Polis Pandit$Elon Musk’s Free Speech Fallacy and the Liver King Fraud
- This post includes paid content that is not available publicly. See the original source.
Souk's Notes - Unofficial Summaries
Technology and FriendsSara Benhamron on Sustainability in Azure [TECHNOLOGYANDFRIENDS 730]
- Microsoft's targets are too ambitious to meet without its partners. Benhamron, a Microsoft industry cloud architect for sustainability, uses the term to mean carbon measurement under the Greenhouse Gas Protocol. She gives Microsoft's commitments as carbon negative, zero waste and water positive by 2030, and removal of all emissions since its founding by 2050, which she says is why it built tools for its ecosystem.
- Carbon accounting is a data problem spanning the whole value chain. She says emissions must be counted upstream (business travel, commuting, purchased goods, capital goods) and downstream, such as an Xbox drawing power in a customer's home, which requires data sharing with suppliers and customers. Measurement comes first because without a baseline an organization cannot know whether its investments move it toward its goals.
- Sustainability Manager is a Power App acting as an emissions CRM/ERP. Benhamron says Microsoft Cloud for Sustainability is vertical offerings inside Azure, Dynamics and Microsoft 365, not a separate data center, on a sustainability data model in Dataverse. Sustainability Manager has data capture through Power Query and connectors, a calculation service from raw consumption to emissions, and reporting against goals; Azure and Microsoft 365 emissions have their own dashboards, with an Azure emissions API announced at Ignite.
- Customer demand far exceeds supply, because sustainability cuts across every industry. She says she is triple or quadruple booked and that more people are needed in her role. Customers already want to act but do not know how, and account teams that budgeted an hour find customers wanting a whole day; her own job is teaching partners carbon accounting and how to extend the product.
- Unified data must come before Microsoft's other three priority scenarios. Benhamron's four are unifying data intelligence, green IT (cloud migration and workload efficiency, backed by sustainability guidance for the Well-Architected Framework released at Ignite), a transparent value chain, and optimizing operations through energy management and smart facilities. She says the first must be tackled before the rest.
Polis Pandit$The Final Push
- This post includes paid content that is not available publicly. See the original source.
Mr. GlobalHow To Get Oil Wells Free
- Operators give away marginal wells to shed plugging liability. Large companies hold thousands of low- or non-producing wells and are liable for plugging every one, so assigning a well relieves them of that cost; he says nine times out of ten they will hand one over if you call the land department and ask. He describes his own Houston oil company, where he is a partner and vice president, as picking up the "breadcrumbs" the big companies leave.
- You need a company and bonding before taking assignments. He lists an LLC, insurance, and bonding: a P5 in Texas and, he believes, a Form 1006 in Oklahoma, where he also cites a surety bond. Bonds are blanket, covering many wells rather than one each. Candidate wells can be found through well-data services and state regulators' sites such as the Texas Railroad Commission and Oklahoma Corporation Commission, then checked on Google Earth or, more often, in person.
- Plug-and-salvage pays when equipment value clears plugging cost. Plugging costs $30,000–40,000, against a pump jack worth around $30,000, tanks around $10,000 each, tubulars $10,000–15,000 and casing $50,000–70,000. His firm looks for wells with at least $100,000 of salvageable material, and he says selling everything can net $50,000–60,000 per free well.
- Some free wells can be recompleted back into production. His team of geologists reviews geological maps and field production to identify recompletion candidates, and he says the firm has returned several donated wells to good production with very little invested, as part of how it is building the company.
Crosstalk SolutionsWorld's Greatest Pi-hole Tutorial - Easy Raspberry Pi Project!
synopsis
The video walks through building a network-wide ad and malware blocker on a Raspberry Pi, ending with pi-hole blocking roughly 221,000 domains, resolving its own uncached lookups through a locally run Unbound service rather than a third-party DNS provider, and serving every device on the network (including separate guest and IoT subnets) via the router's DHCP settings.
Preparing the hardware. The host notes the project is not resource-intensive — he installs it on a Raspberry Pi 4 with 2GB of RAM, calling even that overkill — and says it runs just as well in a Docker container or VM for anyone who can't find a Pi in stock, pointing to a "Pi locator" tool for stock alerts. He flashes Raspberry Pi OS Lite 64-bit with the Raspberry Pi Imager, pre-configuring a hostname, SSH access, and credentials before writing the card.
