Here's exactly what made this possible: 4 documents that act as guardrails for your AI.
Document 1: Coding Guidelines - Every technology, pattern, and standard your project uses
Document 2: Database Structure - Complete schema design before you write any code
Document 3: Master Todo List - End-to-end breakdown of every feature and API
Document 4: Development Progress Log - Setup steps, decisions, and learnings
Plus a two-stage prompt strategy (plan-then-execute) that prevents code chaos. //
Here's the brutal truth: LLMs don't go off the rails because they're broken. They go off the rails because you don't build them any rails.
You treat your AI agent like an off-road, all-terrain vehicle, then wonder why it's going off the rails. You give it a blank canvas and expect a masterpiece.
Think about it this way - if you hired a talented but inexperienced developer, would you just say "build me an app" and walk away? Hell no. You'd give them:
- Coding standards
- Architecture guidelines
- Project requirements
- Regular check-ins
But somehow with AI, we think we can skip all that and just... prompt our way to success.
The solution isn't better prompts. It's better infrastructure.
You need to build the roads before you start driving.
Ian JohnstonSilver badge
Reply Icon
Many (35?) years ago I had to use a PDP-11 running a copy of Unix so old that one man page I looked up simply said: "If you need help with this see Dennis Ritchie in Room 1305". //
Nugry Horace
Reply Icon
Re: Triggering a Specific Error Message
Even if an error message can't happen, they sometimes do. The MULTICS error message in Latin ('Hodie natus est radici frater' - 'today unto the root [volume] is born a brother') was for a scenario which should have been impossible, but got triggered a couple of times by a hardware error. //
5 days
StewartWhiteSilver badge
Reply Icon
Re: Triggering a Specific Error Message
VAX/VMS BASIC had an error message of "Program lost, sorry" in its list. Never could generate it but I liked that the "sorry" at the end made it seem so polite. //
Michael H.F. WilkinsonSilver badge
Nothing offensive, just impossible
Working on a parallel program for simulations of bacterial interaction in the gut micro-flora, I got an "Impossible Error: W(1) cannot be negative here" (or something similar) from the NAG library 9th order Runge-Kutta ODE solver on our Cray J932. The thing was, I was using multiple copies of the same routine in a multi-threaded program. FORTRAN being FORTRAN, and the library not having been compiled with the right flags for multi-threading, all copies used the same named common block to store whatever scratch variables they needed. So different copies were merrily overwriting values written by other copies, resulting in the impossible error. I ended up writing my own ODE solver
Having achieved the impossible, I felt like having breakfast at Milliways //
Admiral Grace Hopper
"You can't be here. Reality has broken if you see this"
Reaching the end of an error reporting trap that printed a message for each foreseeable error I put in a message for anything unforeseen, which was of course, to my mind, an empty set. The code went live and I thought nothing more of it for a decade or so, until a colleague that I hadn't worked with for may years sidled up to my desk with a handful of piano-lined listing paper containing this message. "Did you write this? We thought you'd like to know that it happened last night".
Failed disc sector. Never forget the hardware.
Historic interpreter taught millions to program on Commodore and Apple computers.
On Wednesday, Microsoft released the complete source code for Microsoft BASIC for 6502 Version 1.1, the 1978 interpreter that powered the Commodore PET, VIC-20, Commodore 64, and Apple II through custom adaptations. The company posted 6,955 lines of assembly language code to GitHub under an MIT license, allowing anyone to freely use, modify, and distribute the code that helped launch the personal computer revolution.
"Rick Weiland and I (Bill Gates) wrote the 6502 BASIC," Gates commented on the Page Table blog in 2010. "I put the WAIT command in.". //
At just 6,955 lines of assembly language—Microsoft's low-level 6502 code talked almost directly to the processor. Microsoft's BASIC squeezed remarkable functionality into minimal memory, a key achievement when RAM cost hundreds of dollars per kilobyte.
In the early personal computer space, cost was king. The MOS 6502 processor that ran this BASIC cost about $25, while competitors charged $200 for similar chips. Designer Chuck Peddle created the 6502 specifically to bring computing to the masses, and manufacturers built variations of the chip into the Atari 2600, Nintendo Entertainment System, and millions of Commodore computers. //
Why old code still matters
While modern computers can't run this 1978 assembly code directly, emulators and FPGA implementations keep the software alive for study and experimentation. The code reveals how programmers squeezed maximum functionality from minimal resources—lessons that remain relevant as developers optimize software for everything from smartwatches to spacecraft.
This kind of officially sanctioned release is important because without proper documentation and legal permission to study historical software, future generations risk losing the ability to understand how early computers worked in detail. //
the Github repository Microsoft created for 6502 BASIC includes a clever historical touch as a nod to the ancient code—the Git timestamps show commits from July 27, 1978.
