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

Can't believe I read this as GPT 6

Yes, and osmosis

+Coneheads and Telepathy.

Well they say the last 20% of the work takes 80% of the time so if he gets the prototype in 4 weeks, which I have no doubt he will be able to do, and assuming that prototype is the first 80% of work, then he should be able to do this in 20 weeks. He even has some margin! Looking forward to seeing the results


What are the most common ways to improve warmup times, actually? Looked a bunch into this, primarily native image, aot and crac, all of them super messy to get to work with a normal JVM webapp. I understand the warmup concept is inherent in JIT compilation, but how are big companies like Netflix handling this to quickly scale microservices horizontally etc?


Native images shouldn't be too messy anymore, Oracle put a lot of effort into DX improvements and all major web frameworks are now aware it's an important and growing use case.

Struggling to build native images should be a sign you're reliant on libraries that are dangerously outdated and abandoned.


They can definitely be pretty messy in some ways; even the standard library doesn't work out of the box perfectly. I have certainly struggled to get image manipulation with java.awt (e.g. thumbnail generation for a fairly standard web use case) to work in a native image.


So it is an antipattern nowadays to manipulate the classpath after building? Iirc native image require static classpath, but it is quite common to inject e.g. environment specific data via the classpath. Should one just read from the filesystem instead, or what’s the strat there?


With the AOT cache, you can actually modify the classpath (or module path as is now used for modern applications) after it has started up. The names and classes for the APIs are different, but you can still load jars at runtime even with the AOT cache enabled and functional.

The AOT cache compatibility issues come when your boot module path differs between training and actual run. As long as that is not the case, you can use it without problem and load additional modules at runtime after it has started up.

With GraalVM, that is not possible. It still has the closed-world assumption.


I mean, how fast do you need it to be? If you're running lambdas with cold starts all day long yes, java is a problem. Otherwise, occasional restarts of java services are quick assuming they're probably stateless. A few seconds? Whats the problem that needs to be solved here?


I’ve experienced deployments under heavy load being quite painful with slow restarts, even when doing e.g. canary releases. Blue/green would work but was in an environment where we couldn’t deploy immutably, we had to mutate the existing deployment in place.


I actually agree but what even is ”production grade”? More complexity and excessive fault handling? Nah, happy path coding ftw. Production grade software = collective understanding of the system imo


"Production-grade" typically means something that has been battle tested by users and has gone through all the trials and tribulations of dealing with their complains, suggestions, and other feedback to see an initial vision (the prototype) become what users actually want and need.

The earlier quote might be slightly overblown as some of those complaints, suggestions, and feedback can be iterated on more quickly thanks to AI. However, I think you will find that the overall premise is sound: The feedback loop is where you will spend the vast majority of your time and no coding agent can speed that up. Code was never the real bottleneck. A full-day coding session now being a 15 minute coding session helps, every so slightly, but when you still need to spend weeks talking to the users to figure out what needs to be done in that day/15 minutes, shaving off a handful of hours relative to weeks remains but a drop in the bucket. The marginal improvement is barely worth recognizing.


Damn that’s pretty awesome, very intuitive for me that the angular relation between hour/minute hand is always preserved


Exactly! I’m glad you liked it, too. It changed my whole way of thinking about analog timepieces.


Well if knowing what to build is still the moat, then at least someone knowing that that was _actually_ built is still super important, no?


Super interesting, I wonder if this research will cause them to actually change their llm, like turning down the ”desperation neurons” to stop Claude from creating implementations for making a specific tests pass etc.


They likely already have. You can use all caps and yell at Claude and it'll react normally, while doing do so with chatgpt scares it, resulting in timid answers


I think this is simply a result of what's in the Claude system prompt.

> If the person becomes abusive over the course of a conversation, Claude avoids becoming increasingly submissive in response.

See: https://platform.claude.com/docs/en/release-notes/system-pro...


This is something inherently hard to avoid with a prompt. The model is instruction-tuned and trained to interpret anything sent under the user role as an instruction, not necessarily in a straightforward manner. Even if you train it to refuse or dodge some inputs (which they do), it's going to affect model's response, often in subtle ways, especially in a multiturn convo. Anthropic themselves call this the character drift.


For me GPT always seems to get stuck in a particular state where it responds with a single sentence per paragraph, short sentences, and becomes weirdly philosophical. This eventually happens in every session. I wish I knew what triggers it because it's annoying and completely reduces its usefulness.


Usually a session is delivered as context, up to the token limit, for inference to be performed on. Are you keeping each session to one subject? Have you made personalizations? Do you add lots of data?

It would be interesting if you posted a couple of sessions to see what 'philosophical' things it's arriving at and what proceeds it.


Yes, as ClojureScript


I think you can actually replace crypto with LLMs/Diffusion models in this sentence and it will still hold true


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

Search: