sources keep adding mess, so build governance to manage it. Data literacy matters too: one data point challenged in an executive meeting with no provenance can taint the whole data programme.
• AI feedback loops push organizations toward more of the same. Once a model is deployed it changes the system that produces its training data, so over-reliance narrows toward a single optimum. He argues that organizations need deliberate exploration of options the data does not suggest, such as a geography with no data yet, or they will miss the other "mountaintops".
• Security inside Azure is the headline for enterprise generative AI. Kin and Carta has been a Microsoft partner from the start, and Turner says running OpenAI and Azure models in a secure Azure environment is table stakes, because the model may see every document. If you load every Word document and PowerPoint into an LLM, he warns, you may learn things about your organization you didn't know. He says AI is not right for every problem, but cloud access gives an opening even to a mom-and-pop shop.

2023-10-13 · open the original ↗

• Too many sound devices is the usual cause of lost audio. Hanselman made the video after his son couldn't tell where his sound was going: headphones, Bluetooth, HDMI monitors with speakers and virtual devices (Oculus, NVIDIA Broadcast) all show up as outputs. Right-clicking the taskbar volume icon opens the output picker. To find the right device, he switches between devices while clicking the volume slider to make it beep, or plays music.
• Rename devices and disable the ones you never want. Devices can be renamed and given icons in the modern Sound settings. The older classic dialog is under "More sound settings", or Win+R and mmsys.cpl, and there he used "Don't use this device" to hide the speakers on his three HDMI-connected Dell monitors. His warning: disabled devices disappear, so you have to remember you turned them off.
• Set the default device and the communications default separately. He found NVIDIA Broadcast set as his default communications device for Teams and Zoom and moved that to his headset, leaving his 5.1 speakers as the main default.
• The same controls apply to inputs, so disable stray microphones. On his gaming PC he disabled the front mic jack and his Logitech Brio webcam mic, so a call or recording can't quietly pick the wrong microphone.
• A third-party volume app beats the built-in per-app mixer. He recommends EarTrumpet, a free third-party app from the Microsoft Store by Rafael Rivera and others (he thinks it is also on GitHub). It replaces the taskbar volume icon with per-device and per-app volume, respects devices you have hidden, and still links to the legacy Windows dialogs.

2023-10-12 · open the original ↗

• Proactive monitoring caught the intrusion; snapshots restored it without downtime. In mid-August 2023, Zabbix and Grafana rules flagged the FreePBX firewall being disabled on seven Crosstalk-hosted systems, then changed authorized_keys files; a cron job had added an admin user and keys, called a server in France, and given attackers root, while calls kept working. ClearlyIP blocked PHP execution through the phone apps port within a day or two, and servers were restored from pre-attack snapshots after hours, with customers emailed throughout.
• The likely initial vector was a DEF CON-disclosed FreePBX zero-day. On August 12th 2023, under a week before the hacks, a researcher with the handle "good pseudonym" showed at DEF CON 31 that a phone API could be entered with a registered phone's MAC address, the hard-coded password "login" and a spoofed Digium/Sangoma user-agent, then used for plain command injection into a PHP exec call. The host infers the link from the timing.
• The host argues Sangoma's patches and advice fell short. The researcher says he reported in early May, got one reply and then silence, and demonstrated his exploit still working against Sangoma's patched API. Sangoma's blog cites a September 5th 2023 fix and urges closing phone apps and enabling the responsive firewall; the host's servers were IP-locked or firewalled yet still hacked. A September 28th re-patch was rated "minor", and version 14 and earlier get no fixes.
• Sangoma apparently removed its bug bounty page rather than pay out. Sangoma's blog thanked Systems Research Group, which the host assumes is the same researcher, but as far as he knows paid no bounty; archive.org last captured the page in March, and a Ward Mundy post dated September 17th 2023 still linked it, suggesting removal before Sangoma's September 20th blog post.
• The host thinks FreePBX is stagnating and Sangoma should give it up. He notes no significant update since version 15 in 2019, an aging CentOS 7 base with old PHP and OpenVPN, and a vague FreePBX 17 on Debian; with almost 80% of Sangoma revenue from recurring services, he sees no incentive to compete with its paid products, and urges Sangoma to sell or donate the project, offering to take it over himself.

