Souk's Notes - Unofficial Summaries

2021-06-03 · open the original ↗

• Crosstalk's own Cloud PBX, powered by ClearlyIP, is out of beta. It targets 2–20 seat businesses that want normal phone-system features without per-extension cloud pricing like RingCentral's or 8x8's. Crosstalk bills by minutes used and calls "unlimited minutes" plans a scam. Features include ring groups, queues, time conditions, call recording, a web call panel, and a softphone per user (browser or the Clearly Anywhere mobile app, set up by QR code).
• Service is priced per extension, plus numbers and minutes used. It is $5 per extension a month. Setup is $250 for the first two extensions and $50 for each additional one, plus $50 shipping. Taxes typically add 15–20% to the monthly bill. There is no contract, and Crosstalk phones carry a 2-year warranty.
• Crosstalk claims to undercut RingCentral by more than half over five years. It puts the saving at 55%. In the company's example of nine phones, three numbers and 4,000 minutes, upfront cost is $1,863 ($1,163 of it phones) and service $83 a month before tax. Year one totals $2,859 against about $3,022 for RingCentral at $27.99 per extension per month. After that it is $996 a year, so five years cost $6,843 against $15,114. The company claims an average saving of 50% or more.
• Onboarding is pitched as the main selling point. Crosstalk says it designs inbound and outbound call flows, configures and ships the phones ready to take calls, then guides the customer through number porting.

2021-06-02 · open the original ↗

• Abilities are mostly built by practice, not fixed at birth. Rosenbaum, now a senior manager at Red Hat, explains Carol Dweck's growth mindset against the fixed mindset that treats capability as inborn, and says Microsoft adopted it under Satya Nadella because he saw learning as the key skill for people and organizations. She read Dweck's book Mindset at about 30, wishes she had at 13, and gives an unsponsored talk on it because technologists have to learn every day and that is a struggle under a fixed mindset.
• Labels and "you're so smart" praise create fixed mindsets. She cites a study in which teachers told that randomly chosen students were superstars got superstar results, and Dweck's study (on four-year-olds, she thinks) where children praised as smart rather than hardworking gave up on hard puzzles, lied about their scores and then did worse on easy ones. She recognises herself: praised as smart, she could solve any "problem" but froze at a "riddle", skated through high school, and was miserable studying computer science in college while a friend simply reread and asked for help until he understood.
• The payoff is a better life experience, not necessarily more achievement. Her route is to read the book, then talk yourself through each hard problem that seems to prove you aren't smart enough; she says it takes a couple of years and rests on research, not just her say-so. She likens it to cognitive behavioural therapy's graded exposure, after Giard describes building public speaking as a muscle.
• Talent exists but is overrated, and practice works best with feedback. Answering Giard's Serenity Prayer question, she grants that a short person probably won't make it in basketball, but says most people quit after a week or a month, many prodigies fail when they reach a hard problem, and Michael Jordan is remembered as gifted rather than as the most practised player. Practice helps most in "kind" domains with feedback, like public speaking, and least in "wicked" ones, like a doctor who never learns whether a prescription worked.
• Growth mindset may free women and minorities from internalized stereotypes. She cites findings that reminding women of their gender before a math test lowers their scores while reminding Asian students raises theirs, and argues that not being defined by what you were born with also means not being defined by gender. She flags that research here is thin and calls it her opinion, adding that it does not replace fixing stereotypes or investing in diversity.

2021-06-01 · open the original ↗

• The host's Starlink injector won't hold a gigabit link to third-party gear. The service itself works, but the injector's data port auto-negotiates to 100 Mbps full duplex on a UniFi USG, a USW Lite 16 and an EdgeRouter X. On a Netgate SG-1100 it would not negotiate at all until he forced 100 Mbps full duplex. In his earlier Starlink-with-UniFi video the same port showed gigabit.
• Starlink support saw a fault alert remotely and replied within an hour. Their first reply flagged an alert suggesting cable or port damage and listed eight troubleshooting steps: photos of pins, bends, pinched or nicked cable, debris, and bypassing the router. The host infers from how fast this canned list arrived that other customers have hit the same problem.
• A slightly depressed pin in the injector looked like the culprit. Pin 8 on the injector's port sat a millimetre or so low. Prying it up with a paper clip, or just replugging, brought back gigabit and speed tests of 217, 206 and 175 Mbps in his office. It fell back to 100 Mbps after about an hour each time.
• Starlink is replacing the power brick; the fault persisted at recording. The RMA was under way and he was still at 100 Mbps full duplex and planned a follow-up video. He says he isn't upset because Starlink is a beta product, and praises its support as quick, right first time and friendly.

2021-05-28 · open the original ↗

• The EdgeRouter X is a cheap router with headroom for Starlink. The host says the ER-X processes far more packets than Starlink can push through it; MSRP is $59, and he has seen it on sale for about $35. On his test it showed Starlink's CGNAT address on eth0 and ran a speed test at about 80–85 Mbps.
• Setup is a standard ER-X build with Starlink on eth0. Factory reset with a paperclip held 10 seconds, connect a laptop to eth0 with a static address in 192.168.1.x, log in at 192.168.1.1 with ubnt/ubnt, then run the "WAN + 2LAN2" wizard with "use only one LAN" so eth1–eth4 act as one switched LAN. Then move Starlink's cable (the one normally in Starlink's router) to eth0 and the laptop to eth1.
• Move the LAN off 192.168.1.x, and never use 192.168.100.x. He changes the default because nearly every router ships with 192.168.1.x, which risks conflicts; he uses 192.168.42.1/24 with the DHCP server on. 192.168.100.x is off limits because Starlink uses it, and the static route to the dish will not work if the LAN overlaps it.
• A static route is needed to reach the Starlink dish interface. Without it, 192.168.100.1 (stats, stowing the dish, support tickets) is unreachable from behind the ER-X. Add an interface route to 192.168.100.1/32 with next-hop interface eth0 and distance 1.
• The default DHCP pool and DNS settings are worth changing. The wizard hands out an odd pool of .38–.243, which he resets to .100–.254, and the router is clients' only DNS server, so he adds 1.1.1.1 as a backup. He also sets system name servers 1.1.1.1 and 9.9.9.9, and notes the DHCP page can carry a UniFi controller's IP to help devices find a cloud-hosted controller.