Installing pi-hole and choosing block lists. After an apt update/upgrade and setting a static IP (either a DHCP reservation or a manual dhcpcd.conf edit), he runs the pi-hole installer, accepting the default Steven Black unified block list and enabling full query logging. He then supplements that list using the Firebog, a curated index of block lists graded by maintenance quality, taking the top two "green" (safe, well-maintained) lists from each category. Running "update gravity" afterward raises the blocked-domain count from about 158,000 to 221,000. He cautions against adding too many lists, since over-blocking breaks ordinary browsing and tends to get pi-hole disabled rather than tuned.
Serving the whole network. Rather than configuring DNS on each device, he points his DHCP server's DNS settings at the pi-hole's IP (and a second existing pi-hole as backup). Because his network has separate LAN, guest, and IoT subnets, he also changes pi-hole's DNS interface setting from "only allow local requests" to "respond only on interface eth0" so it will answer queries from those other VLANs. He flags that without a firewall rule, a tech-savvy user can bypass pi-hole by manually setting a device's DNS to something else, and recommends blocking outbound port 53 to anywhere but the pi-hole.
Adding Unbound for private resolution. To stop any upstream provider from logging the household's DNS history, he installs Unbound, drops in a configuration file from the pi-hole documentation, and confirms it resolves queries on 127.0.0.1 port 5335 using the "dig" command. He then repoints pi-hole's upstream DNS setting at that local address and port instead of Cloudflare.
Verifying and upkeep. Testing against an ad-blocking test site, he shows pi-hole blocking 93% of tracking/ad hosts, versus 61% when the same device is switched to plain Google DNS. He closes with routine maintenance: running "sudo pihole -up" to update, and a scripted URL for temporarily disabling blocking (which he's wired to a Stream Deck button) for troubleshooting false blocks across multiple pi-holes.
- The host runs Unbound locally instead of an upstream DNS provider. After installing pi-hole with Cloudflare as a temporary upstream, he installs Unbound (config at /etc/unbound/unbound.conf.d/pihole.conf) so it queries root DNS servers directly on 127.0.0.1 port 5335, then points pi-hole's upstream DNS at that custom resolver so no outside party logs the household's DNS lookups.
- He warns against stacking too many block lists, since over-blocking breaks browsing. The default install pulls Steven Black's unified list (~158,000 domains); he adds the top two lists from each category on the Firebog's "green" (well-maintained, low-breakage) tier, raising the total to about 221,000 blocked domains, roughly a 30% increase.
- A static IP must be set before running the pi-hole installer. It can be a router-side DHCP reservation or a manual edit to the pi's dhcpcd.conf; for networks with multiple subnets (he runs a LAN, guest, and IoT network), the DNS setting must also change from "only allow local requests" to "respond only on interface eth0" so pi-hole answers queries from other VLANs.
- Without a firewall rule, savvy users can bypass pi-hole entirely. A device can simply be pointed at another DNS server, such as 1.1.1.1, so the host recommends blocking outbound port 53 traffic to any address except the pi-hole's own IP, though he leaves the exact router configuration out of scope of the tutorial.
Crosstalk SolutionsFlipper Zero LED Light Strip Control #flipperzero #hacking
- A Flipper Zero can clone an LED strip's remote control. The host uses the Sub-GHz "read" function to capture the remote's on and off buttons for the strip behind his server rack, saves them as "rack on" and "rack off", then replays each from the saved list with "emulate"; he says the same works for infrared-controlled LEDs.
Polis Pandit$Tough, Trump, Crypto Thoughts, and Elizabeth Holmes
- This post includes paid content that is not available publicly. See the original source.
Andrej KarpathyBuilding makemore Part 5: Building a WaveNet
synopsis
Starting from the part-3 character-level MLP (three characters of context feeding one hidden layer, validation loss 2.10), he restructures it into a WaveNet-style hierarchical network — progressively fusing pairs of characters across multiple layers instead of squashing all context into one layer at once — and lands, after fixing a BatchNorm bug and scaling up, at a validation loss of 1.993 on the same names dataset, while also converting the codebase to a small PyTorch-like module library (Embedding, Flatten, Sequential) reusable going forward.
