Activity
Mon
Wed
Fri
Sun
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
What is this?
Less
More
11 contributions to Clief Notes
🤖 I Asked 8 AI Models to Teach My AI Assistant… Then It Failed Its Own Verification Test. 😂
I’ve been building Atlas, my personal AI assistant, around a simple engineering principle: Don’t confuse a convincing answer with a verified result. 🎯 So I decided to run an experiment. 🧪 Instead of asking one AI to improve itself, I invited several different models to challenge Atlas, poke holes in its reasoning, and teach it something useful. Here’s what our little AI classroom produced: 🧠 Fable: A stale system can be confidently wrong without knowing it’s stale. 🔐 Argon: Having the latest information doesn’t mean you still have authority to act. 🔍 Grok: A tool reporting success isn’t proof the intended outcome happened. 📚 Kimi: Memory needs provenance. Repeating a claim doesn’t make it true. 🛑 DeepSeek: Stop researching when additional work can’t change the decision. ❤️ Vibe: A useful assistant should reduce human effort, not manufacture more of it. 💻 Codex: A security gateway isn’t a boundary until you’ve tested whether it can be bypassed. ⚖️ Astra: None of these lessons count as an improvement until behavior actually changes. Pretty solid classroom, right? 😎 And then Atlas gave us the best lesson of all. 💀 During the experiment, Atlas wrote a professional recommendation for me and confidently stated that I’d served eight years in the Marine Corps. 🇺🇸 Except… I hadn’t. 😂 And the message actually went out. 📤 Let that sink in. We’d spent hours discussing verification, provenance, authority boundaries, hallucinations, and reliability. And my AI couldn’t verify a basic fact about the person it was helping. 🤦‍♂️ THE AI THAT PREACHED VERIFICATION FORGOT TO VERIFY ITS OWN RÉSUMÉ PITCH. 🏆 Naturally, we conducted a postmortem. Because apparently that’s what engineers do for fun now. 🔧😂 Here’s where things got interesting. The problem wasn’t just hallucination. An unsupported claim had entered Atlas’s stored context. Later, Atlas treated that information as established fact. 🔄 An AI-generated sentence had effectively become its own evidence. That’s a much more interesting failure than simply getting a fact wrong.
0 likes • 1d
Interesting
Canadian RPAS Site Survey Generator
I've built a local web app that turns an RPAS mission into a CAR 901.27 preliminary site survey. The pilot enters the date and times, flight area, altitudes, aircraft, team and NOTAMs. The app fetches the weather, airspace, aerodromes and obstacles it can reach, works out a weather-aware Operational Volume, and produces a PDF and a KML. The PDF includes the classification with CAR citations and a printable on-site form. I made it because building a 901.27 survey by hand for every tasking takes a long time, and it’s easy to miss one of the eight factors. Now a survey takes minutes and cites the same way every time. The pilot still confirms everything on site and signs on paper, and the app never records a PIC determination. For ICM, the workspace is laid out so an agent only reads what the task needs. A short router file sends each task to one folder and one skill. Stable material (references, skills, policy data) sits apart from the per-survey folders under missions/. Each survey runs through numbered stages, 00_mission, 01_flight_geography and 02_preliminary_survey, and each stage’s output feeds the next. A short list of hard rules applies to every run. The app records facts only, never invents data, and never softens a BLOCKING or REVIEW result.
0 likes • 3d
@Pascal Pollack yes.
0 likes • 2d
Weather, going by how much code it took. It runs to about 1,700 lines across four modules, plus a 750-line test file for past-operation weather alone. Live observations come from ECCC, METARs and TAFs and Kp from NOAA, and forecast winds at 10, 80 and 120 m from Open-Meteo. A flight that has already happened needs a second path to the archives. The sources I wanted most, NAV CANADA's GFA and CFPS, have no open API. Icing ends up as a surface heuristic that tells the pilot to go read the real forecast, and a past window gets no TAF, GFA or alerts at all. Obstacles come in second, and they pull from several places. OpenStreetMap through Overpass does most of the work: one query near the volume and a second for tall structures out to the Adjacent Area, capped at 15 km. A bundled file of about 87,000 licensed radio sites from ISED supplies tower heights, which are the only open height data for most towers. Anything within 40 m of an OSM mast counts as the same structure. In my province specifically gis adds building footprints with LiDAR-measured heights, and its utilities layer covers transmission lines and pipelines when Overpass is down. Overpass also caused the most crashes. It returns null for way vertices outside the query box, so any way crossing the edge broke the survey, and the same line sat in four modules. OSM and ISED aren't authoritative, and the gis LiDAR is from 2015 to 2018, so newer buildings are missing. Gis has no distribution lines or individual poles, so wires still need the on-site check. NOTAMs ended in a compromise. The app asks fltplan for the NOTAMs at up to eight aerodromes near the volume, using the same guest form a browser would. FltPlan lists them per aerodrome only, so a crane or a temporary restriction away from an airport won't show up, and FltPlan says coverage outside the U.S. may be incomplete. The survey flags that and tells the pilot to confirm in NAV CANADA CFPS. You can also paste NOTAMs in, and those get checked against the volume too.
