HNHacker News
TopNewBestAskShowJobs

lsy

3,257 karma · joined April 28, 2015

submissionscomments
lsy··on An In-depth Look at Gemini's Language Abilities
I don't think "accuracy" is going to be the defining feature of which chatbot succeeds. People just aren't using them for tasks where a 3-5 point difference makes the grade, because the difference between 67 and 100 is more important than the difference between 64 and 67. If you can integrate a relatively speedy bot somewhere people can use it conveniently that'll get more usage than a slightly more factual response you have to tab out to.
lsy··on She Gave a Hotel a Mediocre Review. Then Came the Badgering
Ultimately a professional reviewer wouldn't be so easily swayed by this kind of sleazy behavior, but this is the price for having a stack of "reviews" for every gin joint in every town in the world, all written by Joe Public. Ultimately I think the old system was better: for most restaurants or trips, the "cost" of finding out whether you personally like a place is relatively small, and for bigger outlays we had the Michelin guide, Zagat, or what have you.
lsy··on The origin of the law of torture: A cautionary tale
This isn't true under current US jurisprudence: if a law that infringes on a fundamental constitutional right can be shown to be narrowly tailored to, and the least restrictive means of achieving, a compelling state interest, then it passes strict scrutiny and is considered constitutional.
lsy··on Prompt engineering
On the "humans need prompt engineering too": well, kind of. I hope that we are not all working on the precise verbiage to get the plumber to fix our toilet correctly. There are also actions we expect others to perform without any communication whatsoever. Human actions take place in a social web that includes incentives and accountability as well as expectations on the part of interlocutors around what level of detail is required for communication. Sometimes this very clearly breaks down. But the fact that interpersonal communications sometimes require elaboration or precision doesn't negate the differences between an LLM's outputs and the actions of people who are engaged in active coping with the world around them.
lsy··on Prompt engineering
What makes LLMs somewhat unique as a software product is that there is little-to-no separation between input and instruction. In most cases, the user's input can also be considered part of the "prompt". This leads to the well-known "prompt injection" "vulnerability" which is really just a byproduct of the fundamental inability (and indeed undesirability) for the model to distinguish instruction from input (undesirable because the value and flexibility comes from allowing the user, rather than a programmer, to specify an action).

On top of it though, it introduces a sort of disciplinary sloppiness around whether the program can be reasoned about. It's assumed that prompt `P` works for whatever input `I`, but the concatenated `P+I` is really the full input to the program that produces a desired output. But the only way to be confident about the program's behavior is to exhaust the input space, as no `P+A` tells you anything about how `P+B` will behave. This makes it difficult to leverage an LLM in any process where the desired result is 1. unknown and 2. matters. If the result is unknown it's not clear how to determine mistakes or correct them if they're made. And if the correctness of the result matters it's courting disaster to connect it to a process which is not able to be reasoned about. I think that's why LLMs are primarily being used to assist ideation (which is cool!) or "spammy" use cases like third-tier customer service or listicle generation, and haven't yet broken into any use case where they need to be reliable for complex tasks.

lsy··on Police have undermined the promise of body cameras
Body cams were always a band-aid solution on top of over-arming police and granting them almost total immunity from charges of misconduct. The resulting mentality of us-vs-them and mafioso, bullying behavior from police is all down to feeling a cut above “civilians” and being confident in their power to legally murder anyone who’s too resistant.

This behavior also discourages any thoughtful people with integrity from applying, and beats it out of them quickly when they do. The streets would be a lot safer with more citizens performing a genuine public safety role with fewer weapons and chips on their shoulders, but the police lobby essentially holds cities hostage by monopolizing the public safety role and strategically withdrawing from that role in protest whenever an insufficiently deferential policy is passed.

