Souk's Notes - Unofficial Summaries
Scott HanselmanBuilding The Spoiler Talk app with Charnelle Asante
synopsis
Charnelle Asante, a registered nurse and founder of Spoiler Talk, ends the conversation two weeks past a relaunch, with 440 users, a developer in Ghana, a newly hired marketer and no intention of slowing down. Her position is that the idea kept proving itself every year she nearly quit, and that a smaller, safe community is acceptable even as she expects it to grow large.
Six years to launch. She had the idea in 2018 and, after nursing school in 2019, saved for it by working 70-plus hours a week. A team of about 10 in India built the first version, cost more than she expected, and left an app with problems she could no longer afford to fix. Developers she later showed the code to could not follow it, so in 2023 she started over with a developer found through her brother in Ghana.
Choosing and trusting a developer. With no coding background, she let seven candidates read the repository and picked the one who was confident, had a plan and seemed to care. She says caring mattered as much as skill, and she defers to him on technical choices such as React Native.
Getting through Apple. Apple rejected the app repeatedly, raising one issue per submission; the last was the reporting, blocking and filtering a social app must have. A phone call with Apple resolved it.
What the product is. Discussion is tied to a specific book, show, film or club, so people who have not seen it stay out. She ignores competitors, plans author pages and support for independent authors, and is building moderation tools. Books come from an API; TV and films are entered manually for now.
Scale and growth. Hanselman puts the "thousand true fans" idea to her; she goes back and forth between a boutique app and an eventual buyout by an entertainment media company. Growth so far is from building in public to close to 5,000 TikTok followers, many from BookTok, with a target of converting 10%. She is self-funded and waiting on a grant.
- The app survived a failed first build and a full rewrite. Asante, a registered nurse who now manages a healthcare clinic, had the idea in 2018 and funded it by working 70-plus hours a week as a nurse; a pricey team of about 10 in India built version one, then she ran out of money as it developed problems. In 2023 she found a developer in Ghana through her brother, and they rebuilt from scratch after several developers who saw the old code said they could not follow it.
- A non-coder should pick the developer who cares, not just codes. She gave seven candidates access to the repository; she chose the one who was confident about what he saw, had a plan and seemed genuinely invested, where she felt the earlier team was "just slinging code". She lets him make the technical calls (the app is React Native on iOS and Android).
- App Store review rejects one problem at a time; ask for a call. Apple, stricter than Google Play, kept rejecting the app, the last time because a social app must let users report, block, hide or filter others. She says a phone call with a human at Apple explained everything and the app was accepted soon after.
- She wants a safe, structured space more than sheer scale. Conversation is organised around a specific book, film, show or club, so someone avoiding spoilers simply stays out of that room; she builds to her own vision rather than copying Letterboxd or Goodreads, and an admin panel for moderation is in progress. She wavers between a "boutique" community and expecting an entertainment media company to buy it out; Hanselman offered his "thousand true fans" framing (a thousand fans at $100 a year is a $100,000 business).
- Early traction comes from building in public on TikTok. At recording, two weeks after launch, the app had 440 users across iOS and Android, after about 200 in beta; she has close to 5,000 TikTok followers, about half from BookTok, and targets converting 10%. She is bootstrapping and waiting on a grant, has just hired a content marketing specialist in Nigeria, and courts independent authors; books arrive through an API refreshed every 24 hours, while TV and films are still added by hand.
Crosstalk SolutionsThe Return of Dr. Frankenstein's Dream Machine!
- A UK Ultra board fits inside the Dream Machine's drive tray. The host's team mounted a stripped UniFi UK Ultra (the "Swiss army knife" access point) on a 3D-printed plate in the Dream Machine SE's hard drive tray, powered by a short patch cable to a front PoE port; STL files are linked. The trade-off is losing the drive bay for UniFi Protect.
- The second build packs router, access point and camera into 1U. This version keeps the drive tray, wedges another UK Ultra beside the power supply with RP-SMA connectors extended to the front bezel for external antennas, and wires a dissected G5 Flex camera to look out from the front panel, IR night vision included.
- Internal power came from hardwiring two of the switch's PoE ports. The team pulled the modular connectors behind switch ports 1 and 2 and hardwired patch cables to the pins, making those ports unusable from the front. UniFi still shows both devices normally, the UK Ultra at 1 Gbps and the G5 Flex at Fast Ethernet.
- The host frames it as a Halloween stunt, not a recommendation. He says Ubiquiti would not condone voiding warranties this way, notes he bought over $1,000 of hardware for the video, and gives no reason for the builds beyond seeing whether it could be done.
