Fripplebubby

> the model just reflects those answers

I think there is a lot happening in the word "reflects"! Is it so simple?

Does this mean that the model takes on the opinion of a specific lit crit essay it has "read"? Does that mean it takes on some kind of "average" opinion from everything? How would you define the "average" opinion on a topic, anyway?

Anyway, although I think this is really interesting stuff and cuts to the core of what an LLM is, this paper isn't where you're going to get the answer to that, because it is much more focused and narrow.

I don't know what your operating definition of "grifter" is but for me, a grifter is not a company that delivers a product which gains a large adoption and mindshare (ChatGPT) and essentially sets the world on fire. (not an opinion on Altman specifically but OpenAI at large)

I think what they did here was go text prompt -> generate multiple 2d views -> reconstruction network to go multiple 2d images to 3d representation -> mesh extraction from 3d representation.

That's a long way of saying, no, I don't think that this introduces a component that specifically goes 2d -> 3d from a single 2d image.

To be more concrete: you might decide not to learn about RNNs, but still find them lurking in the things you did learn about!

Ah, that makes sense. So, we consider two hidden layers more as "memory" or "buffers", and actually the rule is implemented in just one layer, at least for a single token.

To further problematize this question (which I don't feel like I can actually answer), consider this paper: "Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention" - https://arxiv.org/pdf/2006.16236

What this shows is that actually a specific narrow definition of transformer (a transformer with "causal masking" - see paper) is equivalent to an RNN, and vice versa.

Similarly Mamba (https://arxiv.org/abs/2312.00752), the other hot architecture at the moment, has an equivalent unit to a gated RNN. For performance reasons, I believe they use an equivalent CNN during training and an RNN during inference!

Love this post! Gets into the details of what it _really_ means to take some function and turn it into an RNN, and comparing that to the "batteries included" RNNs included in PyTorch, as a learning experience.

Question:

> To model the state, we need to add three hidden layers to the network.

How did you determine that it would be three hidden layers? Is it a consequence of the particular rule you were implementing, or is that generally how many layers you would use to implement a rule of this shape (using your architecture rather than Elman's - could we use fewer layers with Elman's?)?

Maybe you're being a little tongue-in-cheek about the crypto thing, but, assuming you are at least a little sincere: I think if you drop the benefit of hindsight and change the proposition to "changing careers to one that you are well suited to (apparently) which the market favors" vs "gambling your net worth with no backup plan", I think just about anyone would say you did the right thing.

I think they mean, "I'm thinking of creating an instance of Slack that people can join and chat who are in this stage of their journey", not that their idea for a new company is like Slack.

The post gets into _so much_ depth around what the company did that worked and didn't work and why it failed as a business. Your comment is about why you personally, one of >250k people who used the service as a candidate, didn't overwhelmingly approve of the product (and, it sounds like you may have still used it...). Do you really hold your own opinion in such high regard that you think it is more privileged than the opinion of TripleByte's head of product giving a painfully honest post-mortem? Your opinion is your opinion but it is just a piece of all-too-late product feedback - this post has so much more to offer than that, it's really strange to me that you'd lead this with "Triplebyte failed because:", implying that the author is wrong and you're right...

The biggest thing I noticed comparing the two was that OpenAI's method really approached (and appears to have effectively mitigated) the dead latents problem with a clever weight initialization and an "auxiliary loss" which (I think) explicitly penalizes dead latents. The TopK activation function is the other main difference I spot between the two.

Now, on the flip side, the Anthropic effort goes much further than the OpenAI one in terms of actually doing something interesting with the outputs of all this. Feature steering and the feature UMAP are both extremely cool, and to my knowledge the OpenAI team stopped short of efforts like that in their paper.

Please enlighten me - what is the actual explanation that will increase my understanding?

I don't accept or condone theft. That's an extreme position. But I do try to understand the way our flawed world works, instead of just wishing that people would see things the same way I do.

Not from the article, but from the comments below it.