lsy··on A Camera, Not an Engine
I think the camera analogy is much more apt than the engine analogy, so this article was intriguing, but unfortunately the author falls into the rhetoric they decry when they complain about equating black holes with God, which is to just pull out a flourish like "monkeys at typewriters all the way down" to casually step over the difference between intelligence and mere computation, or to equivocate what is observed in neural networks with "intelligence" at all, and then by way of disclaimer to then dismiss intelligence completely. It would be much more interesting to discuss how machine learning gives us a lens into latent structures in data, and where that could be applied in the future, or even what that learning says about what we don't know about intelligence. And I was hoping that's where this would go, but the article seems to be attempting some grand theory of everything that's unlikely to have a practical application, which makes it difficult to discuss without first blocking out the millenarian aspects.
lsy··on Denied promotion for having a smaller headcount
I don't know, to me this sort of makes sense. I agree with the underlying critique I've seen elsewhere that headcount alone is often considered sufficient cause for promotion in tech — rather than what was created with that headcount. But I also see why a company might have a tough time promoting someone to "Director" who manages seven reports, regardless of their impact on the company. Ostensibly they want someone at that level to be capable of a certain amount of management that goes beyond 1:1s and into value-setting and overall strategy, so that when they want to laterally transfer they can take a director position on another project. And it's pretty hard to prove that with seven reports. The general reaction here is that Copilot made a lot of money etc, so Director is commensurately well-compensated. But probably companies should just be better about rewarding people for work within their current role rather than giving folks senior management titles and responsibilities merely for working on hit projects.
lsy··on Nvidia's earnings are up 206% from last year as it continues riding the AI wave
I don't think it's exactly "luck", the trends are being driven by the technology, which they have done a good job of inventing. For better or worse, low-watt parallel compute is the primary area of innovation in compute hardware, so naturally the areas of innovation in software are more likely to leverage this innovative hardware. In other words, when all you have is a ton of tiny processors, every problem looks like a massively parallel process. It does raise the question of what we are missing because we don't have e.g. super-fast single-threaded processors or whatever new tech exists in an alternate universe.
lsy··on San Francisco Streets with Andrew Callaghan [video]
One issue is the failure of the straight-and-narrow status quo to appeal to normal people. When you can't work an entry-level or low-skilled job and afford housing in the area you live in, or when working the job you'd need is more undignified and unappealing than hustling, you're going to hustle to make things work. What came through in the video to me was desperation for dignity, to be seen as a controller of your own fate. For most of the people in the video their "legal" alternative is to be underpaid and mistreated, to have their free time totally consumed by labor and traveling to and from it, and to have nothing left to show after the basic costs of life. It's no wonder they turn to a criminal life where boldness and recklessness earn you either respect or hatred. Lower rent costs and better jobs would fix most of this issue with no recourse to the legal system.

On the other hand, criminal justice in the US is anything but. Voters have decided — I think correctly! — that mere theft and drug use don't warrant the kind of punitive, criminogenic long-term caging and low-grade torture that constitutes our penal system, but there is simultaneously little appetite for improving conditions or exploring alternative rehabilitation so that those going down this road can be firmly turned around or given better choices. So the current situation is that our society, given two bad options, is experimenting with accepting a somewhat elevated crime rate in the interest of reducing over-incarceration and the mass human suffering it creates.

Ultimately my feeling is that you pay for everything one way or another. The stolen laptop and broken window are coming out of the extra money made selling a home in an area with artificially constrained housing supply. People save on property taxes but spend more on Ubers to avoid the train full of poor and needy people who have been screwed by the same system. IMO creating a spectrum of options, rather than the constant debate between extremes, would let people narrow in more strongly on the best balance.

lsy··on I disagree with Geoff Hinton regarding "glorified autocomplete"
That's the difference though. I know my world model is fundamentally incomplete. Even more foundationally, I know that there is a world, and when my world model and the world disagree, the world wins. To a neural network there is no distinction. The closest the entire dynamic comes is the very basic annotation of RLHF which itself is done by an external human who is providing the value judgment, but even that is absent once training is over.

Despite not having the bird's sense for electromagnetic waves, I have an understanding that they are there, because humans saw behavior they couldn't describe and investigated, in a back-and-forth with a world that has some capacity to disprove hypotheses.