Technology and FriendsChristina Aldan on Changing Careers [EPISODE 830]
- Career transitions strain mental health, so get two mentors. Aldan, who sits on the boards of two mental health nonprofits, says transitions are the most stressful periods of life, including layoffs and promotions from developer to manager without training. She urges a career mentor for networking and resume review and a technical mentor for getting unstuck, and says mentors push people to apply when only, say, 90% qualified. Giard raised women's reluctance to apply without meeting every requirement; Aldan says peer-reviewed studies repeatedly find it, and adds cultural pressure for perfection among the Asian Pacific Islander students she mentors for free.
- Knowing your core values tells you which jobs to refuse. Aldan turned down an entry-level front-end job in Las Vegas 17 years ago when the sites turned out to be pornography sites, and she says values also settle questions like working in the legal marijuana industry. Giard gives his own case: he left an e-commerce build because it sold cigarettes behind an age checkbox a child could tick. Aldan says enforcing such a boundary also builds confidence.
- Protect sleep and health during a job search, or interviews suffer. Aldan recommends at least seven hours of sleep, nourishing food, walks, coffee with a friend and time in nature, so that your nervous system is calm when the interview comes. Searching around the clock on four cups of coffee, she says, makes you interview out of desperation and exhaustion. Both agree that rejection is routine, often comes with no reason given, and does not mean you did anything wrong.
- Rehearsing a hard conversation makes the next one easier. Aldan and her LinkedIn Live co-host Leslie Martinich run a workshop with role-play of difficult conversations and salary negotiation. Her reasoning is that once the neural pathway has been created it only needs reinforcing, so each hard interview leaves you better prepared rather than more afraid. She reminds people they have "survived 100% of your worst days so far."
- Aldan is leaving tech-adjacent consulting to train in myofascial release. Her brand strategy clients, mostly mid-sized tech companies and startups, closed or lost funding during lockdown. After a seminar and retreats of one month and 10 days in John Barnes's style of myofascial release, she was told the advanced classes require a licence such as massage therapist, so she is in massage school four days a week. She says losing an identity built on 18–19 years of marketing work is hard, and she steadies herself with her core values: contribution, growth and community.
Scott HanselmanEPISODE 3 - Scott & Mark Learn To... Use AI and Know AI Limitations
synopsis
Mark Russinovich ends where he starts: AI is a large productivity gain he uses every day, but hallucination and jailbreaks are built into how these models work, so the people deploying them have to design for those failures rather than hope them away. His forthcoming LLM fundamentals paper for Communications of the ACM covers the same ground.
Using AI without knowing the language. Russinovich writes Python and PyTorch with Copilot and says he needs only enough Python to work efficiently. Hanselman, who rubber-ducks with models, asks whether this rots the brain; Russinovich relays that professors now teach fundamentals first and then allow the tools.
Where tone comes from. Claude saying "certainly" against its own published system prompt is, in his account, baked in by alignment training. Grok, which he calls unaligned on safety, shows each vendor deciding what its model will say.
Hallucination. He explains next-token sampling: wrong answers seen in training keep some probability, and an early wrong word pushes the model to justify it, which Hanselman pictures as a marble falling into a gravity well. Longer coherent chats now come from more data, multi-turn training and better alignment. His reliable triggers are a made-up "Russinovich's 10 immutable laws" and requests for citations, which fail even with web search. Omissions count, and groundedness checks catch only one class.
Jailbreaks. His Crescendo technique walks a model from the Finnish resistance to Molotov cocktail instructions using its own answers. Hanselman compares it to badgering a witness.
Guardrails. Hanselman argues barriers belong where speeds are high and that decisions are being made by companies behind opaque models; Russinovich says deployers must know whether their risk is a quiet street or the autobahn.
- Know a language only as well as efficiency requires. Russinovich says he can no longer code without AI and writes a lot of Python and PyTorch for AI research without knowing Python well, learning more only when relying on the model stops being the efficient route. Hanselman likens the worry to older fears that syntax highlighting, IntelliSense and calculators would "rot your brain"; Russinovich relays that university professors have settled on students learning fundamentals first, then leaning on tools such as GitHub Copilot.
- A model's tone comes from post-training, not the underlying technology. Russinovich attributes Claude's habit of saying "certainly", which the published Claude 3.5 Sonnet system prompt of July 12 tells it to avoid, to examples in its alignment training. He calls Grok the only GPT-4-class model without safety alignment, and says each model maker deciding what its model will and will not say risks a fractured landscape.
