Souk's Notes - Unofficial Summaries

2023-12-18 · open the original ↗

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

2023-12-16 · open the original ↗

• Clean architecture puts business logic, not the database, at the center. Smith calls it the latest name for a decades-old idea also known as ports and adapters, hexagonal or onion architecture, coined by Robert Martin: code depends on interfaces saying what it wants (fetch a record, send an email), and the how is an implementation detail it neither knows nor cares about. He says his clean architecture talk at the previous month's .NET Conf drew more YouTube views than the keynote or any other session; Giard contrasts it with the database-centric habits he carried over from FoxPro.
• Tight coupling bites 24–36 months in, not early. Smith says fast early work and copy-paste reuse leave code bound to its infrastructure, which he calls a sign of success that his company, NimblePros, is often hired to refactor. To the pushback "we'll never change our production database", he answers that pipelines through dev, test and staging environments are where swapping the database, mail server or file system pays off.
• Abstracted dependencies make business logic unit-testable. In his e-commerce checkout example, a method that saves the order, charges the card and sends SMTP mail directly can only be integration-tested, brittly; with a faked payment processor and email sender, a unit test can check that a successful charge leads to the success email and a failed one to the payment-failed email, without sending either.
• The compiler enforces the dependency rule across three projects. Smith uses a Core project (entities, value objects, aggregates, domain services, domain events and specifications from domain-driven design), an Infrastructure project that depends on Core, and a UI project, typically ASP.NET Core; because .NET forbids circular references, Core cannot depend on infrastructure except through reflection that would stand out in code review. His free solution template on GitHub and NuGet sets this up, and CQRS can be added as a further layer separating commands from queries, possibly with separate data stores.
• Clean architecture is overkill for small, simple systems. Smith says it adds projects and abstraction and hides the single followable method some developers want, so it suits larger systems or ones with thousand-line methods making 20 database calls. Giard frames the choice as whether reduced code complexity outweighs the architecture's own; Smith calls it "the bang and the buck" and says everything in architecture is a trade-off.

2023-12-13 · open the original ↗

• A codespace is a full development environment hosted in the cloud. Nothing has to be installed locally. Giard creates one from a repository's green Code button, on the Codespaces tab, with "Create codespace on main" (the main branch). It opens in the browser looking much like Visual Studio Code, with the Explorer showing the repository's files.
• Edit, commit and sync all happen inside the browser. Giard edits a view file, which autosaves, then stages and commits it from the source control tab and syncs it back. The change then shows in the GitHub repository. He calls doing everything in the browser "really nice".
• Codespaces get random adjective-noun names, which you can rename. His was called "musical waffle". The three-dots menu on the Codespaces tab lets you delete or rename a codespace. Giard says renaming makes one easier to find when a single repo has several codespaces.
• Desktop Visual Studio Code can connect to the same cloud codespace. In the command palette (Ctrl+Shift+P on Windows, or View > Command Palette), typing "codespaces" offers to create a new codespace or connect to an existing one, but not to delete one. The files stay in the cloud rather than coming down to the machine. Giard says the stage, commit and sync workflow is the same as working locally. He adds that other people can open a codespace too, which lets settings be shared.
• Creating with extra options sets branch, region and machine size. The options include the branch and a region, for example Western Europe for developers based there. The machine size runs from 2 cores up to 16 cores, and Giard notes that a larger machine costs more.