Additional modalities are really just reducible to more kinds of text. That still doesn't exhaust the world, and unless a machine has some ability to integrate new data in real time alongside a meaningful commitment and accountability to the world as a world, it won't be able to cope with the real world in a way that would constitute genuine intelligence.

lsy··on I disagree with Geoff Hinton regarding "glorified autocomplete"
This is of course sometimes true, we take shortcuts to minimize cognitive effort. However, when the situation warrants it we think more carefully about which words to use to achieve goals, or to reach correspondence with the situation at hand. Or we move more precisely and carefully to do something we haven't done before. I've no doubt that an LLM can approximate whatever model of language a person has from their life experience, but I don't think this type of model is capable of active coping, making judgments, or of having accountability to the world it's meant to operate in.
lsy··on I disagree with Geoff Hinton regarding "glorified autocomplete"
The difference is that "the world" is not exhaustible in the same way as Go is. While it's surely true that the number of possible overall Go game states is extremely large, the game itself is trivially representable as a set of legal moves and rules. The "world model" of the Go board is actually just already exhaustive and finite, and the computer's work in playing against itself is to generate more varied data within that model rather than to develop that model itself. We know that when Alpha Zero plays a game against itself it is valuable data because it is a legitimate game which most likely represents a new situation it hasn't seen before and thus expands its capacity.

For an LLM, this is not even close to being the case. The sum of all human artifacts ever made (or yet to be made) doesn't exhaust the description of a rock in your front yard, let alone the world in all its varied possibility. And we certainly haven't figured out a "model" which would let a computer generate new and valid data that expands its understanding of the world beyond its inputs, so self-training is a non-starter for LLMs. What the LLM is "understanding", and what it is reinforced to "understand" is not the world but the format of texts, and while it may get very good at understanding the format of texts, that isn't equivalent to an understanding of the world.

lsy··on On Cultures That Build (2020)
While I agree that much of what is wrong with society boils down to a lack of civic participation and familiarity with institution-building, what I think this article perhaps misses is that social and technological changes militate against the effectiveness of a robust civil society. As conglomerates and governments can expand their reach through communications and management technologies, it becomes more feasible for institutions to manage large jurisdictions at a distance without intermediaries. And that same ability to grow to a large size also prevents smaller organizations from arising and seeing any effectiveness in their efforts.

I also disagree that Silicon Valley represents any kind of solution to this issue. The centralization of power in the Valley is almost more extreme than in the "old world", as multinational companies dominate the industry and destroy competition. Most people will do better for themselves financially by joining the elect of the fiefdoms of FAANG than by fruitlessly attempting to build a smaller company or product, and the big companies keep it that way through massive networks of financialization and regulatory capture.

In order for civic and social institutions, or small self-built institutions, to be attractive to Americans, people who join them have to have some experience of succeeding in their aims, and I think this is less and less the case, as the institutions that already exist suck the oxygen out of the space of potential action.

lsy··on A Zoomer explains her generation's malaise to older generations
The summary is that traditional sources of meaning — marriage, friends, patriotism, career, childbearing, and faith — are now devalued, and cell phones offer in their place corporate solutions that are harmful.

As can be expected, Haidt and the poster argue that solutions are found at the level of the cell phone and limiting access to it, rather than improvements at the social level where the boomer, gen x, and increasingly millennial generations continue to fail to deliver meaning through the aforementioned institutions.

If Gen Z could look to society and especially older generations for

- interpersonal relationships based on cooperation rather than competition

- work that was meaningfully compensated rather than exploitative

- governments that were responsive to their constituents rather than self-serving

- social support systems that reward rather than punish parenting, or

- religious institutions that delivered compassion and moral goodness rather than prejudice and contempt,

then I think it could be argued that folks could find meaning in these institutions instead of behind a phone screen. Until then the complaints about social decay are almost insulting.

lsy··on [dead]
In order to derive a useful conclusion, an experiment typically needs to have some controls and ways of removing confounders. This article describes several merely-nominally-related situations and then bundles them all together to reach a too-neat conclusion.