PyTorch-ifying the codebase. Before touching the architecture, he refactors the ad hoc layer list into proper module classes — Embedding and Flatten alongside the existing Linear, BatchNorm1d, and Tanh — plus a Sequential container that chains them, mirroring torch.nn's own API. This isn't cosmetic: the forward pass collapses to a single call through the model, exposing a real bug where forgetting to leave BatchNorm in evaluation mode (rather than training mode) caused a single-example forward pass to compute a variance over one number, silently corrupting output — a reminder that BatchNorm's train/eval state is an easy source of hidden bugs.
Baseline: just widening the flat network. Simply increasing context from 3 to 8 characters in the existing one-hidden-layer architecture drops validation loss from 2.10 to 2.02, with no structural change — establishing a baseline against which the hierarchical redesign has to earn its keep.
Building the hierarchical (WaveNet-style) network. Rather than flattening all 8 characters into one long vector immediately, he fuses them progressively: a FlattenConsecutive layer groups every 2 adjacent character embeddings into a single unit, which a linear/BatchNorm/Tanh block processes; this repeats, so 8 characters become 4 pairs, then 2 groups of 4, before a final layer produces the prediction — matching WaveNet's tree-like, dilated causal convolution structure, though implemented purely by reshaping tensors rather than by using real convolutions. He shows PyTorch's matrix multiply already supports this: it operates only on a tensor's last dimension, treating any number of leading dimensions as batch dimensions, so grouping character-pairs into a new "batch-like" dimension is just a view/reshape, not a new operation.
The 3D BatchNorm bug. Feeding the now-3D intermediate tensors (batch × groups × channels) into a BatchNorm1D layer written for 2D input caused it to silently compute separate running statistics per group instead of per channel — reducing only over dimension 0 instead of over both the batch and group dimensions. The fix (reducing jointly over dimensions 0 and 1 when the input is 3D) recovers a proper per-channel normalization and gives a small but real improvement, from 2.029 to 2.022 — and he notes this convention deliberately departs from PyTorch's own BatchNorm1d, which places the channel dimension in the middle rather than last.
Results and what convolutions actually buy you. With the corrected hierarchical architecture and its parameter budget matched to the flat baseline, results are initially a wash (~2.02 either way), but scaling embedding and hidden sizes up (to about 76,000 parameters) pushes validation loss to 1.993. He closes by clarifying that real WaveNet-style convolutions wouldn't change what's computed here at all — they're purely an efficiency mechanism, sliding the same shared filters across a sequence inside fast CUDA kernels and reusing intermediate values shared between overlapping positions, rather than recomputing the whole hierarchy independently at every character position as his implementation does.
- Fusing context characters gradually in a tree, rather than all at once in one layer, improves prediction without adding parameters. Mirroring WaveNet's dilated causal convolutions, he restructures the MLP to fuse two characters per layer, then pairs of those, doubling context each step; holding parameter count fixed at roughly the same budget as a flat network, this hierarchical structure still needed a bug fix before it beat the naive approach, but scaling it up (block size 8, then a bigger hierarchical version) pushed validation loss from 2.10 down to 1.993, versus 2.02 for simply widening a flat one-hidden-layer network to the same context length.
- A subtle BatchNorm bug appeared when the same 2D normalization code was fed 3D tensors, and PyTorch never raised an error. Reducing only over dimension zero left the middle "group" dimension untouched, so the layer silently maintained a separate running mean and variance for each of the 4 groups × 68 channels instead of just 68 channels; averaging over both the batch and group dimensions together fixed it, improving validation loss slightly (2.029 to 2.022) — a reminder that broadcasting can make a wrong shape run without complaint.
- Convolutions are purely an efficiency trick over this same computation, not a different model. Applying the identical stack of layers independently at every position of a name like "DeAndre" (one call per position) computes the same values a convolution would, but redundantly — a convolution instead slides shared weights across the sequence inside fast CUDA kernels and reuses intermediate values that are shared between adjacent output positions, avoiding the recomputation his naive per-position loop would require.
