nicklecompte

I recommend clicking through to the links on copyrighted training data - my favorite prompt being "blank-184": https://www.musicbusinessworldwide.com/suno-is-a-music-ai-co...

Also Udio: https://www.musicbusinessworldwide.com/yes-udios-output-rese...

> Despite anxiety from artists, the startups point to the number of people who will soon be able to make compelling, professional-sounding music of their own using at least some AI tools. “There’s going to be an enormous number of folks who there was a barrier to entry, economic or otherwise, that prevented them from entering music,” Sanchez said. “And we think that this is going to enable them to do so going forward.”

"I am sure you artists who made all of our training data are mad, especially considering we're bragging about billions in revenue while you get nothing. But think of all the talentless hacks who can vandalize your work by replacing the lyrics with memes! It's actually pretty selfish of you to harsh their mellow."

I know user agreements are a bit of a punchline, but when you created an account with SO presumably[1] you also agreed to waive certain copyright/ownership claims on your contributions, and in general you voluntarily agreed to benefit SO’s business by contributing there. Same as Twitter, etc. That is not the case for ChatGPT. OpenAI definitely trained it on my personal GitHub account, along with everyone else’s, yet I don’t remember accepting an agreement with GitHub that said any Microsoft subsidiary could use my code for whatever they wanted.

In the literal context of my comment, we can say that when the UN establishes a stock exchange :)

Obviously I am kidding, I meant "via the American economy" in a literal sense, via Wall Street (or Silicon Valley). I agree it makes more sense to view this stuff as a global economy with the US as simply a very big player.

The thing that makes me so sad about this: when I steal an answer from StackOverflow I always put a comment linking to where I got the answer. I could pretend that I do this because it's a good software maintenance practice. Truthfully, I only do it because it's the right thing to do. It's about professionalism and integrity.

Laundering human responses via a large language model not only makes it impossible to acknowledge SO contributors: it encourages people to think GPT figured these things out solely because it's simply so darn clever.

It doesn't help that SO's marketing is encouraging developers to not care about integrity or professionalism:

> provide OpenAI users and customers with the accurate and vetted data foundation that AI tools need to quickly find a solution to a problem so that technologists can stay focused on priority tasks.

Hey buddy, you got priority tasks to focus on. Just let the plagiarism robot do its thing.

In addition to mytailorisrich's comment, this is a vacuous description of Sweden's economy because many of its new billionaires are operating in a global technology and finance economy that wouldn't exist without US capitalism. If a Swedish company lists their IPO on the New York Stock Exchange and the founder becomes a billionaire, they became a billionaire via the American economy, not solely the Swedish one.

I suspect there is effectively a "negative externality" of income inequality, where Swedish residents can enjoy America's horrible shareholder-first corporate culture, without directly screwing over the Swedish poor. To be clear I am not trying to say that the Swedish are exploiting the US[1] and the Stockholm exchange does make plenty of billionaires (Ericsson is listed there). But you can't consider, say, Spotify to be a Swedish success story in isolation.

[1] In the precise same sense that most US corporations are not "exploiting" Delaware.

Apparently it's not a painting:

> The artwork, commissioned in 2016, is rendered in metal mesh by renowned South Korean artist Seungmo Park. Park also created a larger-than-life depiction of Olympic sprinter Sebastian Coe for Nike’s headquarters in Oregon and the façade of Gucci’s flagship store in Seoul.

I noticed the texture looked odd when I zoomed into the photo. I suspect Park did the best he could, but the medium itself seems poorly-suited for the vision.

That's not a fair or relevant comparison. The OP is about ancient Greek music: the choir is singing a song derived from very old sheet music. The link you have is contemporary music inspired by Greek tunings. It's not at all the same thing.

From the article:

  Meikeisha Scott-Parker, out of work as a project manager since January, recently cleared a second-round interview and believed an offer was imminent. But she didn't get the job because she was overqualified and too expensive, a recruiter told her.

