Souk's Notes - Unofficial Summaries

2024-09-29 · open the original ↗

• With scaling ending, gains must come from whole systems. Hwu, a senior director of research at Nvidia who taught at Illinois for 34 years, says Dennard scaling (lower voltage and power and faster switching as transistors shrink) mattered more than Moore's law: for about 15 years code got faster untouched. Dennard scaling ended in the early 2000s, and he sees "at least another two generations" of Moore's law. He expects the next advances in networking, moving from electrical to optical links that cost more than 10 times their copper equivalents, in primitives for synchronising huge numbers of machines, and in cooling, where data centres are "at the cooling limit".
• Research has to target the technology of ten years out. In the 1980s his Berkeley group under Yale Patt bet that abundant fast transistors could find parallelism no processor then detected. Intel's P6 (Pentium Pro), designed by 1993, embodied those mechanisms, and he draws the lesson that ideas take five to ten years to reach hardware.
• Python frameworks throw away parallelism that lower layers must recover. He cites the Python interpreter lock, and GPU work launched from Python frameworks such as Torch as separate, isolated API calls. He expects runtime techniques such as fusing those calls on the fly, long researched, to be deployed "in the next decade or so" because other optimisations are running out.
• He predicts a new processor unit built for retrieving data. People buy processors "reluctantly" to run applications, and he argues deep-learning models bypassed going straight to the data. His proposed unit would tolerate latency and pull, for example, one kilobyte out of four petabytes in real time to feed models. If the industry "does the right things", he expects within a decade a pocket device holding a person's lifetime of memory, and says he will retire when he has one.
• Enough value makes a missing software ecosystem appear. In 2011 an NSF review panel called his GPU-based Blue Waters supercomputer proposal "risky to say the least and probably irresponsible" for lacking math libraries; he answered that GPUs were "a small puppy with big paws". Thirteen years later those libraries exist. He describes the follow-on Delta project as filling a shortage of GPUs for academic researchers, not as pushing the frontier.

2024-09-27 · open the original ↗

• Copilot in Excel requires AutoSave and the Copilot service. AutoSave means the workbook has to be stored in OneDrive. With both in place, a Copilot button appears on the Home ribbon and opens a pane that takes plain-English questions and instructions as a second way of working the sheet, alongside formulas.
• Copilot turns a plain-English rule into conditional formatting. Giard asks it to highlight each player's name where dribbles per 90 minutes beat either interceptions or tackles per 90. Copilot first says what it will do and applies it only when he clicks Apply. The formula it wrote is visible under the conditional formatting rules. Giard says he could have built it himself, but it saved him the trouble and is a head start if he wants to change it.
• Advanced analysis writes the Python for you. "Start advanced analysis" adds a new worksheet holding a data frame that previews the data it will work on. Asked for each region's monthly total value as a line chart, it produced one covering January 2022 to the end of 2023, about 24 months, with east and west regions. The generated Python code is shown and can be edited and rerun, which Giard calls a good starting point.
• Follow-up prompts add a forecast and restyle the chart. Asked to forecast each region's total value for the first six months of 2024 on the same chart, Copilot drew the forecast as a dashed line without being told to, but in a different colour from the historical line. A further prompt (label the forecast explicitly, use the same colour for actual and forecast, and make forecast lines dashed and four times as thick) regenerated the chart as asked.