Technology and FriendsChander Dhall on AI in Azure, Google, and AWS [TECHNOLOGYANDFRIENDS 729]
synopsis
Dhall, CEO of a cloud consultancy, ends up arguing that no cloud is best at AI. Microsoft, Google and Amazon each win on different tasks and change from one version to the next, so for AI services, which are just API calls, he would combine at least two clouds and add domain-specific logic.
Why AI projects fail. Dhall cites failure rates of 83–92% and a Gartner prediction that 85% of AI implementations would fail by 2022. He reads these as projects failing, often for non-technical reasons. He says his firm vets feasibility up front and claims a 100% success rate. One customer, he says, spent a few million dollars running a deterministic algorithm in R and then Python, and concluded the cloud did not scale. He says it was never an AI problem and would suit C# or Java.
Handwriting benchmarks. In his firm's May 2020 tests, Microsoft had 12 errors, Google 20 and Amazon 37, though Amazon did not then support handwriting. Google did best on French and was alone in separating two combined images. In an October 2022 retest Google read more words correctly but made a structural error. Weighted 5 for structure against 2 for words, its score got worse.
Forms and receipts. Microsoft Form Recognizer trained in seconds where the others took about an hour, and it alone returned bounding boxes. Google hit 100% accuracy on 10 samples. Giard notes that a tax return justifies waiting where a coupon does not. Dhall keeps a human review step, which he says cuts the job from 200 people to three.
Real-time action and multi-cloud. Dhall describes a bot that matches a rival's discount automatically, with one-click human approval. Multi-cloud costs enterprises 70–80% more when data moves, he says, but not for API calls.
- Most AI projects fail, often for reasons that are not technical. Dhall shows a slide quoting 83–92% failure rates and a Gartner prediction that 85% of AI implementations would fail by 2022, which he reads as projects not delivering rather than models misfiring. He claims his own consultancy has a 100% success rate on AI projects because it tests feasibility and sets expectations up front. Giard adds that on his own AI projects the customer was told at the start that the data might not support any prediction.
- No cloud won outright on handwriting recognition, and results drift. In Dhall's firm's May 2020 tests, Microsoft had 12 errors overall, Google 20 and Amazon 37, though Amazon did not then support handwriting. On neat cursive Google had 15 times Microsoft's errors, on French Microsoft had about five times Google's, and only Google recognized two pasted-together images as separate. In an October 2022 retest Google's word accuracy improved but it made a structural error, so under his weighting (5 for structure, 2 for word accuracy) it scored worse, and he says users have to keep track of where their cloud is going.
- Pick a form-recognition service by whether speed or accuracy matters. In May 2020 Microsoft Form Recognizer trained in about five seconds where Amazon Textract and Google Document AI took about an hour, and it was then the only one returning bounding boxes with the OCR. Google reached 100% accuracy on 10 samples. Dhall's firm still adds a human review UI, which he says turns a 200-person job into a three-person one. On receipts, none of the three read "lb", and he ranks Amazon above Google because Google made three structural errors to Amazon's one.
- The goal is real-time action, not just real-time analysis. In Dhall's hypothetical, a fast-food chain's bot reads a rival's 20%-off ad with computer vision and automatically posts a deal 5% better. Where approval is required, he would have everything ready for a human to approve with one click.
- Using more than one cloud for AI APIs is "a no-brainer." Dhall says enterprises expect multi-cloud to cost 70–80% more when data has to move, but an AI service is just an API call, so at least two of the three clouds should be used. He would pipeline them and add domain-specific rules, such as correcting "1B" to "lb". Customers barred from the cloud by compliance may have to train their own models instead.
Crosstalk SolutionsDo This Now! Yubikey + Google U2F Setup - EASY!
- U2F hardware keys are the strongest second factor against phishing. The host says U2F cuts the chance of being phished to "practically zero", while OTP apps such as Google Authenticator or Authy, though far better than nothing, can still be phished. SMS is the least secure option, yet Google's two-step setup offers the account's phone number by default, so choose "show more options", then "security key".
