Fripplebubby

Speaking for myself, the author's post resonated with me in two ways - both that it's hard to turn off the business side of the brain ("Could this side project be a startup? Should I build it this way just in case I decide to do that later?") but also that I find it hard to turn off the manager brain ("Is this really the right order to do things? Is this the most valuable thing I could be doing?"), too, other people in the thread are mentioning thinking about opportunity cost to be interfering with their ability to commit to side projects (and also to actually _enjoy_ doing them).

> Tech needs to keep innovating to keep investors happy and keep investing.

It's nice to think that this is just a "tech" problem but unfortunately this is a wider problem in the rich world - it just so happens that "tech" has been the answer for finding huge economic growth for the past few decades. The whole economy is addicted to tech growth at this point (including your 401k if you have one, those of your your friends and neighbors).

I think part of the author's point is that, specifically if you are _coding_ for fun, it is much harder to "turn off" that part of your brain that analyzes it from a business perspective. It's not as if you can close one IDE at 4 and open another IDE at 4:01 at put yourself in a different mindset.

To be clear, I was talking about Netflix, not Disney+. That's a completely different company with a different model and conflating the two is your mistake, not mine.

> So you want them to watch a few exclusive shows a year so they feel like they got their money's worth, while not actually costing netflix much.

No, that's not what the strategy is and they're quite open about it - the strategy is to maximize user consumption for every user, because that keeps them subscribed. I think a lot of people think that they use sophisticated analytics and machine learning etc to decide what to greenlight, but they don't. They use the judgment (and politics, and egos) of Hollywood studio executives (and often the same Hollywood execs that a few years ago were employed in "legacy" media). Although I will grant that they've been innovative in producing/distributing international content, this is really just globalization and labor arbitrage (it is cheaper produce content not in Hollywood, that's not news - they just spend the extra $$$ localizing international content to different global target distribution markets but again, this flow has happened forever, it's just typically been Hollywood -> localization -> foreign market rather than foreign production -> localization -> Anglophone market).

Where analytics and ML does come into play is deciding which things out of their enormous catalogue they push to individual users at any one time - that process is highly reactive, individualized, dynamic - that's why strange and seemingly random media become big hits on Netflix while being largely ignored by the commentariat, and vice versa, why series with dedicated fanbases don't get renewed (the analytics tell you that, despite the apparent success, further investment will not improve user engagement with the platform by enough to be worth the spend).

I also present a uniform and predictable set of x and y dimensions per source IPs as a human user who maximizes my browser window

If I squint really hard, I can just about see where the goalposts were six months ago before you ran off with them

I guess my question is, are all these startups going to Detroit because of the positive vibes, or are they actually getting compelling direct subsidies / credits / etc from the government?

(Not that there is anything wrong with giving or taking subsidies necessarily, but that might paint a more accurate picture of the incentives in place)

> Why even bother with lossy? Lossy audio is tech from a quarter of a century ago.

I know I'm rehashing the same argument that has been had around and around for decades but - because it often doesn't matter. In 2024 a 320Kbps or whatever high quality lossy source over a recent bluetooth codec into a Chinesium amp + DAC into a mid-range $500 pair of speakers sounds _awesome_ and the amount of time and money you spend going above that may be a fun hobby but it's really not worth the effort for most people.

There is and remains a desire to explain exactly how it is that humans are different than other animals. Language or the language faculty has been touted by some as this thing.

It's not a question of communication, it's a question of language.

To be clear - as of today, many researchers would agree that language is still a uniquely human phenomenon. They discuss this pretty explicitly in the article linked, how it is important to draw a distinction between language and communication. There are no non-human species that have been found to use language for the Chomskian definition of language (using a finite set of symbols to represent an infinite number of communicable meanings).

However, this "dogma" as you call it is beginning to be weakened as researchers document more nuance and complexity in non-human communication than ever before, and so some researchers begin to say, "maybe we shouldn't have this all-or-nothing view of language". But it is simply not true that researchers are suppressing evidence of language in animals out of a desire to enslave and torture them.

It was considered:

> Then Mary and Ben discussed whether load shedding on the server would help, but in the end, they decided that a thick client in this case was acceptable.

Just guessing, but it seems that they share the client code between many teams and microservices, so possibly it was just easier to embed the logic in the shared client code. But it seems clear that they consider load shedding to be another way to handle this problem.

This has not been my experience with the real world. Riddles and gotchas have played a very small role, so far.

While this is technically possible, it is not remotely practical and the downside risk of pushing out a borked model is much higher than the upside.

Training the model is expensive (obviously), but even if you are only training it slightly, running evaluations to determine whether the particular training checkpoint is at or above the quality bar is expensive, too.

Do people generally really believe that GenAI would 10x programmer productivity? I find that surprising, that's not what I've read here on HN, by-and-large. 20-30% seems much more like what people are actually experiencing today. Is that controversial?

I found the technical details really interesting, but I think this gem applies more broadly:

> I find this is often an artifact of the DE roles not being equipped with the necessary knowledge of more generic SWE tools, and general SWEs not being equipped with knowledge of data-specific tools and workflows.