2023-10-11 · open the original ↗

• Automating Zoom's clicks made the technology disappear from worship. After the church's first Zoom service on Easter Day 2020, Skeet found Zoom had an SDK and had an alpha of "Zoom and Enhance" running by the next Sunday: one double-click spotlights a person, mutes everyone else and stops any video, or starts the screen share and plays a video. He says the aim of all his work since is that no one notices the technology, with no five-second pause between sermon and hymn.
• The church is its people, so remote worshippers must not be dropped. Skeet planned for stages from April 2020: 10–20% of people in the building, then about 50%, then 80–90% with 10–20% still remote. He argues the pandemic was a wake-up call: churches had been failing the housebound, the ill and those away for centuries, could only have fixed it in the decade or so before, and should now stop.
• He only bought hardware his own software could control. The old analog Allen & Heath mixer had 14 channels and about 150 knobs and buttons, and intimidated volunteers. Skeet replaced it with a digital mixer (audio over USB, control over Ethernet) and pan-tilt-zoom cameras, ruling out proprietary protocols without an SDK, because he found no off-the-shelf system that even controlled a mixer and showed hymn words. The resulting app, "At Your Service", has since gained DMX lighting (red and orange on the reader at Pentecost) and, at the request of another church with Iranian refugees, live transcription through Google Cloud Translate, read on phones with a five-to-ten-second delay.
• His biggest challenge was giving each user everything needed and nothing else. Services are prepared in advance, for example from home, and uploaded to Google Cloud Storage, so the person at the desk on Sunday only syncs and presents. A Stream Deck XL (32 buttons that are also small screens) and a Behringer X-Touch Mini give a glanceable mute panel; both talk only to his software, never to the mixer directly.
• The system fits his own church and does not yet scale. No other church uses it live yet; two are interested and two more are planning AV revamps he will advise on. It is C# and WPF in a private GitHub repo, configured by JSON files with no GUI, and he would like to open-source it if he retires, "give me 10 years". He wrote his own small MVVM framework to learn how WPF works, and says he would use a tested off-the-shelf framework for commercial work.

2023-10-09 · open the original ↗

• The robot was a free review unit, not a paid video. Hanselman discloses that Elecfreaks sent him the XGO Lite CM4 robot dog to review but did not pay him for the video. It runs on a Raspberry Pi Compute Module, which he says is squarely in his interests, along with block-based programming he uses when teaching at schools.
• The hardware is unusually well built for the price. He points to careful packaging and an all-metal body with many servos. It has a front camera doing on-device hand recognition on a 1.5 GHz ARM Cortex, plus a grabber arm he uses by remote control to pick up a red ball. It can walk or trot. Elecfreaks advises not letting the battery drop below 20%.
• Network setup works but should be easier. The dog joins Wi-Fi by scanning a QR code from the app or by connecting to a phone hotspot with a name it recognizes. He recommends a 2.4 GHz IoT network for compatibility.
• The programming tool lacks polish. The XGO Blockly editor is a fork of Google Blockly, Apache-licensed, so Elecfreaks needn't publish its changes, though he'd like them to. It loads from a bare public IP address and talks to the robot over WebSockets across domains. Chrome and Edge blocked that and Firefox didn't. He wants it cloneable to run locally or in a container, or else served with a domain name and SSL certificate.
• He rates it highly despite the steep price. His score is 8 out of 10. At $599 he calls it an investment, "2 Xboxes", and says it is limited only by your ability to program it in Python or blocks. He plans to give it to a local high school robotics team.