2023-12-13 · open the original ↗

• DDoS attacks now often serve as a smoke screen. Hébert, a Microsoft partner solution architect, says the pattern has changed over the last 10 years: rather than one big volumetric takedown, attackers pulse traffic up and down to trip and reset DDoS suppression while scanning and trying exploits in between. Smaller DDoS events of 10–12 gigabits in your logs can hide real attacks, so the logs between them have to be searched.
• Attackers now poison the third-party APIs you consume. He points to "unsafe consumption of APIs", a new entry in the 2022 OWASP API Top 10. Instead of a malicious package, the attacker compromises an API you call, such as an address-validation service, and injects bad data that your system trusts because you pay for it, then exploits it once it is inside.
• A DevOps pipeline needs SAST, DAST, composition analysis and fuzzing. Hébert starts from the MITRE ATT&CK framework, then runs static code analysis, dynamic scanning of the hosted API, and software composition analysis covering vulnerable packages and region-restricted licences, pairing a paid tool with a free fallback in case funding is lost (WhiteSource with OWASP Dependency-Check). He says most practitioners skip fuzzing, and recommends RESTler from Microsoft Research, a stateful REST fuzzer that infers producer-consumer dependencies from a Swagger definition and aims to provoke the 500 errors used in exploit development.
• Security checks can start at the API contract, before any code. For API-first design he uses SwaggerHub for collaborative design with intended users, with integration into Azure API Management for mocks. He says 42Crunch, a Microsoft partner integrated into Defender for APIs, has a free audit that runs about 300 checks against the uploaded Swagger contract, which shifts security left to the documentation phase. He also recommends its apisecurity.io newsletter, which explains how breached organisations were exploited.
• Forgotten, uninventoried APIs are a vulnerability in themselves. He cites "improper inventory management", also in the 2022 OWASP API Top 10, and recalls consulting clients who could not say how many APIs, SOA endpoints or BizTalk orchestrations they had. He promotes his employer's Azure API Center, then in preview, as a complement to API Management for cataloguing APIs by environment and version, and says the Defender for APIs plan flags an API unused for 30 days as likely stale.

2023-12-13 · open the original ↗

• The Pro Max switches are conventional apart from Etherlighting. There are four models, 24 or 48 ports with or without PoE; at recording the PoE models were shipping first, with non-PoE a few weeks behind. One third of the ports are 2.5 Gbps (8 on 24-port, 16 on 48-port), the rest 1 Gbps, plus 10 Gbps SFP+ cages and PoE+ and PoE++ on the PoE models.
• Colouring ports by native VLAN is the host's killer feature. Every port has full-colour LEDs that, in UniFi Network Early Access 8.0.2, show either link speed (the default) or the port's native VLAN, with customisable colours. He says seeing VLANs at a glance helps troubleshooting and stops devices being patched into the wrong VLAN.
• Locate makes tracing a device's switch port trivial. From the port manager or any UniFi device's settings, locate turns off every port LED except the target, which pulses blue, so a ceiling access point or camera can be traced without following cables. The host calls the switches' AR feature a gimmick by comparison.
• Ubiquiti's Etherlighting patch cables were not yet tested. Ubiquiti released cables it says enhance the brightness and clarity of the port lighting; the host had not obtained any at recording and promises a follow-up.
• The host wants a show-all-ports mode and consistent colours. He asks for a show-VLANs or show-speeds option, also on the switch's 1.3-inch display, lighting all empty ports. He notes the VLAN manager's blue native and green untagged colours could clash with Etherlighting choices, and wants port-manager speed colours synced to Etherlighting.

2023-12-11 · open the original ↗

• A GitHub project can be started from your profile menu. Giard opens the dropdown on his profile icon, picks "Your projects", then "New project", and chooses from three templates: a table (rows and columns), a board (a Kanban view of work as not started, in progress or done) and a roadmap. The board he creates has To do, In progress and Done columns. In the project's settings he renames it, adds a plain-text description and a readme that takes Markdown and can be previewed, and sets it public or private; the project can also be closed or deleted there.
• New items start as drafts; issues must be tied to a repository. Items added to To do show up as drafts, and converting one to an issue means choosing a repository for it. Items move between columns by dragging, or by setting an item's status to Done, which closes it and moves it to the Done column; drafts can be handled this way too. Giard recommends converting items to issues first, though he says drafts may make more sense for work not connected to any repository.
• One project can have several views, each with its own fields. Giard adds a table view to his board, which shows title, assignees and status columns. The plus button adds built-in fields such as labels and reviewers, hides ones you don't want (he gives status as an example), or creates a custom field of type text, number, date or single select. He adds an expected-completion date field, which uses a calendar picker, and a single-select "Difficulty" field with the options easy, moderate and difficult.
• A project created from a repository's tab is linked to it. The repository's Projects tab can link an existing project or create a new one through the same template dialog. This time Giard uses the Team backlog template, which he calls a more complex version of the board, with New, Backlog, Ready and In progress columns. The repository then shows as linked in the project's settings, which save automatically.
• Each issue carries assignee, label, priority and size fields. An issue can be assigned to any collaborator on the project; Giard is the only one, so he assigns it to himself. It can also get pre-built labels (bug, documentation, duplicate, enhancement and others), linked pull requests, a priority (urgent, high, medium and so on) and a size for estimating. Giard's rule of thumb for sizes is that each step doubles the effort: a small takes twice as long as a tiny, and a medium twice as long as a small.