2024-09-26 · open the original ↗

• Most data API code is copied CRUD that a tool should provide. Nixon, a Microsoft product manager on the SQL Server team under Azure Data who is in charge of Data API Builder, says the layer between client and database is mostly create/read/update/delete repositories, endpoints, unit tests and CI/CD steps copied from project to project, "almost a third of the code" in an application. Data API Builder (DAB) drops that data API on top of the database instead; he jokes it should be "CRUDE", with execute for stored procedures.
• It is zero code: a JSON file handed to a stock container. The file describes the database, which can be SQL Server, Postgres, Cosmos DB or MySQL, in Azure, AWS or on-prem. The container is a standard Microsoft one that runs on Kubernetes, Azure Container Apps, Azure Container Instances, or inside Static Web Apps, which hides the container for the many users who cannot run one. Clients in any language just call it over HTTP and get JSON back. Nixon says large customers whose policies now forbid talking to the database directly use it, sometimes with a citizen developer and no development team. The config can be built with a .NET CLI tool (dab init, dab add) rather than by hand-editing JSON, and a drag-and-drop configurator is planned for "hopefully sometime next year".
• REST and GraphQL can be served at once from one config. Nixon says this meets REST shops trying out GraphQL as well as GraphQL shops with one REST operation. He argues GraphQL saves mobile bandwidth with one nested query, such as an invoice, its line items and product photos, and can insert an invoice and its lines as one transaction with no custom endpoint. Entities can point at tables, views, stored procedures or NoSQL documents. A Cosmos DB and a SQL Server source can be queried together into one payload, and relationships can be declared even where the database has none ("we can make a crappy database look awesome").
• Security is set per entity in the config, not by the DBA. Each entity lists roles and their allowed actions; his example lets anonymous users create and read but not delete. Policies such as owner ID equalling the caller's user-ID claim restrict users to their own rows. Connections can use passwords, system- or user-assigned managed identities, or secrets read from environment variables, and telemetry goes to Application Insights, with more log sinks planned.
• DAB replaces only the CRUD, and it is free and open source. Nixon says business operations ("launch the missile") still need their own API, possibly still on Entity Framework, which DAB neither uses nor replaces. It has no Microsoft prerequisite and runs fully on-prem for some customers. The repository is at aka.ms/dab, where community sign-ups shaped the team's priorities for the coming six-month planning "semester".

2024-09-24 · open the original ↗

• The MAX BR2 can run two 5G carriers at once. The host describes the Peplink MAX BR2 as a professional-grade cellular router with two onboard 5G modems, plus two wired WAN ports, all manageable together through load balancing, failover or Peplink's SpeedFusion bonding.
• It is built for vehicle mounting, as a mobile network in a box. The host points to a metal case with built-in mounting brackets and a 10–30 V DC terminal block with ignition sense, and suggests emergency vehicles, police cars, ambulances or broadcast vans as uses. Crosstalk says it is a Peplink partner and offers sales and support.

2024-09-23 · open the original ↗

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

2024-09-20 · open the original ↗

• Most users will notice no difference between $100 and $500 mesh. At Costco the host found Deco mesh systems at about $100 for Wi-Fi 6, $249.99 for Wi-Fi 6E and about $500 for Wi-Fi 7, and says the vast majority of users would see no difference whatsoever between the cheapest and the dearest.

2024-09-19 · open the original ↗

• Staying a one-person shop trades polish for time. Finney, a Microsoft MVP and former mechanical engineer of 10 years who builds apps beside a full-time job, says time is his scarcest resource and finding the right hire could take months with nothing shipped. He accepts that design and architecture "suffer a little" as long as the result is good enough, and he picks whichever framework avoids fighting each app's one hard step.
• Data-loss bugs are the ones he will not tolerate. He reasons that large companies ship bugs too, so the fear he prioritises is Ink Calendar users losing their journals. In app messages and review replies he writes "I" rather than "the team" so users know it is just him.
• App store reviews give users no channel for dialogue. Text Grab users file GitHub issues with line numbers. Ink Calendar, found through the store, draws one-star reviews about Store download and update failures that have nothing to do with the app. Hanselman suggests developers review more kindly than consumers, and Finney agrees.
• Store packaging, not coding, is what eats an indie's time. For his React Native app he pays for Expo's hosted service, where "EAS build all" builds, signs and submits to both stores and generates icons. On Windows the toolchain fails midway or at certification, and the MS Store CLI has not worked for him. He defends the stores' cut (70 cents on a dollar) as covering hosting and distribution he could not run himself.
• Giving Text Extractor to PowerToys paid in exposure, not money. He contributed it when a GitHub issue proposed the feature, keeping a narrower scope while Text Grab stayed his more flexible app. Users still come to Text Grab from it, it led to his MVP award, and the team later built that functionality into Snipping Tool, which he calls "inevitable".