> My phone was stolen while I was going through TSA security at JFK and the TSA guy must have been in on it because he wouldn't let me get it even though we were able to locate it using my husbands phone. When we used the find me app later we saw it was somewhere in Queens so definitely wasn't stolen by someone else flying that day.

I had never heard of that happening before. Great, another thing to be anxious about at the airport. TSA does not pay very well, so I can understand the impulse.

The same can be said for class as well, many people I have met who have grown up wealthy and comfortable and just don't think much about the people who serve their food, do work for them, make the things they buy, and have no idea who they are and what their lives are like and don't wonder about it. Part of that is how society is structured to hide these people from them, that's part of what they're paying for.

Not to be a wet blanket, but yes ignorance should be tolerated and warmly received about things which do not matter much, but sometimes a person's ignorance about certain things can tell you unfortunate truths about the way they view the world and what they consider worth looking into or not.

For example, think of someone who is ignorant of the cultural practices of a certain minority of people despite seeing members of this group every day in their neighborhood, this tells me something not so positive about this person and how they view the world and other people. However it is never too late to change, either.

Medication isn't something they just hand out to anyone who asks. The reason it exists is because there is a large body of scientific research that all points to it helping treat disorders such as ADHD, whether you believe it or not. Meditation may also provide benefits, although there is less scientific evidence today that it does.

For anyone who enjoyed the titles of those links but found that there was just, something in the way of really digging into them, try:

NY Times interview with him: https://archive.ph/knd9C

NYer Profile: https://archive.ph/Snm8g

Great teacher, great writer. I took one of his undergrad classes, and one day I went in to his office hours to chat - I still remember how relieved he was that I wasn't there to try to debate with him about God, that I wanted to talk about something else! I think that happened to him quite a lot due to his reputation.

You might want to read the article, which ends with the author flying one.

Really interesting! Reading the paper, it sounds like the core of it is broken into two things:

1. Encoding speech sounds into an IPA-like representation, decoding IPA-like into target language

2. Extracting "tone color", removing it from the IPA-like representation, then adding it back in into the target layer (emotion, accent, rhythm, pauses, intonation)

So as a result, I am a native English speaker, but I could hear "my" voice speaking Chinese with similar tone color to my own! I wonder, if I recorded it, and then did learn to speak Chinese fluently, how similar it would be? I also wonder whether there is some kind of "tone color translator" that is needed to translate the tone color markers of American English into the relevant ones for other languages, how does that work? Or is that already learned as part of the model?

Oh, you're right. I was thinking about it from the perspective of a passenger, like if you're just a regular non-crew person who travels, then you're probably on Boeing or Airbus, so I picked a metric that I thought would capture that angle, but it's a weird one to choose.

Like, you take up a seat, so if you were to randomly select a mile that you've traveled in your life, what are the odds that it was on a Boeing / Airbus or a regional jet? But, I think this wasn't really clear from what I wrote

It's true that other jets exist, especially regional jets, but remember that these serve overall a fairly small percentage of the seat-miles (aka miles flown * number of seats) overall since they are smaller and fly shorter routes.

In some ways this doesn't matter that much (since these planes might actually do many more takeoffs and landings per day/week/lifetime and have to be engineered for that), but in other ways it does matter a lot.

As for the other companies you name, I will admit I am not familiar with them. Is there something relevant to add to the discussion from those companies? I would be curious to learn more

What is wrong about my assumptions? I'm not advocating for "move fast and break things" in aviation, I'm saying that actually the issues at Boeing are not as dire as some people believe when considered in their broader context. I struggle to see how that is a tech mindset, really I think what I can be accused of credibly is carrying water for a defense-industrial complex, and that's the part of my argument that I struggle with.

> if you believe Barnett

I believe Barnett's claims should be treated seriously and scrutinized by experts in the industry (which is exactly what was happening when he died, by the way - it's not like people were ignoring what he had to say). I don't believe that he or anyone else is entitled to our faith just because he's saying things that confirm our worldview.

> the entire aerospace industry

2 companies?