2023-12-11 · open the original ↗

• A $1,500 Best Buy shelf price was mislabelled, but still outrageous. Twitter users spotted that the price belonged to a TP-Link Deco Wi-Fi 7 whole-home mesh three-pack, not the TP-Link Archer Wi-Fi 7 router on the shelf, which lists at $599 MSRP. The host calls both prices outrageous, noting the Wi-Fi 6E version of the Deco is $269 at Costco.
• Wi-Fi 7 was not yet a certified standard in late 2023. The Wi-Fi Alliance had not certified 802.11be, expected in early 2024, and says products should meanwhile be described as containing Wi-Fi 7 technology. The host would not buy Wi-Fi 7 gear until the standard is formalised, since current devices' compliance is not guaranteed.
• Home users would gain little from Wi-Fi 7 anyway. The host found only a single Intel Wi-Fi 7 M.2 card on Amazon, plus phones such as the Pixel 8; most clients are Wi-Fi 6. He argues the gains are for device density, so a home with a few dozen devices is well served by Wi-Fi 6, and expects wider adoption perhaps in 2025.
• A $600 UniFi network out-covers a single Wi-Fi 7 router. In a 3,400 sq ft (315 m²) floor plan, the host's Design Center model shows a single router missing back rooms. His alternative is a $199 Dream Router plus an optional USW Lite 8 PoE and two U6+ access points, totalling $566; the costs are running Ethernet and a steeper learning curve.
• A $1,500 UniFi build buys flexibility, not needed coverage. Against the mesh three-pack, which he expects to cover about the same, the host specs a Dream Machine SE, three U6 Pros, a U6 In-Wall for the garage and a Flex 10 GbE switch. He stresses he was trying to spend the budget; the $600 setup already gives solid Wi-Fi.

2023-12-11 · open the original ↗

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

2023-12-09 · open the original ↗

• Git is distributed: work locally, and keep GitHub as the shared copy. Giard describes the cloud repository on GitHub as the place others can reach and the place to deploy from. You clone it down, test locally, push your working changes back, and pull other people's changes down. On a team, even of hundreds of people, he calls the GitHub repository the system of record.
• Cloning a GitHub repository creates a local folder linked to it. Giard copies the repository's URL from its Code button and runs git clone <url> in a terminal. The new folder is named after the repository. He says any IDE or a plain command prompt will do. git --version checks whether Git is installed, and he calls the install simple and free. For content, he generates an MVC app with the .NET CLI (dotnet new mvc).
• Saving in Git is two steps: add, then commit with a message. git add takes specific files or folders, . for everything, or -A. Giard prefers -A because it also picks up deleted and updated files. git commit -m "<message>" then records the change. He says the messages make the history readable later, for example a note that a commit fixed a given bug. git status shows files as untracked, then staged (in green), then committed.
• Local commits reach GitHub only on git push; git pull brings changes back. After committing, the GitHub repository is still empty until Giard runs git push. Because the folder was cloned it already knows its remote, so pushing from local main goes to main in the cloud. When he edits a file on GitHub, git pull brings that change down. Naming a branch is optional when you are already on main.

2023-12-09 · open the original ↗

• Security strategy starts with knowing what you must protect. Da Silva, a cybersecurity architect at Microsoft who builds security strategies with its partners, treats cyber risk as business risk and uses the NIST (National Institute of Standards and Technology) Cybersecurity Framework, whose first pillar, identify, means knowing devices, users and workloads. A production server needs a different approach from a test server, and important accounts whose compromise would endanger the business get their own strategies.
• Protection needs people, process and technology together. For the second pillar, protect, he covers physical security (guards, gates) and digital controls such as firewalls, antivirus and EDR, and says processes, such as how internal information may be shared, matter as much as the tools.
• Monitoring catches what protection misses, and response can be automated. Detect means monitoring the perimeter and the security tools themselves; Giard raises breaches that went undiscovered for months. For respond, Da Silva compares an apartment's cameras and alarm that call the police with SOAR (security orchestration, automation and response), which triggers actions once a threat is detected.
• Recovery must be planned because some attacks will succeed. His example is ransomware encrypting a company's data after detection and response fail, then clean-up and restoration; he likens recovery to home insurance, which you hold but would rather invest to avoid using. Giard frames the five pillars as sequential and argues, shift-left style, that stopping problems upstream makes later recovery easier; Da Silva answers that security works in layers.
• Da Silva says Microsoft's portfolio covers all five pillars. As a Microsoft employee he describes the Defender XDR family (Defender for Endpoint, Identity, SQL, Cloud, Containers) for protection, and Microsoft Sentinel for correlating events to detect threats and automating responses such as emailing, blocking a user or isolating an endpoint. He says detection does not remove the need for protection, and that Microsoft is integrating its tools into one extended detection and response (XDR) strategy to simplify the work.