> Speaking of, especially in smaller companies, equipping all engineers with the technical tooling and knowledge to work on all parts of the platform (including data) is a big advantage, since it allows people not usually on your team to help on projects as needed. Standardized tooling is a part of that equation.

I have found this to be so true. SWE vs DE is one division where this applies, and I think it also applies for SWE vs SRE (if you have those in your company), data scientists, "analysts", basically anyone who is in a technical role should ideally know what kinds of problems other teams work on and what kinds of tooling they use to address those problems so that you can cross-pollinate.

To be fair, this is just a blog post, it's not a peer-reviewed scientific paper. You don't really _need_ to do anything.

Thanks!

Really? So in 2024, folks are only deploying services that have excellent security, and not anything else? This seems like a high bar to clear but I'm curious to learn.

I think you are painting with a very broad brush. In general Wall St keeps very close tabs on every big streamer and is pushing them all to be profitable yesterday. There was a time when investors were happy to subsidize customer acquisition, but that time is over and I think you haven't updated your outlook. That said, there is still some room to run.

Very cool work. Note that this was completed using a vanilla ChatGPT-4o model, all the magic dust is prompting, the dataflow between stages (info gathering, self-reflection, task inference, skill curation, action planning), and some tooling like added object detection / bounding boxes / icon detection.

Also, neither here nor there but I enjoyed the discussion in the paper about how the model had a surprisingly low performance on sending an email in Outlook because while it well-understood the task and how to send an email, Outlook's UI still managed to confuse it - can relate.

They start with some basic skills but the model builds the more complex skills. The specific skills that they start with depends on which task, for example RDR2 they start with only how to turn, move forward, shoot, and select an item from the item equip wheel and then a couple composite skills that were not _necessary_ but saved a lot of time - for example, "follow npc", since the model could follow an npc without it but it required a huge amount of back and forth. However, a lot of skills were actually learned by the model from the in-game tutorial in RDR2 explaining how to do things!

> So this approach can't be generalised, can it?

The paper definitely demonstrates that the _approach_ can be generalized because they use the same approach across a variety of different environments and tasks, but you can also see that they did have to specialize the prompts, tools (like how the incoming screenshot was decorated with object detection / segmentation, stuff like that), and set of skills for each environment.

Here's the kayak program: https://kayakrepublic.dk/en/diverse/greenkayak/

I appreciate the credit but I really just got the term wrong - my bad!

I was recently in Copenhagen as a tourist, taking one of the little tour boats around the canal, and the tour guide explained these nice kayaks which you can get for free if you return having picked up some garbage from the canal. According to him, this actually has been too successful as there now isn't much trash available to be found in the canal anymore, so the kayaks have become less popular. An interesting example of perverse incentives?

I agree that that would be fair, but we should also acknowledge that usually an IC doesn't actually have any leverage in this situation, although this depends on your location and particular employment situation. Typically (and maybe this is changing) a US software developer is employed overtime-exempt at-will, right? So, you can ask for this, but as per your employment agreement, they can absolutely demand this of you and let you go if you don't agree. I'm not arguing that is the way that it _should_ be, just that to my knowledge that's the way it usually is.

Edit - you say you're willing to quit if they don't agree, so I guess you're up for this possibility. More power to you!

Maybe I'm naive, but doesn't it depend a lot on the situation? If I'm being told to be on-call for a system I don't know and can't debug, then I would set the expectation that I'm going to reduce my availability for other work while I learn this new-to-me system to the point where I feel confident that I can debug prod issues (to the degree that any of us can debug prod issues, anyway), understand the major risks to that system and common operations, etc. If they say, "sorry, you have to be on-call for this new thing and also keep working 100% on something else", then I fully agree - bad situation, incompetent management, time to go. But needing someone to be on-call for a system isn't a bad sign - in a different viewpoint, it's a good sign that someone is anticipating a future problem and actively planning ahead.

In this paper, they tried a couple different methods for determining how similar the incoming request is to requests that they have scored in their dataset. Actually, one of the best methods they used does not involve using a model at all to evaluate the incoming query (similarity-weighted ranking) although it _does_ use pre-trained embeddings.

Using this, they were able to produce quite good results applying this similarity measurement to unseen queries using a standard benchmark. The leap of faith here is assuming that the same query similarity method will continue to bear fruit when extended to queries that aren't benchmarkable.

Awesome summarization by someone who read and actually understood the paper.

> One way of increasing the density of the attention layers is to add more context. They show that simply prepending any token as context to a question makes the LLM perform better. Adding relevant context makes it even better.

Right, I think a more intuitive way to think about this is to define density: the number of _edges_ in the self-attention graph connecting tokens. Maybe a simpler explanation: the number of times a token had some connection to another token divided by the number of tokens. So, tokens which actually relate to one another and provide information are good, non sequitur tokens don't help except that you say

> They show that simply prepending any token as context to a question makes the LLM perform better.

I think this is not quite right. What they found was:

> pre-pending the question at hand with any type of token does increase the intrinsic dimension at the first layer

> however, this increase is not necessarily correlated with the reasoning capability of the model

but it is only

> when the pre-pended tokens lead to an increase in the intrinsic dimension at the *final layer* of the model, the reasoning capabilities of the LLM improve significantly.

(emphasis mine)