derefr

80 ms·

751.
by derefr 2y ago
LLMs are still fundamentally, at their core, next-token predictors. Presuming you have an interface to a model where you can edit the model’s responses and then continue generation, and/or where you can insert fake responses from the m
752.
by derefr 2y ago
No, and they can find porn on their own too. But social media services still have per-poster content ratings, and user-account age restrictions vis-a-vis viewing content with those content ratings. The goal isn’t to protect the children, it
753.
by derefr 2y ago
I would think that the risk isn’t of a human being reading those instructions, but of those instructions being automatically piped into an API request to some service that makes chemicals on demand and then sends them by mail, all fully aut
754.
by derefr 2y ago
So, it’s good at hard-logic reasoning (which is great, and no small feat.) Does this reasoning capability generalize outside of the knowledge domains the model was trained to reason about, into “softer” domains? For example, is O1 better at
755.
by derefr 2y ago
For coding problems specifically, you could get quite far by giving the model a the tool-use of a sandboxed compiler/interpreter (perhaps even with your project files already loaded into the sandbox); and then training the model to tes
756.
by derefr 2y ago
No, the call-to-action being referenced in the article is "stop archiving to hard drives" (and use tape instead, every other industry does.)
757.
by derefr 2y ago
I mean, under the scheme I outlined, the kernel is still going to do that by default. It's not that the NIC's driver is overridden or anything; the kernel would still be reading the receive buffer from this NIC and triggering per-
758.
by derefr 2y ago
The profit margin figure cited for grocers is the margin on their "main business" — i.e. revenue on grocery goods sold minus cost of grocery goods from suppliers. But this ignores other ways grocers make money (at much higher ma
759.
by derefr 2y ago
Presuming that this is a server that has One (public) Job, couldn't you: 1. dedicate a NIC to the application; 2. and have the userland app open a packet socket against the NIC, to drink from its firehose through MMIO against the kerne
760.
by derefr 2y ago
Ignoring the point about managing secrets for a moment, and just speaking about "sharing/syncing .env files" — I would like to suggest that anyone that cares about this, first gets onto using direnv(1) — and getting their col
761.
by derefr 2y ago
> if you can eat it you certainly can put it on your skin That doesn’t always follow. There are (organic, large-molecule) substances that your mucus membranes will protect you from, and which you’ll then digest (denature) and greatly met
762.
by derefr 2y ago
> Real, filtered table beers Who said anything about "filtered"? We're talking about watered-down beer here (or rather, to put a fine point on it, beered-up water.) In other words, grog, but with [whatever spirit is common
763.
by derefr 2y ago
That and you get user observability for free, and support injection in a way that to this day there’s no good way to do in an “app” experience. Presuming your bot requires people to interact with it in public channels, your CSRs can then ju
764.
by derefr 2y ago
> if your object has so many dynamically-computed properties that eagerly collecting them is a performance concern, then it should probably be a Map Should be, but there's so much legacy code that is going to hand you objects. Also,
765.
by derefr 2y ago
How about iterating them to find and flush cache entries with expired TTLs, before they're weak-expired?
766.
by derefr 2y ago
I don't get the problem. The generated code for your example would return the same function pointer + evaluated args list to the wrapping trampoline (i.e. the code that contains the loop), where that loop expects each thing it invokes
767.
by derefr 2y ago
An LLM trained on a given dataset should — at least in theory — "contain" (in a lossless-data-compression sense) a full superset of the knowledge of a Hidden Markov Model trained on the same dataset. I.e. that information is ther
768.
by derefr 2y ago
> I wonder if grandparent was thinking of a different country Nope! I was speaking of offshoring mostly in the services industry, not the manufacturing industry[1]. Things like call centers (as you linked.) Chinese companies like BBK use
769.
by derefr 2y ago
China was actually trying to heavily establish offshoring in the Philippines—but that has seemingly dried up with the increasing South China Sea tensions.
770.
by derefr 2y ago
iSH does it. Not sure how they get away with it, but the dev is around on here so maybe they can explain.
771.
by derefr 2y ago
Re: feature entitlements shared through a shared SSO auth backend and/or Group Container — yes, apps do this. Mostly this is in app "suites" where you can IAP a feature entitlement in one app in the suite, and the entitlement
772.
by derefr 2y ago
Ah, alright. There's nothing actually stopping you from doing this — it requires two things: 1. either a third-party licensing server (and thus some SSO auth system — but just require Apple's own SSO for it and it'll still be
773.
by derefr 2y ago
> They still don't support upgrades/upgrade pricing. What would this mean, exactly? You can sell people a demo→full-version permanent unlock as a one-time purchase, same as you can sell DLC in a game. And you can also have subs
774.
by derefr 2y ago
It would be impossible for .internal domains to be publicly CAed, because they're non-unique; the whole point of .internal domains is that, just like private-use IP space, anyone can reuse the same .internal DNS names within thei
775.
by derefr 2y ago
Reminds me of a proposal I've seen floating around for a Metroidvania game where you lose abilities as you progress (never gaining them back.) IIRC the idea is that it would still be a game with progressive exploration (if not necess
776.
by derefr 2y ago
> Is "musicians who don't even know the genre they'll use professionally yet" a valid market in the first place? Genre is contextual. An instrument can “sound like” one genre solo / when highlighted, yet contribu
777.
by derefr 2y ago
The goal of Unicode is to ascribe semantic, machine-readable/indexible identifiers (codepoints) to “things that appear in what are conventionally considered plaintext documents” so that “text” can be handled in a standard way across sy
778.
by derefr 2y ago
> reduces ANN expressiveness But does it? It’s been my hypothesis for a while that every grad-trained NN is hauling around a lot of “nascent” nodes — nodes that were on their way to being useful, but haven’t received enough input yet t
779.
by derefr 2y ago
But drip irrigation doesn’t scale because you would need to lay + connect + pressurize + maintain hundreds of miles of hoses. It’s high-CapEx. A “watering robot”, meanwhile, can just do what a human gardener does to water a garden, “at sc
780.
by derefr 2y ago
This agent isn't claiming to displace that work; more like act as the contribution equivalent of Tier 1 Support in a customer-service setting (where the actual project maintainers would then be Tier 2 Support). Such agents would mostly

More ›