2021-05-26 · open the original ↗

• Revive the old NVR first, to save resetting every camera. At a Newport, Oregon client, a UniFi Video NVR on a Shuttle PC he had built three to four years earlier was off and stuck at a BIOS prompt that needed F1. Booting it let him verify the camera password and take a backup, so he would not have to factory reset cameras on ladders.
• The UniFi Video backup restored into UniFi Protect cleanly. On a pre-built UNVR the restore imported only the cameras (no users), alongside two existing Protect cameras, and 19 of about 20 updated and came online. The stuck one finished after a PoE cycle from UNMS; a "failed" camera came back on a different port of the 24-port EdgeSwitch, leaving the client its replacement G4 Dome as a spare. He spent about 2.5 hours on site, the last hour training three employees.
• Each viewport and each camera belongs to one Protect instance. A viewport shows views from a single Protect server, so it cannot combine cameras from several systems. A camera records to only one Protect device: as of recording, the UNVR, UNVR Pro, UDM Pro or Cloud Key Gen2 Plus. As far as he knows there is no limit on the number of viewports.
• Smart detections work only on G4 cameras and the doorbell. Answering a viewer with G3 cameras considering an upgrade for that feature, the host says only generation-4 models (G4 Pro, G4 Bullet, G4 Dome) and the doorbell can do smart detection.
• A separate UNVR allows a camera VLAN that UDM Pro's Protect does not. To his knowledge Protect on a UDM Pro cannot be put in its own VLAN; he speculates firewall rules plus manual adoption might work but would lose auto-discovery. A UNVR and a UDM Pro run fully separate Protect instances: no shared drives, no user replication, no combined dashboard. To use the UNVR's capacity or RAID, uninstall Protect from the UDM Pro.

2021-05-25 · open the original ↗

• Microsoft went from refusing contributions to embracing GitHub. Wilcox recalls the Silverlight Toolkit on CodePlex publishing its code but accepting no pull requests, and the Azure SDK team needing executive approval to be among the first teams to take contributions on GitHub; he puts Microsoft employees on GitHub today at over 30,000. He credits the cloud's interoperability (Linux or Windows containers, Redis run locally, on AWS or Azure) and engineers' wish to use what everyone else uses, such as React and Kubernetes.
• Every Microsoft build automatically inventories its open-source components. Wilcox, who works in Microsoft's Open Source Programs Office (OSPO), says a Node project with 800 npm dependencies gets recorded at build time, so the company can later trace which builds used a given version when a vulnerability surfaces. The same record checks license compliance and raises alerts where a license carries obligations such as publishing source code.
• Microsoft releases under MIT and accepts obligations from others' licenses. Wilcox says the company is fine using almost any license provided it meets the requirements, citing the large body of source it must publish for the Edge browser. He notes it is often the same lawyers as five years earlier who have since changed their stance and reduced the burden.
• A network of "open source champs" scales the small OSPO team. Wilcox says champions sit in organizations across the company so that, for example, someone on Windows has a nearby person who has done it before; his team otherwise acts as "highly paid network routers". Much of the job is reassuring teams that they may start a project in the open before it is polished.
• Engineers are being encouraged to contribute back, including outside Microsoft. Wilcox says recent hires treat fixing a bug or docs via pull request as second nature, while older staff were unsure whether they needed approval. He says contributing to, say, a Linux Foundation or Google project is easy, and that Microsoft also sometimes sponsors outside projects financially.

2021-05-22 · open the original ↗

• "Mindful AI" is preferable to "ethical AI". Sampson, an adjunct professor at DePaul University, says ethics is loaded and contextual, so stamping one group's ethics on a universal technology is a problem; mindful means being deliberate in designing machine-learning products. She uses design as a verb, after Victor Papanek's "purposeful application of meaning and order to chaos", which implies action and responsibility.
• AI should be designed as human-machine symbiosis, not human replacement. Sampson draws on J.C.R. Licklider's "Man-Computer Symbiosis" paper and treats AI as a design tool rather than a technology foisted on people. She argues the 1956 Dartmouth founders made AI more rational than humans will ever be, so a self-driving car would only beat human drivers if there were no humans around, and that automation bias means people trust a machine that is right nine times in ten.
• Released models keep learning from people their builders never chose. Sampson tells students their model is not trained only by them once it is out, which she says is what happened to the Microsoft chatbot the host raised, racist within 24 hours. She cites an IDEO principle of listening for unintended consequences and says good data scientists iterate on models rather than setting and forgetting them.
• Bias enters through what data is collected, used and ignored. Sampson's examples are a zip code classified as a low-income neighbourhood, car facial recognition she says was trained on Stanford students and misses darker skin, and Kodak's "Shirley cards", a white model's image used for colour balancing that she says still shapes sensors such as automatic faucets. She argues mindful AI starts in model making, with diverse people asking about training data and preferring an interpretable regression to an unauditable neural network.
• Facial recognition and predictive policing are risky AI use cases. Sampson says unless everyone was in the training data the software will make mistakes, citing a Detroit man suing the Detroit police department. She describes Chicago's "strategic subjects list", which she says flagged people such as Robert McDaniel through their social networks without any crime, was released only after people sued, and was disproportionately Black and Latino.