2023-10-06 · open the original ↗

• Building from the command line works but gives no editor help. Hanselman builds a console app with dotnet new console, dotnet build and dotnet run on .NET 8, then a release candidate. Editing in Notepad gets no syntax highlighting and no IntelliSense, which is why people use an IDE.
• A full IDE shows the project's logical structure, not its files. Visual Studio's Solution Explorer isn't a file view: it shows libraries, analyzers and the target framework. He also shows IntelliCode's starred machine-learning suggestions, IntelliSense and GitHub Copilot. The released Visual Studio wouldn't open the .NET 8 project, so he had to switch to Visual Studio Preview.
• The new extension brings full-IDE features into the lightweight editor. Until then VS Code had only a language server, he says, and showed the raw files, including bin, obj and the project files. The C# Dev Kit, the new official extension, adds Solution Explorer, a Test Explorer and Roslyn-powered code navigation, and installs the base C# extension and runtime with it. It runs on Windows, Mac, Linux and Codespaces. You sign in with the same account as Visual Studio Community or a paid licence.
• His demo covers debugging and tests on a real project. He runs F5 debugging with breakpoints, watches and locals. On his Hanselminutes podcast site he browses project references and package versions, then runs the unit tests from Test Explorer. The same project opens in both Visual Studio and VS Code.
• He says it became generally available that day. It is installed by searching "C#" in VS Code's extensions, and he says it updates automatically. "Today" refers to the recording date.

2023-10-05 · open the original ↗

• Microsoft's OpenAI partnership long predates this year's product announcements. Lussier, who works with ISVs in Microsoft's Global Partner Solutions group (as does the host), traces it from OpenAI's founding in 2015 to CTO Kevin Scott, ex-LinkedIn, introducing Satya Nadella to OpenAI's CEO, then a first big investment in 2019 with an exclusivity agreement putting OpenAI on Azure. In 2020 Microsoft built OpenAI an AI supercomputer with 285,000 CPUs and 10,000 GPUs, top five in the world then but no longer, he says.
• Microsoft's own AI ran in parallel and is available now. Lussier dates Cognitive Services to 2016, when Project Oxford was rebranded, and says an "AI at Scale" division formed in 2020 also worked on large language models, partly with NVIDIA. He hopes OpenAI interest opens conversations about that full suite, noting that Azure OpenAI access is coming in staged rollouts while Cognitive Services need no wait list.
• Copilots over business data will change how users interact with systems. Lussier expects asking a system for "the top customer accounts in the last quarter" in conversation to be transformative, and says partners he works with are building their own copilots and adding roles such as prompt engineer.
• Adopting AI needs organizational change and a responsible-AI framework. Against the CEO who reads a magazine on a plane and wants it in weeks (the host's image), Lussier calls for change management, new skills and architectures, and legal teams rethinking terms. He cites Microsoft's framework of fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability, and on bias gives an AI image set of how Europeans see Americans in which every person depicted was white.
• The meaningful uses go beyond novelty demos. MediaValet, a Canadian digital asset management company, used Microsoft's facial recognition to identify individual chimpanzees across decades of the Jane Goodall foundation's images and videos. Microsoft added Inuktitut to its translation service with the government of Nunavut to help preserve languages, and Lussier points developers to Microsoft Learn's AI tracks and the Azure ChatGPT project on GitHub.

2023-10-02 · open the original ↗

• The Ukraine war will likely end with Putin replaced, the host predicts. High oil prices let Russia fund the war for a long time, but the host argues Russia can damage Ukraine without ever conquering it, given better-equipped, more motivated Ukrainian defenders. On Armenia and Azerbaijan, the host expects more Azerbaijani punitive action, Armenia drifting further from Russia, and the US building ties with both.
• China's local-government debt is the real threat to its economy. The host puts it at $23 trillion and says about 40% of the real-estate market is in default, so developers and lenders lose both principal and interest in a domino effect. The host does not expect a Taiwan invasion soon, instead speculating China may move on Outer Manchuria, which Russia took in 1860.
• The US is drifting toward deflation, not stagflation or collapse. The host expects the Fed to print money once inflation turns negative, sees no housing crash while inventory is low and demand high, and calls the Cuban Missile Crisis closer to world war than today. Saudi Arabia is selling Treasuries to rebuild reserves down over $200 billion since COVID, while China buys other US agency bonds.
• The West must treat Africa better to blunt China. The host calls France the worst colonizer, claiming some former colonies pay about 85% of national income to France for colonial-era infrastructure. Chinese roads, bridges and factories cannot be siphoned off like cash aid, and the host urges the West to follow Japan's approach of aid since the 1970s.
• Officeholders need age limits and money out of politics. The host proposes mandatory retirement at 75 with frequent competency exams, and calls Trump's indictments justified for an insurrection. The host predicts Biden wins in 2024 if he survives, RFK Jr. draws Republican votes, and a Modi third term would risk dictatorship; Citizens United and gerrymandering should go.

