sieve

bcachefs exists outside of any drama. I switched my pools to it from ZFS last year and it is working swimmingly. What ZFS use case do you have that bcachefs does not solve? Only thing I can think of is zfs send/receive and I never used that.

Both bcachefs and ZFS are social problems, not technical. Nothing stops a distro from shipping ZFS as part of their kernel. Except a legal interpretation of whether CDDL and GPL2.0 allow that marriage.

One is not like the other.

We have no idea when the Rāmāyaṇa was composed or who composed it. We traditionally attribute it to Vālmīki and call him the ādikavi and the work the ādikāvyam -- first poet and first piece of poetry respectively. The entire set up presents it as the story of a man named Rāma.

The poet specifically says in the fifth sarga of the Bālakāṇḍa:

तदिदं वर्तयिष्यामि सर्वं निखिलमादितः।

धर्मकामार्थसहितं श्रोतव्यमनसूयया॥ ४

"Now I shall recite the entire thing from start to end. As it relates to dharma, kāma and artha, it should be heard without complaint/malice/envy/etc."

The neo-Western standard of written or epigraphic evidence being the only real evidence was not really that relevant in the ancient world. Hinduism itself is based on the concept of time as a cycle instead of the linearity that is common in Western religions.

So, the story could have received its present form some time in the past 3-4K years, and has been fiddled with by grammar-obsessed scribes across the country over the past 2K years, but it could encode memories from a time long before that. The story has many fantastical elements like the tale of Sagara and his 60K sons and the descent of the river Gaṅgā from the heavens, but it is much more reserved in other aspects. The etymology of the term vānara itself is debatable. Has it been derived from narō vā (is it a man/human)? Or vanam + naraḥ (forest + man = forest dweller)?

I dumped my 24V dual LA battery setup for a single 24V LFP battery for my inverter/UPS backup. LFP supposedly has a 10-15y life based on the 4-5000 charge cycles most cells claim. Sodium ion is on the way as well. My setup gives me about 10-12h of backup for critical loads (my PC, lights, fans).

I have a 5 kWp solar plant that produces ~25kWh per day, but it is grid-tied. I considered a hybrid setup with LFP battery back up, but decided to separate the two systems to avoid a single point of failure.

I have been developing my language+runtime for the past few months, using LLMs, and would like to see someone whip one up quickly, using LLMs, particularly with no PL background.

This has been my hobby for nearly two decades and it is not so easy because there are hundreds of edge cases that must be considered even if the LLM writes all the code.

My dad is extremely irritated with Android TVs as a concept. He just wants to watch something and complains about it taking a few minutes on startup to even show the screen where can can select YouTube.

He does not understand that it is just like a phone. Unless you keep it on standby (which he does not want to do: wastes power), it will boot, slowly, till all the subsystems are up and running. The low powered processors, the low speed ssds, and minimal RAM makes everything worse.

The pricing is opaque for stupid reasons. All the information is actually available for you to calculate your usage exactly.

- You pay $10/mo for Go and get $10 worth of usage per month

- Now, each model they make available has a multiplier from 1.5-8x

- So, the expensive ones like Kimi K3 have a 1.5x multiplier, cheaper ones like DeepSeek 4.1 Flash have 8x

- This gives you usage of $15 and $80 per month respectively

- You can mix and match

- So, you can use 25% of your quota with K3 and 75% with DS41Flash ... usage of $3.75 and $60 respectively

- You divide the two figures by their respective multiplier ... $3.75 / 1.5 and $60 / 8 ... and you end up with $2.5 + $7.5 = $10 which is what you paid

- You should ignore the requests table and go directly to the price/M tokens table where all the actual rates are given

- They control usage at 5h/1w/1mo intervals

- You can use 25% of your monthly quota in a 5h block, 50% in a week, and 100% in a month.

- If you were only using DS41Flash, that is $20m, $40 and $80 worth of usage respectively which is $2.5, $5 and $10 of your subscription figure after accounting for the multipliers.