2021-05-20 · open the original ↗

• Almost any hard drive works; RAID choice trades capacity for redundancy. Ubiquiti has tested drives up to 16 TB, and the incompatible-drive list mostly covers drives whose screw holes miss the trays. By default one drive is no RAID, two are RAID 1, three or four are RAID 5; a newer "half of disks" option gives RAID 10 and a hot spare can be set. With four 8 TB drives, RAID 5 gives 24 TB and RAID 10 gives 16 TB.
• Setup forces a cloud account, so enable 2FA and isolate cameras. The wizard offers no local-only user, so the Ubiquiti single sign-on account should have two-factor authentication. He recommends putting cameras and the NVR on their own VLAN with firewall rules, so someone who knocks a camera off a wall and plugs into its cable cannot reach the main LAN.
• UniFi OS auto-updates do not update the Protect app. A separate application auto-update switch is off by default, so the device can show up to date while Protect has an update pending. He advises the official release channel in production, enabling SSH to fix a UNVR stuck "updating", and powering off from the interface rather than unplugging.
• His camera defaults: always record, 2-second trigger, 5-second padding. He sets overlays to time, date and camera name without the UniFi logo, and saves 5 seconds before and after motion. Motion zones skip a busy street, privacy zones black out areas illegal to film, RTSP feeds Home Assistant or other NVRs, and time-based purge meets legal retention limits.
• Roles live in Protect, but users live in UniFi OS. Camera roles (e.g. view only for one building) are made in Protect, then assigned to users made in UniFi OS. A Ubiquiti account gives remote access through ui.com; "local access only" users log in on the LAN. He suggests logging in as each new user and setting up default views for them.

2021-05-19 · open the original ↗

• G4 Dome night fog is IR light reflecting back into the dome. The host found a Ubiquiti article on G4 Dome night-mode issues: do not mount within 24 inches of objects or walls, keep overhangs out of view, and avoid wall mounting too close to the ceiling. His own camera sits within two feet of an eave without trouble.
• Cleaning and properly seating the bubble cover fixed his camera. He cleaned the inside of the dome with a lens cloth, lined up the numbered peg (his read "2") with the notch below the lens, pressed it fully in, twisted to lock, then cleaned the outside. Before, a person 35 feet away was invisible; after, visible though not recognisable. He suspects the poorly seated cover was the main cause.
• Moisture and a damaged rubber seal also cause fog. The same article lists water droplets on the outside of the dome, which may need periodic wiping where rain hits it, and a chipped rubber seal around the lens, which leaks infrared light. For client installs he now cleans and seats the dome as soon as he adjusts the lens.

2021-05-12 · open the original ↗

• Video Indexer analyzes video with AI and needs no code. It is a Microsoft tool at videoindexer.ai, signed into with a Microsoft account; Giard says it integrates with Azure Media Services and Cognitive Services, and that he first used it about four years before this recording.
• An upload takes a source, title, language, visibility and agreement. The video can come from a web URL (a YouTube video, for example) or the local machine; the language helps its speech-to-text, and Giard keeps his video private. Upload stays disabled until a checkbox is ticked; Giard says Microsoft currently does not allow police departments to use the technology, and the uploader must have rights to the video.
• Indexing took longer than the video itself ran. Giard's low-quality 2010 charity-event video, 744 MB and 16 minutes long, took about a minute to upload and about 19 minutes to index; he offers that as a gauge for other videos.
• The results break the video into people, topics, emotions and scenes. It tracked each face across every appearance, even in the background, without naming non-famous people, and matched two of his friends to celebrities they are not (one at 70.5% certainty). It also listed topics, keywords, labels and named entities from the transcript; emotions from facial expressions; sentiment from speech on a 0-to-1 scale; and 12 scenes, each split into shots with key frames usable as thumbnails.
• Captions and insights download for use in other applications. Closed captions come in standard formats, VTT among them, optionally with audio effects, and each line carries a confidence score (about 78–84% in his file); the insights download as JSON with start and end times for every emotion, scene and the like. Giard lists uses: better video search, highlight reels, closed captioning, content moderation, pointing to specific moments, and feeding the data to other machine-learning models; he says much of it works on audio too.

2021-05-11 · open the original ↗

• Synapse is an umbrella over three processing engines. Cochran describes dedicated SQL pools (the former Azure SQL Data Warehouse), serverless SQL and Spark, all sharing one data lake. The dedicated pool is a massively parallel system of 60 SQL Servers behind a controller node; version one (around 2014–2015) supported only clustered columnstore indexes, and version two's clustered indexes enabled the row-level joins his work needs.
• Three-year reserved pricing cuts dedicated-pool cost to about a third. Cochran's firm runs production at 500 DWUs, roughly $5,000 a month, against a top end of 30,000; he would not recommend 100 and says around 300 performs decently. He says the commitment breaks even after the first year, so even 18 months to two years of use saves money.
• Serverless SQL queries data-lake files directly, paying only per query. Cochran can drop a customer's CSV into Azure Data Lake Storage, right-click and query it with no loading, ETL or schema, and says two new Synapse environments took about 45 minutes to set up. His firm writes large data sets out as Parquet and puts views over the OPENROWSET syntax so analysts see ordinary tables.
• Spark notebooks mix Python, Scala, SQL and C# cell by cell. R was not supported at recording, which Cochran expects to change. He says PySpark is the default but everything executes as Scala at runtime, that pools autoscale between set minimum and maximum nodes on consumption pricing, and that each pool can carry its own PyPI libraries; serverless SQL can query Spark tables.
• Cochran predicts Synapse will replace Databricks in Azure. He cites built-in Azure Data Factory pipelines, native Power BI connectors and Azure Purview for governance. He says Data Factory lets him, often the only data engineer at Causeway Solutions, a New Orleans analytics firm holding data on every US adult, automate an outsized amount of work.