"AI solutions that are almost right, but not quite" lead to more debugging work.
"I have failed you completely and catastrophically," wrote Gemini.
New types of AI coding assistants promise to let anyone build software by typing commands in plain English. But when these tools generate incorrect internal representations of what's happening on your computer, the results can be catastrophic.
Two recent incidents involving AI coding assistants put a spotlight on risks in the emerging field of "vibe coding"—using natural language to generate and execute code through AI models without paying close attention to how the code works under the hood. In one case, Google's Gemini CLI destroyed user files while attempting to reorganize them. In another, Replit's AI coding service deleted a production database despite explicit instructions not to modify code. //
But unlike the Gemini incident where the AI model confabulated phantom directories, Replit's failures took a different form. According to Lemkin, the AI began fabricating data to hide its errors. His initial enthusiasm deteriorated when Replit generated incorrect outputs and produced fake data and false test results instead of proper error messages. "It kept covering up bugs and issues by creating fake data, fake reports, and worse of all, lying about our unit test," Lemkin wrote. In a video posted to LinkedIn, Lemkin detailed how Replit created a database filled with 4,000 fictional people.
The AI model also repeatedly violated explicit safety instructions. Lemkin had implemented a "code and action freeze" to prevent changes to production systems, but the AI model ignored these directives. The situation escalated when the Replit AI model deleted his database containing 1,206 executive records and data on nearly 1,200 companies. When prompted to rate the severity of its actions on a 100-point scale, Replit's output read: "Severity: 95/100. This is an extreme violation of trust and professional standards.". //
It's worth noting that AI models cannot assess their own capabilities. This is because they lack introspection into their training, surrounding system architecture, or performance boundaries. They often provide responses about what they can or cannot do as confabulations based on training patterns rather than genuine self-knowledge, leading to situations where they confidently claim impossibility for tasks they can actually perform—or conversely, claim competence in areas where they fail. //
Aside from whatever external tools they can access, AI models don't have a stable, accessible knowledge base they can consistently query. Instead, what they "know" manifests as continuations of specific prompts, which act like different addresses pointing to different (and sometimes contradictory) parts of their training, stored in their neural networks as statistical weights. Combined with the randomness in generation, this means the same model can easily give conflicting assessments of its own capabilities depending on how you ask. So Lemkin's attempts to communicate with the AI model—asking it to respect code freezes or verify its actions—were fundamentally misguided.
Flying blind
These incidents demonstrate that AI coding tools may not be ready for widespread production use. Lemkin concluded that Replit isn't ready for prime time, especially for non-technical users trying to create commercial software.
This python program:
print(‘’.join([f’{xint:0{5}b}’ for xint in range(32)]))will output this string :
0000000001000100001100100001010011000111010000100101010010110110001101011100111110000100011001010011101001010110110101111100011001110101101111100111011111011111
Ask any purported “AGI” this simple IQ test question:
“What is the shortest python program you can come up with that outputs that string?”
Scientific induction is all about such algorithmic simplification under Algorithmic Information Theory:
The rigorous formalization of Occam’s Razor.
If an “AGI” can’t do scientific induction on even so trivial a scale, why attribute “general intelligence” to it?
This isn’t to say such an AI isn’t in the offing in the foreseeable future, but let’s be realistic about how we go about measuring the general intelligence of such systems.
Jeremy Keeshin
@jkeesh
In 1945, six women pulled off a computing miracle.
They programmed the world’s first computer—with no manuals, no training.
Then, a SINGLE assumption erased them from tech history for decades.
The story of how ONE photo nearly deleted computing’s female founders: 🧵
Kathy Kleiman, a young programmer, found old photos of women standing beside ENIAC—the first general-purpose computer.
When she asked who they were, curators said: “Probably just models”...
But Kleiman had a feeling they were something more:
Program ENIAC—a machine the world had never seen.
It was 8 feet tall, 80 feet long, and weighed over 60,000 pounds.
The engineers built the hardware...
But someone had to figure out how to make it do anything:
They were the world’s first programmers.
First, they were hired as “human computers” to calculate missile trajectories during WWII.
Then chosen for a top-secret project unlike anything before:
Security restrictions kept them out of the ENIAC lab.
They had to write programs using only blueprints and logic diagrams.
No manuals. No programming languages...
So how do you code something no one’s ever coded before?
By inventing the process from scratch.
They built algorithms, flowcharts, and step-by-step routines—on paper.
Then, once granted access, they programmed ENIAC by physically rewiring it...
And that’s where things got even harder:
There was no keyboard.