As an example, the article undermines its own argument when discussing China's strict approach. It talks about what happened after lockdowns were lifted as though it meant that the lockdowns themselves were ineffective, and compares that to the US's poorly-adhered-to stay-at-home orders. But if a country is able to keep deaths low until a vaccine is developed, that decreases the total number of deaths. Whether that militates for lockdowns as strict as China's, I seriously doubt. But the article is so intent on making the (clearly false) point that lockdowns were totally useless that it cannot help but fail to grapple in any thoughtful way with the intensity and form of mitigations.

lsy··on The case against AI everything, everywhere, all at once
Aside from the social risks and centralization of control, I think the rush to convert everything to LLMs puts quite a bit of our progress in computer science at risk should it prove to be successful. Over the last several decades we've built up more complex and precise notation and systems to allow for repeatability, reliability, composability, etc. And large companies are currently attempting to trash all of the field's progress on notation in favor of talking to the computer in English (a rich but highly error-prone language) and then having the computer do impenetrable transformations guided only by pretty-please prompts with the hope that some high percentage of the time the result will be usable.
lsy··on Yoshua Bengio: We need a humanity defense organization
"I have the impression that it might not take a lot to fix what’s missing."

The history of the field is full of people saying this, and then eventually realizing that it does indeed take a lot to fix what's missing. Right now I think we have stumbled on a novel and interesting result, which is that if you statistically collate all of the text in the world then a computer can also generate text that encodes many of the statistical relations present in that text. But I have yet to hear a plausible argument for how this translates to intelligence except that people have an impression that it's "close", never mind that we don't have any stable definition for what it's close to.

lsy··on Advent of Code 2023's new AI/LLM Policy
Maybe I'm a dinosaur but I'm utterly mystified by folks who are interested in making the leaderboard or even solving AoC puzzles with AI assistance. To me it feels like going to a 5k charity run, taking a motorcycle to the finish line, and parading your gold medal around. Like, who are you impressing? And are you even proud of yourself?

I understand the appeal of using the best and fastest means when there are real-world stakes — taking a motorcycle is better than running for many transport scenarios. But the whole point of this is to have fun! Is it fun to hit <tab> over and over?

lsy··on LLMs confabulate not hallucinate
LLMs do not confabulate or hallucinate. They predict the next word in a sequence based on correlations derived from a lot of training data. Sometimes that data happens to be structured enough to statistically generate a word that is fluent and corresponds to what we as readers understand as reality. Sometimes, due to the unclear nature of "truth" and the necessarily incomplete data used to train the model, the word or words that are statistically generated are merely fluent but do not correspond to reality.

The answer to "What are LLMs but humans with extreme amnesia and no central coherence?" is that they are not like humans at all, and only resemble them superficially due to our anthropomorphic tendency to infer a mind upon seeing fluent and roughly correspondent text.

lsy··on Language Models Represent Space and Time
"Multimodal" models only promise an additional layer of indirection and interpretation from the world through sensors into the same conceptual mechanism. An image is converted to tokens, a sensor is converted to coordinates, and it is all fed back into a correlating model. But this isn't adequate either. The famous dress illusion picture has RGB values closer to gold and white, but the actual physical dress was black and blue. The difference between perception and reality is something we don't have a handle on when it comes to these models.
lsy··on Language Models Represent Space and Time
The interpretation of the LLM having a "world model" is a big stretch in terminology. Encoding a coherent or accurate set of spatial coordinates for places is not equivalent to understanding the actuality of world space and the relationship between items in that space. By this definition an accurate spreadsheet of the lat-long of every major city would also constitute a "world model". It shouldn't be surprising to anyone that a model trained on (likely many) such spreadsheets would also encode that coordinate data. What's notable is that it still takes a team of human researchers to plot these raw numbers onto a map and to interpret them as experiential differentiators in physical space.

There's no doubt that an LLM can uncover whatever is structurally included in its training data, even if that encoding is implicit. What's less believable is that the LLM somehow achieves a grounded understanding of physical aspects of experience purely from streams of raw text.