- Register several keys and name each one descriptively. The host registers three, two YubiKey 5 NFC and a YubiKey 5Ci (on a USB-C to USB-A adapter), each named for its skin, so a lost key can be removed from the account without guessing. He says any U2F-capable key works the same way, even a Flipper Zero.
- Keep the ten backup codes, and skip weaker fallback methods. Printed or downloaded and stored safely, the single-use codes still let you in if every key is lost. Adding an authenticator app, voice or text message as further second factors reintroduces phishable routes, so the host advises keeping the account as locked down as possible.
- NFC keys make phone sign-in a tap rather than a plug-in. On the host's iPhone, Gmail's two-step prompt was cleared by holding the YubiKey 5 NFC to the back of the phone. The 5Ci's Lightning connector would also work, but he found NFC more efficient.
- Google domain admins should require two-step verification, then go further. Requiring it beats password-only sign-in, the host says, and for the next level he points administrators to Google's Advanced Protection Program, which requires U2F.
Crosstalk SolutionsNice PTMP Setup at an RV park. #wireless #networking
- A powered pole with a PtMP receiver suits a remote park corner. At the far end of an RV park, the host points to a pole carrying power, a light, a box he takes to be a PoE network switch, two cameras and a receiving antenna on top; he calls it not pretty but done right for covering an area far from everything else.
Polis Pandit$Democrats Outperform, Crypto's Lehman Moment, and Ukraine Liberates Kherson
- This post includes paid content that is not available publicly. See the original source.
Technology and FriendsBuck Woody on Database Security [TECHNOLOGYANDFRIENDS 728]
synopsis
Woody's position is that nobody has to do everything to secure a database, but everyone has to do something. That means the basics first, then a zero-trust review of how applications reach the data. He treats security as a shared "tax" that is worth paying, because people make mistakes no technology can prevent.
Why it matters more now. Woody points to newer professionals who were never taught the fundamentals, more data spread across more places, and nation-state attack tools that have leaked to criminals.
The basics. Patch and upgrade, even when 2,500 custom applications make testing hard. He says upgrading from SQL Server 2005 picks up 17 years of security patches. Most attackers try unlocked doors, so strong passwords and encrypting the sensitive columns turn most of them away. Giard frames this as making a break-in cost more than the data is worth.
Defense in depth and zero trust. Defense in depth is the layered goal, and zero trust is the method: verify explicitly and assume breach. Woody uses a neighbour collecting his car keys to separate identity, authentication and authorization.
Applications and accounts. He says apps that use one super-user login hand everything to whoever breaches it, and that almost every user list still holds former staff.
The course and the Uber story. Woody's free, self-paced course is on GitHub through aka.ms/sqlworkshops. He also teaches it module by module at a Tampa SQL Server user group. He closes on the Uber breach, which went from a phished MFA code to a trojan sent to IT. That, he says, is "wetware", not software.
- Database security fundamentals are slipping as the threat grows. Woody says newer data professionals have had more done for them and are stretched across BI, machine learning and AI, while data now sits in far more places and sophisticated nation-state attack tools have leaked to cybercriminals.
- Do at least the basics; you need not do everything. Woody says most attackers are script kiddies trying doors, so patching, upgrading and strong, changed passwords will send them to the next target, though a nation-state that wants your data will likely get it. He cites one shop with 2,500 custom applications against one server, which made testing a patch hard, and a customer still being upgraded from SQL Server 2005, 17 years of security patches behind; he would encrypt sensitive columns rather than the whole database. Giard frames it as making the cost of a breach exceed the data's value, and offers his own guess that cheap measures protect 90% of the data.
- Zero trust is the "how" of defense in depth. Woody says defense in depth layers protection from the building and OS (down to driver patches) to row, column, at-rest and in-transit encryption built into SQL Server, and zero trust says how: verify explicitly at each step, and assume breach, for example by checking logs for logins from countries where you have no users. His free, self-paced, hands-on course on GitHub, aka.ms/SQLSecurity101 (all his courses at aka.ms/SQLWorkshops), carries a defense-in-depth checklist and ends with a post-breach after-action report; he also runs it module by module monthly at the Tampa SQL Server user group.
