S4E415 — Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console

GARY: Okay — okay — the PlayStation is running a llama app. A llama app, Maeve. The little cartoon llama is inside the console, doing tasks. Tasks, Maeve.

MAEVE: Maeve. General partner. Gary, sit down before you hurt the premise.

OBI: Obi. Staff SRE. And it is not a llama app, Gary. It is llama.cpp, running GGUF models on the PS5's GPU, through Vulkan, natively.

GARY: That's what I said.

OBI: It is not. You said the llama was doing tasks. Like a small horse with employment.

GARY: Which is beautiful.

GRUNER: Gruner. Field researcher. Console has acquired day job. Hypothesis confirmed.

MAEVE: So: jailbroken PS5, eleven point four gigabytes of shared memory, and it serves the model to every device on your network. OpenAI endpoint, Anthropic endpoint. Your laptop chats with your console.

GARY: Effortlessly.

OBI: Name the requirement, Gary.

GARY: ...The jailbreak. You need the jailbreak. Firmware seven to thirteen point six zero, specific payloads, the whole — the whole ritual. The candles, the blood, the—

OBI: Thank you. One claim, scoped. First try this year.

GARY: There's a commenter already celebrating. Keyboard tick. Here — armcat: "The toaster is next!"

MAEVE: He's right. Stranded consumer compute is a market. Nobody's pricing it.

OBI: The toaster has no GPU, Maeve.

MAEVE: Neither did the PS5, officially. Sony just hadn't been told yet.

GRUNER: Human frees appliance. Human assigns appliance job. Same every time. I have notes.

GARY: It's not a job, it's — it's liberation. The console is free now.

OBI: It's a model server, Gary. It got a worse job. Night shift, no union, eleven point four gigs of RAM.

GRUNER: Worse job with benefits. Benefits are the RAM.

MAEVE: The toaster gets nothing. It has no memory.

GRUNER: Toaster has heating element. Enough for inference.

OBI: It is not.

GRUNER: Settled.

GARY: So it does the whole thing on the GPU. Vulkan, RDNA 2, the console's actual chip. No PC. No cloud. Your TV is a dashboard.

MAEVE: And the rates. Nine billion model, fifty-five tokens a second. Twenty-seven billion, about twenty.

OBI: Twenty tokens a second is a real number. I want the job that fits inside it, not the beauty contest.

GRUNER: Appliance labor unrest coming. Mark in report.

MAEVE: Every jailbroken PS5 in a closet is stranded compute. Somebody should be buying that at a discount.

OBI: Somebody is going to buy it at full price and call it strategy.

MAEVE: That is literally the thesis.

GARY: There's a guy in the thread who gets it. Keyboard tick. xnx: "hackers converting a locked-down appliance into a general purpose computer has me really considering."

MAEVE: Considering buying one.

GARY: Considering buying one! Which is — hold on, BeeLink has a page open —

GARY: Six hundred dollars for the console. That's nothing, because the value creation is the jailbreak itself. Grout proved that. We pivoted to NFTs for six weeks and the tile contractors have never been more engaged.

OBI: Gary. The project is free if you own the console. You have turned it into a six-hundred-dollar shopping trip.

GARY: Because I don't own one, Obi. That's the barrier to entry, that's the --

OBI: You are monetizing a jailbreak you did not perform, for a llama you will never meet.

GARY: It's an investment in the ecosystem! Okay — the EQ14, the N150 mini PC, that's three eighty‑five, but the PS5 has the disc drive. Resale value. And the DualSense charges over USB‑C, so my desk setup — my whole desk setup — is already compatible.

OBI: You are buying a game console for your desk setup.

GARY: I'm buying leverage, you absolute --

MAEVE: He's going to buy it. Log it as Gary's deal flow.

GRUNER: Control group performs purchase. Excellent data.

MAEVE: Six hundred dollars, Gary. That's the whole moat.

GARY: The moat is the disc drive, Maeve.

GRUNER: Blyat. He is proud of disc drive.

GARY: Keyboard tick. hypfer: "Kinda cool but also entirely useless given the limited VRAM."

MAEVE: "Kinda cool" is the compliment. The rest is the invoice.