  A few days later, she saw the same position reposted with an advertised wage that was $10 an hour less than originally listed. She figures the company either trimmed its budget or concluded that it could find adequate help at a bargain price.

  "The truth is, had they asked me, I might have taken $10 less because of the situation that I'm in," Scott-Parker says. "The market is so slow right now."

> What bothers her most is the way some employers seem to relish being able to string candidates along. One flew her in for several hours of face-to-face interviews and never followed up.

I am sure a lot - maybe most! - of this "ghosting" really is just rudeness and/or abusing power. But in my anecdotal experience I was always way too busy to do my normal duties along with handling a hiring process, and I ended up stringing candidates along without intending to - in fact I felt terrible about it! Some managers accommodate the work of hiring by taking other stuff off your plate. Other managers dump too much of the hiring process onto you, even if you're already coding for 60 hours a week (my last job really sucked).

That said, note the extremely important caveat tucked at the very end:

> More than three-quarters of the reported job gains in the last month were concentrated in four sectors: government, healthcare, construction, and leisure and hospitality. Jobs in tech, finance, law, and accounting were essentially flat or slightly down.

This seems like a strawman. I am not aware of any serious LLM critics who object to using transformers for machine translation or cleaning up English prose. LLMs are quite reliable at purely technical translation, and decent at editing provided they don't have to summarize anything.

What people object to is using LLMs to generate grist for academic paper mills, or alternatively to outsource peer review. The quality of academic papers is already pretty deplorable! I am quite confident LLMs (and excessive reliance on preprint servers) are making things much much worse.

Beyond the bad acoustics, I think the real issue may have ultimately been the conductor. Those movements are more confusing than helpful, and seemed out of sync with the song.

But like everyone else, I don't want to be mean to researchers moonlighting as musicians. She did the job in terms of depicting what they think the music basically sounded like, and it's not reasonable to study conduction techniques for 4 years just so your team can make a better YouTube video.

I suspect it's legal. The focus seems to be on face and speech recognition and I am sure MSFT's internal metrics on the AI's ability to recognize non-white faces/accents/etc are very bad.

Even with LLMs, the existence of lawyers unknowingly using ChatGPT confabulations in court suggests that police officers will do the same with investigations or surveillance. A cop who takes Sam Altman's dumb tweets too seriously might think Azure's GPT-4 service can look at an evidence report and apply its AI Magic. Probably not something Microsoft wants to defend in court.

That is interesting, but my concern is a bit different than that article - I'm worried about what happens when the astronaut stops. After they are done running, they will be at rest in a low-g environment (unlike this experiment), but their inner ear fluid will have most of the momentum it had while they were running - this is the physiological basis of merry-go-round dizziness.

On earth the fluids quickly settle and the dizziness fades. I wonder if maybe it would take considerably longer on the moon. It is probably something they can adjust to, but I was wondering.

One question I have that I didn't see addressed in the article: how does the inner ear work in low gravity? I wonder if this might make astronauts unusually dizzy on the moon because of differences in the vestibular system.

This article seems like a good survey of the general impact low-g has on the inner ear changes, but I've barely skimmed it: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8595211/

In the past we used the word "sentience" for this, but sci-fi confused "sentience" with "sapience," and now we have four terms that are meaningless, causing pointless arguments because different people interpret them in different ways.

- consciousness

- sentience

- self-awareness

- sapience

(It really doesn't help that 3/4 these are too poorly-understood to even have a good formal definition.)

Yes, it was client-side JavaScript making the range requests, asking for a string of genomic data to render in the browser. It was only to give the scientists a pretty picture :) The EC2 costs were largely ElasticSearch for a different function, which never looked at the data in S3.

> Microsoft has placed a major bet that customers will embrace its effort to incorporate AI into its security product platforms. Microsoft said, after months of testing with customers, that Security Copilot helped security analysts be 22% faster and 7% more accurate.

Good news! I am sure there aren't any deeper issues around Microsoft's ideology, arrogance, or crass financially-motivated reasoning. Otherwise I might start wondering if maybe Microsoft's security AI might not actually work that well.

