dvt

30 ms·

511.
by dvt 2y ago
Function signatures already solve that problem. We have all kinds of functions that take two numbers that mean different things (width/height, x/y, etc.). Branding seems like a solution looking for a problem. I just think it'
512.
by dvt 2y ago
> I'd question whether people even need to know OrgId is a string. I mean, with numbers it's even more confusing (this might be a TS bug?): type SpecialNumber = number & { __brand: "SpecialNumber" }; let
513.
by dvt 2y ago
> And once your own trusted code has made an OrgId, you don't need any runtime checking to see if it actually is an OrgId. Right, and once I have a verified OrgId, I'll just keep using the `myOrgId` variable throughout my code,
514.
by dvt 2y ago
> The primary point of branded types is to allow you to use the type system to ensure that a particular runtime check has taken place. I see. This is kind of cool, though the branding can still be broken via down-the-stream mutations. Wo
515.
by dvt 2y ago
My point is that the API consumer does not guarantee types (say it's REST or something), so the assumption that the string you send it will always be the right type (or call it "format") seems like a bad one. Unless you contr
516.
by dvt 2y ago
While you're right that branding makes passing arguments more ergonomic, I will still always be able to do `as OrgId` or `as UserId` so you need to do have some failure handling anyway, unless you're okay with blowing up in the
517.
by dvt 2y ago
Weird idea, as types in TS are structural by design . If this is something you need, it smells like "runtime checking" not amending the type system.
518.
by dvt 2y ago
> If someone happens to be your child, you're happy to give them positions they don't deserve even if they're completely incompetent. Yeah, I'm being a bit contrarian & spicy for the sake of argument (don't h
519.
by dvt 2y ago
> You seem to be presenting nepotism as if it's a feature when it's obviously a bug. I guess that's where we disagree: in my view, it's definitely a feature. When I have kids, I will 100% be willing to give them oppor
520.
by dvt 2y ago
> People failing with golden parachutes, and others failing upwards into even more lucrative roles. I read stuff like this all the time, but this take is actually extremely reductive. (Otherwise, every moron out there would be making 7 f
521.
by dvt 2y ago
Currently working on an AI product (essentially a OS-context-sensitive agent caller, basically a "smart" Spotlight for Windows/macOS), and from my experience (~2 months of working on this basically full-time), complex prompts
522.
by dvt 3y ago
Yes, I use fastembed-rs[1] in a project I'm working on and it runs flawlessly. You can store the embeddings in any boring database (it's just an array of f32s at the end of the day). But for fast vector math (which you need for si
523.
The OneJoker Project(lcrocker.github.io)
1 points by dvt 3y ago|0 comments
524.
by dvt 3y ago
> I spent it on sweet, stupid things that make life an appropriate journey. Little things with some elegance attached. Things only worth buying because they have little value in the long run. Like life, itself. Feel like I can oddly rela
525.
by dvt 3y ago
Jodorowsky's The Holy Mountain
526.
by dvt 3y ago
> Well, its 2AM, some dork has checked in code which breaks production, and it absolutely positively has to be fixed by 6:00am before the customer comes in. This is a classic XY problem. The issue isn't the data format, it's th
527.
by dvt 3y ago
Being as charitable as possible here, and Rerank 3 might be the bee's knees, but the examples are absolutely awful . Do you really need to use embeddings + a large language model to search for "action" and "Christian Ba
528.
by dvt 3y ago
Candle already exists[1], and it runs pretty well. Can use both CUDA and Metal backends (or just plain-old CPU). [1] https://github.com/huggingface/candle
529.
by dvt 3y ago
> Are they practical and useful, or still facing significant limitations? They are. Working on a product using a fine-tuned Mistral-7B-Instruct-v0.2 model and it's pretty mind-blowing. Works flawlessly on my RTX3090 and serviceable
530.
by dvt 3y ago
> Prompt engineering doesn't feel like an activity that creates sustainable AI advancement. Agreed, it should really be rolled into fine tuning. If you're building a model for PR, for example, it should already be fine tuned so
531.
by dvt 3y ago
This is super cool. Had no idea UEFI APIs were so readily-available and well-documented. Awesome work! Out of curiosity, what was the dev cycle like? I assume you were running the thing in a VM. Did you have to "boot up" every tim
532.
by dvt 3y ago
> that prevents the optimizer from performing constant propagation I understand how it works, and why you would even maybe want this, but this just straight-up seems like a mistake. It breaks const-invariance, which I guess is no longer
533.
by dvt 3y ago
> We’re building some “smart search” functionality for some teams and I start to wonder if a traditional search results list (i.e. sans the LLM, or used only ti rewrite the user query) with the document chunks wouldn’t be better than bli
534.
by dvt 3y ago
I think you may be misunderstanding what fine tuning does. It does not teach the model new knowledge. In fact, Meta has a paper out that argues you only need a data set of 1000[1] to achieve pretty good alignment (fine-tuning) results. (1
535.
by dvt 3y ago
Still a lot of work to be done before that point, but I'll write at least 1000 prompts (some synthetic, some hand-crafted) based on the "arrow notation" I posted here[1]. The fine-tuning itself is actually quite easy[2] as lo
536.
by dvt 3y ago
I've been working on this for the past month, and in its current state it runs some actions, but it's very context-dependent. For example, if I'm on a GitHub page and I say "clone this git repo," it works pretty f
537.
by dvt 3y ago
Whoops, meant to just make it unlisted, fixed.
538.
by dvt 3y ago
I'm very excited about work being done in this area. In fact, I'm working on a product that does exactly this (runs in the background, local LLM, has access to screen-space entities, can take certain actions). It feels pretty ma
539.
by dvt 3y ago
The paper[1] seems promising, is there a fine-tuned model available? Or will we have to fine tune Llama-7b or Mistral-7b ourselves? [1] https://arxiv.org/pdf/2311.05657.pdf
540.
by dvt 3y ago
> Once we have models that can act as full senior SWEs.. the models can engineer the models. This is such an extremely bullish case, I'm not sure why you'd think this is even remotely possible. A Google search is usually more v

More ›