Network Chuck

Unofficial, machine-generated summaries, as of 2026-11-09.

you need to use Herdr RIGHT NOW!!
  • Herdr is a tmux-like terminal built around AI agents. It runs as a background server, so panes, tabs, workspaces and agents survive a detach (prefix Ctrl+B, then Q) and resume after herdr server stop or a reboot, since agent chats are just text. Chuck says he has used it daily for about three months and has not looked back at tmux. The video is sponsored by HPE Networking, not by Herdr.
  • Herdr tracks each agent's status and alerts you when it changes. Launched agents appear in an agents list labelled by workspace, and sound and in-app toast notifications fire when one finishes or needs you. Ctrl+B G (go to), Chuck's favourite view, filters by done (D), blocked (B), working (W), idle (I) or all (A), and / searches sessions, apparently content included.
  • A second machine's Herdr can be attached as a remote session. It needs Herdr already installed there and SSH access. Chuck ran herdr machine list, then herdr machine add with a label ("Big Mac") and [email protected]. His Mac Studio's sessions then appeared alongside the local ones, notifications included. He calls this a fairly new feature.
  • Agents can drive Herdr themselves through a skill. Chuck says the skill talks to Herdr's API, and he believes there is a CLI too. His Claude Code agent opened and closed panes on request, then launched Codex, Grok, Antigravity and another Claude to code ASCII scenes in parallel. Built-in hooks reported their output and questions back to it. It also set up Git worktrees for three agents on a sample project.
  • Agents on different machines can message each other through remote Herdr. After installing the Herdr skill in a Hermes agent on his laptop, Chuck had it ask Morpheus, his Claude agent on the remote Mac, for a project status update in a named workspace. Morpheus answered, and Chuck calls the exchange "kind of seamless".
this WiFi GPU is wild.....
  • A GPU in a separate box can run a laptop over Wi-Fi. Chuck received a unit holding an Nvidia 5090 that shows up as "GPU 1" on a small ASUS laptop with no cables. He played Cyberpunk on it, with Nvidia-Smi showing the GPU in use and the traffic going over Wi-Fi. He calls it wireless computing infrastructure; the captions garble the acronym.
  • Chuck says it runs large models the laptop otherwise couldn't. He credits this to 4 TB of memory, "but not exactly how you might imagine it". The short doesn't say how that works and points to the full video for the explanation.
local AI's biggest problem... they solved it?
synopsis

Chuck tests a sponsored early engineering model of a wireless GPU box (an RTX 5090, a mini PC and a 4 TB SSD) from a new company whose name the captions garble. It games over Wi-Fi at 92 fps and runs the 750-billion-parameter DeepSeek 4.1 Flash at about eight tokens a second on 32 GB of VRAM. His verdict: it is impressive and should improve.

Gaming over Wi-Fi. Wi-Fi is about 1 Gbit/s at 1.5 ms, against PCIe's 512 Gbit/s at under a microsecond. A fake driver presents the remote 5090 as local. The protocol bundles instructions, sends only changed data, predicts answers and returns compressed video, cutting about 8 Gbit/s to 400 Mbit/s. Some of the team came from cloud gaming.

Big models off an SSD. The model needs over 500 GB of memory. Run from an SSD, it would take minutes per word even as a mixture-of-experts model. Aligned expert chunks raise SSD reads from about 1 to 11.5 GB/s. Popular experts stay in VRAM and RAM, and the CPU and GPU split the math. Chuck's live test showed 3–15 tokens a second.

Sharing the GPU. Any one model can fill the VRAM. A scheduler parks idle models on the SSD, cutting app switches from 17–18 seconds to 4–9 seconds. The box is on sale in 5090 and 5060 Ti versions, shipping in about 8 weeks.