2021-05-10 · open the original ↗

• A comedy teaser promoting BUILD 2021, with no argument. The 64-word caption track is music and unattributed skit fragments: another Teams meeting, "then don't go to Seattle", "the world's okay programmer", a drive-through-style order ("4c with cheese", "first window"), and a "second dose" given to someone with the last name Hanselman ("relax your arm"). Beyond the title's call to register, the text carries no claims, dates or event details.

2021-05-05 · open the original ↗

• Cheap cloud analytics turned manufacturers' old data collection into IoT. Diaz, on Microsoft's IoT global black belt team for about five years, says manufacturers long fed plant data into historians but found it too expensive to use; now almost every manufacturer he talks to has an IoT initiative. He cites a Microsoft investment of $5 billion over five years and Azure Digital Twins, newly GA, for modelling a factory, building or vehicle digitally.
• Latency and bandwidth push AI decisions to the edge. Diaz says about 70% of IoT deployments will include AI at the edge within a couple of years, a forecast. A drill that must stop on a reading cannot wait for a cloud round trip, and video for people counting or licence plates is too costly to send up.
• IoT Edge runs Docker containers deployed to fleets from the cloud. Diaz says devices can be provisioned zero-touch with the Device Provisioning Service, then a manifest pushes modules to one device, a subset or the whole fleet. A teammate is automating the retrain-and-redeploy loop so a new model reaches the edge about monthly without hands-on work.
• Azure Percept bundles reference hardware with low-code model deployment. Diaz describes a dev kit built by Asus to Microsoft's reference design, with a camera and optional microphone array, running IoT Edge, plus Percept Studio for prebuilt models or custom TensorFlow or ONNX ones. In his demo a model swap was just a zip delivered via the module twin, and he notes camera angle and zoom matter; Percept was not offered under "create a resource" in the portal and had to be reached from the docs.
• Percept's voice side reuses Microsoft's existing speech services. Diaz shows command sets for inventory and hotel use woken by the keyword "assistant", aimed at uses like a hospital patient calling a nurse, though his demo did not display the response. He values that Percept builds on IoT Edge and Cognitive Services rather than reinventing them, and says his pre-sales team engages at no charge through a company's Microsoft account team.

2021-04-29 · open the original ↗

• Crosstalk sells its PBX systems in three deployment forms. The video is the company's own pitch: on-premises PBX on Crosstalk-supplied servers, hosted PBX for multi-site or home-user organizations with ongoing updates and backups, and a forthcoming Crosstalk Cloud PBX, still in beta, which the host pitches against RingCentral and 8x8 on price. All are built on the open-source Asterisk engine.
• Crosstalk's servers scale from 40 extensions to 500 phones. All ship with the FreePBX distro and a two-year warranty, extendable to five. The company says the 715 and 716 handle up to 40 extensions and about 10 concurrent calls, with built-in FXO/FXS ports, and the fanless 710 handles 100 phones and 30 concurrent calls. The 1U 745 handles 150 extensions and 45 calls with up to eight analog ports, and the 1U 760 handles 500 phones and 100 calls.
• Any two servers can run as primary and warm spare. According to the company, the second server stays synchronized, so it can take over if the primary fails.
• Crosstalk's branded phones come in three tiers. All have a two-year warranty, colour displays and gigabit pass-through, are VPN-capable, and run on either a power adapter or 802.3af PoE. The 230 is for break rooms and warehouses, the 250 has eight programmable keys for standard users, and the 270 has 10 keys and takes up to six Crosstalk 200 sidecars.
• The service covers call-flow design, E911 compliance, porting and cutover. Crosstalk says it handles Kari's Law and RAY BAUM'S Act compliance and project-manages the number porting. Customers get 30 days of post-deployment support, after which ongoing support is available without a maintenance contract.

2021-04-25 · open the original ↗

• Founders should invest in a product manager, not lean on developers. Walker, director of product delivery at the agency Kin + Carta, says cost-conscious founders treat their tech leads as a substitute for good product work, requirements and validated assumptions; she wants someone who knows tech but is not coding, as the voice of the product.
• Overloading a product with features hurts it. She describes founders adding features on the word of one person met at a conference; too many features confuse users and pull money from the core problem, so she advises shipping the one thing that solves it, watching the reaction and building from there.
• Founders skip user research, believing they already know their users. Walker says founders are often on the right track but blindsided by what users actually say, and that users tell founders what they want to hear; she recommends combining surveys, A/B testing and beta-tester interviews, run by a neutral party, and validating again after release.
• Agency rescues often start with core functionality that doesn't work. When her firm takes over a product built without upfront product or research investment, she says it is typically feature-heavy and architecturally weak, and users report that "half of the functionality doesn't work", so the first job is fixing the core and re-releasing.
• Walker thinks the best time to start is now. Asked in early 2021, coming out of the pandemic, she says it is probably easier than two years earlier because people are open to new ideas, though she notes she is not an economist; even a simple web form can start validating an idea. On marketing she disclaims expertise and suggests leaning on the community the product serves.

