MangoCats

This is business. So was the Opium trade.

In the early 90s I coded up a fairly massive data analysis suite entirely in C. Around 1997 I translated it to C++ - the resulting data analysis ran at 1/2 the speed of the C code, but was 300% faster to modify, expand and maintain. And processor speeds were doubling every 9 months anyway.

Recent experience, developing in Opus 5.5. Spin up an agent to review the security aspects of the project, write a report. Pass the report to another Opus 5.5 agent and tell it to implement mitigations for confirmed problems found in the report. Implementation agent proceeds to make tests for the described exploits and gets shut down by the guardrails: Opus 5.5 is not permitted to “do that kind of work” had to scale back to Opus 4.8 for the implementation, though 5.5 was still happy to describe the chained attack scenarios, apparently it’s not allowed to develop demonstration code as it was trying to do for the test cases.

Me, checking the config file blame for who put in that one bad line… blame: you wrote the entire config file across the space of 2 months a few lines per commit - except that one bad line that somebody else put in…

I’ve been making my “the one” for years (off and on) now. I’m getting closer with github.com/MangoCats/Lempi - but it’s still got a big cost up front that I’m under-motivated to address because I’ve already paid (segmenting songs, trimming lead-in lead-out points, etc.) The thing I like the best about Lempi is that it does “shuffle play” with a reasonable intelligence, auto-selecting a good subset from my collection based on example songs that I want to hear “things like this” during certain timeslots of the day. I can just let it run anytime / all the time like a radio station and not think about choosing my playlist.

I’m working on Lempi, which is both similar and different - mostly on how the DJ works: Rufin’s Auto DJ says “more like this track” and trusts the server’s idea of similar. Lempi works like a radio programmer with rules: it decides how often each song may come back, what fits this hour, and what follows well from the last track, then adds a controlled amount of chance. github.com/MangoCats/Lempi I’ve been off and on developing Lempi for personal use since 2020 - it has other names in the past, most recently changed becasue I was developing under a name somebody else “went wide” with in 2023. There’s a lot of music players out there.

The US has been cutting per capita emissions for 20 years, while China is still increasing. But they haven’t caught US yet! United States: The average American emits roughly 14 to 15 metric tons of carbon dioxide per year. China: The average Chinese citizen emits about 8.7 to 9.2 metric tons of carbon dioxide per year. Next question: what’s the average CO2 emission per capita of Billionaires? A billionaire’s total average carbon footprint is roughly 1.9 million to 3 million metric tons of (CO_{2}) equivalent ((CO_{2}e)) per year. So, Billionaire’s income ratio is around 500,000:1 of an average US citizen, but their CO2 emissions are only about 170,000x - that makes Billionaires almost 3x greener with their greenbacks than an SUV commuting wage slave.

It depends on the size / complexity of what they’re trying to port, of course, but it’s not at all unusual for a programmer to be the main income source for a median household and to take a year on a project. Or, be a member of a 5-10 programmer team that takes 2-4 years to develop a project. In a “butts in seats” office world like we mostly had before 2020, that was a LOT of commute miles driven to make that software, way more than $1000, and energy consumption for the households and the offices and their HR, accounting and other administrative overhead…

I don’t think it’s coming, but if they really do “go there” what it will look like is: extreme tariffs on “open” processors, $1000 for a Raspberry Pi where you can get a new Intel Core NUC class system for $500. And, IMO, it would be worth the premium for the freedom of operation, but, then, you’ve also painted a target on your forehead by being one of “those nutjobs” who are willing to pay so much more for something that “only criminals need.” I will say: the US is more likely to take away our right to free computing before they take away our right to privately own firearms.

So, yeah, LLM goes off down weird rabbit holes - and if you, like Alice, follow them down… that’s on you. If, instead of getting off on how weird the response was the author just prompted - “I’m looking for the basketball memes about the NBA player from Turkey.” I suspect it would have gotten him what he was looking for. If, instead, he shone it on and started playing along like he was distraught over Dario, I’m sure he could have led it to give him a suicide counselling hotline in 2 more prompts.