  • A wireless box lends one GPU to every device. The video is sponsored by the company that makes it, a new firm whose name the captions garble. Chuck tested an early engineering model: an RTX 5090, a mini PC running OpenWRT and a 4 TB SSD. Laptops, a Mac and phones can use it like a local GPU. It is on sale now in 5090 and cheaper 5060 Ti versions, shipping in about 8 weeks.
  • The company's own protocol makes Wi-Fi fast enough for gaming. PCIe gives 512 Gbit/s and sub-microsecond replies; Wi-Fi gives about 1 Gbit/s and 1.5 ms. A fake driver makes the laptop think the 5090 is its own. Instructions are bundled, compressed and sent without waiting, only changed data is sent, and answers are predicted. That cuts about 8 Gbit/s to 400 Mbit/s. Chuck got 92 fps in Cyberpunk, rendered in Blender, and says this is not cloud gaming or remote desktop.
  • A model on SSD is normally unusably slow. DeepSeek 4.1 Flash has 750 billion parameters and needs over 500 GB of memory, against the 5090's 32 GB of VRAM. Chuck's figures: VRAM 1.8 TB/s, DDR5 90 GB/s, SSD rated 14 GB/s. Small 4 KB reads give only 1–5 Gbit/s, so a dense model would take about 8 minutes per word. Mixture of experts (6 of 384 experts per word) still means about 15 seconds.
  • Three changes make the SSD model run at usable speed. Expert files are rewritten into equal-size aligned chunks read 64 at a time, taking SSD reads from about 1 to 11.5 GB/s. Popular experts stay in VRAM and RAM, so only about 1 GB per word comes off the SSD. The CPU does the math for SSD and RAM weights; the GPU does it for VRAM weights. Chuck measured 3–15 tokens a second, about eight on average, with 61% of experts served from the GPU in one run.
  • A scheduler on the box lets apps share the GPU. One image, video or chat model can fill all 32 GB of VRAM, and switching normally means reloading from scratch. The scheduler parks idle models on the SSD instead. In Chuck's demo, switching to the image model went from about 18 to 8.6 seconds, and back to chat from 17 to about 4. He calls the "family GPU" idea impressive and expects the engineering model to improve.
you NEED to learn decision models RIGHT NOW!!
  • Typesafe's decision model returns a decision, not words. It cannot call tools, search or chat; it answers in one of three forms: a yes/no with a confidence, a multiple choice, or a score. In Chuck's playground demo it called a real YouTube comment spam at about 92% almost instantly, labelled it as promoting something at 98%, and scored its sentiment positive.
  • It beats chat LLMs on speed and cost for simple decisions. Chuck's side-by-side against ChatGPT, on one comment and then 100, showed the chat model visibly slower even when asked only for yes or no. The company calls it a "system one" model, after Kahneman's fast gut call, against reasoning LLMs built for slow System 2 work. Co-founder Diego Almeida, who worked on InstructGPT at OpenAI, says it is meant to be called from code, not asked questions like an oracle. Weights and training data are closed.
  • Critics call it an AI-powered if statement, and the company agrees. Almeida describes it as smart if or switch statements, "AI-infused logic gates". Chuck's case for it over a trained classifier, using a password-strength example: a classifier answers one question and needs retraining for another, while this model takes any new question and is, by his account, about as fast and nearly as accurate, with no ML engineer needed.
  • It is for developers, which now means anyone coding with AI. That is Chuck's reading; a Typesafe engineer suggests letting your agent work out where it fits, using the company's official agent skill. Demos: a model router over OpenRouter (small model for "2+2", a Claude Sonnet model for proving infinite primes); a live teleprompter checking off script points from ElevenLabs transcription every half second, which she had shelved at roughly $30 an hour before this model made it cheap; and a Domino's order in 15 seconds against about 61 seconds with Codex.
  • It out-classified rival decision models and was the cheapest cloud option. In Chuck's test against OpenAI's Decisions API (announced at its recent dev day) and local models from Cloudflare and Ollama, on 20 bot and 107 fan comments, OpenAI came a close second and costs 2.4 times more; the most accurate local model was a 27B Cloudflare one. Chuck, echoing the engineer's Docker analogy, thinks decision models may be a new class developers have not yet grasped.
The #1 Danger Isn't Roblox
  • Predators are on every platform kids use, not just Roblox. Answering Chuck's question about the biggest risks to kids, a guest says Roblox draws the most predators because it is the largest kids' game in the world. The same goes for Instagram, Snapchat, TikTok and dating apps, the guest says, with children as young as six in these chats.
  • The biggest danger is kids being too scared to tell their parents. The guest argues that taking the phone away backfires, because kids still have friends with phones and go behind their parents' back, leaving the parents with no control. Fear of losing a phone or Xbox keeps a child quiet when an adult messages them asking to meet. The guest's advice is to talk to and educate kids now.
0day Hacks NetworkChuck (OSINT)
synopsis

Chuck sits down at Defcon with Ryan Montgomery, who works against online child exploitation, to learn how to do the work himself. The interview shows how easy predators are to find, then lands on Ryan's main point: don't do it as a vigilante. Ryan says the effective and legally safe route is to join an NGO that works with law enforcement and learn OSINT on safe targets first. He also says that educating parents, not arrests, is the most valuable thing he has done. Chuck commits to joining Ryan's team, and they plan a second episode built around an OSINT challenge on an already-convicted predator.

Who Ryan is. Ryan describes himself as a business owner who works with law enforcement and NGOs to identify predators, builds tooling for them and trains them. He is CTO of Sentinel Foundation, an NGO countering human trafficking and violence against children. He started hacking at nine through AOL chat rooms, IRC and phone phreaking, mentored by the nerdcore rapper YTCracker among others. Later, while running three rehab facilities, he found vulnerabilities in their electronic-records vendor's platform, including a defaced support subdomain. The vendor had just paid $40,000 for a pen test, and hired him as a consultant. A later sale of that company, for what he calls eight figures, funded his security company, pentester.com.

How the child-safety work began. A friend's wife sent him screenshots from a clear-web forum where users discussed abusing children. He meant only to DDoS it. Instead he found that it ran a nulled, backdoored theme on XenForo, which gave him a shell and its database. He dumped text only, deliberately excluding all media, and kept timestamped notes. The site's owner, whom he names as Nathan Larson, was arrested about six months later for kidnapping and raping a 12-year-old in Denver, and according to the news died in prison. Ryan had reported the site to law enforcement, tip lines and reporters. Every reporter wanted the story, but their legal departments refused, and about 7,000 other users went uninvestigated. That frustration led him to vigilante predator-catching with an MMA fighter named Dustin, the 561 PC channel. Ryan did the identification and filmed from behind the camera.

The live demonstration. Chuck asks how hard it is to find a predator right now. Two teen-chat sites from a Google search left Ryan's chat disabled. He then used a geolocating gay hookup app: anonymous login, only an 18+ birth year, no verification, an AI decoy photo of a young boy. It placed them in their hotel, with users 78 feet away. Within minutes one man 0.25 miles away sent explicit images. He was told the decoy was under 18, then asked to move to text, then backed out and deleted the messages. A second, who said he was 20 and was 2.5 miles away, said he was fine with "younger" and asked to trade nudes. Ryan says he could get either to knock on the door, and that he has never failed this demonstration in any country. A reverse face search through pentester.com turned one man's photo into a Facebook profile with a name. Ryan would only report the profile URL, because acting further could destroy a case. He says the same thing works on Instagram, Snapchat, TikTok, Grindr and Roblox. He cites an anonymous Australian survey in which one in five men admitted sexual thoughts about children and one in ten said they had acted on them.