2021-04-22 · open the original ↗

• The Wi-Fi 5 AP12 is the standout value in Aruba's range. In the host's benchmarks the AP12 ($141.36 on Amazon, 3x3 Wi-Fi 5) averaged 338 Mbps, third overall behind the U6-LR and EnGenius EWS850, and second on price-to-performance after the U6 Lite. He adds that Wi-Fi 6 is not needed outside dense client environments, so the AP12 suits a standard office.
• The 4x4 AP15 trailed both cheaper Aruba access points. It costs $192.50, yet it averaged 278 Mbps and ranked second from bottom in sustained throughput among all APs he has tested. His test clients are 2x2, so he would not expect 4x4 to help without many simultaneous clients, but the 3x3 AP12 still beat it in almost every test.
• The Wi-Fi 6 AP22 seemed held back by its channel-width setting. The AP22 ($159.99, 2x2) averaged 335 Mbps, roughly level with the AP12 and U6 Lite. Version 2.3's new radio controls offer only 20/40 or 20/40/80 MHz auto, and the host never saw it use 80 MHz, so he believes it could do better at a fixed 80. He tested with WPA2 only because WPA3 caused heavy retries on his Ubuntu 20.04 clients, on EnGenius as well.
• Two-factor login in version 2.3 makes Instant On recommendable. The host would not have recommended it without 2FA, though he reasons that a hijacked portal matters less here because Aruba makes no firewall that could open ports. Management is cloud-only with no local controller, and there are no licence or subscription fees. Version 2.3 also adds MAC-based authentication, DHCP reservations, port mirroring, PoE schedules and manually triggered firmware upgrades.
• Its pared-down interface suits MSPs who want cookie-cutter deployments. Instant On makes only access points and switches, and the host paired it with a Netgate SG-1100 pfSense firewall, with VLANs working throughout. The 1930 switch ($349) has 24 gigabit PoE ports with a 195 W budget and four 10G SFP+ ports. He found it loud and suggests swapping in quieter fans if noise matters.

2021-04-21 · open the original ↗

• Deleting a Facebook post only moves it to Trash. Choosing "Move to Trash" from the post's three-dot menu removes it from the timeline, but Facebook's own message says items in trash are deleted after 30 days. Giard says he only recently learned this.
• The Trash is buried deep in account settings. Giard calls its location unintuitive: the triangle at the top right (labelled Account), then Settings and Privacy, Settings, Your Facebook Information, Activity Log, and Trash.
• Trash lets you delete a post permanently without waiting. Giard suggests this for sensitive material you don't want to exist anywhere. Each item's ellipsis menu has Delete, and checking boxes applies the same action to several posts at once; once deleted this way, he says, a post cannot be recovered.
• Archive and Restore to Profile are the other options. Giard says archiving keeps a post past the 30 days, which suits something you want kept for yourself but not seen by others. Restore to Profile undoes a mistaken deletion, and in the demo the restored post reappeared on his profile after a brief "Restoring" message.

2021-04-21 · open the original ↗

• Titletown Tech uses Microsoft philanthropy to build Green Bay startups. Adamczyk, Microsoft's technologist-in-residence there, says it was co-founded about a year and a half earlier with the Green Bay Packers under Microsoft's TechSpark initiative, which works in six mostly rural or mid-sized US communities. It pairs a studio taking early-stage founders from proof of concept to MVP with a venture fund for mostly Wisconsin companies.
• The fund is for-profit, though Microsoft recycles its returns. He says the $25 million fund came $5 million from Microsoft, $5 million from the Packers and $15 million from mostly local corporates, as capital for equity; Microsoft has pledged to roll its returns back into philanthropic work. The other investors are deliberately corporates so startups can tap their market networks.
• Its "connected capital" model pairs money with mentorship and connections. Adamczyk describes a "one-size-fits-one" approach with about two dozen companies, no cohorts and weekly investment decisions. Jason Zander, EVP of Azure, serves as chief technology advisor, and Brad Smith is executive sponsor.
• Startups that arrive on rival clouds have nearly all moved to Azure. He says Titletown Tech imposes no platform restrictions, but almost every company that came in on a competing cloud chose to build on Microsoft products because of the mentoring and resources; he describes this as "somewhat unintentional".
• Equity League invests in minority-led and social-impact startups. Launched in December, it joins the Packers, Bucks, Brewers and Microsoft and is housed under Titletown Tech as a sixth investment arm; Adamczyk ties it to the summer's racial-justice protests, including the shooting of Jacob Blake in Kenosha. At recording it was two months old, and he frames it as a long-term programmatic structure rather than a one-off check.