I think the paper engages in some equivocation here, as the abstract differentiates between "an enormous collection of superficial statistics" and "a coherent model of the data generating process", without admitting that the first should well imply the second. But that doesn't then further imply a "world model" in the sense that we understand it as sentient beings. For us, an internal model is useful, but when our model or texts don't agree with the actuality of the world, the actuality of the world takes precedence. For an LLM there is no distinguishing between its trained representations and an actual exterior place, or even any sense that an exterior space exists.

lsy··on Study Finds Hybrid Work Improves Mental Health Compared to Remote or In-Office
The study finds a statistically significant correlation but that’s hard to disentangle from all the factors that make it possible for people to work remote or hybrid. It also would be interesting to see how this result varies across number of days spent in the office. My guess would be that seeing coworkers 1-2x a month would get most of whatever social benefits hybrid has if any.
lsy··on Chomsky explains why nobody is a moral relativist
But what you are saying is just that the “range” of morality is handed down by nature. And the idea that this range is “naturally” limited is too often used, as it is in your comment, to make an argument from nature’s authority in favor of certain applications of power that may or may not be good for people. What I think relativism does is restore accountability to the individual and the society for moral decisions, rather than appealing to natural or theological authorities to absolve people and institutions of responsibility for their choices.
lsy··on Chomsky explains why nobody is a moral relativist
It's weird to hear Foucault described as "amoral" when he himself claimed to be a moralist [1].

What is meant by moral relativism in this case I think is merely that morals do not come to us handed down by God or Nature as Chomsky claims, but are a constant choice made by people over a potentially infinite range. But to assert that the morals don't come from God is not to say that morals are not valuable or worth having.

As Foucault puts it, "What is good, is something that comes through innovation. The good does not exist, like that, in an atemporal sky, with people who would be like the Astrologers of the Good, whose job is to determine what is the favorable nature of the stars. The good is defined by us, it is practiced, it is invented. And this is a collective work."

1. http://www.critical-theory.com/read-me-foucault-interview-in...

lsy··on New AI experiences across our family of apps and devices
I feel like this is missing the point of the kinds of interactions they're trying to simulate with AI. People like when actual celebrities appear in their interactions (like recording turn-by-turn directions) because it's fun to imagine the celebrity actually saying these things. The variety of stickers, memes, and emojis are fun because you or your friend had to make a creative leap from something they wanted to say to the particular sticker, or even use a common sticker in a creative way. When everything can be generated at a click I suspect the value largely evaporates. Are we all going to be enthralled by sending increasingly specific autogenerated stickers to each other? What is the value of the social interaction when anyone in the chat could have created the hiking "merit badge"? Does it still hold any charm to hear a celebrity's voice when you know it was generated by a computer?
lsy··on Fandom can't decide if leaked songs are real or AI-generated
The behavior of the sellers described in the article makes me think it's all but certain the songs are AI-generated. Imagine telling a jeweler a diamond looks fake, and they first pull out their handy diamond faking machine to show you how bad a fake diamond looks in comparison to what they're selling, then curse at you and close the store.
lsy··on Leaked Google pay data shows how much employees report making
Does BI not have anyone on staff familiar with statistics? The map with “minimum salaries” seems to just prove that a more populated office location will contain a broader range and more extreme outliers of self-reported salaries. A previous article on this topic claimed that Software Engineer had the highest max salary but there are probably just more people with that title at the company than any other title.
lsy··on Proofs based on diagonalization help reveal the limits of algorithms
def unc(algstr) { if (algstr(algstr) == 0) return 1 else return 0 }

unc(unc) -> what does this return?

lsy··on Critical periods for language
Regarding the drop in learning speed after school, I have a pet theory that primary and secondary education are less about teaching people specific details of educational topics, but rather giving them a wide range of meaningful vocabulary in their native language to talk about those topics and others. As soon as you enter the workforce or even a college education, your range of meaningful topics to discuss dramatically narrows as you focus on a specialty or profession.
← PreviousPage 5 of 9Next →