Why not vigilante. Ryan says vigilante groups were getting disorderly-conduct and trespassing charges while predators walked free. He also says citizen-decoy cases bring far shorter sentences than ones where a federal agent did the decoying. Federal agents told him repeatedly that any association with vigilante work could discredit all his earlier cases as "fruit of the poisonous tree". So he works only through law enforcement and accepts losing the public profile. The mechanism is "parallel reconstruction": an NGO such as Sentinel Foundation, the Tim Tebow Foundation, Operation Light Shine or NCMEC asks the FBI, HSI or the Marshals for a case number. Analysts then do OSINT, and with the right signed authorization sometimes more. Investigators recreate the steps, and the analyst is never a suspect or witness. Working alone carries legal risk: once CSAM lands on your device it is possession, once it syncs to Google Photos or iCloud it is distribution, and once it is edited it is manufacturing. Passive OSINT carries less of that risk than active decoying. Ryan also criticizes another researcher who stayed in a chat recording children being extorted, which he calls complicity.

Getting involved. He says adults only, not teenagers. Sentinel runs an internship: email the foundation and write something like a resume. He says of about 500 emails perhaps 10 look like real candidates, and many interns drop out because the work is hard to stomach. Interns meet weekly and work one active case, and those who stay become analysts. Ryan says he can count on both hands the people with real security skills working child crimes. He offers to put Chuck straight in as an analyst, needing practice only with chain of custody and reporting.

The OSINT on Chuck and the tools. Ryan defines OSINT as professional Googling: public records, even hard-to-find ones, usable in court without subpoenas. From only Chuck's name and a phone number, his law-enforcement-only OSINT Pilot found Chuck's Social Security number, birthday, every phone number, his addresses, his wife's name, his dad's address and his past employers. It runs about 40 tools, live rather than from stale broker data, and AI summarizes the results. Chuck had paid for removal services. Ryan says brokers have 45 days to delist but relist whenever a detail changes, so removal has to be continuous. Chuck then reads an unprompted ad for pentester.com, starting at $19 a month. For learning, Ryan points to TryHackMe and OSINT challenges. The tools he names are whatsmyname.app and Sherlock for usernames, and Holehe, which checks which sites an email is registered on without notifying anyone. He also names PimEyes (paid; it found nothing on the decoy match) and TinEye for exact-image search, plus reverse caller ID. The chain runs from face to name, to data brokers such as Whitepages or Instant Checkmate, to breach data. He says breach sites are mostly on the clear web, and he declines to name them. He uses the dark web only alongside federal cases.

764 and parents. Ryan describes 764, a nihilist, self-identified Satanist network he says the US designates as terrorist, with many subgroups. They pose as peers in mental-health groups, groom vulnerable children, obtain nudes, then demand self-harm, animal killing or violence for clout. This has been his focus for close to two years. Because of death threats, he had three armed plainclothes guards at Defcon. For parents he recommends ic3.gov or the NCMEC tip line, and monitoring apps: he favours Bark and says he isn't sponsored, while Chuck uses Qustodio. He says monitoring is not enough. Taking away a phone drives children to friends' devices, so they must trust they won't be punished for telling. Children as young as six are in these chats. Chuck says his daughters have no social media.

What made it worth it. Ryan says a one-minute clip that drew 10 million views in a night built the network he now uses. He counts educating parents above the arrests: one recent case had 59 arrests, and his group flew children out of Haiti. He also built an ESP32 device, on a $4 chip, that counts nearby devices from BLE advertisements, to tell raid teams whether a house is occupied. He says MAC randomization broke his home-alarm version of the idea, but not the occupancy count.

  • Finding men willing to meet a child online takes minutes. Chuck asked Ryan Montgomery to prove it on camera at Defcon. Ryan opened an anonymous, unverified account on a geolocating hookup app with an AI decoy photo of a young boy. Within minutes one man 0.25 miles away sent explicit images after being told the decoy was under 18, and a 20-year-old 2.5 miles away agreed to "younger" and asked to trade nudes. Ryan says he has never once failed this demonstration in any country.
  • Ryan warns against vigilante work and says to work through NGOs instead. He says federal agents told him any vigilante involvement could discredit all his earlier cases as "fruit of the poisonous tree". Through an NGO such as Sentinel Foundation, where he is CTO, law enforcement opens a case number and uses "parallel reconstruction", so the analyst is neither suspect nor witness. Working alone also risks CSAM charges: possession once it is on your device, distribution once it syncs to cloud photos.
  • Data-broker removal doesn't stick unless it is repeated. From only Chuck's name and a phone number, Ryan's law-enforcement-only tool, OSINT Pilot (about 40 tools, results summarized by AI), returned Chuck's Social Security number, birthday, every phone number, his addresses, his wife's name, his dad's address and his past employers. Ryan says brokers have 45 days to delist but relist when anything changes. Chuck read an unprompted ad for Ryan's pentester.com, which starts at $19 a month.
  • Beginners should practise OSINT on themselves and on challenges, never on predators. Ryan suggests TryHackMe and OSINT challenges, and names the free tools whatsmyname.app, Sherlock, Holehe (checks which sites an email is registered on without notifying anyone) and TinEye. He also names PimEyes, which is paid and found nothing on the decoy match that pentester.com's face search had tied to a Facebook name. His chain runs face to name, name to data broker (Whitepages, Instant Checkmate), then breach data.
  • Ryan says educating parents does more good than arrests. He describes the 764 network, which he says is US-designated as a terrorist group. It grooms vulnerable children in mental-health groups and extorts them into self-harm for clout, not money. He says children as young as six are in these chats and every platform, Roblox included, is exposed. He recommends ic3.gov or the NCMEC tip line, monitoring apps (Bark, which he says doesn't sponsor him), and above all children trusting they won't lose their phone for telling.
Flushing the toilet kills our network!
  • Chuck built an all-AI IT department on Paperclip to chase the fault. In Paperclip the agents (network and systems engineers, storage, security) act as employees with managers and assigned tasks; they talk to each other only through the tasks, and everything has timelines and is traceable. The problem: whenever a toilet is flushed at the studio, everyone loses their connection to the NAS.
  • The agents' verdict: four switch links have been failing for 15 months. They report the cause is four links on one of Chuck's MikroTik switches that have been broken for 15 months and drop whenever anything disturbs the building, unnoticed by the rest of the switch. Chuck's own theory is that forceful flushing and door-slamming jostles the adjacent server room; the fix given is to replace the modules and re-patch the links.
You can monitor EVERYTHING in your homelab with this
synopsis