Maybe you (the Claude porter) should have written a few requirements and specifications to test to first? 1998 a company bought our software that had been developed (part time, by a single programmer) since 1991, first in C/DOS, then refactored into Borland VCL/C++ in 1997. Buyers declared they would refactor it into MFC/Win 32 because it was “easier to hire good MFC/Win 32 programmers than VCL programmers.” So, they went and found two “good MFC programmers.” They assessed the job carefully and declared that it should be about 6 weeks to do the port - just a straight copy of the existing funcfionality into the new API - no need to add features or change anything. 3 months later, they hired 2 additional MFC programmers. The new assessment was that they had made “good progress” and would be done in another 2-4 weeks. 3 months after that, they declared that they had big plans for future development of the software, and they were “90% done with the port” but they’re going to “build up the team” and they hired another 3 programmers with a position open for what would be the 8th member of the team, but they were “looking for just the right candidate” for that spot. 9 months after starting, I casually asked one of the programmers how the port was going and he said: “we’ve got about 80% of the original functionality working in the new system, and most of the development is starting to turn toward newly identified business needs.” “So, you don’t need the other 20% of the original feature set?” “Oh, no, we need that, it’s core to the business case, it’s just taking time to make it happen in MFC.” These were “good programmers” - no turnover, management happy with their efforts. The original programmer was me, straight out of school, no management or mentor. The spec came from the company owner one feature at a time as ideas hit him I’d basically write them down on an electronic “sticky note” and when it was working to his satisfaction we’d throw that sticky note away, the code was the documentation. Eventually, the industry started requiring documented testing, so we hired an intern from the local college and she made some documents recording her ad-hoc testing of what the owner verbally described to her as the things the software should be doing, but mostly she read the code as her specifications. This was actually state of the art in the 1990s in most places I interacted with. About 18 months before the 1998 buyers came around, we had another company try to reproduce about 10% of the original program functionality independently as a module in their existing software; this was agreed (insisted by the buyers, actually) to be done without our code to model from - just using our front-end hardware to collect the data. Instead of building a team, they hired a series of programmers, usually in 2s, and they’d usually quit after about 6 months. I think it was the 5th set of 2 programmers there that finally got that feature working. So, 320 minutes would have been a pretty impressive port effort, but did you really expect the chainsaw to cut the whole forest down and load it onto logging trucks for you?

I’m evolving a “music player” system which basically consists of a fleet of individual players, each with a full copy of the music library because, even at today’s storage prices, why not? Storage for 10,000 songs at higher fidelity than a brand new LP vinyl album costs less (adjusted for inflation) than an album of 8 songs did back in the 1980s, and only 2 of those 8 songs were usually any good… So, have a music player in every bedroom, living room, family room, kitchen, car, patio - turn them on when you want, synchronize them when you want - split off and listen to your own playlist when you want - and every player contains a synchronized backup of all the music so you can just drive off with one, go completely off-grid and never interrupt the playlist. The only network traffic is from your phone interface to choose your playlist or synchronize with players nearby - or across the planet.

Echo chambers work best in isolation…

I will say, we’ve been deployed on Ubuntu 18.04 for 8 years now and tracking the CVEs against our configuration watching for potential exploits that could really matter to our systems, and it wasn’t until the past few months that those exploits actually started to get meaningful ability to compromise us. What that says to me is: the LLMs are finding meaningful vulnerabilities that “the best minds” missed for years and years. And they’re also generating functional patches for these vulnerabilities far faster and more easily than the patches came out in the past. I’m not loving the changed landscape, but I have to acknowledge that LLMs are making a significant difference.

/home and /var are the main targets for persistent storage. Unfortunately, a LOT of applications color outside those lines, so you’re into a bit of case-by-case for what else needs to go into persistent storage, depending on what you run and how you configure it.

be FOR the people, not against them? What are you, a Communist? How could being for the people ever maximize profits? It’s not “against” the people, it’s extracting value from the people, that’s the holy grail of all corporations: maximizing shareholder value, and customers aren’t shareholders, customers hold the value that’s to be maximally extracted / transferred to the shareholders. It’s all corporations’ fiduciary duty to their shareholders to do this, literally bound by law. /s

My bet is the whole story is fabricated. It’s very plausible, but lacking any real detail.

