HNHacker News
TopNewBestAskShowJobs

lblume

488 karma · joined November 15, 2024

submissionscomments
lblume··on Training language models to be warm and empathetic makes them less reliable
Empirically, there seems to be strong evidence for LLMs giving factual output for accessible knowledge questions. Many benchmarks test this.
lblume··on What does it mean to be thirsty?
> And no, there wasn't anything medically wrong with her.

Do you mean physically? The behavior surely sounds compulsive...

lblume··on Monero appears to be in the midst of a successful 51% attack
https://xcancel.com/p3b7_/status/1955173413992984988
lblume··on Try and
Prescriptivism is inherently limited to situations where near-unanimous consent exists between all speakers.
lblume··on Try and
I fail to imagine how one could use "try and" in a resume, without actively trying to.
lblume··on Yet Another LLM Rant
This is interesting, because I wouldn't be too sure about that. Whether I am able to play chess well exclusively depends on the strengths of my opponents, because there is no absolute baseline. If society somehow decided that what you were writing wasn't "music" anymore, you would be the only person left stating you had that skill.

I believe that most claims about one's own skills do come from outward judgement and interpretation of one's actions, not introspection. The only thing humans have is, to radically appropriate the jargon, a way longer context window (spanning over an entire lifetime!), together with many ways to compress it.

lblume··on Yet Another LLM Rant
Accuracy certainly does play a role, but this in itself is not sufficient for preventing an infinite regress – how does one determine the accuracy of a source if not by evaluating claims about the source, that themselves have sources that need to be checked for accuracy? Empirical inquiries are optimal but often very unpractical. Reputation is accuracy as imperfectly valued by society or specific social groups collectively.
lblume··on Yet Another LLM Rant
The position of Kant does not align with the direction modern neuroscience is heading towards. Current evidence seems to prefer decentralized theories of consciousness like Dennett's multiple drafts model[1], suggesting there is no central point where everything comes together to form conscious experience, but instead that it itself is constituted by collaborative processes that have multiple realizations.

[1]: https://en.wikipedia.org/wiki/Multiple_drafts_model

lblume··on Yet Another LLM Rant
How is this relevant? Modern LLMs don't rearrange words at all in a meaningful sense of the word, and are certainly better than just using random chance to sample tokens.
lblume··on Yet Another LLM Rant
With chronic liar being a severe anthropomorphization by itself due to assuming some level of intent by the LLM, correct?
lblume··on Yet Another LLM Rant
It seems that on some level, you have to in order to not just constantly reflecting upon your thoughts and researching facts. Whether you trust a given source should surely depend upon its reputation regarding the validity of its claims.
lblume··on Yet Another LLM Rant
Expected Value?
lblume··on Yet Another LLM Rant
Do you really know what you don't know? This would rule out unknown unknowns entirely.
lblume··on GPT-5: Key characteristics, pricing and system card
Doesn't seem too likely, world knowledge is very valuable.
lblume··on LLM Inflation
Ambiguity is resolved via provided context, but just as with conversations, this context may be severely underspecified.
lblume··on Do LLMs identify fonts?
Why would you even need LLMs for that? Notwithstanding context, finding text that fits into a given bounding box is already perfectly doable via a classical algorithm (in this case e.g. based on dynamic programming).
lblume··on Tesla must pay portion of $329M damages after fatal Autopilot crash, jury says
I tend to agree, however the government is not an unincentivized incentivizer. By being able to impose such fines, the government is potentially itself incentivized to not prevent these accidents for they potentially cause this kind of revenue.

There are ways to mitigate this, such as forcing the government to use these revenues in a way that is relevant to the issue at hand, i.e. creating safety jobs, strengthening control authorities, or something else.