2024-09-19 · open the original ↗

• Copilot in OneDrive summarizes selected files without opening them. In OneDrive on the web, Giard selects one or more files under My Files and chooses Summarize from the Copilot menu, which returns a summary of each file after a few seconds. He shows that follow-up questions can then ask for more detail or about something specific in the files.
• A Compare option builds a side-by-side table across up to five files. Again without opening them, Copilot lays out differences as a table. In his demo of a Word document and a PowerPoint deck it gave the title, creator, purpose, target audience, methodology and expected outcome, and it took follow-up questions.
• Files that share a format give a more useful comparison. Giard says the Word-and-PowerPoint pair worked only because the topic matched. Three of his old status reports from 2015, all with the same sections (past two weeks, next two weeks, upcoming, issues), came back as a table set out by those sections. He calls it a "very simple interface" for analyzing and comparing files quickly, which he says works because Copilot integrates directly with OneDrive.

2024-09-19 · open the original ↗

• Newsom is attacking high gas prices his own legislation caused. He mocks the special legislative session and Newsom's claim that oil companies charge Californians more, contrasting about $2.50 gas in Oklahoma with about $5 in California.
• Newsom's refinery storage mandate would take a decade to build. California uses over 900,000 barrels of gasoline a day. Catherine Boyd of the Western States Petroleum Association put each tank at $35 million, with land, environmental studies and permits taking about 10 years, by which time Newsom says gas cars will be gone.

2024-09-18 · open the original ↗

• Utilities raise rates to fund capital spending, not fuel costs. His example is Alliant Energy in Iowa, serving about half a million people, which seeks higher natural gas rates despite near two-year historic-low gas prices. It earned $700 million last year, about $1,400 per customer, raised dividends 21 straight years, and wants grid, line-burial and renewables costs covered.
• Taxpayers already subsidize the same buildout. He cites federal data as of 2022, before the full effect of the Inflation Reduction Act, showing renewables get over seven times oil and gas subsidies. The second-largest energy subsidy is end-use help with utility bills. He expects bills to keep rising.

2024-09-17 · open the original ↗

• Managed identities hand secret management to Azure, shrinking leak risk. Rodriguez, who works at Microsoft supporting enterprise Azure customers, says you set the permissions and Azure creates and rotates the tokens underneath. A connection string or access token can always leak by human error, and the question is the blast radius and the work to clean it up. He notes that GitHub has become better at blocking token uploads. Giard adds that his employer scans his repos and emails him about checked-in secrets, and that Teams blocks messages containing connection strings, which Rodriguez says can be tricked.
• Service principals, the mechanism underneath, are what make this painful by hand. Rodriguez compares them to on-prem service accounts. Their client secret expires after a maximum he is unsure of ("something like six months, nine months"), and he says Azure gives no reminder, so the first warning is users complaining. He says he relearned how to create one three or four times because he used them so rarely.
• The managed identity only covers Azure-to-Azure calls. For on-prem, AWS or Google resources you are back to managing secrets. His approach is to keep those in Azure Key Vault and read them with a managed identity. He warns against apps holding the Key Vault's own password, since that one leak exposes everything in the vault.
• Code passes a credential object instead of a connection string. In C# you construct DefaultAzureCredential and pass it to the service client. It walks a list of credential sources, and he recommends disabling the unused ones for performance. Alternatively, ChainedTokenCredential starts empty and you add only the sources you use. Giard calls the chained form more secure; Rodriguez says only that it fits zero trust better. Locally, it picks up the developer's Azure CLI, PowerShell or Visual Studio sign-in, so the developer personally needs the same permissions the deployed app's identity has.
• The up-front learning makes development easier afterwards. Rodriguez says each developer can stand up their own environment and grant themselves access, so nobody asks in team chat for an updated connection string and nobody is tempted to commit one. He blames Microsoft for early Azure samples that used connection strings. He says learn.microsoft.com documentation now defaults to the managed-identity version of the code, though not everywhere yet.

