spion

19 ms·

181.
by spion 3y ago
The system as described is not susceptible to prompt injection: - The tool-using-LLM never sees data, only variables that are placeholders for the data. - A post tool-using-LLM templating layer translates variables into content before passi
182.
by spion 3y ago
What does it mean to be aware of our environment? LLMs can produce knowledge about themselves and the environment. Would placing a LLM in a body that has sensors which convert the current environment into a text stream make LLMs concious?
183.
by spion 3y ago
> Although precise definitions are hard to come by, intuitively we all know what consciousness is. This is the key problem. To me it looks like most of this article is conflicted about what the definition exactly is: - basic conscious ex
184.
by spion 3y ago
If other tools handling the details just works for you, thats pretty amazing (and kinda surprising). Last time I was working on JS tooling there was a smorgesboard of various choices and it looked like most are opinionated just enough to no
185.
by spion 3y ago
Which worst-case scenario would you choose? a) the trademark getting invalidated b) creating a chilling effect which robs the community of momentum on meetups, books, tutorials or even just really enthusiastic fan/technical channels li
186.
by spion 3y ago
So a GCC implementation of Rust cannot call itself The GCC Rust compiler. Not sure that that's reasonable.
187.
by spion 3y ago
Is there a lawyer that agrees with that statement? A reminder that the Debian project found it necessary to fork Firefox into Iceweasel for a very similar reason.
188.
by spion 3y ago
Then why is the foundation adding language that signals their intention to sue people who use the trademark in domain names, learning material, packages and so on?
189.
by spion 3y ago
> someone having a trademark doesn’t mean that outsiders aren’t allowed to say or write the name. Did you read the new trademark policy?
190.
by spion 3y ago
There is a new proposed trademark policy that bans the use of the "rust" trademark in domain names, package names and educational material (unless the educational material starts with a disclamer denouncing affiliation with the Ru
191.
by spion 3y ago
There is still an opportunity for Qt to win in this area - they should make it really easy to create (well documented) bindings for new languages (that integrate with their runtimes)
192.
by spion 3y ago
Almost every single programming language ever invented has to reinvent more or less the same ecosystem of libraries from scratch. What is the best estimate of how many hours of work that is per language?
193.
by spion 4y ago
Its not an interface though: public abstract class Reader Did Java add multiple inheritance or am I missing something?
194.
by spion 4y ago
Is that the modern Java equivalent of the single-method interface in Go? I can see its an abstract class - what do you do if you want to implement both Reader and Writer?
195.
by spion 4y ago
Kind of, but not quite. There is also: - compiling to a single binary (I guess jpackage fixes this) - saner / less elaborate / more ergonomic interfaces to implement and use e.g. compare `io.Reader` ( https://go.dev&#x2F
196.
by spion 4y ago
If you look at the replies you'll see the huge variety of responses people have to this, which is absolutely fascinating. I think this might actually be different for different people.
197.
by spion 4y ago
Yes, we're shelling into k8s containers now
198.
by spion 4y ago
This is very likely the issue, as ChatGPT AFAIK is strongly fine-tuned to give answers right away rather than ask follow-up questions.
199.
by spion 4y ago
Its not pointless. Proxies (such as istio), caches or middlewares (like auth and validation) should be able to set errors without interfearing with the main application errors. The namespacing scheme allows that.
200.
by spion 4y ago
The namespace is great - it will be useful for proxies, auth middlewares, caches and other intermediaries that can intermix with the main application.
201.
by spion 4y ago
> The article is, frankly, not very interesting. You change the inputs and get different outputs? What a surprise! Here is what this argument amounts to: "Humans are, frankly, not very interseting. You change the things you tell the
202.
by spion 4y ago
> And honestly, excluding counting-related answers is pretty arbitrary. It sounds like you've decided on your beliefs here, and are just rejecting answers that don't fit your beliefs. Just because LLMs are known to be flawed in
203.
by spion 4y ago
Describing things that LLMs can't do isn't proof we understand them. I can describe things that human brains cannot do without understanding how human brains work.
204.
by spion 4y ago
You are conflating the counting problem with an inability to do anything with data it doesn't have. For its ability to invent new languages and translate to and from them, see https://maximumeffort.substack.com/p/i
205.
by spion 4y ago
I have. I'm impressed in terms of how incredibly helpful it is, but its clear its far from being a replacement yet. One example its ability to do problem-solving when diagnosing a complex issue: 1. It repeatedly ignores described const
206.
by spion 4y ago
Interesting how this story got immediately pushed away from the first 200 items on the front page (where we have items 2 days old with 50 poins and 50 comments).
207.
by spion 4y ago
Its mind-bogging amount of work to write correct algebraic types that make sense for most JSON and YAML config files. Pretty much every property is optional, but not really as some combinations are allowed while others not. If you are lucky
208.
by spion 4y ago
Why do you think jsonnet is quirky? It has pretty normal components overall: local bindings with lexical scope, imports, functions. Certainly more pedestrian / closer to most general purpose languages. The only strange bit may be thunk
209.
by spion 4y ago
This works for some languages but not others (some typed languages are particularly ill-suited to the problem of writing CLI tools to dump heterogenous dictionaries and/or objects with many optional fields)
210.
by spion 4y ago
This is how NNs are trained, not how they work. This is a good analogy that illustrates the problem with not making the distinction https://twitter.com/nearcyan/status/1632661647226462211

More ›