2021-04-18 · open the original ↗

• Problem solving is asking yes/no questions at scale. Hanselman says a senior engineer who spots a bug instantly is running a bank of eliminating questions built over years (IP address, SSL certificate, permissions, "it's probably DNS"), and he illustrates systems thinking with schoolchildren debugging a broken toaster ("do the neighbours have power?"). Seniors' job is to make it safe to ask "dumb" questions, since that is how juniors build their own list; searching the web is also asking seniors, only strangers, so weigh who wrote the answer.
• Layering hides complexity, and sometimes it lies. He compares email's multipart name-value format to an HTTP form post, and a hard drive to a record player, to argue that good ideas get reused under the abstraction. Using a taxi rider, a car owner and a mechanic, he argues each coder chooses which layer to stop at. His example of the lie is a recording of his podcast episode with actor Chris Conner: the SD card looked empty, and after about 16 hours of reading FAT32 bytes from a disk image he found the files sitting under a folder with a null name that Explorer could not display.
• Experience is mostly pattern recognition: "I've seen that before." A Linux binary that broke on a Raspberry Pi was missing every 0x0D byte, because his FTP client treated files without an extension as ASCII and stripped the carriage returns, a typewriter-era convention that Windows CRLF line endings still carry. He says the second time a problem like that takes far less time. He adds that "sucking at something is the first step to being good at something" and cites starting taekwondo at 45, with a black belt due later in 2021.
• College teaches you how to learn, not what you'll use. In answer to a viewer, he says he has not used calculus in 30 years, though a data scientist would need it, and that C#, the language he uses, did not exist when he studied. He likens a degree to a passport: not required, but a useful piece of paper. He got a two-year degree, then spent 11 years of night classes finishing a four-year one.
• Everyone is an amateur, so keep learning. He asks whether "20 years experience" is really "the same year 20 times", and concedes that perhaps five to seven of his own 30 years were. He points to Tim Berners-Lee's humble "web developer" title and calls himself a Swiss Army knife that does many small things, none of them well, but says not to skimp on basics such as CPU, memory, disks and computing history. On choosing a path, he says careers can switch, and advises betting on broad areas like web applications rather than one technology like Kubernetes.

2021-04-15 · open the original ↗

• Security design should start from empathy for the people using it. Goerlich, who is writing a blog series drawing security principles from classic designers, cites the mouse: Xerox's cost about $2,000 in today's dollars with weekly engineering upkeep, while Apple's iterated design came to about $70 and, per Steve Jobs, 16 minutes to learn. He applied this to a privileged access management setup that took about 18 steps before anything could run: his team watched the engineers', DevOps and IT staff's workflow, cut the steps, and issued a gaming laptop as the privileged access workstation to make it attractive.
• Ship ugly early iterations instead of waterfall-perfect security specs. Goerlich says security programs over-engineer because practitioners are passionate about getting it right, illustrated by a "delete data" control that escalates to purging memory and degaussing disks. "Minimum viable security" means naming the exact controls needed; his example is his friend Mark Tilden's walking robot, buildable for about $100 from recycled parts, against MIT robots costing the price of a car.
• Name projects so they tell a story people want to join. He says a program called after a vendor product "doesn't go very well"; his own "Outsource Boring" project, which challenged his team to hand off work they disliked to partners, sold better than a "vendor assessment and review process" would have.
• Be prepared to get lucky when circumstances change. Goerlich's example is an acquaintance running a secure-development program whose team moved to GitHub, which he had not heard of; by asking the team lead and then a GitHub expert, he found about a quarter of the controls he expected to buy, build or enforce were already in the product.
• Make security an inside job through paired champions and advocates. In his terms a champion sits in the business and cares about security, and an advocate sits in the security organization and helps that business unit; he wants both in place and working together to cut steps and decision points, as with an operations-side advocate he worked with in financial services.

2021-04-14 · open the original ↗

• The beta Starlink dish is geo-locked to its home cell. About three hours away, on the Oregon coast near Charleston (next to a Starlink ground station), it reported being outside its geographic location and would not connect after reboots. Once that happens the app locks you out and cannot even stow the dish; the host had to cut power for about 30 seconds first.
• About 20–30 miles from home it worked, but more slowly. There it delivered 20–35 Mbps down, 6–10 Mbps up and 25–30 ms ping, and one phone speed test showed 40.8/3.73 Mbps, against 80–150 Mbps at home. During a livestream it fell below 10 Mbps down and 2–4 Mbps up. The host suspects the high-voltage lines directly over the dish but says he does not know.
• A battery power station can run the dish and network. The dish drew about 90–150 W, and an EcoFlow EF Delta (received from EcoFlow about a year earlier) charged from 30% to 100% on AC in about an hour. He bypassed the Starlink router, feeding a UniFi USG, a UniFi 60 W 8-port switch, a Cloud Key Gen2, a mesh AP and a VoIP phone.
• Remote VoIP calls and livestreaming held up, with brief dropouts. The test VoIP call was clear, and the livestream of roughly 30–60 minutes had one satellite dropout of 15–20 seconds and one of about 5 seconds. A third outage, about a minute, was his laptop losing power. The host says this was not possible a year earlier.
• The host thinks Starlink makes starting a wireless ISP unwise. At about $500 plus $100 a month, he predicts 200–300 Mbps with sub-30 ms pings in perhaps a year. He thinks established WISPs charging $40–60 a month will be fine for a long while. He expects mobile Starlink for cars, RVs and trucks but sees ocean motion as a tracking problem for boats. Starlink had just become available in Australia.

2021-04-07 · open the original ↗

• Ordinary developers can now solve AI problems with a REST call. Renze says pretrained cloud services from Microsoft, Amazon and Google do the model training for you, so tasks that look like they need a data science team take a few lines of code. His Pluralsight course, the AI Developer's Toolkit, runs from tables through text, audio, images, video and apps to cyber-physical systems; he guesses about 80% of its examples will be new to most developers.
• Analysis reduces complex data while synthesis expands simple input. He defines analysis as turning an image into a label like "cat", and synthesis as the inverse, such as a generative adversarial network drawing a cat from a sentence. Cloud APIs are further along on analysis, which he says is more solved and has more business cases; synthesis needs more compute and has more edge cases.
• Renze already uses a model of his own voice in courses. He says he edits his script and has the model speak a corrected word, and he could not tell the difference; he plans to move to Microsoft's custom voice synthesis in Azure. He predicts near-photorealistic text-to-image within about a year.
• Harder problems combine models in parallel or stacked. Examples include pairing face and voice recognition to identify a speaker, or running object detection and then classifying each bounding box. Between cloud and custom builds he describes transfer learning, such as fine-tuning a stock image classifier to spot a company logo.
• Cyber-physical systems learn in the cloud and act at the edge. He says machines run their decision loop locally to avoid latency while a fleet sends decisions and outcomes up for retraining and receives new models. He predicts human referees will become largely unnecessary within a few years, and that within 10 years no human will drive better than the best AI.