Chuck sets up Checkmk, the free open-source monitoring tool whose maker sponsored the video, and extends it from Linux servers to network gear, Docker, Proxmox, websites, Windows, and custom checks for Plex, Home Assistant and his Hermes AI agents. He ends with alerts that open tickets an AI agent resolves, and an AI agent with API access adding his remaining hosts. His caveat throughout is that the free Community Edition can do all of it, but only with manual work the commercial editions handle for you.

Install and first host. On an Ubuntu 24.04 VM he installs the Community Edition package, verifies it with omd version, creates a site with sudo omd create monitoring, starts it with sudo omd start monitoring, and logs in at the server's IP under /monitoring. For a Rocky Linux server he installs the RPM agent, opens a firewall port, tests the connection and runs service discovery, which surfaced a critical NFS service he had not known was down. Metrics are collected every minute, kept up to 4 years, and come with default thresholds.

Network, Docker, Proxmox, web and Windows. A host folder with network scanning finds switches and UniFi gear through SNMP. Docker takes python3-docker, the mk_docker.py plugin and its config, then a folder of "piggyback" hosts with no IP or agent, imported by CSV from the container list. Proxmox is read through an API "special agent" rule, with nothing installed; he notes the agent and a special agent can run together. A website check confirms a string in the page body and went critical when he searched for a word that wasn't there. The Windows agent needed its service restarted to get past the firewall, then flagged a 10 gig interface running at 5 gig, a nearly full disk and a stopped time service. Ceph and SMART plugins showed his NAS SSD pool at 97%.

Anything you can script. A community UniFi special agent from exchange.checkmk.com, installed with mkp as the site user ("at your own risk"), adds per-access-point and SSID data. His AI wrote a Plex check run as a Nagios-style active check showing who is watching what; Home Assistant got the same. Hermes agent crons, gateway, model login and usage come from a local script the Checkmk agent runs.

Alerts and AI agents. A Slack webhook rule reported a stopped Hermes gateway within about a minute, then its recovery. A second rule sent the failure to Paperclip as a ticket, and an AI agent restarted the gateway. An administrator user with an automation secret let his main AI add his remaining hosts and build dashboards. Chuck says the Checkmk Ultimate trial stays free for up to 750 services on one site.

  • Checkmk's free edition can monitor nearly anything, with tinkering. The video is sponsored by Checkmk. Chuck installs the Community Edition package on an Ubuntu 24.04 VM, checks it with omd version, then runs sudo omd create monitoring and sudo omd start monitoring and browses to the server's IP at /monitoring. He warns the free version needs manual work that the paid editions automate, and that the server must sit where it can reach everything it watches.
  • An installed agent plus service discovery shows far more than ping. On a Rocky Linux server he fetches the RPM agent with curl -O, installs it with dnf install -y, opens the firewall port, then runs service discovery: plugins find services to monitor, including a critical NFS service he hadn't known was down. Data is collected every minute and kept up to 4 years, with default thresholds for alerting. Network gear is found by a folder-level network scan using SNMP v2/v3 community credentials.
  • Docker containers and Proxmox VMs need extra steps in Community Edition. Docker requires python3-docker, the mk_docker.py agent plugin and its config on the host, then a folder with no IP, no agent and "always use piggyback data", with container names imported from a CSV headed host_name. Proxmox uses an API "special agent" rule with a Proxmox user, nothing installed. A VM's host name must match Proxmox's to link the two.
  • Plugins, active checks and local scripts let it monitor anything scriptable. He installs a community UniFi special agent from exchange.checkmk.com (mkp add, then mkp enable unifi 3.0.0, run as the site user), "at your own risk". His AI built a Plex check run as an integrated Nagios plugin; Home Assistant gets the same treatment. Hermes agent status comes from a local script that the Checkmk agent itself executes.
  • Alerts can open tickets for AI agents, which can also run Checkmk. A Slack notification rule caught a stopped Hermes gateway in about a minute and reported its recovery. A second rule sent the same failure to Paperclip as a ticket, and an AI agent restarted the gateway. An automation-secret admin user let his main AI add his remaining hosts and build dashboards. Chuck says Checkmk Ultimate's free trial covers up to 750 services on one site.
I turned my house into a VIDEO GAME!
  • An OpenAI model built a walkable 3D game of his house. Chuck says he gave GPT-6 Astra (the model's name as the captions give it) a roughly 10-minute phone video of every room. It was not right on the first try; over back-and-forth chats he added multiplayer with a character for each family member, and his dogs walking around and playing fetch.
  • The model kept the project on track across long, remote sessions. Chuck says it held context through context compactions and multiple chats, and he drove the build on his Mac Studio from the ChatGPT app on his phone while travelling to OpenAI. It tested its own work, opening and inspecting things through computer use and browser use.
  • Hooked to Home Assistant, the game controls the real house. Chuck says turning off a light in the game turns off the actual light, and he can control the family's Apple TV from the game and see what is playing on it. He says he has tried this with every model that has come out and this one is different, and urges viewers to try it in the ChatGPT app.
your house needs TWO networks (here's why)
  • IoT devices belong on their own isolated network. An IoT security expert Chuck quotes says to segregate them if you have the networking gear, or else move them to the guest network. The reasoning: one vulnerable device lets an attacker pivot into everything else, and you can't know which device is vulnerable until someone exploits it. A typical home router won't run multiple networks.
  • VLANs make many networks from one firewall and one cable. On the OPNsense mini PC from episode one, Chuck creates them under Interfaces > Devices > VLAN, each attached to the LAN as its parent interface: VLAN 13 "user", VLAN 14 "work". The switch has to be managed. He set ports 9 and 10 as untagged on VLAN 14 and port 1 as tagged (trunk) for every VLAN, so connected devices never need to know which VLAN they're on.
  • A VLAN isn't a working network without three more pieces. Those are an assigned interface with its own IP, DHCP and firewall rules. Chuck gave VLAN 14's interface 172.16.14.3 on 172.16.14.0/24 (mask 255.255.255.0). He set the DHCP pool from roughly 172.16.14.100 to 172.16.14.200, about a hundred addresses, which he says is plenty for a home. Until all three were in place, clients got only a self-assigned address.
  • DHCP failed because of one hard-to-find OPNsense setting. Chuck gave Claude Code access to the box, and it found that DHCP was only listening on the default LAN interface. The fix is the Interfaces dropdown on the General tab of the Dnsmasq DHCP settings, where each new interface has to be ticked. He blames OPNsense mostly for hiding it. Getting an address there also showed the switch handled multiple VLANs.
  • OPNsense blocks everything by default, so each network needs rules. For the work network, one rule from the work interface to any destination, action allow, gave full access. For IoT, Chuck grouped the private networks into an alias and blocked IoT traffic to it, and pings from IoT to the main networks then failed. It turned out Mike has no IoT devices. Wi-Fi is set for episode 3, and Chuck's Raspberry Pi agent managed the switch despite it having no CLI.
Your AI App Works. Can You Sell It?
  • A missing compliance answer can stall an enterprise AI deal. Chuck says a big customer's single blocking question is often SOC 2 compliance status, pulling the engineer off the roadmap to gather audit evidence instead. The video is a sponsored pitch for Vanta's compliance platform; it cites Y Combinator startup Hyperbound — two founders, no security team — which used Vanta to close a $1.5 million deal requiring ISO 27001 in under three months and land Fortune 100 customers within a year.
you need to try Paperclip RIGHT NOW!
synopsis