2023-10-02 · open the original ↗

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

2023-10-01 · open the original ↗

• The productivity–pay chart overstates how high the minimum wage can go. On calls for roughly $20 an hour, economist Christopher Clarke says the chart deflates output with the GDP deflator but wages with CPI, and the gap shrinks on a common index. The highest inflation-adjusted federal minimum, in the 1960s, equals about $11–13 today, a safe range; labor's falling income share is real but dates from around 2000, not the 1970s.
• The Fed would fight deflation hard if it arrived. The host has watched Chinese producer prices fall about 4% and consumer prices turn slightly negative; Clarke expects the Fed to cut and print, as during 2009's brief deflation. He says 2008 bank rescue loans were largely repaid with interest while homeowner relief was too small, a lesson applied in the COVID stimulus.
• Official inflation data are generally sound; Argentina is the exception. With headline inflation at 3.7% and core falling, Clarke points to Alberto Cavallo's Billion Prices Project, which applies government methods to scraped online prices. It found statistical agencies broadly reliable, except Argentina in 2007–2013, whose official inflation appeared to be about a third of the actual rate.
• Immigration, including low-skilled, is a long-run net gain. Clarke says about 80% of studies find a net benefit, through entrepreneurship and greater productive capacity that eventually eases prices, and that H-1B visas run out by April. Research on Indian and Philippine migration finds sending countries gain too, such as more Philippine nursing schools; New York's migrant costs stem from delayed work authorization.
• Monopsony, not the labor theory of value, explains underpaid workers. Clarke says the 1870s marginal revolution showed competitive wages equal workers' marginal product; Joan Robinson's 1933 work on employer market power explains pay below productivity more clearly than Marx. He attributes union decline mainly to shrinking manufacturing employment rather than Reagan alone, and notes Nordic economies innovate despite near-universal unionization.

2023-09-28 · open the original ↗

• Schools should ask when to adopt AI, not whether. Rexach, who works on Microsoft's Philanthropies team in Chicago and taught secondary school through Teach for America, expects AI to change society as the dot-com era did, and argues that withholding it from students does them a disservice because the workplace they enter will not be free of it. She likens the resistance to the early pushback against Texas Instruments calculators; Giard adds internet search engines to the comparison.
• Responsible use means using AI to sharpen students' own thinking. Against the concerns she has read about (academic dishonesty, students skipping the think time of the writing process, data privacy), she offers a Hispanic Heritage Month unit: students draft their own questions for an imagined interview with Cesar Chavez, then compare them with what ChatGPT or Bing Chat suggests, to practise "the art of powerful questions".
• AI fits cross-disciplinary projects ending in a capstone artifact. Her example compares fan engagement for the Cubs and White Sox, combining math analytics with social-science research on neighbourhood demographics and transit, while teaching students to vet the sources behind a claim.
• AI use will become another citation, policed by plagiarism tools. She points to Turnitin-style tools schools already buy and supposes they may now filter for ChatGPT or Bing Chat output; she expects citing AI use to sit alongside MLA or APA references, as is already discussed for images made with Bing Image Creator.
• Teacher buy-in depends on professional development and hands-on time. She wants teachers trained through summer and institute-day sessions and given time to "play in the sandbox" with the tools against their unit plans, and argues AI can also take rote work off teachers; Giard's example is drafting a parent newsletter.

