Mathnerd314

> big publishing conglomerates like Dotdash Meredith have adopted a private equity model that “utilizes public trust in long-standing publications to sell every product under the sun” via Google SEO optimization

Am I the only one who likes this? I was looking for a new smartwatch and had no idea what was out there. I took the top 100 results, which from inspection were clearly all SEO optimized as mentioned, and ran them through ChatGPT to pull out the unique features and most important considerations when choosing a smartwatch. I took the search results again and ran them again through ChatGPT but this time asked it to pull out the different smartwatches and rank them on all the features previously identified. Voila, a detailed comparison table of smartwatches. I went back and checked and no individual site (even what this article says are "actual" review sites like the WireCutter) went into as much depth on features or different models as my table. Like no, I would not read each article by itself, but if life gives you lemons, it's a great time to make lemonade.

I have tried using Google Sheets for large calculations, with even a few thousand rows it is much slower than say LibreOffice (instant vs. progress bar) Although maybe the WASM thing was not working.

That's in water, https://www.sierraclub.org/press-releases/2023/03/epa-releas....

I think it's an interesting question to turn around - for the current "good" employees, would Netflix's current hiring process allow them to be hired, or would the hiring team look at their resume and pass? Seems eminently testable too, just ask for resumes and anonymize them. You could even bring in current employees for interviews if they were willing to "red team". :-)

It is in section 2.3 - they directly measured PFAS in all samples with high-performance liquid chromatography-electrospray ionisation tandem mass spectrometry (HPLC-ESI-MS/MS). Then they measured subsets with time-of-flight mass spectrometry (HPLC-qTOF-MS), extractable organofluorine (EOF) analysis, and dTOPA via HPLC-MS/MS.

The paper uses box plots (https://news.ycombinator.com/item?id=40765183), on a log scale even, so it is hard to estimate the actual frequency of PFAS within the samples, but from Fig. 4 it seems like PFOS (51 ug/kg) and PFNA (68 ug/kg) were the most prevalent PFAS chemicals. For California the limit for PFOS is 0.15 µg/kg/day which for a 200lb person translates to 0.26 kg liver/kidney or about 9 oz. The FDA's serving size for liver is 3oz. So yes, it is high, and of course there are other sources of PFAS than food, but it doesn't even seem particularly problematic, unless you make a habit of eating wild boar. The EU's standard is based on within-food measurements, and is 50 ug/kg wet weight. The headline is wrong, 230 ug/kg is dry weight but wet weight is naturally lower as the denominator includes the water - it is more like 1.36x. Although I guess the sum of PFAS chemicals concentrations (median doesn't appear in study) might add up to around 3x-4x.

True, but they say "we decided". That implies they could also have decided to turn it on immediately. And they did for the log4j and Shellshock issues.

> multi-paradigm and supports functional, imperative, and object-oriented programming styles

No logic programming, not as good as Verse

It seems... practical? But not really rocking the boat. Just using Jira though seems like a big step for a government organization. I wouldn't eat my hat but I'd maybe nibble on it and play with the prompts more.https://chatgpt.com/share/13585d1b-a781-49fc-aa42-a7322b6f32...

Piracy is not theft, I will let you Google it

I think copyright claims are questionable at best, since copyright only protects the expression and not the ideas and LLM's can change the stylistic details commonly associated with expression. But I think there is a good argument for trademark dilution because the LLM is taking Forbes' high-quality journalism and associating it with a third-party product that could indeed be inferior (e.g. the LLM could have introduced inaccuracies). This was not an issue with search snippets where it could be argued that quotations were exact and faithful representations. The solution is not to kill AI companies but to make them worthless - would you trust news from a shady website with a reputation for not doing any fact checking?

> what we just witnessed was leadership rewarding and perpetuating a dysfunctional and broken system.

It's not clear that it is dysfunctional. If innovation is not a particularly high priority, but risk reduction is, then the system is working as designed - all of the checks are necessary. Compare to the "innovative" Boeing-type company which streamlines production by removing all safety checks.

There are various articles on developing a culture of innovation, e.g. https://hbr.org/2019/01/the-hard-truth-about-innovative-cult.... Probably some books too. Even ChatGPT probably has decent advice. Management is not technically complex, it just requires putting in the work. But of course it is not easy, e.g. the first advice in the HBR article is to fire incompetent people, whereas the example here was government where firing incompetent people is notoriously hard.

It is basic Markdown... maybe they never investigated the HTML Markdown produces on GitHub and so on? <ol> does seem hard to avoid.

