The host, Dave Barger and YouTuber Techno Tim spent three days wiring Cerro Gordo, an 1800s silver-mining ghost town at 9,000 ft in California. They left it with wireless coverage across town, a backup WAN, and a 10 Gbit fiber link 700 ft down a mine shaft. Ubiquiti sponsored all the network equipment, TrueCable donated the outdoor Cat6, and Infinite Cables donated the fiber.
The core moves. The plan put the core on the General Store, but it had no power. The host moved the 4U rack, 1U UPS, UniFi Dream Machine SE and a Rocket Prism with an omnidirectional antenna into the Gordon House instead, and he later judged that the better site. Internet comes from a wireless ISP in Lone Pine through a solar relay pole. The crew left that relay alone ("if it ain't broke") because a Starlink Mini now gives a second WAN.
The Bunk House and the hill. Until then, guests' Wi-Fi had come from an indoor AP hanging out of a window. The Bunk House got a NanoBeam, a U6 Mesh Pro outside and a U6 Plus inside. It was also the only place with line of sight past the tailings pile, so the crew ran 300 ft of shielded direct-burial Cat6 up the hill to a NanoBeam aimed at the Hoist House. They painted visible gear brown with Plasti Dip so it would blend in.
The mine shaft. The old coax could carry Wi-Fi down the 900 ft shaft but couldn't power cameras or more APs. The crew chose a hybrid: 244 m of armored six-strand OM3 fiber to an Enterprise 8 PoE switch at the 700 level, and pre-terminated Cat6 for the UK-Ultra APs at 100, 200 and 900 ft. A G5 Flex camera watches the water source. Brent, the owner, installed the run about three weeks later, and the switches' SFP+ ports gave a 10 Gbit link, which the host calls overkill.
The smaller sites. At the recording studio, a NanoBeam's PoE-out port powers a mesh AP over one cable. The General Store got a Switch Ultra 60W, a hidden NanoBeam and two G5 Flex cameras, running off an extension cord until permanent power goes in. Brent's house got a NanoBeam and a U6 Mesh Pro bridged with PoE injectors. Dave measured about the ISP's full 50/40 Mbps inside the tin-sided church, so they skipped it and kept the spares.
The owner's view. Brent recalls a 1–2 megabytes-per-second hotspot and driving to Lone Pine to upload videos. He names the hotel and future cabins as the next needs.
Dante Lex, CEO of Onboardbase, argues that secrets management fails because the secure path is less convenient than the insecure one. His product therefore assumes developers will make mistakes and hides secrets from them, while making sharing and leak-catching automatic.
Positioning against Vault. Lex markets Onboardbase openly as a simpler, cheaper HashiCorp Vault. He likes Vault but says it needs dedicated security engineers, and that naming it brings in customers who already know what they want.
Why secrets leak. Lex blames convenience. Developers keep secrets in text files, Notion and Slack, and he has lived through a breach. His design target is a tool as easy as pasting into Slack.
Seeing nothing, sharing deliberately. In his account, secrets are injected in real time and stay invisible to engineers. Service tokens are scoped per environment, device tokens can be revoked, credentials rotate after a set number of uses or days, and shared secrets travel as links that self-delete minutes after viewing. He wants .env files gone.
SecureLog and assumed negligence. The new product scans console logging and React server components, then masks or strips leaked secrets before they reach production. Lex dismisses the worry that it breeds laziness: he says mistakes are certain. Hanselman calls this falling into "the pit of success".
Testing and friction. Lex says each feature gets a week of security checks and that his team uses the product on all its own work. He calls friction the one controllable variable, and claims setup takes five minutes.
Lin Qiao, CEO and co-founder of Fireworks AI, argues that generative AI in business is heading toward small, customized open-weights models. In her view, the hard problem is serving them fast and cheaply, tuned to each customer's workload.
Leaving Meta. After five years taking PyTorch to production at Meta, Qiao saw enterprises such as Walmart, Disney and John Deere lacking the software, hardware and teams to adopt AI. She judged helping the whole industry the bigger impact.
Why inference. Inference scales with end users, training only with researchers. She calls training a poor business because models are superseded weekly and hardware generations arrive several times a year. Latency and cost are what break products, and cost can bankrupt a startup that finds product-market fit.
FireOptimizer. On top of Fireworks' own PyTorch build, kernels and caching, the optimizer works like a database query planner. Each customer gets a personalized quality–latency–cost curve and chooses a point on it. Speculative decoding pairs a small and a large model and pays off only when the small model's hit rate is high. The optimizer aligns both models to the workload.
Safety. Qiao says model providers such as Meta carry most of the guardrail work. Because industries and regions differ, she argues customers need to customize their own safety rules.
Small models. She predicts that models of a billion parameters or fewer, customized, will handle narrow business tasks. Hanselman adds his own preference for small local models over giant ones.
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.
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.
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.