You could also say that the amount is insignificant, but that could of course change with every lawsuit, and it of course accumulates. Or one could speculate that the interests are not really monetarily aligned at all (e.g. prisons), or that the judicial system is independent enough to stop propagation of these incentives. I think it is still needed to consider and try to controlledly align these motives between the relevant actors.

lblume··on Tesla must pay portion of $329M damages after fatal Autopilot crash, jury says
In most jurisdictions, a corporation is a juridical person[1]. When not explicitly mentioning natural persons, whether a corporation is a "person" is thus ambiguous.

[1]: https://en.wikipedia.org/wiki/Juridical_person

lblume··on Tesla must pay portion of $329M damages after fatal Autopilot crash, jury says
Density is fundamentally incompatible with car-friendly cities and a large amount of road. In general, the denser a city becomes, the more people will go by foot, cycle, or use public transport, especially in Europe.
lblume··on Tesla must pay portion of $329M damages after fatal Autopilot crash, jury says
The difference is that the tube driver (assuming there even is one) can't just decide to take a slight turn from the planned route, while the car driver really might.
lblume··on The Math Is Haunted
Correct. The only way to not subject oneself to (economic) irrationality is to model your beliefs about the world using probabilities and using Bayes to update in light of new evidence.
lblume··on Fast
Correct. This is also known as a rebound effect [1], or, specifically with regard to technological improvements, as the Jevons paradox [2].

[1]: https://en.wikipedia.org/wiki/Rebound_effect_(conservation)

[2]: https://en.wikipedia.org/wiki/Jevons_paradox

lblume··on Fast
> you have to assume

We don't have to assume. We know that JavaScript is slow in many cases, that shipping more bundles instead of less will decrease performance, and that with regard to the amount of content served generally less is more.

Whether this amount of baggage every web app seems to come with these days is seen as "necessary" or not is subjective, but I would tend to agree that many developers are ignorant of different methods or dislike the idea of deviating from the implied norms.

lblume··on Fast
The fact that this article and similar ones get upvoted very frequently on this platform is strong evidence against this claim.

Considering the current state of the Web and user application development, I tend to agree with regard to its developers, but HN seems to still abide by other principles.

lblume··on Fast
At least in theoretical computer science, often, but that's another matter entirely.
lblume··on It's time for modern CSS to kill the SPA
Yes, SPAs should store simple user state in the URL as well.
lblume··on Nullable but not null
In that case we read it entirely differently — the last paragraph explicitly mentions inferring "Nullable fields that never hold nulls are often the result of incomplete migrations or forgotten cleanup steps". The author also proposes a script "to identify cases where the field could safely be changed to non-nullable". But alas, we cannot do that with full generality, due to there being a big difference in intent that cannot be captured purely by counting nulls, and surely not by only calculating percentages.
lblume··on Nullable but not null
Sure, if one just leaves a column nullable due to negligence, one should check the actual requirements and enforce them to make invalid states unrepresentable. The author still makes too strong of a claim that becomes detached from the migration aspect, insinuating that one can just empirically test the database to check whether this is the case, to which I disagree.
lblume··on When photography was born, fascination, obsession, and danger followed
A counterpoint: photography enabled integrity and authenticity of images, which fueled journalistic credibility, in general becoming one part in fostering a high-trust society. AI image generation leads to these advances being slowly eradicated as generated images increasingly cannot be differentiated from unfabricated evidence. In this light, photography was socially progressive, while AI generation of photorealistic images could be seen as regressive.
lblume··on Nullable but not null
> But a field that is nullable in the schema and never null in practice is a silent lie.

This seems to be the central claim. But just as lies require intent, so does database design to some degree.

A column that is nullable but never null does not conclusively say anything, really. That's like saying a birthday column that is never before 1970 in the current data should be restricted to years after that date.

A nullable column signals that the data may be left empty, which is entirely different from that column actually having to be empty with the current data. Is-ought distinction, the potentiality of null ("nullability") is not to be confused with the actuality of current vacancy ("null values"). The programmer extracting data from the database must check for null if the data is nullable, not just if is null now as a principle of robust fault-tolerant code.

← PreviousPage 3 of 10Next →