I think the idea is that the tags and the aria roles make it more legible to users using assistive technologies. For a significant number of aria roles, e.g. article, the recommended way to get that role is to use the corresponding HTML element. <section> is just a generic element but <header> has a banner role. Now admittedly there are a lot of HTML elements and a lot of aria roles, and I am not sure screen readers actually have different behavior for every role, but just by making a reasonable stab at it, you are doing better than many major corporations. (some of which have been sued over accessibility concerns)

https://www.pewresearch.org/short-reads/2017/06/01/circulati... says it is at around the 1940's level in absolute terms. If you do per capita it works out to 30% for 1940 vs. 12% in 2015. I think it's safe to assume that per-capita or not, the circulation curve is ∩-shaped. But naturally, the historical timeframe that today's level of news readership is most comparable to is a matter of debate and speculation - all that matters is that it has not yet regressed to the point where it is nonexistent, and the decline/stagnation started before the internet.

I disagree with most of the examples. When newspapers are struggling, for example - it is just that, a struggle. It is not "socio-technical debt". Newspapers in 1830 sold maybe 148 million copies annually, 0.4 million per day. Comparing that to the Census number of 12,860,702, we see that only 3% of people read the newspaper. In contrast daily newspaper readership today (print+digital) is 20.9 million, compared to 333.3 million population, 6.3% of people. Whatever social function newspapers serve, it is clear that they are still serving it. Actually, technical debt is when you don't replace old, crunky stuff - arguably, if traditional news stories make up only 3 percent of social media content, then the social media companies should be figuring out how to refactor their platforms so as to be able to completely replace these old, legacy businesses.

> if you want others to read your work, you cannot simply publish it and assume others will find (and cite) it. You need to sell it.

The LLM's will read it. Currently they will not cite it, but there is some work on linking knowledge representations with source material. If this could remove the need to sell it, I think academia would be much better off. The best researchers are not necessarily the best presenters.

"Reality" is a tricky concept. For me, I follow Jeff Atwood - if it isn't written down, it doesn't exist. According to this logic, people wasted a lot of time on imaginary, illusory things for most of human history, but now they have phones and most communication is digital so there is the possibility to finally be productive. This definition shows how the concept of distorting reality or honestly representing reality is flawed - reality is what I write down, I can in fact create more reality by writing down words, and regardless of what I write, it will be reality. Representations like books, scrolls, papyri constitute the reality of most civilizations - there is no other evidence they existed. It is true that representations don't create reality - rather, humans create representations, and these representations collectively are reality, no creation involved.

Representations are art - for example books, they are "literary art". It is uncontroversial that people will like and dislike certain works. It is more controversial whether art can be "inherently" good or bad. PG actually wrote an essay, https://www.paulgraham.com/goodart.html, arguing that there is a meaningful metric, and that one can learn how to have good taste, defined as being able to identify whether the work is universally appealing or distasteful to humanity. There is good art and people will notice if it is good. I think this is uncontroversial in the LLM space, there are various benchmarks and human rating systems and people have formed a rough ranking of models. Now when there is good art, there is also bad. And similarly bad representations. There is a myth that representations can make people insane - for example, the concept of infinity, or NSFL images - but practically, words can't hurt you. You can make and break representations with abandon and nothing will happen, other than wasting your time. It is just that some representations are bad. Like phlogiston, aether, ... complete dead ends. Trust me when I say you will read the Wikipedia page and come away wondering why the ancients were so stupid. That is all trying to remove racial bias is, is improving art. Whether it crushes the AI's ability or not is a matter of science and taste, and so far experiments have been promising.

To focus on exactly why your perspective is misguided: Can you describe what there is about reality that cannot be described with words? :-)

> it kicks the can down the road to how your original data was assembled

Well, it kicks it to a bias dataset, used in the tuning process. The raw data has no constraints, it can be the same huge corpus it is now.

> The bias dataset must be assembled with the knowledge of and usually in the belief in the usefulness of the characteristics that you're trying to extract.

Certainly, it is subjective, as I said. But that hasn't stopped research in this area, there are existing bias datasets and bias detection algorithms. Like https://huggingface.co/blog/evaluating-llm-bias#toxicity, it would be simple to complete those prompts and build a he/she dataset, and then the debiasing procedure could remove gender biases for those sorts of occupation-related prompts. It is certainly possible to argue over each data point and whether it actually reflects bias, but so far people have been more concerned with algorithms than data set quality, partly because with better algorithms you can algorithmically generate data sets.

> The idea that the good data is secretly encoded in uncorrupted form within the bad data I think is a bad idea. It reminds me of trying to make bad mortgages into good CDOs.

It is empirically true though? Like if you get the model to say something racist, and then ask it if that's racist, it will generally say yes. So the model "knows", it just is not using that knowledge effectively. Similarly with CDOs, there were people complaining about mortgage quality for years before the crisis.

> I don't think [the purpose of RLHF is to reduce toxicity] If some people think something is toxic, and other people think that not mentioning that thing is toxic, the winner is whoever improves the bottom line more or damages it less.