It would be so much simpler if they credited 1M microdollars per $1 of subscription and quoted the per M token rates for each model in this currency.

But then Claude/Codex/StepFun and all the other subscription services are far more opaque.

The only difference is timing. You can call it groups or channels or whatever else you want. You can expose the same data via two different methods.

- irc.example.org

- news.example.org

could be the same thing. IRC would poll. NEWS would not.

- Moderated newsgroups if you really want to chat.

- You can change the back end and store everything in an sqlite3 database.

- Spam is what killed usenet as a discussion medium. Topic-specific, moderated google groups are still active and see traffic.

- People keep reinventing the same things over and over again.

It will be when done. I have a lot of stuff in the works and switch between them based upon my current interest. I like plotting and writing stories and Gemma is one of the nicer writers.

The goal is to get the 31B working on a 16GB card. Which means weights have to be < 4 bits. The 9070 XT has wmma support for int4. Which doubles compute performance on pp compared to i8/fp8. tg is limited by bandwidth.

> Using a GPU has, once, caused a driver reset that took the desktop with it.

Only once? This has happened to me so many times with LLM-assisted development. I was writing an OCR engine and pathological shapes would cause OOM that would take the display out. On other occasions, the LLM would try to load 24GB of bf16 weights onto a 16GB GPU without understanding the consequences.

Have to encode this in logic: keep some space reserved for the display, and check before allocation. But some errors are bound to slip through.

---

Inspired by DwarfStar4 and Strata, I am trying to build my own gemma4-specific inference engine designed to run on gfx1201.

It is a lot of work. You cannot simply vibecode one in two days.

They are a subscription service.

Subscriptions like these make money by relying on the fact that most people do not try to (or have the time to) extract every last dime out of their plan. The eighty people who do not use up their quota cover for the twenty that do.

That and deals with providers to use spare capacity I guess.

Yeah, I barely manage to exhaust the $60 worth of tokens. I have to think of new things to use up $1500 worth.

Thanks to competition from open models, there is a limit to how much prices can be increased though.

Have you tried OpenCode Go? $10/mo is what I am willing to spend because most of the stuff I do is hobbyist. And OpenCode Go gives you $60 worth of MS Contributor 1.3 usage for that $10. Starting today, they actually provide $80 worth of DeepSeek 4.1 Flash usage if you want to use that.

Unless you are running a swarm of agents, it is extremely difficult to exhaust these plans in a month.

We humans have the habit of falling into the trap of anthropocentric views and explanations of the universe and everything within it. The history of AI is the history of moving goalposts. I don't particularly care if they are conscious or not. Humans have a history of treating even actual, conscious, intelligent beings as chattel or food. So I don't really see why our attitudes would change if we somehow realized that LLMs are conscious.

I have been thinking about this for a while in the context of the quest for AGI. If all we want is something to take over some parts of our intellectual labor, then LLMs are already there. They can communicate with you in a language you understand and do stuff that you command them to. Often, on their own.

AGI will have its own telos. An LLM does not (at least I refuse to believe that it does).

So, yeah, you could think of LLMs as programs processing data. But there is definitely some emergent behavior you notice that emerges when the LLM runs. How you interpret all this depends on your view on consciousness and intelligence.

There is a lot of terminological confusion in the field. The fundamental difference is about how the "AI" learns. And who, if at all, supervises the learning.

Some systems are fully formed AI. Often based on heuristics. Humans have encoded the rules into the system. Others do limited learning based on pattern recognition. Still others learn based on tagged data. And LLMs don't need even that. In a stream of N tokens, the trainer has access to K+1 when it is training a model that knows 1...K and can check if the prediction was correct.

AI literally refers to any artificial thing that feels intelligent to us: does some thing a simple bash script will not. Your rules based chess engine is AI. The bot you compete against in single player Age of Empires 2 is AI. DeepBlue was AI.

ML is where things become interesting because we teach machines to learn.

NN is where we borrow terminology from neuroscience and create "neurons" that learn.

DL is where we stack neurons deeply in layers.

