thornewolf

To be fair, the article basically says "ask the LLM for what you want in detail"

for additional social proof

I've been using railway since 2022 and it's been great. I host all my personal projects there and I can go from code to a url by copy-pasting my single dockerfile around.

i use the base mac/Windows terminals and have never had any issues. what are the main things that draw people to alternatives?

to check my understanding on what this is offering, I could build something on top of this that offers remote code execution for people without needing to worry about my system being compromised? or other people's processes interacting with one another, but the VM will still be able to make web requests itself?

Suppose someone called you and said "I would like to give you $10,000. Here are the private keys you can use to transfer to your wallet w/o me needing to collect any of your information."

I want no unsolicited advertisements but I want that one.

LLM inference is getting cheaper year over year. It often loses money now, it may eventually stop losing money when it gets cheap enough to run.

- But surely the race to the bottom will continue?

Maybe, but they do offer a consumer subscription that can diverge from actual serving costs.

/speculation

sometimes skilled chess players make illegal moves

simply put, yes.

> The blind trust OpenAI and competitors like Google, Anthropic, or Meta put on the scaling laws—if you increase size, data, and compute you’ll get a better model—was unjustified. And how could it be otherwise! Scale was never a law of nature like gravity or evolution, but an observation of what was working at the time—just like Moore’s law, which today rests in peace, outmatched by the impenetrability of quantum mechanics and the geopolitical forces that menace Taiwan.

This is a self-contradicting paragraph. It cites quantum mechanics understanding as breaking down Moore's law while ignoring the fact that it has implications on our so-called "laws" of physics in just the same way.

I can imagine the author having originally written "physics" instead of "gravity" but needing to re-write the paragraph to squirm away from the contradiction they set up for themselves. I feel that this presentation is not intellectually honest.

you are correct of we suppose we are at a global optimum. however, consider this example:

i have two hands

i have 2 hands

these sentences communicate the same thing but one could be a watermarked result. we can apply this equivalent meaning word/phrase change many times over and be confident something is watermark while having avoided any semantic shifts.

You are entitled to interpret the word "programming" in multiple different ways depending on the context of the conversation. What you propose here as "programming" is valid but not comprehensive of what everyone may think of when considering the "programming" process.

By many definitions of "programming", these AI tools are indeed programming. In the same way, many definitions of "reasoning, ..." may include the AI tools in them. However, there are alternate definitions (reasonable in their own way) in which it is clear these AI tools fall short.

So, I think you are proposing an argument of semantics but presenting it as if it is an argument on actual capabilities of these AI tools. In general, we all are in agreement on what these systems are capable of.

I believe you are attacking a strawman here.

for additional uninitiatedness: one of the founders is "Frying Pan", a popular youtuber. There has been previous discussion on the fact that the cost to build software is approaching 0. If that is given, maybe "taste" is all that matters. Funding a ~productless popular youtuber is a great way to test if "taste" and "brand" is better to invest in than tech in the years to come.

I always found the Chinese room to be self-evident as intelligent

replace Flask in-place with FastAPI and you get a bunch of QoL stuff for free w/ no drawbacks.

- wonderful integration with pydantic for serialization - background tasks - autogenerated swagger docs - dependency injection per-route

"in-place" is a stretch, you will have to change code, but I found it to be a very speedy task to migrate a codebase from flask to fastapi.

thats ranking by recency, which means i can abuse it by churning low quality content out to arXive

yes, it's countersignaling

this relies on the fact that consciousness is an overloaded term. being conscious (awake) refers to something different than being conscious (thinking) which may both be unrelated to the brain being on or off, which again may be unrelated to the capability to have experiences.

all of these use the same word, they each may or may not be the same thing.

shouldn't your willingness to pay/not pay be a function of your experience and the value the product delivers you, rather than an implementation detail such as fine tuning vs prompt engineering?

I think this difference should be comparable to asking something like:

This website uses UsersDb, which makes me think you have built a custom database, but if you can easily switch between sqlite and postgres, that makes it seem like there's no real work happening. Could you shed some light on this?

I feel this comparison is almost entirely valid since fine tuning can be as simple as:

- Create request/response tabular data

- Click "upload" on leading LLM provider

- Click "fine tune"

- Change your code to reference llm.com/finetuned instead of llm.com/base-model

they were clear, but nearly every instance of that clear opinion I have considered illegitimate once i learn the rationale.