- Hallucination is inherent and cannot be driven to zero. He explains it from autoregressive next-token sampling: trolling in training data gives a wrong answer some probability, and a model that opens with "no" will invent a justification for it; omitting key facts from a summary counts too. Reliable triggers he gives are asking for "Russinovich's 10 immutable laws" (they are Scott Culp's laws of security) and asking for a paper reference, which fails even with web search; groundedness-checking APIs in Azure AI Studio remove only one class of error.
- Jailbreaks steer the model with its own earlier answers. His Crescendo technique asks what weapons the Finnish resistance made, then how they made them, getting Molotov cocktail instructions without ever naming the request. Hanselman compares it to a prosecutor leading a witness and to journalists who coerce a model and then report what it said.
- Guardrails should match the stakes of the deployment. Hanselman argues for barriers where traffic is fast, not in quiet neighbourhoods, and says no one would put a VM on the open internet without a firewall; Russinovich says anyone deploying an LLM must know which kind of risk they are running. His LLM fundamentals paper, covering hallucination, indirect prompt injection and jailbreaks, was forthcoming in Communications of the ACM at recording.
Polis Pandit$If Kamala Harris Loses, These Are the 3 Main Reasons Why
- This post includes paid content that is not available publicly. See the original source.
GCastCreating Azure AI Services [GCast 187]
- Azure AI Services is the renamed Cognitive Services. Giard says Microsoft rebranded Cognitive Services as Azure AI Services "not long ago"; the categories include speech, search, vision and language, among others.
- A single-category service needs only the Basics tab. Giard creates a Speech service from the Azure portal's Create a Resource button, filling in a resource group, region (East US), name and the standard pricing tier; networking, identity and tags are optional. The Computer Vision service uses a near-identical dialog, plus a checkbox acknowledging the terms.
- Each single service gives a REST endpoint and its own studio. The resource exposes a REST API endpoint called with one of its keys, and a link to Speech Studio or Vision Studio, with playgrounds and low-code or no-code tools such as real-time speech-to-text.
- One "Azure AI services" resource covers all the categories. Searching for Azure AI services turns up several similarly named entries, and Giard points to the one with a specific icon; creation took about 35 seconds. Its overview blade has a tab per service (Speech, Content Safety, Computer Vision, Language, Translator and more), each with its own endpoints. Speech has three endpoints: speech-to-text, text-to-speech and custom voice. A link opens Azure AI Studio, which has the same playground tools.
- Giard's rule: one service individually, several through the combined resource. He says the combined service makes the most sense when an application calls several services, and a single individual service when only one is needed.
Scott HanselmanThe Xbox Adaptive Joystick with Kaitlyn Jones
synopsis
Kaitlyn Jones, senior product manager for Xbox accessibility, makes the case that adaptive gaming is a kit players assemble for themselves, and that the new $29.99 Adaptive Joystick extends that kit cheaply using standards people already own. She ends on the view that the work will never be finished and should reach every piece of Xbox hardware.
The hub. Hanselman, who used the Adaptive Controller after shoulder surgery, calls it a construction kit. Jones explains its 19 3.5 mm ports, two unmapped wildcard ports that accept a joystick, driverless recognition as a normal controller, and up to three remap profiles saved on the device. Makers have built pillow buttons for children who dislike clicks.
The joystick. It has one thumbstick and six buttons, is ambidextrous, works alone or alongside other controllers, and has a 1/4-20 tripod thread so it can be mounted for chin, foot or wheelchair use. Price was a primary goal because the community already pays for many assistive devices.
Toppers. Xbox Design Lab now generates free STL files for custom thumbstick toppers across the joystick, Elite and standard controllers; Elite support was a priority because many players with disabilities use it.
How she got here. As an occupational therapist she found games more engaging than standard exercises, and her father's nonprofit, Warfighter Engaged, hacked controllers for veterans. A Microsoft engineer reached out, hackathons followed, and that work fed the Adaptive Controller. Her father is stepping back at year's end.
What comes next. Jones sees no end state, wants software such as head tracking and voice integrated with the platform (Hanselman discloses an investment in Cephable), and points to Reddit, Makers Making Change and Logitech's adaptive kit. Both argue ageing players need gentler modes as well.
- The Adaptive Controller is a hub, not a finished controller. Jones, a senior product manager for Xbox accessibility, describes 19 labelled 3.5 mm ports that externalise every standard button, plus two unmapped "wildcard" ports, X1 and X2, that can take a 10K potentiometer joystick. It appears to a machine as an ordinary controller with no drivers and stores up to three remap profiles on the device itself, so they travel to a friend's console; Hanselman, who used one after shoulder surgery, calls it "a controller construction kit".