2024-09-17 · open the original ↗

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

2024-09-12 · open the original ↗

• Battery life comes from an efficient Arm ISA plus dedicated accelerators. Van Doorn, a Qualcomm senior vice president, cites 19 hours for the Dell XPS and 26 hours for HP on Snapdragon X Elite. Running Recall on the NPU costs about 30 minutes of battery against 6 hours or more on a GPU. He credits Arm's later start with shedding x86's 40-year legacy, which he notes is also x86's strength.
• Qualcomm had to unlearn an IP company's caution to court developers. He says he joined two years ago and wrote an internal "open manifesto": a cell-phone business with a handful of customers does not scale to the thousands of developers in PC, XR, IoT and auto, so they need "no touch or low touch" enablement without licence terms. Legal and licensing had to change too. Engineers enabled Linux on X Elite outside the product plan and open-sourced it, and Qualcomm worked with Linaro and Microsoft on LLVM for Windows.
• The PC part is not a phone chip. It shares building blocks such as CPU cores with phone parts, but memory channels, accelerator counts and the NPU's processing elements are scaled per target. He says the PC chip is too big and too expensive for a phone.
• NPUs win on power but cost the most developer effort. NPUs compute in integers, which need far fewer transistors than GPU floating point, but models trained in floating point must be quantised, which he calls tricky and slow. Efficiency runs NPU over GPU over CPU, and ease of bring-up runs the other way, so teams pressed for time to market start on the CPU. He reports Llama 7B and a 1 to 1.5 billion parameter speech-to-text model running on the NPU.
• Local AI will run hybrid: on-device first, cloud as fallback. He gives privacy, response time and power as reasons, since cloud GPUs are expensive and power-hungry. Models of 7 billion parameters run fine on device, while 80 billion is "more tricky". He says the architecture could shift again if one large model gives way to many small ones under an orchestrator, and warns that hallucinations mean users still need expertise.

2024-09-10 · open the original ↗

• Developers should keep the web controlled by its users, not corporations. Stahl says the issue is not corporate control as such but control by someone other than us; on Mastodon, open-source W3C-based software anyone can host, each instance states its own moderation policy, and a user who dislikes it can move all their content and social graph to another instance or their own. He urges POSSE (publish on your own site, syndicate elsewhere), suggesting self-hosting Mastodon from ready-made container instances on Azure under your own domain, so every piece of content has a canonical URL; Giard adds that the $150 monthly Azure credit in an MSDN subscription would cover it.
• ActivityPub can describe any stream of activities, not just social posts. Stahl describes ActivityPub, a W3C spec he puts at version 2.0, as actor–activity–entity messages ("David posted a photo", "Barry liked it") that are easy to extend, with ActivityStreams, in his account, defining the APIs; a Mastodon instance is an implementation of those APIs plus a data store. He floats, without being sure it makes sense, publishing his car company's field advocates' activities (such as fuelling a vehicle) in that format and showing them on a big board through a Mastodon client.
• Fediverse apps interoperate, so cross-posting to every service becomes unnecessary. Mastodon (microblogging), PixelFed (photos) and WriteFreely (blogs) speak the same protocol, so a Mastodon user can follow a WriteFreely blog directly. Stahl says his static GitHub Pages blog could post to his Mastodon account as a GitHub Actions step, or have its own presence (Cognitive Inheritance) that his account follows and boosts; Giard, who copies and pastes each post to Mastodon, Facebook and LinkedIn by hand, summarizes this as push and pull models.
• The Fediverse could save democracy because nothing rewards outrage. Stahl argues engagement-driven corporate platforms profit from anger and push users to extremes and misinformation, while Fediverse feeds are chronological, with no algorithm, no universal policy and no single owner who can boost or silence voices. Governments cannot remove it either: a group whose instance sat on Afghanistan's .af domain was banned by the Taliban, who control the DNS, but could move the instance elsewhere, which he says suits organizing direct action.
• It probably won't, through inertia and missing money, unless technologists act. Stahl calls this a warning rather than a prediction: network effects favour incumbents, and though each upset (Brazil blocking Twitter after its owner refused to appoint a legal representative there) brings an influx, he has an order of magnitude more contacts on Twitter, Facebook and LinkedIn than on Mastodon, and the software is largely a hobby with no financial incentive. He asks listeners to contribute code, reviews, documentation, moderation or money, to build on the open APIs as developers did on Twitter's before it closed (his own idea is an activity-stream bot to "radiate intent" to coworkers privately), and points to a governance report, published "last month", backed by the Digital Infrastructure Insights Fund, that studied which Fediverse governance policies work.