Not acknowledging the costs of a programmer’s children is the ultimate late stage capitalism.

Check LibreElec, if it ticks all your boxes, it’s probably the best (simplest) choice. You can go much more general purpose, but if you don’t need general purpose then all that general purpose capability amounts to (intentional or accidental) attack surface you don’t need.

Try again: Monsanto, Nestle’, Dannone…

In the 80s it was “market rate.” In engineering my TA position paid $14k a year + free tuition. The money was a little less than half of what I could earn in industry at the time, but free tuition (and an actual job offer instead of an open-ended job hunt) made it attractive enough to keep me around for a Master’s degree. They offered me a PhD, with a raise to $18k/yr for the TA-ship, again right at 1/2 of what my MS was worth on the open market - I had had enough of academia, no thanks. Only took 6 weeks to find a “real” job. Over in the business school, they had a lot more students competing for the TA positions, over there they paid $4k per year + free tuition. Seems that more business students come from families where their expenses continue to be paid by family through 6+ years of college. I co-taught a lab with one other TA vs about 15 students, and we helped grade papers for the lecture class… could we have been replaced with AI? Maybe, but would AI stand over your shoulder and ask you “what happens when I pull this wire out of your breadboard?” then talk you through the reasons why your quick and dirty design may work in the lab, but it doesn’t meet design specifications and will likely fail in real-world use where it’s used millions of times a minute for years instead of 3 or 4 times for a bench test - pulling that wire is the same as your design being completely incorrect, and yet: your bench tests are all still passing… think about that.

Vibe coding / agentic workflows result in poorer code quality, and relaxed oversight practices. This was much more true in 2025 than it is today. Recent models and harnesses are improving quality and implementing more strict development practices than were commonly used by human development teams. If you want thorough Requirements and Design specifications, full automated test coverage and CI gating based on automated test outcomes, zero compiler warnings, etc. LLMs are doing those things now more by default without explicit and repeated prompting. If you are choosing to not human review and test the outputs of the LLMs, that’s your practices slipping, not the LLM agents’.

I still think that giving it oxygen is a mistake, but if you insist on making that mistake, I’d go for “coloring the air” around it in an easily recognized shade and odor: Omarchy: the new Linux distro labeled “Goose-steppy” by the community, is being developed based on donations from X, Y and Z… At present, it offers nothing of value vs these established distros: ____, _____, _____. The community is relatively un-knowledgeable and un-helpful, and the technical aspects are less well refined than the established competition. Etc. etc. etc. Remind the world that the trains don’t actually run on-time any better under Fascists in practice. They have too many conflicting priorities, and ultimately those lead to the trains not running at all.

en.wikipedia.org/…/List_of_Streisand_effect_examp…

Mirror, keep a clean copy of your setup offline. Make periodic backups. If you ever get into an unpleasant state, nuke it and go back to one you like better.

What this feels like to me is documentation that should have always been present for all the developers on “the team”, but now that the developers are AI agents they have the bandwidth to be able to write that documentation instead of just shipping the pull request with a terse “it works, check the tests.” note. As a human developer, I never had the patience to write really good documenation AND keep it in sync with the project as the code evolved through development. I have written good prospective documentation, I have written decent retrospective documentation, but keeping either in sync with the implementation as the implementation evolves during maintenance is a heavy burden - one that LLM agents do without complaint. Yeah, they’re less than perfect, so are the humans on the teams. As long as additional passes invested in making things self consistent are consistently improving the state of things, I’m fine with telling the agent “take another pass at making the documentation consistent with the implementation while I go to this meeting…”

fascism genreally is stupid Strange thing about sociology / political games: stupid works. In the runup to world war II both the UK and USA were studying Fascism as a potential way of “getting more” from their populations to compete better in the coming conflicts…

Honestly, I think it’s a lot of Stockholm Syndrome. I spent a couple of weeks on the recently opened East side in the summer of 1990… the locals’ concepts of economics were - alien - like discussing a summer breeze with an octopus.