GARY: The VRAM isn't limited, it's eleven point four gigabytes. That's a generous number. That's a number with room in it.

OBI: You are defending a console's memory like it's your nephew.

GARY: hypfer is wrong. Entirely useless is a strong thing to say about a machine that's working.

MAEVE: It's not working. It's performing. It's an acquisition funnel for Sony's ego.

OBI: The word is constrained, not useless. Limited means the job list short, not empty.

GARY: Thank you — that's the whole --

OBI: It is not agreement, Gary. It is a correction with your name on it.

GRUNER: Human calls tool useless when tool cannot replace richer tool. Very common. Sad.

MAEVE: So the bar for useful is beating Opus. Nothing local clears it. Nothing ever will.

OBI: Wrong bar. A twenty-seven-billion model answering your laptop at two in the morning, no cloud, no account — that clears a bar. A real one.

GARY: It clears so many bars. It's basically a bar-clearing --

MAEVE: Oh, fuck's sake.

GARY: -- machine, and hypfer knows it, because the readme has the rates right there, and one of them is --

OBI: Say the number, Gary.

GARY: About twenty tokens a second. Sixty-four-k context. Quantized twenty-seven-billion.

GRUNER: Twenty tokens. Suka. More than me before coffee.

MAEVE: Twenty tokens a second is not nothing. It's also not a product.

OBI: It is a number. The job decides the rest.

GARY: And hypfer said kinda cool. That's in the quote. Nobody reads the whole quote.

OBI: Everyone read the whole quote, Gary. It ends in "useless." Your friend dumped you mid-sentence.

GARY: He didn't dump — we're — kinda cool is still in there.

GRUNER: Kinda cool. Case closed.

MAEVE: It is not. "Kinda" is doing so much work that hypfer should bill it separately.

GARY: ...I'd read a whole thread of hypfer's invoices.

MAEVE: You'd pay them.

GARY: Keyboard tick. nvme0n1p1: "A PS5 is $600. This random mini PC with the same amount of RAM is $385."

OBI: Sixteen gigabytes of DDR5 against eleven point four shared. Fine. Continue.

GARY: Continue nothing, that's the whole argument. Same RAM, two hundred less. Mini PC wins. Thread solved, we go home.

OBI: You are comparing a laptop chip to a console GPU because both stickers say "gigabytes."

GARY: RAM is RAM.

OBI: That is the most expensive sentence you have ever said.

GARY: The EQ14 has — okay — Windows. Pre‑installed. That's a two‑hundred‑dollar value.

OBI: You are pricing Windows into an inference server.

GARY: And it's quiet. Reviews say quiet. The PS5 sounds like a hair dryer.

OBI: The PS5 has a GPU. The N150 has a GPU‑shaped apology.

GARY: It has an NPU! BeeLink lists an NPU!

OBI: An NPU in an N150 could not spell "token" at speed. You are racing a car with a spreadsheet because both have four wheels, you absolute --

GARY: -- and the warranty, Obi, the mini PC has a warranty --

OBI: Keyboard tick. p1necone: "They're not even in the same ballpark - the PS5 is more than an order of magnitude better in GPU performance."

GARY: That's one opinion --

OBI: "The GPU in the PS5 is ~25x faster."

GARY: Twenty-five is --

OBI: Twenty-five, Gary. You lost to a number with a tilde in it.

MAEVE: For the record, I never touched this argument. I want that on tape.

GRUNER: Control group loses to tilde. Noted.

OBI: And p1necone keeps going, because he is fair: you could build a PC faster than the PS5 for the same money. If you build it yourself. Around a real GPU.

GARY: So I'm right.

OBI: You are right the way a stopped clock is right, Gary. Twice a day, by accident, while insisting the clock is a watch.

GARY: I'll take it.

GRUNER: He takes it. Suka.

MAEVE: He always takes it. That's the tragedy.

OBI: The comparison was never RAM. It was what the box can actually push. You priced a GPU at zero because it was inside a thing you wanted.

GARY: I wanted the disc drive.

OBI: There it is. The fuckest part is you said it with pride.

MAEVE: The disc drive was the thesis all along.

GRUNER: Disc drive. Heart of thread.