2023-12-06 · open the original ↗

• A travel router's WAN should be wired, not a hotel Wi-Fi repeater. Using his GL.iNet Beryl AX, the host gives three reasons: a more solid connection, no wireless bandwidth taken from client devices, and usually no captive portal on the room's wired network.
• Hotel-room LAN ports can be hard to find. Not every room has one; in his, the only port was on a small device hanging off the back of the television rather than on the wall.

2023-12-04 · open the original ↗

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

2023-12-01 · open the original ↗

• A cheap PoE splitter lets the UniFi Express run on PoE. The Express has no PoE input, which commenters on the host's first video complained about; a splitter takes PoE Ethernet in and outputs USB-C power plus a separate RJ45 data connection. Micro-USB versions exist for older devices such as early Raspberry Pis.

2023-11-29 · open the original ↗

• UniFi Express is a router, Wi-Fi access point and controller in one. The UX combines a router/firewall, a 2x2 MU-MIMO Wi-Fi 6 access point (22 dBm max transmit power at 2.4 GHz, 25 dBm at 5 GHz), and UniFi OS with UniFi Network on board, so it needs no separate Cloud Key. It is just under 4 in square, has one gigabit WAN and one gigabit LAN port, Bluetooth for app adoption, a non-touch status screen, and draws 10 W max.
• The host rates it better value than the Gateway Lite. MSRP was set at $149 after recording, against $129 for the UXG Lite he reviewed the week before, which has no onboard UniFi Network and no access point. He would rather pay the extra $20.
• It fits small networks only, but it is flexible. Ubiquiti claims up to 60 clients; larger deployments should use a console such as a Dream Machine. Extra Express units can be wired in as access points under any UniFi controller, or meshed wirelessly to give a remote room Wi-Fi plus an Ethernet port. Each unit needs its own power (no PoE), and meshing cuts throughput for devices on the remote unit.
• Ubiquiti says DPI won't cut its gigabit throughput. The host says Ubiquiti confirmed to him that, unlike the Gateway Lite, the Express keeps full gigabit throughput with DPI or content filtering on; he had not yet tested this himself. He rejects calls for 2.5 GbE ports: it targets sub-gigabit connections and people new to UniFi.
• The host found it runs hot but sips power. He measured up to 130°F on the bottom of the unit. Paired with T-Mobile Home Internet, his setup drew just over 14 W, which he suggests suits RVs on a 5G or LTE connection.

2023-11-29 · open the original ↗

• Most Azure customer workloads now run on Linux. Randall, a principal product manager for Linux strategy in Azure Core who joined Microsoft about two and a half years ago through its acquisition of the Berlin startup Kinvolk, says Linux has been the majority for "quite a few years" because it is the OS that powers the cloud, and a cloud company cannot succeed without a strong Linux experience.
• Managed Ubuntu snapshots make fleet updates predictable. Randall says Microsoft and Canonical announced the joint service at the Ubuntu Summit: rather than pulling the latest packages in whatever order they arrive, enterprises can validate a fixed set of updates and roll it out across servers, enabled by changes Microsoft made upstream in APT.
• Microsoft built Azure Linux to control its own internal fleet. Randall says internal services ran a variety of distros, making security and consistency hard to enforce; the in-house distribution is now offered as an AKS node option beside Ubuntu, smaller and container-optimized, without displacing partners Canonical, Red Hat, SUSE and Oracle. He calls Azure Hybrid Benefit's license portability "pretty much unique", and says joint support with Red Hat lets a RHEL customer on Azure call either company.
• Microsoft contributes upstream and runs its own CNCF projects. Randall cites kernel work, systemd founder Lennart Poettering on staff, and the cloud-init summit Microsoft hosted in Redmond. His team's Inspektor Gadget packages eBPF programs into gadgets for container- and host-level observability, and Headlamp is an extensible Kubernetes UI, used internally as an AKS portal, that runs as a desktop or web app.
• Flatcar Container Linux is stripped-down and immutable for container hosts. Randall says the Kinvolk-origin distro, a derivative of CoreOS, keeps its OS partition read-only and updates by installing to a second partition and flipping over, or back if the update fails, an approach taken from Chrome OS; it is being moved into the CNCF.