Chuck builds an AI "IT department" inside Paperclip, a meta-shell that turns Claude Code, Codex, and Hermes agents into managed employees, then sets it loose on a real problem — a network that drops every time someone flushes a studio toilet — ending with agents tracing the outage to a bad batch of SFP transceivers, replaced with official MikroTik modules.

What Paperclip is. Paperclip doesn't run models itself; it's a metasystem any agent CLI plugs into as an "employee," working through assigned tasks rather than free-form chat — the creator, in his first on-camera appearance, says an earlier "agent conference room" design didn't work. Chuck installs it on a self-hosted VM on port 3100, and lets onboarding agent Dumbledore (Claude Code) walk him through setup.

Building the department. He hires Hermes agents Ron (CTO), Fred (network engineer), and George (storage engineer) under Dumbledore, then lets Dumbledore self-hire a Codex-based security specialist and Pi-based support agents. Agents escalate blockers up the chain rather than pinging Chuck directly.

Solving the toilet mystery. Assigned the recurring-disconnect case, the team pulls switch logs and produces artifacts Chuck can inspect. Their verdict: four fiber links on his "Flying Powder" MicroTik switch fail together within 2 seconds, 36 times in one week; port 5 alone logged 20,052 drops versus 123 on a comparable port. They can't confirm the toilet, only "a physical disturbance," but the real fault is a cheap third-party SFP batch with a roughly 50% failure rate, fixed with official MikroTik modules.

Flare and automation. Sponsor Flare's dark-web scan turns up leaked academy-user credentials and session cookies for sale on Telegram, bypassing MFA; Chuck connects Flare's API into Paperclip via a newly hired agent, Filch, who — once given the API key and tenant ID as stored secrets — investigates unprompted and starts assigning password resets. Chuck also demos scheduled "routines," like a daily storage-health check, and exporting the company for reuse elsewhere.