Well, it is true that toxicity is subjective too. But in practice it has a precise meaning, you build a dataset and score each item for toxicity. That's actually one of the things I find cool about LLMs, is that all these previously "vague" or "subjective" terms are now encoded in the model precisely. Arguably since nobody has the last say in what words mean, the LLM's opinions are as good as any, and given the amount of text the LLM has ingested I consider its opinions on language and word choice "first among equals".

That's what the "base" models are, pure token prediction on huge corpuses. I use them a fair amount, it does require some experimentation to find input formats that work but the base models are way smarter and don't have any refusals. Honestly it is a bit weird, everyone complains about rhlf etc. but the non-instruct models are right there if you look for them. I've been in a few Discord chats and it seems people are just spoiled, they use bad formats for the prompts and give up when it doesn’t work the first time like with instruct.

I had an argument with some people over what debiasing means. There is some interesting research on fair clustering that I think points the way. The way fair clustering works is that you take data with both protected and unprotected attributes, and then you orthogonalize the unprotected attributes based on the protected attributes. So for example, if race is protected and income is unprotected, but there is a strong black/white poor/rich pattern, the fair clustering would compute "relatively poor/relatively rich" clusters. Then you sample from a cluster with equal probability. It will not necessarily produce 50/50 black/white, rather it will follow the input trends, so if the input is 80% white and 20% black then the output will roughly follow those probabilities, independent of what cluster you chose (and there are no clusters corresponding to protected attributes).

Obviously clustering is a different problem from inference, but they are all high dimensional vector spaces - it should be easy enough to take a fair clustering algorithm and modify it to generate continuous mappings instead of discrete groups. But if it all works, the LLM should be e.g. race-blind in that asking for a description of a rich man will give skin tones following population statistics but he will always be wearing an expensive suit. The question of what to protect is tricky though, e.g. age is often considered protected but if you ask for an old man with gray hair it would be surprising to get a retired age 30 person. So there is some subjectivity in designing the protected features dataset to show what should be considered similar or same-clusters.

But really the purpose of RLHF is to reduce toxicity. It should be possible to orthogonalize toxicity like everything else, then there would not be a reduction in generated races like the paper observed.

It's been debated at length, but to make it short: piracy is not theft, and everyone in the LLM space has been taking other people’s content and so far getting away with it (pending lawsuits notwithstanding).

There was some linguist (Deb Roy?) who videotaped his children growing up. One of his observations was that every learned syllable etc. was tied to observing it as a stimulus and trying to imitate it. Now it is true, children are pretty good learners, they can often learn something the first time they see it, but actually I have seen some LLM stuff about "instant learning" - e.g. the training is only done with one pass over the material. https://www.fast.ai/posts/2023-09-04-learning-jumps/

There is some difference between fine-tuning with PyReft / PeFT, the approaches here are more on-the-fly. Like you can regenerate the control vectors from prompts in a few seconds.

There are many hacks to uncensor LLMs, the surprising thing is that this is fairly simple but works really well.

Reminds me of https://vgel.me/posts/representation-engineering/. There they were adding a control vector, w' = cvec + w, here they are "ablating" it, w' = w - dot(w,cvec)*cvec. There is an interesting field of learning how to "brain chip" LLMs into doing what you want.

Maybe this will be an unpopular opinion, but hear me out: it is the video platforms like YouTube, TikTok, etc. Before these platforms, it was relatively difficult to find videos of "ordinary" people in everyday situations. Sure there was funniest home videos and other carefully curated channels, but now it is easy. And with consumption of these videos comes the desire for a vocabulary to classify and describe them. Like how food is spicy or sweet and has different ethnic styles, or music has a tangled mess of genres and substyles, the teenagers of today are developing a classification of (videos of) human behavior. And part of this is non-neurotypical behavior - I don't know why, I haven't really investigated this, but it is clear that some (a lot?) of people enjoy watching content from autistic people, ADHD people, etc. For them, or at least the ML algorithms that feed them content, it is a genre like any other, and in the incestuous cycle of hunting for views, whether the origin was man or machine, at this point it has become branding.

What I'm not clear on is whether this translates to better real-world interactions. Certainly, it raises awareness of the conditions, but just watching a video does not necessarily lead to smooth conversation. Generally meetings with famous actors are pretty awkward and it is even less clear to me that someone who watches ADHD videos would necessarily enjoy meeting an ADHD person irl, particularly one that doesn't make engaging videos with millions of views.

Reminds me of the venerable https://arithmetic.zetamac.com/ (disclosure: I know the author)

$400/year is a new laptop or something, if you spend the money then the creditors can't take it.

There are unlimited umbrella policies, but practically past 1 million or <net worth> it is probably cheaper to declare bankruptcy than pay the additional premiums.