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.

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.

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.

- 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.

- irc.example.org

- news.example.org

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

- 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.

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.

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.

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.

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

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

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.

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

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.

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 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

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.

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.