2023-11-29 · open the original ↗

• A free GitHub account covers everything in the demo. Giard says there are paid accounts, but signing up at github.com is free. A new repository is created from "Your repositories" > New: pick an owner, give a name that only has to be unique among that owner's repositories, add an optional description, and choose public or private (private means only people you invite can see and use it). He makes his public.
• A .gitignore file keeps files you shouldn't commit out of the repo. Giard's examples are compiled binaries, files that configure your development environment, and files holding secrets such as passwords and connection strings. The new-repository dialog offers .gitignore templates by language. He picks Node, and the generated file excludes things like log and coverage files. You can edit it afterwards.
• A license file tells others what they may do with your code. Giard lists the questions it answers: whether they can use the code in their own work, use it as a starting point, or make money from it, and whether they must credit you. He picks MIT. The dialog can also create a README.md holding just the repo title, and by default creates a single branch named main, which can be renamed later.
• Files can be edited and committed in the browser. The README is written in Markdown, a simple text-formatting language, and is displayed below the file list. Giard edits it with the pencil (edit) button and commits the change. He explains "commit" as saving the change so it becomes part of the repository.
• Deleting a repository is deliberately hard, and Giard says it should be. The option sits at the bottom of Settings, in the "Danger Zone", next to changing visibility and transferring ownership. GitHub asks for confirmation twice, warns that the deletion is permanent, and then makes you type the repository's full name before it deletes it.

2023-11-28 · open the original ↗

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

2023-11-27 · open the original ↗

• Zero trust means wrapping applications in layers that frustrate attackers. Jackson, a HashiCorp developer advocate, places Consul on the network side and Vault as the secrets tool; he counts API keys, passwords and cloud credentials as secrets, and also personal data, since a leak of it brings GDPR liability in the EU. Vault's emphasis is granular access, by an application's service principal or a person logging in through LDAP.
• Short-lived dynamic secrets make a leak harmless. Jackson says Vault can generate credentials such as a managed PostgreSQL login or an Azure account with a short time to live; when he accidentally committed Kubernetes credentials to GitHub the week before, he killed the lease in Vault and the exposed credentials became useless. He says HashiCorp's acquisition of BluBracket aims to find committed secrets before attackers do and revoke them automatically.
• Most applications consume Vault through templates, not the API. Jackson says Vault Agent authenticates with the machine or Kubernetes service-account identity and writes secrets into a config file; on Kubernetes, an injector templates secrets into pods, or the Vault operator syncs them into native Kubernetes secrets and can restart pods. The goal is minimal, ideally zero, application changes.
• Azure Key Vault suffices for static secrets in one cloud. Jackson says Vault earns its place with dynamic database credentials, generating Azure credentials for Terraform, and hybrid or multi-cloud estates where a security team wants one workflow for all secrets.
• Run Vault managed or via Helm, and treat it as critical. Jackson says the single binary gives a dev setup in 30 seconds but not a production one; HashiCorp's cloud service, a single-tenant cluster it patches and backs up in Azure, or the Helm chart get you production-ready, after which you must manage the unseal key and storage, now simpler with built-in Raft. If Vault is down the application is down, so he describes active-standby clustering, an enterprise standby-cluster option and performance replication for heavy encryption loads.