- Applications that log in as one super-user are the worst pattern. Woody says it is very common for an app to keep its own users and permissions and act in the database through one privileged login, so whoever gets that password has everything. Nearly every customer environment he reviews still lists users who left, one five years earlier.
- People, not technology, are the weakest link. In Woody's account of the Uber breach, attackers texted an employee posing as IT, got her MFA code, then sent IT a screenshot carrying a Trojan that spread through an administrator to the whole company — social engineering, not tech. He calls the state of database security "good but in peril", because security is a tax that does not sell products and IT staff are being laid off.
Crosstalk SolutionsFlipper Zero vs. UniFi Access - Don't Make This NFC Mistake!!
- Official UniFi Access cards are far more capable than cheap NFC tags. Per the data sheet, the starter kit's 20 cards are ISO 14443 Type A MIFARE DESFire EV2 with 4 KB of memory, room for a unique ID plus a cryptographic key, and the "UniFi Access proprietary security protocol"; they cost $2–4 on Amazon. NTAG215 tags (as in Nintendo amiibo) hold 504 bytes, no room for keys, and cost 35–65 cents.
- UniFi Access accepts almost any NFC credential, cheap tags included. On his test setup the host enrolled and opened the door with an official card, a YubiKey and an NTAG215 tag, but warns that being able to use any NFC type does not make it advisable.
- A Flipper Zero clones an NTAG215 tag and opens the door. Reading the official card and the YubiKey gave an "unknown ISO tag" whose UID emulation did nothing; the NTAG215 was identified, emulated and granted access. The host's advice: never use cheap NFC tags for access control.
- The host argues Ubiquiti should reject insecure cards outright. He asks why Access cannot detect that a tag cannot run its security protocol and either block it or accept only official cards, as Ubiquiti already restricts UniFi Protect to its own cameras. He admits he does not understand the card technology well and asks viewers to explain.
Crosstalk SolutionsIs This a Good Idea?? - UniFi EV Station
- The host expects businesses to need managed workplace EV charging. Over the next five to ten years he expects many more employees charging at work, which raises who pays, how per-employee use is tracked or capped, how outsiders are kept off, and how electric delivery fleets are managed. He thinks that makes a real market for a product like Ubiquiti's EV Station.
- EV Station's announced specs are in line with other smart chargers. It lists at $799 in Ubiquiti's Early Access store, and the host notes EA pricing is subject to change. From the announcement: an SAE J1772 connector (Teslas need an adapter), 50 A charging shown, making it Level 2, a 10.1-inch touchscreen, a 16.4 ft (5 m) cable and a NEMA 3R weatherproof rating.
- It is managed through UniFi Connect with three access modes. The screenshots show plug-and-charge (open to anyone, he assumes), authorized users only, and disabled. The host speculates a price-per-kWh field to bill employees could follow, and spots SSO logins (Amazon, Microsoft, GitHub) and a "UniFi Talk" button he guesses is a softphone in the Connect app.
- He likes the idea but wouldn't bet a parking lot on Ubiquiti. Given Ubiquiti's history of launching products and then abandoning them, he would not spend thousands or tens of thousands filling a lot with EV Stations on the gamble they are still supported in three to five years, and he thinks the company may be ahead of the curve.
Polis Pandit$The Midterms Are Upon Us...
- This post includes paid content that is not available publicly. See the original source.
Scott HanselmanWhat If 2 with Randall Munroe on the Hanselminutes Podcast 865
synopsis
Randall Munroe, author of What If 2 and the webcomic xkcd, ends up describing his work as disciplined curiosity: answer the absurd question quickly, follow the rabbit holes it opens, state plainly how sure you are, and go to whoever actually knows, whether a librarian, a cookbook or an astronaut. He says he still finds it hard to imagine a better job, because xkcd's format lets him draw whatever he is into that day.
Uncertainty as a habit. Munroe, a physics graduate, explains his frequent disclaimers as a way of making uncertainty explicit. He checks a constant in two sources, would check a third if his life depended on it, and wants readers to add their own doubt rather than treat him as an authority. He does the arithmetic in Mathematica, whose unit-aware calculation he calls "the dream", and chooses precision to suit the question: his figure of a couple of millimetres of rock for the weight-loss chapter barely moves whatever definition is used, although small density differences matter when two masses are subtracted.
