namibj

18 ms·

301.
by namibj 1y ago
Basically you can generate the next two tokens at once in the same matmul, and rollback to one-at-a-time when your generation said you guessed wrong (as that will mean the second of your pair you generated was generated based on revoked con
302.
by namibj 1y ago
> Requiring money upfront would classify the platform as an e-money institution, which is highly problematic from the legal perspective. What if you just reserve it on the card?
303.
by namibj 1y ago
After looking at Cases 4, 9, 23, 33, and 61, I think it might be suited to take in several wide-angle pictures or photospheres or such from inside a residence, and output a corresponding floor plan schematic. If anyone has examples, guides,
304.
by namibj 1y ago
`tailscale serve` is easy. Set appropriate permissions/credentials to authenticate your ChatGPT to the MCP.
305.
by namibj 1y ago
While it sounds likely true for the US, it's the opposite in Germany: likely due to societal expectations on "creature comforts" and German homes not being framed with 2x4's but instead getting guild-approved craftsmen
306.
by namibj 1y ago
`rclone mount` an `rclone crypt` over a Cloudflare R3 backend of `rclone`? Or if it's sufficiently often "off/idle", take 3 USB HDDs (1~2 years ago I bought an iirc WD MyPassport 5TB for very similar workload) into a RAI
307.
by namibj 1y ago
rclone. Just use rclone if you need to turn object storage semantics usage into an encrypted mount. It doesn't do well with non-object-storage access patterns but we're not putting an sqlite database on it here so that should be f
308.
by namibj 1y ago
They don't need to for the location to be reasonable; you can use the grid connection with reverse power flow to feed the data center just fine. Leave enough to feed the coolant pumps in an outage though.
309.
by namibj 1y ago
As a sort of TL;DR: it's basically like selection sort but instead of selecting the sort-compliant choice among the remaining ones, you ask the RNG and pick a random one.
310.
by namibj 1y ago
I thought Btrfs includes a fix by reserving about half a gig up front for that purpose/prevention?
311.
by namibj 1y ago
If Easter doesn't work for you, you can come to Evoke instead, that's (late?) summer in Cologne and not quite as big. Definitely welcoming though.
312.
by namibj 1y ago
I have fond memories of squeezing so deep exposure stacking out of the auto/adaptive-HDR-bracketing script in CHDK on my old "IXUS 100IS", that the AFAIK still CCD had severe blooming around the window in the scene. Still gre
313.
by namibj 1y ago
Materialize.com runs on those techniques; not sure how far this has gotten into the query planner yet, though. It's been a while since I looked at those details for it. I'm pretty sure some datalog (adjacent?) but otherwise quite
314.
by namibj 1y ago
Both reasons together :D But yeah you can snatch traversal statistics from your index accesses as you're executing and if those hit planner-provided thresholds you go do what the planner said to do/check in case of that specific t
315.
by namibj 1y ago
Tbf it said could, not would: the potential for watching the win is what entices; without game fixing it'd be good sportsmanship to not shit on a person's favorite if one wants anything from said person.
316.
by namibj 1y ago
There's a reason semiconductor manufacturing is so highly automated, and it's not labor cost. Humans err. Computers only err when told. But they'll repeat a task reliably without random mistakes if told what to do by a compe
317.
by namibj 1y ago
In which cloud can I book a machine with a guaranteed (up to general uptime SLA) end/termination time that's fixed for both?
318.
by namibj 1y ago
You mean like e.g. Nvidia Maxwell? (There's decent 3rd party documentation from nervana systems from when they squeezed all they could out of f32 dense matrix multiply, at the time substantially faster than Nvidia's cuBLAS library
319.
by namibj 1y ago
Given that Nvidia Maxwell / Pascal (mostly GTX 900 / GTX 1000 series) had a bit for each ISA read operand slot that said whether to cache that register file access for reuse by a subsequent instruction, and ARM and RISC-V have thu
320.
by namibj 1y ago
Sure atomic reads need to at least check against a shared cache; but atomic writes only have to ensure the write is visible to any other atomic read when the atomic write finishes. And some less-strict memory orderings/consistency mode
321.
by namibj 1y ago
Lean does feasibly let you shard out verification work; i.e., you can take a huge proof tree, chop it into an assortment of independently-proven branches (trivial if it's a true tree), let those branches be verified in parallel by a si
322.
by namibj 1y ago
Also, for long haul, you won't get by without dispersion compensation, so the lower noise from the higher signal levels getting into the amplifiers helps you. There are limits to the amplifier output power due to how tightly single mod
323.
by namibj 1y ago
a and b are about 3rd party safety. C.f. gun safe mandates.
324.
by namibj 1y ago
No.
325.
by namibj 1y ago
Yeah, I'd expect anything solar thermal to scale worse than current standard n-topcon glass glass PV panels.
326.
by namibj 1y ago
I was prepared for a chair, not..... _that._
327.
by namibj 1y ago
In practice you're actually hitting further problems because you don't have those synthetic top-k tasks but rather open-domain documents and queries to support. And if you hope to get better than "just" having the top-k
328.
by namibj 1y ago
You're conflating two things that shouldn't be: (1) the utilization factor over the obsolescence-limited "useful" life of the hardware; (2) the short-term (sub-month) training job scheduling onto a physical cluster. For
329.
by namibj 1y ago
They're VERY suitable for such trivially deferrable workloads as (much of) AI (specifically, LLM pretraining and other similar training).
330.
by namibj 1y ago
There's NCA (drones), NMC (laptops, phones, many electric cars), LFP (stationary/grid), and LFMP (many new electric cars; slightly more expensive but higher current variant of LFP).

More ›