- The new Adaptive Joystick reuses standards people already have. It is a one-thumbstick, ambidextrous device with six buttons that works alone on PC or console, with Controller Assist (formerly Copilot) or plugged into the Adaptive Controller. Its 1/4-20 tripod thread lets it be mounted for chin, foot or wheelchair use, the same logic as the headphone-jack ports; due early next year at recording, it is priced at $29.99 because users already spend heavily on medical and assistive devices.
- Adaptive Toppers let players 3D-print their own thumbstick grips. Announced in August through Xbox Design Lab, the tool offers about six topper shapes, adjustable height and width, and a free STL file for the Adaptive Joystick, the Elite or a standard Series X|S controller. Jones says Elite support mattered most because many players with disabilities use it; Hanselman notes prints take minutes and a library can make them.
- The product came from outside hackers brought into Microsoft. Jones was an occupational therapist who used games in treatment, and her father's nonprofit, Warfighter Engaged, an all-volunteer 501(c)(3), soldered adapted controllers for veterans with limb loss. A Microsoft engineer saw its Twitter feed, hackathons followed, and those partnerships fed into the Adaptive Controller; her father plans to stop custom designs at the end of the year, hoping commercial products like the joystick fill the gap.
- Accessibility work has no end point and is spreading to standard hardware. Jones wants inclusive features in every future Xbox device and hopes tools combining head tracking and voice can integrate with the platform; Hanselman discloses he is an investor in one such tool, Cephable, free for individual use. She cites Logitech's adaptive kit, with 12 buttons and boards for $100 against about $60 for a single assistive button before it, and both argue that ageing players need easier modes too.
GeoHussarChina Produces over 3.5 million stem grads every year, the unitedstates produces 800,000 to 900,000.
- China out-trains the US in STEM by sheer volume. He cites over 3.5 million Chinese STEM graduates a year, more than four times the US figure; 2 million in science and engineering in 2020; 1.4 million engineering graduates in 2021, six times the US. He also says a Chinese eighth grader's math matches an American high-school senior's, and puts China's average work week at 48 hours against 36.4 in the US.
- America still leads because it imports talent and rewards dissent. He gives three reasons: the US takes in about 100,000 of the world's best STEM graduates each year, still has the best universities (the Chinese still send students there), and its R&D market pays for revolutionary ideas. He argues innovation requires questioning authority, which an authoritarian system suppresses.
- China's own landmark skyscrapers show it trusts Western quality. He says the Shanghai Tower, China's tallest building, was designed by the American firm Gensler, as were its second, fourth, fifth, sixth, seventh and tenth tallest. He reads this as China relying on Western firms for its symbols of prominence.
- He warns the US edge may not last. His opinion is that US public education is poor and that a religious movement is "keeping students dumb" in Republican-leaning states, citing science denial on evolution, climate change and vaccines, which he says China's system does not tolerate.
Crosstalk SolutionsPeplink SpeedFusion: Bond Starlink with this Internet Game-Changer!
- Bonding beats load balancing because sessions survive a link failure. Peplink sponsored this video, and Crosstalk is a Peplink gold partner that sells and installs it. The host explains that load balancing pins each client to one link, capping its throughput and dropping calls if that link dies; SpeedFusion splits each session's packets across all links through an encrypted VPN to an endpoint that reassembles them.
- Smoothing trades extra bandwidth for loss-free live streams. It sends duplicate packets over separate links, so a packet lost on one arrives on another. The host cites a client streaming sports events from a portable Peplink kit with several Starlinks and LTE backup, and "hot failover" keeping VoIP or Zoom calls alive when one link drops.
- Setup through SpeedFusion Connect takes a few clicks. On a Peplink B One 5G (MSRP $599; a B One without 5G is $299), the host picked Seattle, San Francisco and San Jose endpoints from a latency-based suggestion list, avoiding the suggested Vancouver to keep traffic in the US. Dragging WANs between priority levels switches between load balancing and failover, and Wi-Fi can be used as a WAN.
- Bonded throughput is capped well below the router's WAN speed by default. The B One 5G routes 1 Gbps without SpeedFusion, but by default bonded links reach 400 Mbps aggregated unencrypted and 200 Mbps encrypted; paid Peplink plans raise throughput and data allowance. The host's Starlink Mini typically tests 50–150 Mbps down and his Gen 3 80–300 Mbps, and he says bonding evens out that fluctuation, upload especially.
- Crosstalk has deployed Peplink in remote and mission-critical sites. The host describes an NDA-bound TV production in remote Northern Canada given a PACE plan: dual Starlinks, Iridium satellite, LTE SIMs from several carriers and local wireless ISPs. He also says Peplink has bonded up to 12 Starlinks on Royal Caribbean cruise ships using a Balance 2500 EC.