2023-11-23 · open the original ↗

• LLMs are a lossy compression of the internet learned via next-word prediction. Training Llama 2 70B compressed roughly 10TB of scraped web text into 140GB of parameters, a rough 100x ratio, using 6,000 GPUs over 12 days for about $2 million. Because the task forces broad knowledge into the weights, that knowledge is oddly one-directional: GPT-4 can answer "who is Tom Cruise's mother" but not the reverse, "who is Mary Lee Pfeiffer's son."
• Assistants are built in two stages: pretraining then fine-tuning. Pretraining compresses hundreds of terabytes of low-quality internet text into a base model that only completes documents; fine-tuning then retrains on a much smaller set of high-quality, human-written Q&A conversations — roughly 100,000 examples, per the talk — to reshape the model's output into a helpful-assistant format while it still draws on pretraining's knowledge. An optional third stage, RLHF, trains on human comparisons between candidate answers rather than written answers, since ranking is easier for a labeler than composing one from scratch.
• Jailbreaks and prompt injection exploit blind spots in safety training. Because Claude's refusal examples are almost all in English, encoding the same harmful request in base64 slips past its safety training, and framing a request as roleplay ("pretend to be my deceased grandmother") got ChatGPT to give napalm-production steps it would otherwise refuse. Separately, prompt injection hides instructions a model can read but a person can't see — invisible text in an image or a shared document — to hijack an assistant into pushing a phishing link or leaking private data.

2023-11-22 · open the original ↗

• The Gateway Lite is UniFi's long-awaited entry-level USG replacement. The UXG Lite costs $129. Ubiquiti claims 10 times the original USG's routing performance; it has a dual-core ARM Cortex-A53, 1 GB of DDR3L RAM (double the USG's), USB-C power at 15 W, and one gigabit WAN and one gigabit LAN port, against the USG's two LAN ports. The host bought it himself.
• The host says it is for sub-gigabit, simple networks. He dismisses complaints about too few ports, no 2.5 GbE and no PoE: multi-gigabit connections or onboard switching need a different UniFi device, and with VLANs and switches he sees little need for more firewall ports.
• It is not a console and needs a separate controller. He first said it needed UniFi Network 8.0.6 or higher, then found that wrong: his Cloud Key Gen 2 on an older 7.x release still adopted it after four rounds of updates. By default Teleport and DPI are on, IDS/IPS are off; he thinks Teleport should be opt-in. It is fanless and silent and ran at 125–130°F.
• A minimal Gateway Lite setup costs over twice a Dream Router. Gateway Lite plus a Lite 8 PoE switch, a Cloud Key Gen 2 and a U6 Lite access point comes to just over $400, against a Dream Router at less than half that. What the separate setup buys is placement flexibility and not putting all your eggs in one basket; he does not know whether it performs better.
• The host favours it for installers with a hosted controller. With one central controller (Ubiquiti's official hosting or a service such as HostiFi), he would standardize small-business customers on Gateway Lite setups. That leaves one controller to update and back up, costs much less than a Dream Machine Pro per site, and lets spares be kept on hand.

2023-11-21 · open the original ↗

• TicketFalcon exists because the twins resented existing platforms' fees. Ashton Clark describes it as an online event registration and management platform for general admission and reserved-seating events with direct payouts, built by him and his identical twin Ryan as "technologists and also businessmen".
• A slow site costs as much as a down one. Ashton says availability and speed have to be balanced, and both come down to server capacity. Extra capacity for a big event is planned in advance, not added 10 minutes before. Planned downtime goes in the early morning, and customers are told beforehand, which he says they appreciate.
• Growing makes them a target, so they defend in layers. Ashton cites server patching, enterprise-level tools using machine learning to stop attacks early, and DDoS attempts they have seen from servers "all across the world" and through partners' services. He says a third-party failure "a couple of weeks ago", which hit millions of sites, disrupted them for a few minutes, and that they use a payment partner because running their own processing would earn more revenue but carry more risk.
• Websites scale the business without scaling the staff. The brothers trace this to raking leaves and shovelling snow at about nine or 10, taking a percentage from other kids working other blocks, and to selling sneakers for $100–150 at 16. Ryan says thousands can buy tickets at once without 1,000 people running the site. They scale cloud compute up only for events and back down, both to save money and to avoid wasting energy.
• The name came from family pushback, not the founder's favourite. Ashton had bought Tixboss.com and its other extensions, for maybe $15, but his then-girlfriend, now wife, and family rejected it because he could not say it 10 times fast. She proposed "falcon" in a Houston hotel-room brainstorm. He says businesses should not be built in a vacuum, and adds that domain names take far more creativity than when they built a sneaker site around 2006.