I am trying to make sense of all this myself.[1] So I am building a curriculum for it as I learn. So much of the teaching material in the wild revolves around torch. Makes it useless if you want to peek inside the actual machinery.

[1] https://bhashika.org.in/logs/2026/an-ml-curriculum

This is a knowledge problem. You can fix it by pointing the model to documentation (if it exists). Otherwise the model will give you the next best guess

Software has rarely been the moat. Or file formats. You have always been able to reverse engineer them. The problem, always, has been network effects.

I can build an entire, fairly useful, spreadsheet app over a weekend. But can I send my "expenses.cells" files to my accountant? Will it work with the Excel/Google docs he uses?

AI can build or reverse engineer anything as long as you are motivated enough to do it.

> The people complaining most loudly about RAM costs are the people who want to run local models

This is a tiny percentage of the population.

Samsung is cutting phone production because of RAM prices.[1] The consumer market is badly affected: budget phones, laptops, general electronics.

The budget segment of sub $100 devices in India has been almost wiped out. Manufacturers cannot afford to spend 50% BOM on RAM+storage. Unless employees are getting a 15-20% wage rise this year, I expect a similar situation in most places.

Between the engineered conflict in the ME triggering O&G price rises, and stratospheric RAM pricing, the situation is pretty bad.

[1] "There is no profit even if we sell"…Samsung to cut smartphone production by 30% (https://www.mt.co.kr/en/tech/2026/10/08/2026100709554237233)

I do it for the LE certificate. But "cheap" is subjective. Domain registration and management should not be as expensive as it is. Nothing the registrar or registry does could be worth $8/10/200 a year with guaranteed 5-10% price increases every year. It is a grift.

Work is JVM. So not much space there. But I did manage to incorporate a tiny LISP-like language into my principal application to make some configuration tasks easier. Everything outside it, I can control.

I get your usage. If you never spend enough time in slow mode, there is no point moving the code elsewhere as the savings do not justify the labor.

My issue is that I do not want to use or maintain code in 10 different languages. A single language, even one that is 3-5x slower than C would work. My only asks are that it should:

- be performant

- have basic infrastructure (cryptography, database etc)

- be aesthetically pleasing to read and write

> Any serious person should understand what CPU bound tasks are vs what IO bound ones are.

This is a decision fork only when you are stuck using a slow language. I don't have to do this --- at all --- when using C/Java and so many other languages.

Using Python as glue and writing extensions in C/Rust is a choice. But I'd rather write everything in C23 by that point. Or move to a faster language if syntax/aesthetics is an issue.

Further, you often do not have a choice on what platform you get to work with. Someone might have built an expensive-to-replace website using PHP. And then you have to build HipHop to speed it up.

Yes, the workflow matters. And it is one reason why Python is popular in the ETL and other processes. You are essentially offloading computation to efficient code while Python simply orchestrates. You probably spend one percent of the time inside Python itself, so it is pretty convenient.

I normally have one "tooling" language that I write all kinds of stuff in. And it goes beyond glue. So performance matters to me.

re: testing, I am pretty old-fashioned. I don't test everything. Only stuff that is critical or complex enough where I suspect breakage in the future. Also, I think borrowing ideas from Ada/Haskell/Clean etc for a new language could solve a lot of these pain points, now that LLMs are writing most of the code.

I looked at it briefly, but did not use it. I actually made an effort to use PyPy, but the issues I mentioned probably ensure that it will always lag behind CPython as far as language support is concerned.

The situation is tricky and the friction is considerable. It is not like JS where v8 made the existing language blazing fast with no changes to the language itself.

Would rather use something like Nim with Python-like syntax.

I have been using Python for the past two years for various things. But it is criminally slow. I joke that using Python on your modern 2020s CPU upgrades it to a Pentium 4 from 2004 running code compiled with C. And the 2004 version might still be faster.

Yes, some of the syntax is nice. But the rest of it, the scoping rules, the obstinacy around the lambda syntax, the venv/pip nonsense, path resolution etc is a hot mess. Were it not for astral tooling like uv, I would have abandoned it in a few weeks.