Rabbit holes and librarians. A typical chapter is one research sprint, often a single day, and then digressions. Niagara Falls through a McDonald's straw comes out at about a third of the speed of light within minutes. Checking the flow rate, though, led to a US–Canada treaty with separate day and night minimums and one compliance officer per country. Six hours on overlapping Great Lakes agency websites found nothing; librarians at the Harvard Law Library, already helping him with another question, got the officers' names and biographies from the commissions themselves.
Experts, footnotes and corrections. He uses the steak-from-orbit chapter of his earlier book to show the method. He rebuilt an old NASA Flash calculator for leading-edge heating, then gave up on heat-flow equations for a cookbook. Unable to model how a raw steak tumbles at hypersonic speed, he left a footnote asking for help. Two University of Manchester doctoral students with spare wind-tunnel time ran the test: the surface seared and ablated, the inside stayed raw, and the result looked far worse than his drawing. He welcomes such corrections, and says fact-checkers catch most algebra slips and later printings fix the rest.
The astronaut and the T-Rex. Interviewing Chris Hadfield for his book How To, Munroe escalated landing scenarios hoping Hadfield would hang up, but he never hesitated. A later text about the What If 2 cover, a T-Rex on a jetliner, drew the same calm: fine near the centre of gravity, like the shuttle-carrier 747; nose down for negative G to dislodge it; and on saving the animal for science, "not my concern".
What a star smells like. A friend's random text led to six hours on olfaction. Patients whose olfactory bulb was in a radiation beam report bleach, burning hair and ozone, which he puts down to ionised mucus. Stellar plasma's free hydrogen ions would taste sour, and a chemist agreed this explains the sour taste reported in a lab criticality accident.
- Every number carries uncertainty, so precision should match the question. Munroe says his disclaimers make his uncertainty explicit: he checks a value in two sources, would check a third "if I was going to trust my life" on it, and does not want to be cited as the authority. His "two millimeters of rock" for the 20-pound weight-loss question is deliberately loose, since sand versus bedrock moves it only between about 2.5 and 3.2 centimetres in a related calculation; he says density choices only become critical when two masses are subtracted.
- Easy answers open rabbit holes, and librarians beat six hours of Googling. Pushing Niagara Falls through a McDonald's straw takes about a third of the speed of light, but checking the flow rate led him to a US–Canada treaty setting daytime and night-time minimums and appointing one compliance officer per country. After hours on overlapping Great Lakes agency websites, the Harvard Law Library's librarians got the officers' names from the commission itself.
- The practical expert often beats the theory. For the steak dropped from orbit, he abandoned engineering heat-flow equations for a cookbook, and footnoted that he did not know how raw steak tumbles at hypersonic speed. Two doctoral students at the University of Manchester later tested a steak in a hypersonic wind tunnel: as he predicted, the outside seared and ablated away, the inside stayed raw, and it looked far more ragged than he drew it.
- Test pilots answer absurd scenarios with flat, correct priorities. Chris Hadfield told him a T-Rex on a jetliner roof is fine if it stays near the centre of gravity, citing the 747 that carried the Space Shuttle; to dislodge it he would push the nose down for negative G, dive or roll. Asked about saving the dinosaur for science he replied "not my concern", which Munroe ties to the pilots' order of aviate, navigate, communicate.
- A star would smell like ozone and taste sour. Munroe draws this from cancer patients whose olfactory bulb was in a radiation beam and reported bleach, burning hair and ozone smells, likely from ionised mucus; stellar plasma's free hydrogen ions would trigger sour receptors. He says, and a chemist he consulted agreed, that this explains the sour taste reported by a victim of a lab criticality accident with a bomb core.
Crosstalk SolutionsHow the Internet Saved Florida - 2 Days with the ITDRC
- ITDRC volunteers restore communications after disasters. An ITDRC leader describes the nonprofit as about 3,300 tech-sector volunteers, set up in 2008 after a government-sponsored post-9/11 effort never took off. About a month after Hurricane Ian it had opened about 130–135 sites in southwest Florida, moving from search-and-rescue support to shelters and community recovery, and had over 80 active US deployments.