2024-09-09 · open the original ↗

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

2024-09-05 · open the original ↗

• AI serves human desires and cannot supply them. Washington, a .NET programmer, Microsoft MVP and author of books on AI, says LLMs seem alive only because they manipulate language and humans think in language. His grand-niece at one year old, who wanted her block back, was "so much more advanced" than ChatGPT. AI will not take jobs, but "the people who use AI may", like a person with a sewing machine.
• Whole books from AI fail; single paragraphs succeed. Asked to write a whole book, the model produces work he grades F, against Hanselman's B-minus, and feeding in two books to get a third "will never ever... work". His free, open-source AI Story Builders writes one paragraph at a time that he then edits. He reports about 1,200 words in a 15-minute session, which he says was not possible before.
• Tight grounding in each prompt keeps paragraphs on the rails. Without it, a flashback to a five-year-old has "little Timmy" driving to an appointment. The app uses retrieval-augmented generation to put character attributes, age at that point, location, season and weather into the prompt. It also extracts new characters and attributes from paragraphs he writes, but he stresses a human still builds the database.
• Fine-tune for behaviour, never to add content. In what he calls his own opinion, which others dispute, new information belongs in RAG, while fine-tuning should only change behaviour such as writing style. After about three chapters, a button in the app produces a fine-tuning file. He uploads it to OpenAI and gets back a model in about an hour that writes closer to his style.
• Plain text files are enough to store embeddings. To keep the app free without paying for databases, he saves embeddings as text files. The Blazor MAUI app in the Windows store keeps them in app storage. The Blazor WebAssembly version zips them into browser local storage, and both run the same code.

2024-09-04 · open the original ↗

• Pull-request reviews serve two ends: quality gating and knowledge sharing. Dixon, a director of web and mobile development at the agency VML, says the gate keeps changes within a project's architectural constraints. Her examples are rejecting an unmaintained third-party library where the rule requires maintained, unit-tested ones, and catching a React dependency array that triggers an infinite loop. She prefers "better practice" to "best practice", because best is relative.
• Reviewing active pull requests is how she onboards fast. She says developers typically take six to eight months to know a codebase, but her agency contracts run three to six months. She reads the pull requests because that is the code being actively worked on, and that is how she finds things like helper functions. She did not write the codebase of over 30,000 lines that her team asks her about; she read as much of it as she could.
• Skipped reviews cost more time than they save. She says a pull request should rarely live beyond about 48 hours, and longer ones point to process problems. Her example is a car-service scheduling client with three entry points: the dealer's website, the owner's portal and an app, including in-vehicle. They share React components, so a size change made for the in-vehicle or phone screen silently broke desktop. She counts work that breaks production as not done. She also wants changes small: about 10 files at most, two or three for a single component, because 88 changed files take her a long time to review.
• Reviews go faster with a team mindset and agreed norms. On one team she introduced an experiment in which everyone unblocks the team first thing each morning, before stand-up across time zones, by reviewing open code. She wants a few designated reviewers who rotate, so the burden does not fall on the lead or architect. She asks whether a solution is good enough within the project's constraints rather than her favourite, and links to the existing helper rather than just rejecting a rewrite. Authors need not accept every suggestion and should discuss them. The other reasons people skip reviews, she says, are reviewing outside their own stack and juniors afraid to review seniors.
• Her review routine starts from the ticket and reads code bottom-up. She uses a PR template that links the ticket ("if there's not a ticket for it, why are we working on it?") and checks that the change solves that scope. She may pull the branch and run it locally, since "works on my box" is not acceptable for enterprise customer-facing sites. She opens the smallest files first and ticks GitHub's "viewed" box, then reads from the bottom up, a habit from English essays, so her brain does not fill in missing words. She traces logic such as a continue button's validation and checks accessibility, such as close buttons without labels and carousels without arrows. She also checks that the design system's styles and the existing custom button are used rather than new ones.

2024-09-02 · open the original ↗

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