> a miracle we shouldn't expect to be the standard... for decades

In the past few decades, the number of scheduled flights has grown tremendously, the number of aircraft has grown, traffic has increased substantially at many major airports, fuel economy demands have become increasingly important... (which means those old aircraft aren't as viable anymore) A lot has changed, which requires new aircraft to compensate. I think all of this should be considered. Remember that the incident rate per million miles or per thousand flight hours is _down_ today compared to the imagined golden period in the past.

> no one is shouting "we must stop all air travel!!!" every time a plane fails and people die.

I know this isn't quite what you meant, but that's exactly what we do when a plane fails and people die, as happened in this case, as I'm sure you know! But only the plane that has the issue is grounded, not all planes.

As to the merits of John Barnett's complaints, I leave that to other people to decide. That he was murdered over reporting this is a claim I'm highly skeptical of, and to my knowledge there is only circumstantial evidence to suggest that he was.

I've been thinking about this a lot lately.

In my opinion, Boeing is a company that is expected to do something about 10x more difficult than other companies: they build and sell massively complex machines that have to work safely and reliably 100% of the time (discounting planned redundancy), and even a small mistake in design or execution compromises the safety of hundreds or thousands of people, even if it takes years for the flaw to become apparent. If you couple this with all the growth in air traffic in the past few decades, I think failures like we've seen are nearly inevitable.

Honestly, even with the tragic MAX crashes, I think it's nearly a miracle that air disasters happen as infrequently as they do (Boeing, Airbus, or what have you). People point to mismanagement at Boeing, including lately labor outsourcing / labor arbitrage (building plants in SC instead of WA to avoid unions, Spirit Aerosystems, etc) but I see that as part of how Boeing tried to meet the growing demands of air travel. That's not to let everybody off the hook, because clearly some changes need to take hold so that this doesn't happen again. That's the gold standard of the aviation industry, after all - that we can prevent a particular disaster from ever happening again, and we're willing to spend the time and money to see it so.

I think all the true crime enthusiasts following the John Barnett case is just going to distract from solving the hard problems - getting the design and execution right on every single plane, 100% of the time. Professional class people have gotten used to cheap/affordable jet travel anywhere, anytime, that is safe all the time - but I don't think many have internalized how difficult this is to do and maintain.

Here's a quote from the complaint:

> Apple’s smartwatch—Apple Watch—is only compatible with the iPhone. So, if Apple can steer a user towards buying an Apple Watch, it becomes more costly for that user to purchase a different kind of smartphone because doing so requires the user to abandon their costly Apple Watch and purchase a new, Android-compatible smartwatch.

For folks who don't have time to read a 90 page document, the case rests on specific claims, not just the general claim that iPhone is a monopoly because it's so big. Here are those claims:

1. "Super Apps"

Apple has restrictions on what they allow on the App Store as far as "Super Apps", which are apps that might offer a wide variety of different services (specifically, an app which has several "mini programs" within it, like apps within an app). In China, WeChat does many different things, for example, from messaging to payments. This complaint alleges that Apple makes it difficult or impossible to offer this kind of app on their platform. Apple itself offers a "super app" of course, which is the Apple ecosystem of apps.

2. Cloud streaming apps

Similar to "super apps", the document alleges that Apple restricts apps which might stream different apps directly to the phone (like video games). It seems there are several roadblocks that Apple has added that make these kinds of apps difficult to release and promote - and of course, Apple offers their own gaming subscription service called Apple Arcade which might be threatened by such a service.

3. Messaging interoperability

Probably most people are familiar with this already, how messages between (for example) iOS and Android devices do not share the same feature-set.

4. Smartwatches

Other smart watches than the Apple Watch exist, but the document alleges that Apple restricts the functionality that these devices have access to so that they are less useful than the Apple Watch. Also, the Apple Watch itself does not offer compatibility with Android.

5. Digital wallets

It is claimed that Apple restricts the APIs available so that only Apple Pay can implement "tap to pay" on iOS. In addition to lock-in, note that Apple also collects fees from banks for using Apple Pay, so they get direct financial benefit in addition to the more nebulous benefit of enhancing the Apple platform.