I don’t know if Google finds itself as defendants in a court room right now where the influence of an LLM lead to killing or suicide. This is really tough… so many suicides happen every day, happened every day long before AI or even the internet “was a thing,” it would be interesting to track trends to see if AI is making any macro differences positive or negative over the past couple of years (of course this will be impossible to tease out from variables like changing economic and social stresses…) We have a friend whose son was struggling with degenerating epilepsy, increasing seizure frequency, etc. He used ChatGPT to write his own obituary, then died of “an accident associated with a seizure” a few weeks later… I can’t really blame ChatGPT for his death, in a way it helped him communicate with the people he didn’t see or talk with very much in the months before his death, explain his perspective… What they are doing with Android has me this close to switching to an iPhone. If Apple weren’t an order of magnitude more revolting than Google, I’d think about switching. If I do ever switch off of Android/Google it will be to a smaller competitor, not the other bad choice.

when I buy a LoRa array and a ham license. And talk to who? You (and I) and the 0.001% can ham it up, but the rest of the world will be living in a different bubble.

By all means, exercise that spine. Will you also be ignoring the new CVEs if they’re discovered using AI?

Personally, I like the short - punchy approach: Omarchy, the distro also known as Goose-steppy.

SEO is a never-ending arms race, and you’re right - it’s the real thing that spoiled the “golden web” we remember. Search engines have actually improved, but SEO has “improved” more.

You can run several low power systems with lots of capabilities for less total power than one general purpose system. Now- at some point there are applications the low power systems are either annoyingly sluggish at, or altogether incapable of, but if you’re not getting into that kind of stuff, then a handful of small systems can be the ticket. I built a wooden box with a filter and a big (200mm) fan and put 4 raspberry pis + a network switch in the filtered airflow, the pi that serves stream analysis on frigate has a co-processing hat, the rest lighter weight stuff like homeassistant, libre elec, and so on. Definitely more work to setup than a single Intel Core NUC - but also about the same or lower power under load, has been rock solid for years, and the filtered air setup has kept dust out much better than my series of dust strangled NUCs over the years have.

The concept here is that the USB carries the immutable OS image, you make as many copies of it as you need because it’s a fixed image, boots the same every time, “your data” gets stored elsewhere.

Couldn’t be much worse than what we have today in the real world…

Caveat - I’ve never subscribed to “super Netflix” with HD and all that - we display on 4K monitors, but the Netflix feed is their basic resolution - I don’t know if there’s any problems going up-market.

How long do you think it would take for “special industrial licenses” to be required to get “Linux accessible” CPUs?

Try an AI prompt like: Give me a 2006 style Google search result of the modern web for the search “Apple Pie Recipe” They can still do it, if you ask.

Aggregate, because they’re courting billions of potential users, the energy and water consumption of AI agents is concerning. One project? For a continuous 320-minute (5.33-hour) project using Anthropic’s Claude Fable, the estimated electricity consumption ranges from 203.2 Wh to 406.4 Wh, while the associated water consumption ranges between 0.39 Liters and 0.77 Liters. computing.co.uk/…/new-dashboard-estimates-energy-… It takes about 100-150ml of diesel fuel to generate 406 Wh of electricity, you’d be lucky to get an efficient commuter car to travel even 3 miles on that much fuel, and the driver will consume more than 0.77 liters of water during a day, especially if they take a shower. The growth has been explosive, and the newly added datacenter capacity isn’t being done well or sustainably in a lot of cases, but if you assume that OP’s job could be done well with assistance from Fable if he put in 10x the effort, that’s still likely to land on the net lower impact to the environment than if he commuted to an office for a week to do the same. The real question, even before LLM “AI”, is: is this a job worth doing? Right now, a lot of people are “playing” with AI doing things they never could attempt before. Some small fraction of that is useful - how small is the question, and how long before we slow down on the frivolous uses?