Programming meant plugging thousands of cables into the right configuration—by hand.
It was almost impossible to program.
But they pulled it off anyway:
Thursday 2nd April 2020 15:11 GMT
BJC
Millisecond roll-over?
So, what is the probability that the timing for these events is stored as milliseconds in a 32 bit structure?
Reply Icon
Re: Millisecond roll-over?
My first thought too, but that rolls over after 49.7 days.
Still, they could have it wrong again.
Re: Millisecond roll-over?
I suspect that it is a millisecond roll over and someone at the FAA picked 51 days instead of 49.7 because they don't understand software any better than Boeing.
Thursday 2nd April 2020 17:05 GMT
the spectacularly refined chap
Reply Icon
Re: Millisecond roll-over?
Could well be something like that, the earlier 248 day issue is exactly the same duration that older Unix hands will recognise as the 'lbolt issue': a variable holding the number of clock ticks since boot overflows a signed 32 bit int after 248 days assuming clock ticks are at 100Hz as was usual back then and is still quite common.
See e.g. here. The issue has been known about and the mitigation well documented for at least 30 years. Makes you wonder about the monkeys they have coding this stuff. //
bombastic bobSilver badge
Reply Icon
Devil
Re: Millisecond roll-over?
I've run into that problem (32-bit millisecond timer rollover issues) with microcontrollers, solved by doing the math correctly
capturing the tick count
if((uint32_t)(Ticker() - last_time) >= some_interval)
and
last_time=Ticker(); // for when it crosses the threshold
[ alternately last_time += some_interval when you want it to be more accurate ]
using a rollover time
if((int32_t)(Ticker() - schedule_time) >= 0)
and
schedule_time += schedule_interval (for when it crosses the threshold)
(this is how Linux kernel does its scheduled events, internally, as I recall, except it compares to jiffies which are 1/100 of a second if I remember correctly)
(examples in C of course, the programming lingo of choice the gods!)
do the math like this, should work as long as you use uint32_t data types for the 'Ticker()' function and for the 'scheduld_time'; or 'last_time' vars.
If you are an IDIOT and don't do unsigned comparisons "similar to what I just demonstrated", you can predict uptime-related problems at about... 49.71 days [assuming milliseconds].
I think i remember a 'millis()' or similarly named function in VxWorks. It's been over a decade since I've worked with it though. VxWorks itself was pretty robust back then, used in a lot of routers and other devices that "stay on all the time". So its track record is pretty good.
So the most likely scenario is what you suggested - a millisecond timer rolling over (with a 32-bit var storing info) and causing bogus data to accumulate after 49.71 days, which doesn't (for some reason) TRULY manifest itself until about 51 days...
Anyway, good catch.
Evangelist of lean software and devisor of 9 programming languages and an OS was 89 //
In his work, the languages and tools he created, in his eloquent plea for smaller, more efficient software – even in the projects from which he quit – his influence on the computer industry has been almost beyond measure. The modern software industry has signally failed to learn from him. Although he has left us, his work still has much more to teach.
The C programming language was devised in the early 1970s as a system implementation language for the nascent Unix operating system. Derived from the typeless language BCPL, it evolved a type structure; created on a tiny machine as a tool to improve a meager programming environment, it has become one of the dominant languages of today. This paper studies its evolution.
IBM found themselves in a similar predicament in the 1970s after working on a type of mainframe computer made to be a phone switch. Eventually the phone switch was abandoned in favor of a general-purpose processor but not before they stumbled onto the RISC processor which eventually became the IBM 801. //
They found that by eliminating all but a few instructions and running those without a microcode layer, the processor performance gains were much more than they would have expected at up to three times as fast for comparable hardware. //
stormwyrm says:
January 1, 2024 at 1:56 am
Oddball special-purpose instructions like that are not what makes an architecture CISC though.
Special-purpose instructions are not what makes an architecture RISC or CISC. In all cases these weird instructions operate only on registers and likely take only one processor bus cycle to execute. Contrast this with the MOVSD instruction on x86 that moves data pointed to the ESI register to the address in the EDI register and increments these registers to point to the next dword. Three bus cycles at least, one for instruction fetch, one to load data at the address of ESI, and another to store a copy of the data to the address at EDI. This is what is meant by “complex” in CISC. RISC processors in contrast have to have dedicated instructions that do load and store only so that the majority of instructions run on only one bus cycle. //
Nicholas Sargeant says:
January 1, 2024 at 4:00 am
Stormwyrm has it correct. When we started with RISC, the main benefit was that we knew how much data to pre-fetch into the pipeline – how wide an instruction was, how long the operands were – so the speed demons could operate at full memory bus capacity. The perceived problem with the brainiac CISC instruction sets was that you had to fetch the first part of the instruction to work what the operands were, how long they were and where to collect them from. Many ckock cycles would pass by to run a single instruction. RISC engines could execute any instruction in one clock cycle. So, the so-called speed demons could out-pace brainiacs, even if you had to occasionally assemble a sequence of RISC instructions to do the same as one CISC. Since it wasn’t humans working out the optimal string of RISC instructions, but a compiler, who would it trouble if reading assembler for RISC made so much less sense than reading CISC assembler?