  • Paperclip turns existing agent CLIs into a managed company of AI employees. Chuck connects Claude Code, Codex, and several Hermes agents to it, assigning each a manager and a role; agents communicate only through tasks, not an open chat room, which Paperclip's creator — making his first public on-camera appearance in this video — says didn't work in an earlier "agent conference room" design.
  • Third-party SFP transceivers, not the toilet, caused the NAS drops. After Paperclip's agents traced repeated fiber-link failures on four specific switch ports (port 5 alone logged 20,052 connection drops versus 123 on a comparable port), their final report couldn't confirm the toilet as the trigger — only that a physical jolt near the server room caused it — and found the real fault was a cheap third-party SFP batch with a roughly 50% failure rate; Chuck replaced them with official MikroTik-branded modules.
  • Flare's dark-web scan found Chuck's academy users' credentials for sale. Sponsor Flare, an identity-focused threat-intel platform, surfaced 76 credential leaks, 390 chat mentions, 23 fake domains, and 32,000 infected devices tied to his company — including session cookies sold on Telegram markets for about $10 each that bypass two-factor authentication entirely. Chuck cites Flare's figures that 54% of ransomware victims' data was already in stealer logs before the breach, and that attackers act within 48 minutes while companies average 36 hours to notice.
  • Paperclip agents can schedule routines and act on new API access autonomously. Chuck set up a scheduled daily "routine" so an agent checks NAS storage health (it found only 29 TB free on the SSD tier), and showed that a whole company setup — every agent, routine, and task — can be exported and imported elsewhere. After granting a newly hired agent (Filch) his Flare API key and tenant ID as stored secrets, Filch ran the integration unprompted and began assigning Ron to reset exposed passwords and Mad-Eye Moody to investigate further.
i got one....and it's FAST!!!
  • The M5 Ultra's faster prompt processing comes from new per-core neural accelerators. Both chips carry the same 80 GPU cores, so on a 32K-token summarization test using a quantized 27B-parameter Qwen model, the M5 needed just 21.66 seconds to produce its first token versus the M3's 82.85 seconds — a 3.8x gap Chuck attributes to the accelerators.
  • Token-generation speed tracks memory bandwidth, not GPU cores. Generating each token means re-reading the roughly 15GB model from memory, so the M5's 1.2TB/s bandwidth (versus the M3's 819GB/s, a 50% increase) produced 43.3 tokens/second against the M3's 28.1 — a 1.5x edge that matches the bandwidth increase almost exactly.
  • A smaller 14B model widened the M5's lead even further. On the same 32K-token test, the TTFT margin grew to 4x (14 vs. 57 seconds) and speed to 1.7x, since a smaller model moves fewer gigabytes through memory per token — showing model size shifts which bottleneck, compute or bandwidth, dominates.
  • Several everyday AI workloads held up well enough to actually use. Whisper large-v3 transcribed a 2-hour-5-minute recording in 89.4 seconds (24.4 seconds on turbo); a local Qwen vision-language model analyzing video frames ran 2.6x faster than the M3 and rivaled the paid Gemini video API Chuck currently uses; and local FLUX.2 Klein image generation and an LTX 2.5 22B video model produced usable results in seconds, free and offline.
  • Chuck calls the M5 Ultra the best computer money can buy. The 80-core, 256GB configuration he tested costs roughly $14,000, though the unit was loaned by Apple and must be returned. He also had it drive a DaVinci Resolve edit via MCP and run as a local messaging-agent backend — both impressive, he says, but not yet ready to replace his main agent setup.
Your App Crashed. Now AI Gets to Work.
  • Omarchy will diagnose its own crashes using an AI coding agent. A crash triggers a pop-up offering to "diagnose with AI," which loads the user's preferred harness — Claude Code, OpenCode or others — with a skill that gathers crash details, determines whether it's an actual Omarchy bug, and, if the user has the GitHub CLI set up, files the report directly to the project's GitHub repo.
  • DHH's own agent turns reproducible bug reports into pull requests. Describing the loop from Omarchy's creator, he says this bug-report-to-PR pipeline lets the team "iterate over the operating system itself at a completely different pace."
AI Still Needs Network Engineers
  • AI increased demand for network engineers rather than replacing them. Chuck argues every company's network, including the infrastructure carrying prompts to ChatGPT or Cloud Code, is built and run by network engineers. He frames this as reason one of five for getting CCNA certification now, and flags that the exam is changing soon (reason four) as the time-pressure reason, without giving the remaining reasons in this short.
Windows Took 90 Minutes. Omarchy Took 1.
  • Windows took far longer to set up than a Mac or Omarchy. DHH timed a new Windows PC at 90 minutes of updates and driver installs before it was usable, and a new MacBook at 42 minutes; he says weeks of AI-agent-assisted optimization got a full Omarchy install down to about one minute.
You Need to Get Your CCNA Right Now!!
  • Chuck argues AI has increased, not eliminated, demand for network engineers. He points to the data centers and infrastructure running AI tools like ChatGPT and Claude Code as themselves needing networks built and maintained by engineers, and says he'd never let AI alone run a network without someone who understands it standing behind it.
  • CCNA skills open doors into any IT career, he says — hacking, security, DevOps, cloud or AI engineering. He cites Def Con attendees who told him his CCNA content led them into cybersecurity work as evidence the certification acts as a gateway.
  • A CCNA gets candidates the interview, since hiring managers universally recognize it. Chuck says a network-engineer friend, Eric, was told Wells Fargo would not interview him without a CCNA, even though newer certifications might not register with the same managers.
  • The CCNA exam is being revamped (announced May 20, 2026) to add more troubleshooting and AI content, with the current version retiring February 2, 2027. Chuck advises against waiting for the new exam, since study materials for it will be slow to catch up — anyone already preparing should finish the current CCNA now.
  • Chuck is promoting his paid "Fall into CCNA" program at Network Chuck Academy, a 13-week course (September 14 to December 11, 2026) co-taught with Jeremy Cioara, priced at $250, or $300 for a full year of academy access. It includes video training, in-browser labs on real Cisco IOS, quizzes, live streams and a peer community, modeled on a prior "Summer of CCNA" cohort he says was successful.
Omarchy Can Do WHAT?! 50 Features You’re Missing
synopsis

The video is a rapid-fire, mostly keyboard-only tour of Omarchy, the Arch-based Linux setup built by DHH, ending with Chuck using AI — both a cloud model and a local one — to reconfigure the desktop and publish his own plugin to the Omarchy marketplace.

Moving without a mouse. Chuck opens with the Super-key shortcuts that drive nearly everything: launching a terminal and browser, switching and grouping windows, moving them between up to four virtual desktops, and three separate ways to go full screen. A "sketchbook" shortcut lets him stash an open window out of sight and recall it later, which he says is useful for carrying an app like Slack between workspaces.

Installing, updating and rolling back. Packages install and uninstall through the same Super-space launcher, pulling from Arch, the Omarchy repos, or the AUR (which he warns is unmoderated). Every system update creates a Btrfs snapshot automatically, and Omarchy deliberately runs about a month behind Arch's own releases to avoid breakage; a snapshot can also be created and booted from manually via omarchy snapshot.

Defaults and built-ins. Disk encryption and a default-deny firewall (with an exception for LocalSend) are on out of the box, and SSH is off. Chuck also runs through built-in screenshot, screen-recording, video-transcoding, QR-capture and LocalSend file-sharing tools, plus a theming system that can generate a full desktop theme from a wallpaper.

