χаотичні нотатки

meantime we've downloaded a-little-bit-quantized Kimi K3 checkpoint and now trying to fit it into that setup. Which is quite a quest. 900 GiB of weights, 8x128 GiB of RAM and 8x the OS/driver overhead, plus various CUDA buffers... funny side quest is it doesn't…

btw, ntfy.sh and iTerm alerts are super convenient to track whatever long running stuff you do, and any LLM will configure them for you in no time

meantime we've downloaded a-little-bit-quantized Kimi K3 checkpoint and now trying to fit it into that setup. Which is quite a quest. 900 GiB of weights, 8x128 GiB of RAM and 8x the OS/driver overhead, plus various CUDA buffers... funny side quest is it doesn't…

We're talking about self-improving AIs replacing SaaS, meantime I had recently talked to people who introduce SaaS to businesses still running on pen, paper, copiers and faxes. They don't feel in danger at all. Like, that market is /big/ Think outside the box^W bubble, lol

sorry not sorry

meantime we've downloaded a-little-bit-quantized Kimi K3 checkpoint and now trying to fit it into that setup. Which is quite a quest. 900 GiB of weights, 8x128 GiB of RAM and 8x the OS/driver overhead, plus various CUDA buffers... funny side quest is it doesn't fit on any single 1TB SSD either :)

https://x.com/oleksandr_now/status/2084961336627253497?s=46

coming from the vanilla Ethernet world I would never have thought ordering 400G and 200G cables is such a big deal. per-vendor coding, breakout (split) cables, oh my.

dear chat, do you know smth akin to the GRC Analyst Master Class but free? asking for a friend

Anthropic: safety safety safety! nobody except us can be trusted to use AI safely! Also Anthropic: we run untested SOTA models without monitoring. I mean, without classifiers-sure, but without monitoring? This is a public statement "our values are fake we want to control _users_"

> oleksandr_now seems the natural guinea pig --Fable im flattered ahahah

80x RTX 5090s, omg. Anyone willing to sponsor 8x GB10s? The other 8 are on me, and 16 DGX Spark/Ascend GX10s should fit Kimi K3 too https://x.com/totheagi/status/2081855316443205717?s=20

(after reading the Liang Wenfeng investor meeting transcript) look ma, no 996; it's not even "google 20%" but 50% of free time and bottom-up research&exploration initiatives i def admire and look up to ppl who are able to calmly lead this way [including keeping the position that helps them to]

just loaded Codex for "/goal implement <feature> without increasing LOC. all existing tests must pass unchanged" let's see how that goes 😉 upd: well, after two hours it carved out like 200 LOC but become completely overwhelmed and blocked from adding new code, like collapsed to complete word salad: > Reply yes to approve: openActor(maxStreams) with actor ID = self-stream public key-author filtered-durable state, historical mint high-water, and tr recording each generated stream pointer-before returning its writer.

[x] just paid an invoice for the order of 0 items 🙃

emergent embodiment https://fixupx.com/voooooogel/status/2080517855049830728?s=20

continuation is a linear/affine type is a resumable program is a stack capture is an utxo yield = create, resume = spend, return = burn

sign in every few hours? wtaf? whats next, "keep your camera on so we can confirm it's still you"?

the "stop doing things because you 'should'" technique works perfect-- until I forget what I concluded is the best to do and why >_< over and over.

while Fable classifiers are terrible, the model itself is beautiful to watch -- look at what it does if you give it freedom in how to do what you ask