2023-09-26 · open the original ↗

• Any OpenVPN-capable proxy service can run inside UniFi Network. UniFi's VPN client tab takes the provider's .ovpn file plus account username and password, covering devices like a Roku, game console or smart TV that can't run VPN apps. The host demonstrates with Private Internet Access, which he says he has used for about five years, and links it through his affiliate link (quoted at $2.03 a month plus four months free); the video is separately sponsored by ZimaBoard.
• Generate the PIA config near you; streaming servers remain untested. In PIA's OpenVPN configuration generator he picked OpenVPN 2.4 or newer and the Linux platform (the only one he tested) and chose US West for latency. Hulu warned him off from the standard US West IP; PIA's "streaming optimized" US West and US East servers are meant for that case, but he says he has not tested them.
• Traffic routes send chosen devices or domains through the tunnel. A device rule routes all traffic from one or several selected devices (his iPhone) over the VPN client interface, which he confirmed by matching the public IP on whatismyip.com to the tunnel. A domain rule (privateinternetaccess.com, netflix.com) applies to all or selected devices, and other sites keep the T-Mobile Home Internet WAN IP.
• A dedicated VLAN and SSID makes every connected device private. He creates a network at VLAN 53 with its subnet's third octet matching the VLAN ID, a DHCP pool from .10 to .254 leaving .2–.9 for static devices, an SSID on top, and a route sending all its traffic through PIA. Routing and firewall changes can take a minute or two to take effect.
• The OpenVPN link is fickle, and new networks are open by default. "Test and save" never stops spinning; he waits a minute or two, leaves the page and returns to see the connection up, and re-applies the client when it drops. Every UniFi network can reach every other one, so he advises block-by-default firewall rules, covered in his previous VPN video rather than here.

2023-09-25 · open the original ↗

• Platform engineering complements DevOps rather than replacing it. Leonard, a partner solution architect at Microsoft and Giard's colleague, says that once an organization scales past one DevOps team, team A may build on Jenkins while team B picks Azure DevOps, leaving separate stacks that do the same job; platform engineering gives developers one unified platform instead.
• One platform team ends duplicated effort and uneven tooling expertise. That team supplies CI/CD, runtime (Kubernetes or otherwise), monitoring and observability consistently, so compliance and security fixes are applied once; in his role before Microsoft he saw security staff going from DevOps team to DevOps team with the same issues. Standardizing on one tool instead of GitLab, GitHub and Azure DevOps side by side saves cost. Asked by Giard whether one team can scale across a large enterprise, he says no single person or team can do it all, so the group's responsibilities must be defined up front and kept to.
• Self-service templates let anyone deploy a standard Kubernetes cluster. In a previous role his team wrote infrastructure-as-code templates behind a custom self-service portal, where developers could request a cluster, its version and add-on features, tweak it within set limits, and get it deployed without knowing how to deploy Kubernetes. He says they built it themselves a few years ago because nothing on the market did this.
• The approach is not limited to Kubernetes. He says the published discussion ties it to Kubernetes because of containers and microservices, but the same narrowing applies to event handling or databases: trimming a catalog of perhaps 50 database providers to around five that meet business requirements. He points to a CNCF white paper on platform engineering as a starting resource.
• Start nimble with one slice, not a big-bang rollout. He argues that imposing too much change at once makes enemies; begin with, for example, the CI/CD pipeline with security injected on top so teams no longer handle it, show the value, then expand.