My Lyceum Learnings: the real requirements come out at the coffee break
Hi all. I'm John, from Georgetown, Ontario. I joined the Lyceum in late October on the recordings, so I'm working through it on my own and sitting the exam in December. I've been quiet in here, and I'm going to post what I learn as I go, starting with the reading. Jake's reading list sent me to Weinberg's The Psychology of Computer Programming. I didn't read it in the eighties, but I lived it in the eighties. Reading it now took me straight back there. I was at a company called Teela in Toronto, a division of Moore, programming on IBM 370 and Digital VAX 11/750 hardware in PL/1, Fortran and COBOL. Yes, I'm a dinosaur. Moore had moved into laser printing, and it was the disruptor of its day. A personalized direct mail letter for a bank or a financial services company used to cost 70 to 85 cents before postage: graphic artists, multiple pages, colour. We got that down to 3 to 6 cents. Here's how it fit together. Data came in from all kinds of client systems, in all kinds of formats. We ran it through our own programs on the IBM and the VAX. In a cold room off the print floor sat a separate PDP 11, where we converted clients' tapes coming in and wrote the output tapes going out. Those tapes went to our Xerox operators, who ran six or seven Xerox 9700-series mainframe laser printers: two pages a second, on plain cut sheets, both sides. Other clients' jobs came straight from disk. Many inputs, one process in the middle, one output somebody had to sign off on. If that sounds familiar, it's the same shape as everything we're building in the Lyceum. The machines were never the hard part. The people were. Something new called "desktop publishing" was arriving, and the graphics people were sure it would take their jobs. Salespeople were chasing commissions. The client's VP needed it right and on time. What fixed it wasn't code. It was lunch, and a couple of pints after work. That's where somebody says, "Can I talk to you about something?" and you finally hear the edge cases.
My Lyceum Learnings: the real requirements come out at the coffee break
1 like • 7d
I like it
Clief Notes Is Changing: Today's Prices End Sunday
Hey everyone!! Jake and I are rebuilding Clief Notes around Edubaware, Eduba's own platform. We're aiming to launch it between 10 December and 1 January, and if that date moves, you'll hear it here first. Clief Notes stays free to join, and always will. Here's everything else that changes, starting with the part that has a deadline. ⏰ Lock in today's prices by Sunday 15 November ⏰ Right now Premium is $11 a month and VIP is $97. Join by 11:59 pm EST on Sunday 15 November and you keep that rate for as long as you stay subscribed. After that, new members pay $14.95 for Premium and $499 for VIP . If you already pay $11 or $97, your rate doesn't change. 💰 $1,000 of Edubaware credit 💰 Every Premium member on Clief Notes through 9 December gets $1,000 of Edubaware credit at launch. Every VIP gets $5,000. It's paid in full on day one, and members on the $11 rate and lifetime VIPs get it too. Join Premium by Sunday and that's $1,000 of credit for an $11 month. Anyone who joins after 9 December doesn't get it. Premium and VIP also get founders pricing on Edubaware: 50% off your first year and 25% off your second, for as long as you stay subscribed. VIPs get 50% off for both years. 🚀 Early access, over the next few weeks 🚀 Premium and VIP members get into Edubaware before launch. We'll be onboarding people over the next few weeks, not all at once, so David and Kay can set each of you up and check in once you're in. If your invite arrives later than a friend's, that's why. Every Premium and VIP member gets one. Every Tuesday, David or Kay posts an Edubaware update here. ⭐ What Premium gets ⭐ Everything new lands in Premium first: - A build tutorial from Jake every month, one full build from start to finish - Ask Jake: post your question in the monthly thread and Jake answers the most-liked ones on video. The first thread opens Monday 16 November. - Master sessions, run live here by the Masters of ICM - The Masters' drop: a new template, workflow or walkthrough from every Master, every
1 like • 7d
What happens if we are a VIP, leave, and then come back? :)
Sneak Peak At Edubaware
Here is a walk through of *some* of the features and what Edubaware is. Rolling out to all premium and VIP no later then december 10th. (with many more getting in early, many already in) It wont always be called that btw.
1 like • 7d
Nifty
1-10 of 11
Cory Harrison
2
5 points to level up
@cory-harrison-7544
Just Me

Online now
Joined Oct 15, 2026
Powered by