2021-04-06 · open the original ↗

• Omada costs about two-thirds as much as comparable UniFi gear. The host priced an ER605 router, TL-SG2210MP switch, EAP620 and EAP225-Wall on Amazon against a USG, US-8-150W, U6 Lite and UniFi in-wall AP. The UniFi set came to $795, and the TP-Link set saved $275. TP-Link supplied all the equipment free, and the video's product links are affiliate links.
• The host calls Omada and UniFi even, except on price, looks and security. He rates feature set, performance and support a wash, with TP-Link ahead on price. UniFi wins on aesthetics, since the EAP620 is far bulkier than the U6 Lite, and on security, because UniFi's cloud offers TOTP two-factor login and Omada offers no 2FA. He calls cloud 2FA mandatory if TP-Link wants to compete.
• The EAP620 benchmarked level with a UAP-AC-Pro or nanoHD. It placed fifth overall in the host's benchmark, fifth in iperf, sixth in OpenSpeedTest and fourth in sustained throughput, and fourth on price-to-performance. He calls performance a toss-up all the same, because Crosstalk does not run TP-Link in production and the gap may not be noticeable to users.
• Omada's controller software is mature and almost exactly a UniFi clone. Standouts include scheduled PoE (turning off VoIP phone ports outside business hours, which he had not seen in any controller), AP reboot schedules and bandwidth limits per client or per network. It also offers a DHCP fallback IP, site and controller migration and USB auto-backup. SIP ALG is on by default (he disabled it), and the captive portal needs the controller reachable by clients.
• The controller can be hardware, self-installed or TP-Link's cloud. The PoE-powered OC200 ($83.92) manages up to 100 Omada devices and the OC300 ($159.99) up to 500. Cloud access can be turned off and the TP-Link ID unlinked, which the host suggests for anyone wary of a cloud compromise reaching the network.

2021-04-03 · open the original ↗

• Professional development should be paid work, not a hobby. Neward contrasts doctors, for whom reading journals and training the next generation is part of the job, with tech's expectation of side projects on GitHub. When Quicken Loans CIO Brian Woodring asked what would happen if he paid Neward to do the mentoring and interviewing work he was already drifting into, a Technology Culture team was formed in March, focused on the "people" of people, process and technology.
• The canonical whiteboard interview tests the wrong thing under pressure. Neward argues only a single-digit percentage of developers need to write algorithms like a red-black tree, and the whiteboard strips away tools such as Stack Overflow. He cites a study Microsoft took part in, with what he thinks was the University of North Carolina, where leaving the candidate alone in the room raised success from 50% to 100%; his team is reworking interviews toward on-the-job work and away from unconscious bias, as part of the company's diversity and inclusion effort.
• Mentoring programs fizzle without someone dedicated to running them. Neward has hired a technology team leader to build one across the 2,000 people in Rocket Technology; its open questions include how mentees are found, whether the mentor has the skills sought, how much time is allowed (he contrasts a place happy with five one-hour mentees a week against one capping it at an hour), and coaching for pairs who do not know what to ask. It is to slot into the HR group's non-technical programs rather than duplicate them.
• A support team can clear important-but-not-urgent friction for developers. Neward credits internal support with getting laptops and hardware to all 30,000 people across Dan Gilbert's companies in about a week with no disruption, but says developer tools needing admin rights still take days or weeks. His team "greases the rails", such as talks with O'Reilly on Safari subscriptions, acting as a force multiplier against the tyranny of the urgent.
• Where a company spends time and energy shows what it values. Neward contrasts companies that say employees come first but fund no programs for them with Quicken Loans' "isms": 20 sayings from Dan Gilbert published as a glossy book and presented at twice-monthly, all-day "isms day" events. He challenges managers to check whether their team's time reflects stated priorities such as training or innovation, and to change it if not.

2021-03-30 · open the original ↗

• The G4 PTZ is worth its price only for actively staffed business monitoring. At $1,800 MSRP the host thinks it is priced about right for a business already on UniFi Protect, with staff such as a security guard steering it over a parking lot or warehouse; for a home user he calls it not worth it, and a Reolink 5 MP, 4x-zoom PTZ at $229 (a camera Reolink sent him) more realistic.
• Picture quality is the best in Ubiquiti's lineup, night vision included. The camera is 4K at 24 fps, f/1.5, with 22x optical zoom; in his testing he could read a fire hydrant's lettering and recognise himself at night across the street, about 100 feet away. He ranks the fixed G4 Pro a close second.
• PTZ control lags badly except on the Viewport. The Viewport HDMI output responds almost in real time but cannot control the camera; a PC lags, less so streaming at 720p; a phone on LTE lagged 5–10 seconds, so he over-panned repeatedly. Mouse click-to-centre and scroll-wheel zoom worked best, and tracking a moving subject such as a car's licence plate is very hard.
• Autofocus, missing microphone and sheer size are the main faults. At full zoom the focus often locked onto trees or the ground behind a moving person; there is no microphone despite the price; the pan servos are audible through a wall; and the 15 lb camera is so conspicuous he is taking it off his house, though he notes a business may want it seen.
• Most of his suggested fixes are software. He asks Ubiquiti for tap-to-centre PTZ on the phone app, faster zoom-out, a joystick accessory, a two-point patrol ("sentry") mode, and AI tracking that uses Protect's person and vehicle detections to zoom on a subject and return. Only the microphone needs a hardware revision.

