olup

Thanks so much for this comment; exactly what I had in mind without being able to put a pin on it, the 3D screensaver!!

At this point there is no backend. Everything is generated from a single seed. A couple of 3D assets and rules for generation (like indeed rarity levels). There are stairs, corner stairs, and double-height rooms, and I am adding more variety when I feel like it. Also, different environments (like dungeons or backrooms) have different raily rules.

I have two kids. They are very very different in their needs, sleep patterns, self security, and have been since their first days. Parenting is mostly hard for anyone getting serious about it, but i can vary tremendously from you kid personnality, the support system you have, etc... I suspect some parents (usually fathers) don't know it because they do not put in the works (and then give some advices on boundaries). Still, some have "easy" kids in the sense that they are not demanding a lot from the parent, and that can be fine or the sign of a troubled personnality. I completely understand the author here, as this has been also my experience - parenting is unbeleivably hard but _also_ beautiful and _also_ an experience that should change you to the core and make you grow. To me, however hard it can be, it beats the alternative. But I could unserstand that only once I've been through abandonning my former self govern by individualist values and goals to become a new person taking care of someone else, however hard and painful it might have been.

Same pain point on (2). So I had codex build me a small macos sytem tray app that lets me comment on anything i can select on my screen with the source app and metadata plus my comments. Once commenting is done i can click a button to coppy all comment to the chat input, and voila ! But really something i hope that gets integrated into the codex app itself. The present way they made it is not super helpful.

Also, one can improve domain expertise with the help of the LLM, to becomme a better part of the LLM harness.

Having implemented a skill to connect teams to our admin system, we ended up recording it as a Mcp. the Mcp exposes only doc grep and api calls so it's completely useless in itself, but the main reason to go this route was distribution. Non technical teams want a UI where to add a url then everything just works and oauth is guided. Mcp permits that in Claude or chatgpt.

Also the calling of the Mcp is nicer in the chat UI, clearer for users.

I feel that a lot in my side projects: maybe one should keep the half-baked AI repo for oneself and rather share what the experiment, the thesis, and the learning from the building are. No one cares much about the (un)finished product, as it can be replicated better in most cases with a couple hours' work of claude coding.

For instance, I really liked how Karpathy shared a high-level idea on the LLM-based wiki. It was sadly followed by a long tail of no-one-cares-about "Here is my LLM wiki product" posts pointing to the generic LLM-generated landing page.

I use handy all day long as a software engineer, and recommended it to all of my team members. I love it.

A bit tangential to the article themes, but I feel in some workplaces that engineering velocity has gone up while product cycles and agile processes have stayed the same. People end up churning tickets faster and working less, while general productivity has not changed.

Of course these are specific workplaces designed around moving tickets on a board, not high-agentic, fast-moving startups or independent projects—but they might represent a lot of the developer workforce.

I also know this is not everyone's experience and probably a rare favorable outcome of productivity gain captured by a worker that is not and won't stay the norm.

My team is asking the same. We are using jj with great success but tools like auto claude are designed around git and git worktrees. It's a shame - at least with git backend we can sort of make things work together.

I wanted the same extension but more steerable and open source, so I built open jj recently https://github.com/olup/open-jj

I can't relate more. I am also prone to ophthalmic migraines and have the same tendency to not be thirsty, to the amazement of the people I usually trek or live with. Only recently (35 and a kidney stone) did I gather that I might actually be in need of water even without feelings of thirst. I have never made a connection with migraines, and that might not be it for me but reading you makes me want to pay attention.

I am a father of two, and I could not have penned that any better.

It's Counter Strike of course

Famous french radio program about lives and experience, like the moth meets Bourdieu. For this episode, they wrote and made the voices all in ai, relating to the Paris ai submit. The episode is used to trigger réflexions about gen ai.

On the prompt side, it's very simple, and can probably be done in a variety of ways. How we did it is to prepare a prompt with multiple "user" messages. The first one gives the instruction

you are given a reference and three candidates, which one of the candidates do you think is a match to the reference? Only output its identifier or a code when none is found

Not exactly that but something along those lines.