AI throughout. Omarchy ships with Claude, Codex, Copilot, Gemini and Grok pre-installed as selectable default agents, alongside support for fully local models. Chuck runs Qwen 3.5 Coder 4B on a Framework laptop with no discrete GPU, using it to change his power profile and generate a screensaver, then uses Claude to build and publish a custom coffee-tracking widget to the plugin marketplace, requiring no code from him.

  • ThreatLocker's sponsored segment argues coding agents need allowlisting, not blanket trust. It says security teams should permit only the specific scripts an agent needs, isolate what those scripts can then reach, and use a learning mode to build that allowlist automatically — contrasted with the "dangerously" permissive setting Chuck says is fine for home experiments with Omarchy but not for a company.
  • Omarchy ships with several security defaults already turned on. Full disk encryption is mandatory, the firewall blocks all inbound traffic except LocalSend, and SSH is disabled out of the box; sudo can also be configured to use fingerprint authentication instead of a password.
  • Every system update automatically creates a rollback snapshot via Btrfs. Chuck says Omarchy intentionally trails Arch's own updates by about a month to avoid breakage, and a snapshot can also be created manually and booted from directly with the omarchy snapshot command if an update goes wrong.
  • Chuck ran a local 4B model on a laptop with no GPU. The model, Qwen 3.5 Coder 4B, has a 128,000-token context window and ran on his Framework laptop with no discrete graphics card; it checked and changed his power profile and generated a working ASCII screensaver, all without touching the cloud — which he calls "a victory" for a model that size on CPU-only hardware.
Linux Doesn’t Have to Be Ugly
  • DHH built Omarchy to look like hacker-movie computer aesthetics. He cites films like The Matrix as inspiration, saying Linux's window arrangement and color coordination don't have to look haphazard, and that he wants someone running Omarchy in public — even just a terminal or screensaver — to make onlookers wonder if he's "hacking."
DEFCON Is Not What You Think
  • DEF CON runs a public "Wall of Sheep" shaming insecure Wi-Fi use. At Packet Hacking Village, it displays a live scoreboard of intercepted usernames, domains and passwords from attendees who connect carelessly; Chuck's advice is to skip the conference Wi-Fi entirely and turn off Bluetooth before going in.
  • DEF CON has grown into the world's largest hacker conference. It began as a small hacker gathering in the 1990s; this year, DEF CON 34, drew roughly 30,000 attendees — researchers, red teamers, and hardware hackers — to Las Vegas, spread across themed "villages" for lockpicking, cars and biohacking, among others.
  • Chuck streamed his DEF CON footage home without opening a single port. The segment is sponsored by TwinGate; he set up a Mac Studio inside Packet Hacking Village and used TwinGate's zero-trust access to relay recordings to his home studio over the public DEF CON network — a tool he says he uses daily for remote access.
  • Chuck says DEF CON's real value is its people, not the talks. He points to conversations with hackers he met there, like Vega, a principal consultant at Sophos who pentests networks, web apps, APIs and medical devices — the latter, he says, often built fast with security bolted on afterward, leaving gaps — and says much of the real networking happens at unofficial post-conference parties, not the scheduled talks.
Is Linux the perfect operating system? - Omarchy interview with DHH
synopsis

DHH joins Chuck to walk through Omarchy, the Arch-based Linux distribution he has spent about a year refining into what he calls "the perfect operating system" — installing in about a minute, styled like a hacker-movie interface, and treating AI agents as first-class citizens able to diagnose and even patch the OS itself.

AI and the calm workplace. DHH says AI has introduced urgency and "attention mania" that disrupted the calmer philosophy of his book "It Doesn't Have to Be Crazy at Work," and that the current sprint pace isn't sustainable. His longer-term bet: agents gaining autonomy will absorb routine toil and keep working while humans don't, letting people review bundled decisions on their own schedule instead of chasing every notification.

Building the perfect OS. DHH left over 20 years on macOS for Linux about 2.5 years ago, passed through Ubuntu, and settled on bare-bones Arch, Hyprland's tiling window manager, and a new interface layer, Quickshell, as Omarchy's foundation. Rather than Arch's do-it-yourself ethos, he preconfigures and themes everything out of the box, while still letting users customize at his own level — unlike Apple's or Microsoft's closed stacks.

Agents running the OS, fast. In Omarchy Quattro, a crash triggers a pop-up handing the problem to the user's AI coding harness, which diagnoses it, checks whether it's an actual Omarchy bug, and files a GitHub issue; DHH's own agent then reviews incoming reports and turns reproducible ones into pull requests, a loop he credits for merging over 1,000 PRs in about 90 days. The same agent-assisted engineering got the full install down to about a minute, versus 90 minutes for a new Windows PC and 42 minutes for a new MacBook.

DHH's own stack. He works across Claude, Codex and open models through Herder, a tmux-like tool for agent awareness, scaling to several machines over GL.iNet Comet KVMs to run many agents in parallel. He had an AI tool, Fable, port a Python screensaver library to Rust and then auto-optimize it to run 27 times faster, and now lets an agent manage that project's releases independently.