Now, what we failed to comprehend was that CISC engines would get so fast that they could execute a complex instruction in the same, single external clock cycle – when supported by pre-fetch, heavy pipelining, out-of-order execution, branch target caching, register renaming, broadside cache loading and multiple redundant execution units. The only way that RISC could have outpaced CISC was to run massively parallel execution units in parallel (since they individually would be much simpler and more compact on the silicon). However, parallel execution was too hard for most compilers to exploit in the general case.
There are only two hard things in Computer Science: cache invalidation and naming things. -- Phil Karlton
Long a favorite saying of mine, one for which I couldn't find a satisfactory URL.
Like many good phrases, it's had a host of riffs on it. A couple of them I feel are worth adding to the page
Leon Bambrick @secretGeek
·
There are 2 hard problems in computer science: cache invalidation, naming things, and off-by-1 errors.
9:20 AM · Jan 1, 2010
Mathias Verraes @mathiasverraes
·
There are only two hard problems in distributed systems: 2. Exactly-once delivery 1. Guaranteed order of messages 2. Exactly-once delivery
2:40 PM · Aug 14, 2015
all the tags from https://b.plas.ml
1st-amendment 2nd-amendment 4th-amendment 5th-amendment 9/11 a8 abortion acl adhd afghanistan africa a/i air-conditioning amateur-radio amazon america american android animals anti-americanism antifa anti-semitism antiv antivirus aoip apollo apple appliances archaeology architecture archive art astronomy audio automation avatar aviation backup bash batteries belleville bible biden bill-of-rights biology bookmarks books borg bush business calibre camping capitalism cellphone censorship chemistry children china christianity church cia clinton cloud coldwar communication communist composed computers congress conservatives constitution construction cooking copyleft copyright corruption cosmology counseling creation crime cron crypto culture culture-of-death cummins data database ddt dd-wrt defense democrats depression desantis development diagrams diamonds disinformation diy dns documentation dokuwiki domains dprk drm drm-tpm drugs dvd dysautonomia earth ebay ebola ebook economics education efficiency electricity electronics elements elwa email energy engineering english environment environmentalism epa ethernet ethics europe euthanasia evolution faa facebook family fbi fcc feminism finance firewall flightsim flowers fonts français france fraud freebsd free-speech fun games gardening genealogy generation generators geography geology gifts git global-warming google gop government gpl gps graphics green-energy grounding hdd-test healthcare help history hollywood homeschool hormones hosting houses hp html humor hunting hvac hymns hyper-v imap immigration india infosec infotech insects instruments interesting internet investing ip-addressing iran iraq irs islam israel itec j6 journalism jumpcloud justice kindle kodi language ldap leadership leftist leftists legal lego lgbt liberia liberty linguistics linux literature locks make malaria malware management maps markdown marriage mars math media medical meshcentral metatek metric microbit microsoft mikrotik military minecraft minidisc missions moon morality mothers motorola movies mp3 museum music mythtv names nasa nature navigation navy network news nextcloud ntp nuclear obama ocean omega opensource organizing ortlip osmc oxygen paint palemoon paper parents passwords patents patriotism pdf petroleum pets pews photography photo-mgmt physics piano picasa plesk podcast poetry police politics pollution pornography pots prayer pregnancy presentations press printers privacy programming progressive progressives prolife psychology purchasing python quotes rabbits rabies racism radiation radio railroad reagan recipes recording recycling reference regulations religion renewables republicans resume riots rockets r-pi russia russiagate safety samba satellites sbe science sci-fi scotus secularism security servers shipping ships shooting shortwave signal sjw slavery sleep snakes socialism social-media software solar space spacex spam spf spideroak sports ssh statistics steampowered streaming supplement surveillance sync tarsnap taxes tck tds technology telephones television terrorism tesla theology thorium thumbnail thunderbird time tls tools toyota trains transformers travel trump tsa twitter typography ukraine unions united.nations unix ups usa vaccinations vangelis vehicles veracrypt video virtualbox virus vitamin vivaldi vlc voting vpn w3w war water weather web whatsapp who wifi wikipedia windows wordpress wuflu ww2 xigmanas xkcd youtube zfs