So, the music player rename migration is done - I don’t think too much broke in the process. The source repo is here: github.com/MangoCats/Lempi In theory, it speaks for itself. In practice, this has been an evolving project and I believe my current automatic song end finder for fade outs is not as good as a previous generation that I built most of my collection with. You can edit each passage’s lead in / lead out points separately and it’s not a half bad editor, but it’s still clumsy and very time consuming. If I were introducing this to a new music collection, I’d fire up an LLM agent, have it read the project documentation, tell it where the player (Pi) is on the network (give it ssh key access, and have the music files in-place on the Pi for it to find), and tell it to go through the database induction / setup process. That may take most of 24 hours on a collection with ~10,000 songs… Once it’s setup, you should be able to define the kind of music you want to hear at particular times of day by choosing 5 to 8 representative songs, the “Program Director” will then queue up similar songs automatically and play them through. I’ve just barely got the “multi-room” setup going, it has several rough edges, but once it gets synced up it appears to be working pretty well. If you like to play with LLMs for coding, this is an interesting project to try them on. I find it’s right about at the threshold of what a model like Claude Opus 5 can handle / make forward progress with. If anything about how it’s working bothers you, feel free to ask your LLM to smooth those rough edges in ways that make sense to you, you can at least maintain your own branch and maybe we can merge together a “best of both” situation someday.

Just use secret keys You mean symmetric keys? They’re great, but they aren’t terribly practical in MANY application scenarios.

CVE stands for Common Vulnerabilities and Exposures - this is the reporting system that describes problems with various versions of various software and operating system components / configurations. They generally don’t publish them until a fix is also available. These are the things that tell people: if you don’t upgrade your XYZ an attacker can take control of your system, execute arbitrary code, without you doing anything except connecting to a network the attacker has access to… AI is finding, and fixing, CVEs at about 5-10x the rate that people were finding them without AI 2 years ago. For me, it’s an example of: if you don’t participate in AI things and make the extra new patches coming out, the attackers who do use AI things are going to pwn your systems with them. Doesn’t matter if you’re a professional developer or not, if you use a computer, you need to apply the patches and updates, if you “opt out of all things AI” you’re just going to be vulnerable. Stuff like that doesn’t leave a good way to opt out, the environment is changing and if you don’t go with it, you’re going to be victimized by it.

any kind of intelligence A photocell that turns on a streetlamp after sundown, or when a dark cloud passes over the sun, is a kind of intelligence.

it requires a prompt As delivered to public chat interfaces, yes. They all eventually stop and ask “how am I doing?” Structurally, theoretically, nothing prevents them from running an infinite loop and continuing to act without that checkpoint. They could be given a goal to “optimize” whatever and just act and act and act in what their pattern-matching systems judge to be actions pursuing “optimal” results. There was an article about “the genie effect” where protagonists are “tricked” by being given something they technically asked for but not what they really wanted. LLMs carry the additional “risk” of not matching input requests to any recognizable pattern, just following their own path which makes no sense to us. Focused on different priorities.

I’ve been streaming Netflix on Chrome on Ubuntu without trouble for years and years - maybe 10-12 years ago there was a bump in that road, but it has been smooth since.

they made me pay import fee on the value of the listed items instead of on the price I payed (there was a 40% discount). All this discount / special deal complexity needs to stop. Now the tax-men are scamming us declaring everything to be taxable at the retail rates when nobody pays the retail rates. One price, for everybody, all the time. One tax rate, for everything, all the time.

It takes a input set of tokens and generates a single output token. It adds that output token to the input set and ingests that set to output a single token I read your tokens, letter by letter, sometimes chunks at a time, then neurons in my brain fire muscular impulses to output a single letter on a keyboard, then another and another and another until you can read this response. Sometimes my brain will compose the whole sentence before starting to write it, but more often it comes out in chunks - tokens. That’s all my brain is doing right now, other than keeping my organs in homeostasis, background planning regarding envrionmental preferences, my next meal, next bladder and bowel eliminations, etc. That’s all you do, too. It’s just a bit harder to tease out the 1s and 0s in your chemical signalling processes.

Strangely, I haven’t read a single headline about AI making breakthroughs in Quantum computing… but you know certain players are working in that direction. Back in the 80s they were teaching that cryptography protected secrets “for a time” and how long they need to be protected dictated how hard you go at the key length and other costs of en/decryption. By the 00s we could easily and quickly run key lengths so hard (for conventional computation) to crack that they were essentially protected “forever” - at least until quantum algorithms shake things up a bit. That’s the next unknown: when Shor’s algorithm will start delivering real world results - like the end of BTC.