I hate thinking about memory outside of very specific workflows, so I am willing to accept a 2-3x penalty over C for cleaner and shorter code that follows the happy path. But anything beyond that and you are wasting energy and people's time.

For Python to work as something other than a glue language where we write all critical code in C/Rust and call it from inside Python, something like PyPy is almost mandatory. But the design of the language, the exposed innards, and the need to maintain backward compatibility makes optimization a chore.

I did some benchmarking of Python against C, Go, Rust, Node, LuaJIT etc in preparation for my runtime.[1] My observations based on some stats:

- You can blindly replace C with Rust for most tested workflows. There is very little performance difference outside of compiler speed. (C23 is a nice language though.)

- Go is 2-3x slower than C

- Node and LuaJIT are 3-8x slower on average compared to C

- Python is 60-70x slower. It can get far, far slower in certain cases.

You can run your own benchmarks. I think you might end up in the same ballpark.

[1] I have been interested in reverse engineering, hobbyist compiler/VM development etc for a couple of decades and wondered what is the point of ranting for two years about it without doing anything concrete. So I decided to build a runtime and statically typed language borrowing stuff from Python and Erlang/BEAM. But, even with LLMs implementing my ideas faithfully, the last 20-30% is a never-ending process which means I get bored and move on to something I can finish in 3-4 days.

It is written in TypeScript, not JavaScript.

TS probably has the most expressive type system of any language that I have used, and you can develop at lightning speed without fighting the borrow checker or anything else. The ability to share the exact same code across the front and back end, and encode API contracts in the type system, is a superpower for webapps built with node.

Whether an LLM writes the code or not is besides the point. What matters is that the code should be testable, and understandable. TS wins on both counts, like most sane alternatives.

I personally do not program using any language that does not have the ability to specify types.

> How will skill building happen when LLMs and coding harnesses seemingly provide all the answers at one's fingertips?

They don't provide all the answers. And are often wrong/mistaken. But they also produce stuff faster than you can consume it. This is great for when you know exactly what you want and can cheaply test the output and guide the LLM to the right solution. It is expensive (in developer hours) in other cases.

I have been building a new PL+VM this year and had to pause development on a few occasions because I have not had the time to go through the implementation because there are so many modules. But smaller tools (5-10KLOC), you can judge fairly quickly.

So, the answer to your question is: people need to discover the strengths and weaknesses of the tool on their own, decide what kind of expertise they want to attain, and if it is worth the effort.

My concern is something fairly boring but important nevertheless: your bills/invoices, receipts, credit/debit notes, contract notes, trial balances, ledgers/statements etc.

Internally, you may use whatever precision/scale you want, and compute using doubles or decimals or integers. But I want to see 100.01 - (37.29 + 41.63) = 100.01 - 78.92 = 21.09 in all of these statements.

Accounting has the concept of materiality. But you still want the figures to tally to the last cent.

One way to do this is to separate the computation/storage and presentation layer: use integers to store values as cents for computation/storage and only convert to dollars+cents on display. But then someone might suddenly demand three digits after the decimal point and you cannot go about changing every stored value just because the logic changed in one part of the application. So, using decimals is the safer choice.

I prefer integers for my plain text ledger software though because the format is in my control.

> Most interfaces we use on a day to day basis are the opposite, they're designed to be accessible

I read this and then I think of glass doors with handles on both sides that I don't open till I read the instruction on the affixed plate: PUSH/PULL. The door forces this decision on every single conscientious user. These things are everywhere, including software interfaces.

All great points.

We installed a 5 kWp plant last month. Grid-tied, of course. Our daily usage is in the 12-25 kWh range depending on the season and the plant produces ~12 kWh on the worst days (~35 on the best, so far).

We have been doing our cooking almost exclusively on induction for the last 6-7 years, and I have been trying to get others to switch as well. However, it is difficult to convince people of the benefits. Some Indian cooking requires exposing the food directly to the flames, so piped/bottled gas cannot be completely eliminated.