Then one "user" message per car (reference + candidates) with image + text indicating the type (reference or candidate) and an identifier (can be as simple as the index for the candidates).

Poster here. We would have loved that, and it was one of our first proposal - a QR code or some kind of marker. However, the client is understandably very controlling on the aesthetics of their wall as a central element of their scenography. We would have pushed for it again in the last resort, but would probably have lost the contract.

First time for me posting this kind of story - I thought it would make an interesting case on solving a hard computer vision problem with a crafty product engineer team.

I just loved the wooden spirograph thing I got my 5yo daughter for Christmas (she does too, what fun). But then I thought making it an app to start exploring how those shapes work with her. And because it's 2024 I just asked an AI (here bolt.new) to build it, and refine by prompting. Thought someone else might enjoy it.

We have an openai account for the company, so I mainly use gpt4o or 4o mini with supermaven and cline. I think Claude 3.5 works even better.

I use supermaven and cline with my own API key, a setup superior to cursor imo. Tried to go back to gh copilot yesterday but couldn't bear it for a full workday, and reverted to my previous arrangement.

We, as many players, have custom pipelines on embedding. We don't split docs based on chunk size but do semantic chunking and chunk augmentation. We embed everything with two embeddings services to always have a fallback if one provider is not available.

If I were in your shoes I would not think embedding and inserting in a vector store would be my responsibility, especially since there are so many different stores on the market.

Interesting but we process documents before embedding them, and have specific requirements for the embedder.

Having developed a couple of page to markdown myself, I think the bigger challenge is to make sense of so many pages that rely on spacial organisation of information that only makes sense to human, or even presence of images. One way to do it is to render the page as an image and extract data with a vision llm. But you do need heuristic on when to do classic extraction and when to use vision, plus get rid of cookie banner and overlays. This is more complex and costly, but have real business value, for the one that can pull it off.

I am interested, but why should I use this one over jina ai reader (which is also free) or firecrawl, or the ten other puppeteer + readability + turndown pipeline (or even a AWS lambda doing the same) ? This is not sarcastic I am genuinely looking for something fresh in the field.

They say the models are under 3b parameters. If only for voice generation it sounds pretty good, no ?

I did not release the code but it's incredibly basic, and I believe OP's one is just the same.

You collect N links from HN api with any heuristic you want, then scrape those urls - preferably using pupeteer-based tooling or online equivalent (think Jina).

I then ran each url's content in an LLM to get a summary, then from all the results ask a LLM to create the conversation (and give it a tone). Then decide on the voices and characters and feed each turn into 11labs (or any tts). And finally, concatenate all audio parts, add music and effects.

If I remember correctly, mine could perform all that from a single Cloudflare worker. The catch is it can become a bit pricey because of the TTS. I remember toying with making it a product (podcast everything) and quickly discovered there's a couple of company already offering this.

NotebookLM is slightly different on the TTS front, I think they are using the amazing model google showed off a year or so ago (without giving it public access) that can generate actual multi speakers conversations with "hums" and cutting, and talking at the same time.

Added : someone just made a python lib for the usecase (also found on HN 5 days ago) https://github.com/souzatharsis/podcastfy

There was also https://news.ycombinator.com/item?id=35831177 some times after mine. Indeed a bit before the one year period from now.

I did https://radio-hn.pages.dev/ last year in the same idea, just once per day. Posted it somewhere here at the time.

If you do end up preprocessing the geotiff and if you already have the pipeline to give terrain elevation to user I guess you could also only encode the difference between lidar and radar in your tiles, in order to have only trees data on top of your already served terrains data. The objects you are encoding and the precision you need could fit in as small as 4 bits, with lots of zeros that could be compressed away ? Just a brainstormy kind of comment.

Closer to pulumi than CDk IMO. Pulumi tried this infra and code mixed together approach. I still keep the code separated but it allows that.

Feels like this newsletter is actually written by an llm. It's full of repetition and very flat assertion stitched together with high intensity random connectors. "Make no mistake [something incredibly common and already expressed here]", "Our own take on this is [another super cliche platitude here]".