This must be a regression bug in AWS's internal system. At a past job (2020) we used S3 to store a large amount of genomic data, and a web application read range requests to visualize tiny segments of the genetic sequence in relevant genes - like 5kb out of 50GB. If AWS had billed the cost of an entire genome/exome every time we did that, we would have noticed. I monitored costs pretty closely, S3 was never a problem compared to EC2.

It also seemed like the root cause was an interrupted range request (although I wasn't fully clear on that). Even so that seems like a recent regression. It took me ages to get that stupid app working, I interrupted a lot of range requests :)

Trying to come up with a grand unified theory of infidelity is just setting yourself up for heartbreak and confusion. People cheat for tons of different reasons under tons of different circumstances. A very common circumstance is called "alcohol."

No: it knows what "Jesus" "feet" "washing" "disciple" etc. look like but it doesn't know what they mean. If it understood what they meant it wouldn't have any trouble parsing an unambiguous prompt.

The system will "assemble them in the way the author is requesting" if you use an extremely detailed prompt: https://news.ycombinator.com/item?id=40217685 But the reason this prompt is necessary is that the system does not understand human language with the same depth as a toddler: AI needs way too much extra help. The problem is not that these things are poorly drawn: they are semantically incorrect.

It is like the "horse riding an astronaut" example which art generators still routinely screw up[1]. AI knows what "astronaut" and "person riding a horse" look like, so it can draw "astronaut riding a horse" just fine, using shallow reasoning. But since they don't understand what "horse" "riding" or "astronaut" mean, and no clue what "horse rides a person" looks like, their shallow reasoning fails.

But a child who actually understands human language would probably be able to draw a horse riding an astronaut - I know I could and I suck at drawing. It wouldn't be nearly as clean or detailed as DALL-E 3, but it would be semantically correct. The difference is that children understand what horses and astronauts are, and can use physical reasoning to imagine how a horse might ride an astronaut piggyback/etc. Art generators do not understand what horses are and are incapable of physical reasoning.

[1] https://garymarcus.substack.com/p/horse-rides-astronaut-redu...

I would never purchase this myself, but it seems most people are buying it as a toy or trinket, not as a productivity device.

I can't judge them too much considering how much actual grownup money I've spent on Legos...

Maybe look at it another way: ask GPT to complete the following

  Bananas are yellow due to a

  Bananas are yellow due to an

In the first case it might respond

  Bananas are yellow due to a pigment called bromelain.

In the second case it might respond

  Bananas are yellow due to an organic compound called bromelain, which is a yellow pigment.

So in either case GPT could have picked "a" or "an" without any impact on the semantic meaning of its response. In the extreme case, you could see the LLM operating according to a dumb heuristic:

  The token following "due to" is "a" with 55% probability, "an" with 45% probability.

In reality it is of course more sophisticated than this. But this dumb heuristic would explain the behavior.

And if you didn't actually include any facts about bromelain in the pretraining data, LLMs absolutely could autocomplete this with something about "an optical illusion." GPT-3 made factual mistakes like that pretty routinely, but I recall it figured out the grammatical rules of "a" and "an."

I don't think the concept actually needs to be pre-activated as you said, though I agree with faabian that this "preactivation" probably does happen in some implicit/emergent sense.

I haven't read the paper in full detail yet, but I do have a minor editorial comment: while the appendix L.2 was satisfactory, I thought the condensed argument in 5.2 was a bit too sloppy. In particular,

  H(X) + H(Y) = H(X | Y) + 2I(X ; Y) + H(Y | X)

  By discarding H(Y | X) - which appears again when predicting at the following position - we observe that 2-token prediction increases the importance of I(X ; Y) by a factor of 2.

The argument about "discarding" was not clear to me - if you're predicting the third token Z, then shouldn't H(Y | X) be contained in the implicit context C, and therefore can't be freely discarded? I don't think this argument was clarified in the appendix. But this is mostly about presentation, I wasn't so confused as to doubt the gist of the argument.