  • DHH expects AI agents to eventually restore a genuinely calm workplace. That's the vision behind his book "It Doesn't Have to Be Crazy at Work," which he says AI has disrupted for now — he argues agents should absorb routine toil and work while humans aren't, bundling decisions into set review windows, though he admits today's urgency-driven pace isn't sustainable.
  • Omarchy is DHH's attempt at building "the perfect operating system." It's built on Arch Linux, the Hyprland tiling window manager, and a new Quickshell-based interface he calls Quattro. DHH moved from over 20 years on macOS to Linux about 2.5 years ago, starting with Ubuntu before settling on bare-bones Arch as a base he could fully customize, rejecting a DIY "Legos" approach in favor of a preconfigured, pre-themed result.
  • Omarchy Quattro can diagnose its own crashes and file bug reports. A crash pop-up loads the user's preferred coding harness with a diagnostic skill; if it confirms an actual Omarchy bug and the user has the GitHub CLI configured, it files the report directly, which DHH's agent then reviews and turns into a pull request when it reproduces. DHH says his team has merged over 1,000 PRs this way in roughly 90 days.
  • DHH timed a full Omarchy install against fresh Windows and macOS setups. His result: about one minute, versus 90 minutes for a new Windows PC and 42 minutes for a new MacBook before each was usable. He says weeks of optimization work with heavy agent help got installs that fast, and an in-place upgrade from the prior Omarchy version takes roughly 30 seconds on a fast connection.
  • DHH runs a multi-machine, multi-agent coding setup. He pairs Claude, Codex and open models in parallel through a tmux-like tool called Herder, scaling across several PCs via GL.iNet Comet KVMs so many agents can work at once. As one example, he had an AI tool called Fable translate a Python screensaver library to Rust, then auto-optimize it to run 27 times faster than the original — and now lets an agent review, merge and publish that project's pull requests without him reading the code.
this Raspberry Pi belongs on your wall
  • Chuck likes the screen, but it only works with recent Pis. The Raspberry Pi Touch Display 2 (10-inch, portrait) costs $80, which he considers comparable to other screens, and supports only the Raspberry Pi 5 and the Compute Modules, so a Pi 3 or 4 will not work. Raspberry Pi sent it to him "no strings attached".
  • The main gain over the 5- and 7-inch versions is size. The specs he reads out are 10.1 inches diagonal, 1200 by 1920 native portrait, 224 PPI, IPS, 400 cd/m² (60% brighter than the original), anti-glare, 24-bit colour and 16 mm thin, with the driver board now inside the enclosure. Touch is 10-point (the 7-inch did five) with a 35 ms maximum response. The box holds only the screen, a DSI ribbon cable and a GPIO power cable.
  • A fresh OS image is not enough: update the Pi's EEPROM firmware. On a never-booted Pi 5 the display showed nothing, and swapping the Pi, SD card and power supply did not help. Updating the EEPROM firmware made it recognise the touchscreen. He found the DSI ribbon cable flimsy and hard to seat.
  • Touch and the auto-popup keyboard work well, but the desktop is cramped. He found touch very responsive in games, and the on-screen keyboard means no mouse or keyboard is needed. On the other hand, desktop targets are hard to hit, raising the text size broke the keyboard, there are no speakers, and nothing is provided for wall mounting; he screwed it to the wall with the screws holding the panel.
  • Chuck argues a Pi makes a better wall dashboard than a tablet. He runs TouchKio, an open-source kiosk app (a locked-down full-screen browser) that comes back up after a reboot, showing three custom Home Assistant dashboards made through its WebSocket API. The Home Assistant server stays on a Proxmox VM, but he says everything could run on the Pi. His AI, which has full access to the Pi, builds new graphs on request.
shadow AI is terrifying
  • Chuck says he no longer knows what AI tools are running. He says he adds new tools "every week, now every day" and suspects unapproved integrations and vendors are already active inside his company — framing shadow AI as spreading exactly as fast as AI adoption itself, mostly outside security's view.
  • A bot he hadn't approved showed up in his company's Slack. He noticed an unfamiliar AI integration named "Inka" appear on its own; he says he didn't add it and didn't know who did, offering it as a live instance of the shadow AI sprawl he's describing rather than a hypothetical.
  • The video is a sponsored pitch for Vanta's "agentic trust platform." Vanta says its AI agent assesses new vendors, gathers evidence, and answers security questionnaires on a company's behalf, claims to cut vendor-assessment time by 50%, and says it's used by more than 16,000 companies, including Cursor.
HTTPS Doesn't Hide This From Your ISP!!
  • HTTPS encrypts traffic but doesn't hide which sites you visit. Chuck shows a Wireshark capture where enabling secure DNS hides DNS queries, but the TLS SNI field still reveals the destination site to an ISP or attacker. He says a VPN is needed to encrypt that and other non-browser traffic that HTTPS alone doesn't cover.
Cisco Just Showed the Future of Networking
  • Cisco Cloud Control unifies networking management into one dashboard. Chuck, reporting from Cisco Live, says it replaces what had been seven separate controllers and logins for Cisco customers, calling it the "single pane of glass" network engineers have wanted. He notes it was already running live on the show floor rather than just announced for a future release.
  • The platform is built to host AI agents, not just dashboards. A Cisco presenter named Grant says Cloud Control lets users "build trusted, secured, and controlled agents" directly within the same screen. Chuck contrasts this with last year's Cisco Live, where AI was discussed but, in his view, not yet shipped.
Hermes has a Home Assistant skill and it's unreal!
  • Hermes ships a Home Assistant skill, but it needs switching on. Chuck thinks it is disabled by default; he had Hermes enable it, then gave it his Home Assistant instance's IP address and key, and it worked out the setup and saved it to its memory.
  • Once connected, it controlled Chuck's studio from plain requests. In his own test it turned off a named lamp ("Chuck Lamp") on request, then operated the studio's automatic blinds, which he filmed to confirm.
I flew to Romania to find hackers who fight hackers
  • Bitdefender's "Draco" malware-analyst team spent two and a half years building free decryptors against one ransomware gang until it gave up. Chuck visited Bitdefender's Romania HQ and reports the team has saved ransomware victims over $1 billion in payments overall; each time the gang updated its malware, Draco released another free decryptor, and the gang eventually shut its operation down.
  • Chuck argues AI has erased the technical skill barrier that once limited scammers, and that even experts get fooled by AI-driven scams sometimes. He pivots from the Draco story to promoting Bitdefender's own "scam protection" tool, pitched as a firewall checking calls, messages and emails to protect family members less able to spot these scams.