2023-09-25 · open the original ↗

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

2023-09-23 · open the original ↗

• Vector databases search by meaning and similarity, not exact content. Ekpo, a Microsoft partner solution architect covering data and AI workloads, says they convert text, images, audio or video into numbers, enabling image search on e-commerce sites and copyright detection on music services. For text, keyword search (the BM25 algorithm used by Lucene, Solr and Elasticsearch) misses "David left to get a meal" when the query says "food", whereas vector search matches the meaning.
• The index format depends on accuracy needs and available memory. Exact matching needs a sequential, non-approximating scan, as in facial recognition for passports or criminal investigations; inverted file indexes split data into buckets for memory-limited devices like phones; HNSW (hierarchical navigable small world) uses much memory but retrieves fastest, which he likens to zooming out of a map from Seattle and back in to a Florida street.
• At very large scale, use inner product on normalized vectors. He says Euclidean distance suits grouping similar items, while cosine similarity and inner product cover most cases; for scans across billions of records, normalizing vectors lets inner product reach the same result in fewer computational steps than cosine similarity.
• Production builds should use managed vector stores, not self-hosted ones. He names four first-party Azure options: Azure Cognitive Search (his first choice, offering hybrid keyword and vector search over HNSW), Cosmos DB's MongoDB API, its PostgreSQL-based API using the pgvector index, and Azure Cache for Redis. Partners including Elastic, MongoDB Atlas, Astra DB, SingleStore, Solr and Pinecone also offer one; he says self-hosting suits a proof of concept but diverts a production team from its business.
• His "RSVP" mnemonic describes retrieval-augmented generation. Reasoning engine, input signal, vector database, prompt template: the question is vectorized, matching chunks are retrieved into the prompt, and the model answers from facts it was never trained on, such as a partner's private insurance plans. His example is a rule allowing exercise only below 70 degrees, which the model can apply only once it is told it is 45 degrees.

2023-09-18 · open the original ↗

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

2023-09-17 · open the original ↗

• Communication, not raw technical skill, sets technologists apart. Woodruff, in the field more than 28 years since 1995, says speaking and writing separate people in tech, calls himself not a particularly smart developer, and says he honed communication in consulting in Atlanta in the late 1990s. He argues everyone is a salesperson, selling themselves in interviews, to clients and to their own teams to build trust, and prefers "human skills" to "soft skills"; he says the heads-down "lone wolf" developer flourished in the 1970s–80s but agile and the emphasis on teams forced that type to change.
• Get involved in the tech community early. He says his first regret is joining late: he started a .NET user group in Lexington, Kentucky around 2005, at 35. He began speaking after seeing Mark Miller at a Delphi user group in Atlanta, despite a childhood speech impediment, speech therapy until 13, and avoiding public speaking until about 30.
• Go deep on something likely to last 20 years. Recalling career talks with a Michigan State computer-science freshman, he says that as a first-generation college student he had no model for a post-college career, and advises students to talk to people about career paths and find an area they genuinely care about. He warns that Silverlight experts saw it gone in five years; Giard counters that they still learned XAML, and that his own FoxPro start taught him SQL and relational databases.
• Pick three new things to learn every January. Past picks include F#, which he learned but has not really used, and speaking practice; this year (2023) they are Rust and the Mac operating system on a new Mac Studio, and he could not recall the third. He also urges attending conferences without speaking and reading business books and biographies, having just finished one of J. Robert Oppenheimer.
• Never be the smartest person in the room. He says he sits in the lower half of most rooms of developers, so associating with developers he looks up to elevates him; in turn, those who received help climbing should share what they learn and help the people coming behind them.

2023-09-12 · open the original ↗

• Teleport is one-click remote access, but only for phones and tablets. It reaches internal services (Pi-hole, Home Assistant, a Synology NAS) and routes public-Wi-Fi browsing through the home connection without a proxy-service fee. Admins see their sites in the WiFiman app via Ubiquiti single sign-on; other users need a per-user link that expires in 24 hours if not accepted, and Windows PCs cannot use it.
• The VPN server suits other devices; the host recommends WireGuard. He calls L2TP older and less secure, and says WireGuard has replaced OpenVPN as lighter and faster yet as secure. Manual advanced settings allow a custom subnet and DNS (Active Directory, Pi-hole); create one client per device and send the config file, since he found the QR code hard to scan even on an iPhone 13. Double-NAT setups may need firewall ports opened.
• Teleport and VPN-server clients see every internal network by default. He calls this backwards: build an RFC 1918 IP group (192.168.0.0/16, 172.16.0.0/12, 10.0.0.0/8), add a LAN In drop rule from it to itself above the predefined rules, then pairs of accept rules above that for each allowed path. VPN subnets (Teleport 192.168.3.0/24, WireGuard 192.168.4.0/24) need IP groups because they are not listed as networks; his pings confirmed the result.
• VPN client makes the UniFi gateway a client of an outside server. Traffic routes or a separate VLAN can send chosen traffic to an employer's OpenVPN server or to a proxy service. The host says Private Internet Access, which he uses personally and links through the description, "100%" works with it, and defers the full setup to a later video.
• Use site-to-site VPN across vendors, Site Magic between UniFi gateways. Site-to-site joins mixed hardware with a pre-shared key (he once linked an EdgeRouter to a Cisco RV345P), and he points readers to Crosstalk's own Rogue Support service for it. Site Magic, UniFi's SD-WAN, needs UniFi OS 3.1+ gateway consoles or a UXG Pro with a Cloud Key on Network 7.4+, excludes the old USG, needs one public WAN IP and one owning account, and caps groups at five sites pending UniFi OS 3.2.