Obviously the real answer is "AI doesn't understand what 'Jesus' or 'feet' actually mean."

But I did a quick Google Images search of Jesus washing feet, and noticed that a lot of Western art depicts Jesus in a very humble manner: plain clothing, face turned towards the ground, and typically without adornments like halos, cherubs, etc. But most artistic depictions of Jesus glorify him. So I speculate that art generators were not able to learn that the Jesus who washes peoples' feet in some paintings is the same Jesus who sits on a throne in Heaven in other paintings.

Maybe "Jésus washes the feet of his students" would have gotten better results :)

To be clear there’s a distinction between the quality of the results and the determinism of the results. If a low-precision LLM is wildly stochastic but the variation is mostly linguistic rather than factual or deductive (e.g. coin tosses on synonyms or presenting independent facts in a different order), then there’s not really a contradiction.

AFAIK the determinism side of floating-point precision hasn’t been well-addressed, but it’s been a while since I skimmed those papers.

I guess the root cause of my claim is that OpenAI won't tell us whether or not GPT-3.5 is an MoE model, and I assumed it wasn't. Since GPT-3.5 is clearly nondeterministic at temp=0, I believed the nondeterminism was due to FPU stuff, and this effect was amplified with GPT-4's MoE. But if GPT-3.5 is also MoE then that's just wrong.

What makes this especially tricky is that small models are truly 100% deterministic at temp=0 because the relative likelihoods are too coarse for FPU issues to be a factor. I had thought 3.5 was big enough that some of its token probabilities were too fine-grained for the FPU. But that's probably wrong.

On the other hand, it's not just GPT, there are currently floating-point difficulties in vllm which significantly affect the determinism of any model run on it: https://github.com/vllm-project/vllm/issues/966 Note that a suggested fix is upcasting to float32. So it's possible that GPT-3.5 is using an especially low-precision float and introducing nondeterminism by saving money on compute costs.

Sadly I do not have the money[1] to actually run a test to falsify any of this. It seems like this would be a good little research project.

[1] Or the time, or the motivation :) But this stuff is expensive.

I had a very similar comment last month - albeit more ignorant and less helpful: https://news.ycombinator.com/item?id=39957153

Basically, none of these agentic / MoE / etc papers have actually compared their results to the naive baseline: since these are nondeterministic programs, Randomized Algortihms 101 tells you that if the probability of success is sufficiently high, you can improve performance simply by running the algorithm multiple times and taking the majority/plurality result.

So is MoE or agents actually more effective than doing it the dumb way? AI Snake Oil says "no." Truly bizarre that dozens of researchers didn't even ask! It made me feel like I was missing something.

It's because floating-point arithmetic isn't deterministic, which becomes salient when (speaking loosely) the difference between likelihood of two different tokens is less than the precision of the FPU.

I am not sure to what extent this effect has been quantified.

This wasn't a "metric." It was a test to see whether or not this LLM might actually be useful to me. Just like every other LLM, the answer is a hard no: using this chatbot for real work is at best a huge waste of time, and at worst unconscionably reckless. For my specific question, I would have been much better off with a plain Google Scholar search.

I wasn't talking about "state of the art LLMs," I am aware that commercial offerings are much better trained in Spanish. This was a thought experiment based on comments from people testing GPT-3.5 with Swahili.

> You can definitely query knowledge only learnt in one language in other languages.

Do you have a source on that? I believe this is simply not true, unless maybe the pretraining data has enough context-specific "bridge translations." And I am not sure how on earth you would verify that any major LLM only learnt something in one language. What if the pretraining data includes machine translations?

Frustratingly, just a few months ago I read a paper describing how LLMs excessively rely on English-language representations of ideas, but now I can't find it. So I can't really criticize you if you don't have a source :) The argument was essentially what I said above: since LLMs associate tokens by related tokens, not ideas by related ideas, the emergent conceptual relations formed around the token "cat" do not have any means of transferring to conceptual relations around the token "gato."