GeoHussareconomy headed into a new supercycle in 2025? #econ #inflation #global #interestrates #finance
- Economists foresee a US supercycle with growth and higher inflation. He relays unnamed economists' claim that a sustained expansion lasting longer than a normal business cycle may be coming, making the last 15 years look "boring": strong GDP growth and high employment, but also higher oil, gas and commodity prices and higher inflation. Their prediction is that interest rates will not fall below roughly 3–3.5%, amid instability driven by Russia and the Middle East.
- He expects infrastructure spending to be justified as national security. His own guess, by analogy to the 1950s, when record infrastructure funding was won on national-security grounds, is that industrial manufacturing and infrastructure will be increasingly tied to security interests, as with the CHIPS Act and the domestic chipmaking push.
GeoHussar#Germany steers western #europe economy.And it’s in trouble #europe #russia #energy #eu #european
- Germany is Europe's most productive manufacturer per head. He calls it the world's fourth-largest economy and fourth-largest manufacturer, with about $110,000 of manufacturing output per person versus $8,000 for the US, $6,500 for Japan and $3,100 for China, whose flat output he calls a bad sign.
- Germany generates far more GDP than its population or resources suggest. Per his chart, it produces 22% of the whole continent's GDP with 15.25% of the population and 5% of natural resources, while Russia, with over a quarter of the population and half the resources, contributes 8.7%. The clip cuts off as he turns to Germany's troubles.
GeoHussar#greenscreen #Germany steers the western #european economy. And its in trouble
- Germany is Europe's most productive manufacturer per head. He calls it the world's fourth-largest economy and fourth-largest manufacturer, with about $110,000 of manufacturing output per person versus $8,000 for the US, $6,500 for Japan and $3,100 for China. Per his chart it produces 22% of the continent's GDP with 15.25% of the population and 5% of resources; Russia, with over a quarter of the population and half the resources, contributes 8.7%.
- The German economy has stagnated and may enter recession. He puts GDP at about $4.45 trillion, stagnant or contracting since COVID, and says German economists expect a contraction by year-end; two consecutive quarterly contractions would mean recession.
- Germany's fiscal position is sounder than America's. He cites debt-to-GDP of 63.6%, far less COVID-era money printing than the US, and inflation at 1.6%, though he warns a deeper downturn could tip it into deflation.
- Chinese competition, weak demand, energy and demographics are the drags. He names cheap Chinese goods flooding markets, weak global demand hitting an economy he says is 20% manufacturing, energy disruption from over-reliance on Russian oil and gas after the war, and Europe-wide population decline.
- Some economists expect a German recovery in 2025. He relays this forecast and infers that a German recovery would lift the whole EU, while leaving the outcome open.
Polis Pandit$Our Critical Thinking and Democracy Crisis
- This post includes paid content that is not available publicly. See the original source.
GeoHussarThe #usnavy just picked up a new trick that terrifies #russia #china #iran and their buddies #navy
- The US Navy can now reload missiles at sea. He reports a tested at-sea missile reload, done in sea state 4 (6-foot seas, 20 mph winds), handling a 25-foot missile canister. Previously ships out of missiles had to return to port, sometimes thousands of miles away, since side-by-side replenishment covered only food and fuel. The clip ends mid-sentence.
GeoHussarThe #usnavy just picked up a new trick that terrifies #russia #china #iran and their buddies #navy
- The US Navy can now reload missiles at sea in a combat zone. He reports a successful test in sea state 4 (6-foot seas, 20 mph winds) with a 25-foot missile canister. Previously ships out of missiles had to return to port, sometimes thousands of miles away; side-by-side replenishment covered only food and fuel.
- Houthi attacks revived a 1990s concept that had been shelved. He says the idea was tried in the 1990s and dropped for lack of demand, then made top priority on January 1 after Houthi attacks on Red Sea shipping. The Navy says fleet-wide rollout comes by 2027; he expects earlier, arguing its estimates are usually conservative.
- He argues sustained fire undercuts China and the Houthis. In a Taiwan scenario, destroyers could keep firing without returning to ports in Japan or South Korea, and the Houthis could no longer wait out ships' resupply trips from their bunkers. He reads this as leaving Iran unable to cover its proxy.
- US logistics, not parades, is the real deterrent. He contrasts Russian, Chinese and Iranian military parades with quiet US logistics, claiming those three have never sustained overseas deployments, while the US has moved 200,000–500,000 troops 7,000–8,000 miles within weeks and kept them supplied for years.