2023-09-11 · open the original ↗

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

2023-09-10 · open the original ↗

• Easy AI APIs tempt partners to skip production architecture. Richter, a Microsoft partner solution architect and Giard's colleague, says engagements come in two steps: deciding what capability, model and prompts to build, then making it secure and well-architected for production, which is his team's part. Calling the OpenAI API directly works, he says, but is not the best long-term choice for cost, security and reliability.
• Security spans data, application and network layers. He says data sent to Azure OpenAI is not used to train the model and the model is not shared with other customers; applications can connect through managed identities so no keys or passwords are handled; and the service can be locked to VNets.
• Not every AI problem should go to a large language model. Guessing a name's gender or other simple classification can go to Azure Cognitive Services or a domain model, repeated prompts can be cached, and API Management can route requests and give cheaper models to free or bronze tiers. A partner asking OpenAI to approve loans should instead train a machine-learning model on its own approved and rejected loans to get a confidence score.
• Reliability and scale require endpoints spread across Azure regions. Because the app depends on someone else's API, he recommends redundant endpoints in active-active or active-passive setups via API Management, Traffic Manager or Front Door, with failover automated. He says scale limits apply at regional level, so four or five services in one region share them.
• The hardest part is choosing a use case that repays the cost. He calls going to production not too difficult if the well-architected framework is followed, but says adding AI is expensive, so the use case and its ROI come first. He puts API Management in front of the model partly because Azure OpenAI's logs omit the prompt text, and capturing prompts shows how users actually use the feature.

2023-09-08 · open the original ↗

• Every iPhone user should install iOS 16.6.1 immediately. The host reports that Citizen Lab found a zero-click vulnerability being used to deliver NSO Group's Pegasus spyware: a malicious image sent by iMessage infects the phone without being opened, exposing location, texts, camera, microphone and keylogged passwords. iOS 16.6.1 patches it.

2023-09-04 · open the original ↗

• Fabric unifies the whole analytics chain as one SaaS product. Llovet de Casso, a Microsoft partner solution architect in data analytics and Giard's teammate, describes the offering, in public preview since its announcement in May, as covering ingestion, storage, transformation, warehousing, data science and reporting without wiring together Data Factory, Synapse Analytics, Azure Machine Learning and Cognitive Services. Giard notes it is unrelated to Azure Service Fabric.
• It goes beyond bundling existing services, he says. Giard summarizes Fabric as connecting existing tools; the guest calls that "mostly right" but says the services are also optimized to work together and share one interface, so a user can move from data factory to data science without learning new menus.
• Fabric is built on the Power BI web service. Users sign in with their existing Power BI account and can start a free trial during the preview; he says it is meant to feel familiar both to Power BI users and to those who know Data Factory and Azure ML concepts.
• Low-code comes through visual pipelines now, Copilot later. Transformations normally written in SQL, such as aggregates and group-bys, can be done by selecting icons; natural-language Copilot help that suggests SQL queries has been announced but was not yet available at recording.
• OneLake keeps a single Parquet copy usable by any engine. SQL, Spark and Power BI all read and write the same Parquet files, so data is neither duplicated nor locked to one compute engine; shortcuts point at data in S3 buckets or existing Azure Data Lake storage without copying it. He says the new Direct Lake mode, reading straight from the lake rather than importing or querying a source, made filters and visuals respond quickly in his own test on a dataset he puts at about 50 million rows.

2023-09-04 · open the original ↗

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