- The Fort Myers Beach marina hotspot runs on solar, Starlink and CBRS. The trailer has four 300 W panels and batteries, Starlink as primary and a Meraki CBRS link from a high-rise point-to-multipoint antenna as backup. The open SSID is throttled to 5 Mbps down/5 up with layer-7 filtering against streaming and torrents. Its batteries had stopped charging and were below 10%; a volunteer pulled one battery and used a generator to reach about 50%.
- A community hub got Wi-Fi, 5G, a printer and VoIP. The Pine Island American Legion serves 500–800 meals a day, so ITDRC covered the parking lot and food area for a computer cafe (FEMA forms, email, photos) on a 5G modem. It replaced the flooded printer, added a Polycom VoIP phone, and left a card with a site ID for logging tickets with central control in Texas.
- The host calls Starlink a game changer for disaster response. It recurred at every site: a medical tent with Starlink and a Meraki MR86 (another group withdrew a second dish), and the Cajun Navy's base with extra SSIDs for other organizations. The setups break often: an access point lost power during nearby electrical work and needed a generator, and one site's equipment had vanished.
- ITDRC's mobile command truck combines cellular, satellite and radio interoperability. Mobile Technology Unit 4 has multi-carrier cellular, a 1.8 m VSAT dish as secondary, a 50-ft tower with Wi-Fi and cameras, and 20 roof antenna slots. It carries VHF, UHF, aircraft, marine, 700/800 MHz and 2 m radios, a gateway that cross-patches VHF to 800 MHz with phone patch, and a team awareness kit tracking staff and CRM-fed sites.
Crosstalk SolutionsFlipper Zero vs Tesla Charge Port #tesla #flipperzero
- A saved Flipper Zero signal pops open other people's Tesla charge ports. The host sends a saved "Tesla" Sub-GHz signal in a Costco parking lot and opens the charge ports of two Model Ys, calling it a harmless but fun prank.
Polis Pandit$We're Back
- This post includes paid content that is not available publicly. See the original source.
Crosstalk SolutionsReal World WiFi Install for Eagle Paragliding
- The router was chosen because it also records the cameras. Eagle Paragliding, a Santa Barbara school opening a remote southern Oregon property, needed coverage at a main house, a warehouse and a camping area; Starlink feeds a UDM Pro, picked by the client because it also runs UniFi Protect, fitted with a 4 TB Seagate SkyHawk drive. U6 Lite APs cover interiors, U6 Mesh APs the outdoors, alongside G4 Pro and G4 Bullet cameras.
- Point-to-point wireless bridges tie the three sites together. Two GigaBeam-to-GigaBeam links connect house, warehouse and camping area; the host likes them because the 60 GHz link (about 1.6 Gbps planned) has a 5 GHz fallback he expects to carry 300–400 Mbps in bad weather. On site the links read 82.5% link potential with 1.4 Gbps capacity, and 84.5% with 1.3 Gbps at 70.8 m; he stops at "rock solid" rather than tuning toward 100%.
- The larger switches were chosen for passive PoE, not port count. The GigaBeams need 24 V passive while the APs are 802.3af, and the US-8-150W does both; the warehouse unit ended up using seven of eight ports. The camping area's USW Lite 8 PoE does only 802.3af, so an INS-3AF-I-G converter powers its GigaBeam. Network is 192.168.64.x on DHCP, with the GigaBeams statically set at .10–.13.
- Speed tests came close to full Starlink speed everywhere. The host measured 221/13 Mbps at the main-house driveway, 68.7/11.9 at the warehouse, 80.6/5.52 at the camping area through both GigaBeam hops, and 62.3/8.06 about 50 yards from the camping AP. APs were left at defaults, to be fine-tuned remotely if problems appear; one U6 Mesh for a future building arrived too late and will be shipped.
- The host still recommends UPS units and Ethernet surge protectors. The area gets severe weather and lightning, and the sites had surge strips but no UPS; he wants grounded UPSs for backup and clean power, and grounded Ethernet surge protectors on every outdoor device, of which he brought only one. Crosstalk, which did the design and install, offers both as services.