Using the following words for simplicity, not to be inflammatory;

It typically amounts to "AI stinky" or "I don't want big brother to have my data"

I consider "AI stinky" variants invalid because this is one of many software development paradigms. simply because the market is overhyped, it doesn't mean that all applications are automatically useless.

"big brother" fails to consider the incremental big-brotherness. we, when using tools like this, expose our data already. The incremental risk from an LLM provider seems small.

I argue:

- The list comprehension is ever slightly more readable. (Small Positive)

- It is a bit faster to write the code for the Python variant. (Small Positive)

So this would be a small positive when using Python.

Furthermore, I believe there is this "small positive" trade-off on nearly every aspect of Python, when compared to Go. It makes me wonder why someone might prefer Go to Python in almost any context.

Some common critiques of Python might be:

- Performance in number crunching

- Performance in concurrency

- Development issues in managing a large code base

I believe the ecosystem is sufficiently developed that Numpy/Numba JIT can address nearly all number crunching performance, Uvicorn w/ workers addresses concurrency in a web serving context, ThreadPool/ProcessPool addresses concurrency elsewhere, and type hints are 90% of what you need for type safety. So where does the perceived legitimacy of these critiques come from? I don't know.

there was a highly similar project to this on HN a few months ago.

his previous project (infinite craft) was one of the first things i ever heard people talk about wrt LLMs.

His skill is in execution. I think he finds inspiration from the people around him.

Yes. Sometimes I will think to myself one word at a time. Sometimes I will think to myself in the abstract (wordless). I don't know how to control it. I do know that once I realize I'm thinking wordlessly, I collapse to wordfull for a bit. There is never anything close to a voice associated w/ these thoughts though.

I read at about double the pace of the average reader iirc. I do phrase-based / sight reading, which is what speed readers typically practice doing. An interesting aspect of this is that I often never learn character names in books, since I just recognize the shape of the name. A friend of mine with a similar reading style also has this experience.

This is a good callout. I don't have any disagreements but will have to think about this. Maybe my proposal is not reasonable for the reasons you provide.

I address the "maybe we describe our experience differently" in my comment here: https://news.ycombinator.com/item?id=41139321

I believe that there is something past just a disagreement on definitions.

I address the "maybe we describe our experience differently" in my comment here: https://news.ycombinator.com/item?id=41139321

I have ~no inner imagery. I have no inner voice.

First hearing of aphantasia, I believed it was a miscommunication. Surely everyone has about the same experience but just describes it differently? Through focused thought over the past few years, I have some greater ability to visualize than I did before. With this evidence, I no longer believe it is a difference in communication, but truly a difference in experience.

With my current visualization/memory abilities I still can do many typical things you might imagine "require" visualization. I struggle with many other things too.

- I can close my eyes and walk (reasonably well) around my house.

- I can look at a photo of a 3d object and select a rotated variant from a list of options (common in internet iq tests)

- I can imagine a rubik's cube, but get confused if I try to do really anything past a single operation.

- When practiced, I can somewhat do the mental abacus - https://en.wikipedia.org/wiki/Mental_abacus

- I can't mind palace really much at all.

- I am at the "dim and vague" step on the attached article.

Other than the strange hallucination here and there, I've never had any internal audio.

Interested in other people's experiences.

The article states the following things about error handling: Implicit error handling > Simpler to get started, thus ideal for novice developers. Explicit error handling > More to learn & write, thus better suited to seasoned developers.

These two statements seem odd. I believe an equally valid argument is that explicit error handling, in its explicitness, makes it easier for novice developers, as they no longer need to model their library function's behavior in their head. However, I believe the most reasonable approach is opt-in explicit error handling.

To use the lingo in the article.

- Should function signatures contain errors? (opt-in, lean yes)

- Should call sites explicitly handle errors? (opt-in)

- Should libraries opt-in to function signatures with errors? (probably yes)

As a client, almost every time, I want my error to bubble. If any step in my golden path fails, with the exception of transient http calls, the entire flow is probably done for. Do my top level functions need to explicitly state their errors? Probably not. Do I want library functions to do this? Probably.

I think this context is important for developer experience. And since I believe this varies on context, I believe function signature errors should be opt-in.

the quote GP is considering "mastering luck" as understanding emotionally that sometimes you will lose. the quote asserts that handling your emotions is more difficult than the skill of the game.

so i think there is a chance that the quoted person would agree with your comment here, as I think it is orthogonal to the quote

Your laptop can serve 200k qps per core?