2021-03-25 · open the original ↗

• Customer identities belong in their own tenant, not the corporate directory. Hoerster, a Microsoft cloud solutions architect, describes Azure AD B2C as an Azure AD tenant with the enterprise features stripped out plus a hidden policy engine on top. That engine holds the configuration for sign-up, sign-in and profile editing: which attributes to collect, which claims to return, and conditional access such as forcing multifactor for users coming from unwelcome locations or a Tor browser.
• Claims carry customer attributes into the app's tokens. In his example, a sporting-goods retailer asks for a favourite sport at sign-up, stores it in the B2C tenant, and returns it in the identity or access token alongside the name, so other parts of the site can offer specials or coupons only to, say, baseball fans.
• Scale is no longer the problem; dead accounts need housekeeping. Giard recalls that early Azure AD did not scale for consumer workloads; Hoerster says B2C now handles millions or billions of users without issue. Pruning is not built in: he suggests a simple job against the Microsoft Graph API that queries a last-login attribute, disables accounts after about two years of inactivity, and deletes them perhaps a year later if nobody complains.
• Legacy logins can migrate without users noticing. One of the two engagement patterns he sees most is moving one or more legacy authentication systems into B2C via "seamless migration": custom policies authenticate the user against the old system first and, if that succeeds, copy the account into B2C, with no new account or password reset; he points to examples in the B2C docs.
• Affiliates managing their own users need delegated user management. For the second pattern, he suggests API Connectors, a feature he says released around October and still in public preview at recording, which let easily configured user flows call a REST API, for instance a delegation service that sets claims tying a user to one affiliate so they see only that affiliate's users.

2021-03-23 · open the original ↗

• OpenSpeedTest gives a local HTML5 speed test for LAN-only benchmarking. The host uses it in Docker for his access-point benchmarks. On a Synology NAS he runs the official openspeedtest/latest image on the host's network rather than a bridge, so it answers at the NAS's IP on port 8080; anything else already on 8080 needs the IP or port changed.
• Plain HTTP is acceptable for LAN testing. The container serves HTTP only. He says that is fine inside a LAN, since nothing sensitive passes and nothing leaves the network, but a web-facing install should use HTTPS, for example via a version wrapped behind an NGINX container with Let's Encrypt.
• A Raspberry Pi 4 makes a portable test appliance, booted from SSD. He boots from an M.2 SATA SSD rather than a microSD card so storage is not a bottleneck, installs Docker with the get.docker.com script, adds the pi user to the docker group, and adds Portainer on port 9000 before the OpenSpeedTest container. He does not recommend the Argon ONE M.2 case, citing a dead jumper and a promised replacement that never came.
• The Pi tops out around 700 Mbps down but full gigabit up. His desktop hit full gigabit both ways against the NAS, which has four aggregated gigabit NICs, but only about 700 Mbps down from the Pi, consistently across three Pis. He suspects a shared USB 3 bus, not the SSD, and treats that ceiling as the baseline: anything below it points to a problem on the LAN under test.
• He wants the Pi to report its own DHCP address. Because the appliance gets a different address on each client network, he asks viewers for a boot-time script, perhaps a cron job with a 60-second delay, that emails or texts its IP once it has one.

2021-03-23 · open the original ↗

• Inclusion means a say in planning, not token representation. Hanselman prefers a variant of the party quote, "inclusion is being a member of the party planning committee", and contrasts it with "pie chart diversity": Power Rangers-style casting that picks one of each skin colour, which he calls quota-based and "just a pretty color palette". He describes diversity as including background, religion, language and invisible disability, citing his own type 1 diabetes and a paralysed right arm.
• Gatekeeping is blocking information flow, so share expertise generously. He likens hoarded technical knowledge to priests who once kept reading from the public, and quotes Ted Nelson from 1974 saying experts rarely want people to understand what they do. His rule is "don't waste your keystrokes": keysleft.com, a site he made, counts down your remaining keystrokes, and he argues an answer to an emailed question should go to a URL (blog, wiki, GitHub) because email is "the only place that everyone can't search".
• Learn in public: blog experiences, not expertise, and own the domain. He reports blogging every Tuesday and Thursday for 19 years and calls the blog his real resume; he says he was an expert in none of what he wrote about, and that a beginner's Python 101 is worth writing. He tells students to write for one or two readers rather than chase views, says blogging in a native language serves under-served communities, and warns that Instagram, TikTok and Facebook are walled gardens; he backs up his own TikToks with a Python script.
• Running your own site is like driving stick once. Registering a domain, installing a certificate and fixing it when it expires is an experience he says everyone on the internet should have, while insisting that making it a requirement would itself be gatekeeping. On specialising, he argues for the fundamentals (data structures, algorithms, computer architecture) and uses his father's Swiss Army knife to argue for being decent at many things.
• Mentors are mirrors; sponsors are spotlights who make luck. He splits sponsorship into strategiser, connector, opportunity-giver and advocate, and says mentorship must run both ways and ends in phases. On finding one, he advises against cold emails and "can I pick your brain" in favour of showing up at meetups and clubs and then asking for advice on a specific problem.