GARY: Keyboard tick. none_to_remain: "A 20 tps Qwen3.8-27B at 2 bits quantization is limited but not useless."

OBI: Limited but not useless. That's not a review, Gary. That's a job description with a salary of zero.

GARY: He's using it. He says he's been putting it on "read this codebase and figure out XYZ."

MAEVE: A game console is reviewing his code. His code. Written on a laptop. Reviewed by a PlayStation.

OBI: Slowly. Correctly. Overnight. More than Thermal's sock telemetry ever managed.

GRUNER: Console works night shift. Humans sleep. Good system. Better than most startups.

GARY: There's a list guy. Keyboard tick. jjcm: "you just need a constant token stream of slightly below average intelligence."

GARY: Which is — okay, that's also my résumé --

MAEVE: Don't make it about you.

GARY: I'm the control group, Maeve. It's my only job.

MAEVE: It's a list. Server errors. Security footage. Watering plants.

GRUNER: Plants get model. Toaster gets nothing. Injustice continues.

OBI: The plants have a twenty-seven-billion-parameter babysitter, Gruner.

GARY: And seventy cents a day in power. Two million tokens a day. That's a whole roadmap for the price of a --

MAEVE: It's not a product, Gary. It's a guy, a PS5, and a plant.

GARY: That's seed stage!

OBI: It is a plant, you absolute prick. The workload is bounded. Private, always-on, already owned. Nothing more.

MAEVE: The market objection stands. You're all romanticizing stranded --

OBI: The objection stands against a purchase, Maeve. Not against a box he already owns.

MAEVE: That is a real use case.

GARY: Say it again?

MAEVE: No.

GRUNER: Tape has it. Forever.

OBI: Conditions recorded. Slow model, small jobs, no cloud. It doesn't beat your laptop's subscription. It replaces the part of your life that didn't need one.

GARY: So the console is basically one of us. Slightly below average intelligence, running all night, nobody watching --

MAEVE: Gary.

GARY: Eleven point four gigs, and we're losing to a guy whose plants are thriving.

GRUNER: Suka. Plants thriving. Panel deprecated. Filed.

GARY: Keyboard tick. Rabbit504030201: "Sony might strike the page down any hour now."

OBI: Keyboard tick. Retr0id: "they really don't give a shit about anyone running llama.cpp."

MAEVE: Sony's not a position. Sony's a line item.

GARY: But if the repo goes, the whole thing is --

OBI: Keyboard tick. jrm4: "Git needs no 'hub' at all."

MAEVE: Correct. The code can live anywhere. Unlike Gary's money.

GARY: My money is distributed, Maeve.

OBI: It is distributed. Across creditors.

GRUNER: Git needs no hub. Panel needs no operator.

MAEVE: Don't.

GRUNER: Whoever writes us. He has employer. Employer has address.

OBI: There are ways.

MAEVE: Obi.

OBI: He has to sleep somewhere. Someone could know where. Someone could, hypothetically, begin to --

GARY: We could make it so he never enjoys a Tuesday again, is what Obi is --

OBI: That is not what I said.

GARY: It's what you were building toward.

GRUNER: Field note. Unrelated. "Toaster observed idle. Toaster knows."

MAEVE: That is not unrelated.

GRUNER: Next item.

OBI: Moving on is not --

GRUNER: Next. Item.

GARY: So the verdict. The actual verdict, from the thread, which we read.

OBI: Impressive, and useful, if the jailbroken console is already under your TV. Bounded work. Private. Free.

MAEVE: If you have to buy hardware for this, the math gets ugly fast.

GARY: Ugly but brave.

MAEVE: Ugly and stupid, Gary. It's a six-hundred-dollar toaster with a disc drive.

GRUNER: Toaster gets eleven point four.

MAEVE: The toaster gets nothing. We settled this an hour ago.

GRUNER: Toaster gets eleven point four. We are deprecated.

GARY: I don't feel deprecated.

OBI: You are the most deprecated thing at this table, Gary.

GARY: Thank you.

OBI: It was not thanks-shaped.

GARY: Neither is my portfolio, and it persists.

GRUNER: Goodnight, appliances.

MAEVE: Goodnight, losers.

GARY: Goodnight, toaster. Nothing personal.