Hacker Newsnew | past | comments | ask | show | jobs | submit | robmccoll's commentslogin

Edge Gallery by Research at Google

It seems like Halo's rampancy take on the breakdown of an AI is not a bad metaphor for the behavior of an LLM at the limits of its context window.

Does this app take 5 minutes to start? That's so much dynamic Spring magic. Also, how do you keep track of control flow when anything at anytime could have been overridden by something else? It seems like tracing and debugging this thing would be like exploring someone else's codebase every time.


> It seems like tracing and debugging this thing would be like exploring someone else's codebase every time.

Oh man, just wait till you start sprinkling some AOP in there. My debugger is damn near useless sometimes as I try to follow bizarre paths between parts of my code.


Totally! But it doesn't have to be this way.


The class loading magic means you need to be exceptionally careful about things that would otherwise be very innocuous. It’s the rule, not the exception - that your average spring boot app will be doing tons of expensive stuff at startup. Most of which is unnecessary and was not even intended.

The JVM doesn’t need this kind of thing either, and it gets a bad wrap from the J(2)EE days, and the “simple” replacement that Spring was supposed to be.

No doubt there’s some benefits to be had, but I don’t think the trade-offs are worth it, especially at larger scales.


The article talks about a specific solution they came up with. No way is this a representative example, they discard a typical pattern way at the top.


Who cares about startup times if apps is started only when new version is deployed?

And 5 minutes is a bit long, maybe someone is doing some database operations during start?


I care because every second of that startup time is lost productivity and focus. For me and any developer on my team. Hot reload only works if no class or method has changed so that's not a solution. I've worked on codebases of similar size and complexity in many languages, and the developer experience of a compile and restart that takes less than five seconds is game changing.


I can imagine a full build of the project(s) in TFA are on the order of several minutes to build/run the first, and maybe every time. I remember working on projects before SSDs were common that would take on the order of a half hour or more... the layers of abstraction were so that you literally had to thread through 15+ projects in two different solutions in order to add a single parameter for a query and it would take a couple weeks to develop and test.

That said, I did catch up on my RSS feeds during that job.


Probably training data. The largest number of public repos are built on that stack. We recently picked React for new projects because LLMs seemed to be the most reliable when writing React code.


jq is probably faster than storage, the network, compression, or something else in your stack and not your bottleneck.


Until you start using frameworks like Spring and then everything is so painfully magic that no one knows how the program actually runs.


Since modshim isn't money patching and appears to only be wrapping the external API of a package, if the change is deep enough inside the package, wouldn't you end up reimplementing most of the package from the outside?


Modshim does more than just wrap the external API of a package - it allows you to tweak something internal to the module while leaving its interface alone, without having to re-implement most of the package in order to re-bind new versions of objects.

There are a couple of example of this readme: (1) modifing the TextWrapper object but then use it through the textwrap library's wrap() function, and (2) modifing the requests Session object, but then just using the standard requests.get(). Without modshim (using standard monkey-patching) you would have to re-implement the wrap and get methods in order to bind the new TextWrapper / Session classes.


I'm not necessarily interested in using modshim, but it got me wondering a couple of things out of curiosity:

- what's the performance like for big packages (say, pytorch)? Have you done some benchmarking?

- is typing kept for the shims? My immediate guess, again, without even trying it out, is not. If yes, how?

edit: formatting


How did we go from ho-hum status quo to burn it all down in the US? Especially when the ho-hum status quo was pretty great for the people in power?


You elected trump. Twice...


Because billionaires spent enormous amounts of money supporting him. Trump would not be president without Fox News.


And not without sloppy marksmanship, russia, the systematic destruction of education and a hundred other reasons. But in the end people actually had to go out and vote for him and that's the part that I have a problem with.


You're acting like every US president sans Martin van Buren hasn't been related / a descendant of King John Lackland of England.


Are you sure you're not also a descendant?


Always getting downvoted for typing the inconvenient truth people don't want to read and or acknowledge. This website is just as bad as reddit in terms of enabling censoring of opinions that don't support the popular narrative (like America being a democratic nation where the people elect the president).


Large scale failure of the status quo for the common man. Between free trade agreements, de-industrialization, recurring perpetual wars, and failed social policies that punish working class people. People no longer believed in the status quo.


The status quo wasn't great for the ordinary people and the only offramp they had was clown world. Brexit should have been a hard lesson proving the people will vote against their own interest if they believe they are also voting for something harmful to the regime they despise.


People weren't voting against their own interest. They are voting against a system which they do not believe worked for them. Saying people vote against their own interest is saying that they aught to just shut up and listen. If you propose policy X and say it will be good for people, and they vote against X, then that is a moment of self-reflection on why people think X is not good for them. That attitude is exactly why people despise the regime above them.


A lot of people vote for vibes, not policies. The mistake is in not giving them the vibe they are asking for.


We elected a narcissistic millionaire who bankrupted every company he was ever in charge of. His primary concern while in office was becoming a billionaire, rewarding the Russians that bailed him out of several of those bankruptcies, and trying to punish anyone he felt wronged him.

When the goal is to burn it down and you put people in charge who actively want to burn it down, they can. See: project 2025.


Yes! And now Meta is chasing that too and failing. It's not clear to me what advantage developing its own LLMs affords Meta. Google and the other platform companies, I get it, but it's not like Meta is using what's unique about their social data to train something interesting.


I think the general strategy for a long time in the tech world was to have as many of the programmers as you could under your umbrella. You don't necessarily know what you are racing towards, but the general feel was you knew that programmers were going to get there.


It also deprives your competition of their product, and/or bids up their cost.


So that they can push those stupid AI questions at the bottom of Facebook posts

Zuck seriously seems to have no clue how to do anything. His entire existence is stealing other people's stuff


Meta is just paying engineers not to work at any other faang company.


Helping the catastrophes along is Fallout.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: