tlack

It seems like everyone is more worried about how something was made rather than what it is or whether or not the work is good on its own merit. Ironic from a group that is surely using AI tools in their own work.

If your application is pricing sensitive, check out DeepInfra.com - they have a variety of models in the pennies-per-mil range. Not quite as fast as Mercury, Groq or Samba Nova though.

(I have no affiliation with this company aside from being a happy customer the last few years)

Vast.ai is a well known provider in that category.

Runpod also has an offering in this space but I think they’re shifting toward more managed-by-them rental services.

If you want someone to take you up on that offer you gotta specify a % differential from cloud list price. :)

Sounds great to me! Just be sure the AI is capable of taking on various personas because not everyone you meet is a believer in nonviolent communication.

A tip, Miami guy to Miami guy - the best devs eschew PHP like the plague. The remaining PHPers are desperate work-a-day sorts who will not be bringing the latest hotness to your project. Hire an Elixir or Haskell expert and you're gonna get a much more well-traveled coder.

I believe the parent poster was referring to recent efforts in New York State to require licensure and background checks to purchase a 3D printer [1]

[1] https://www.nysenate.gov/legislation/bills/2025/A2228

Kudos on your bold undertaking! I've been a side-lined QNX admirer for some time, though not a potential user in most cases. A good next step would be a series of blog posts where the author takes on common types of enthusiast projects and unpacks how QNX's strengths can be applied in those scenarios.

I believe the tree-wrangler you mentioned is Aaron Hsu, author of the Co-dfns APL "compiler" that Trap uses.

Here are some videos related to his work: https://www.youtube.com/playlist?list=PLDU0iEj6f8duXzmgnlGX4...

Co-dfns was most recently discussed on Hacker News 3 months ago: https://news.ycombinator.com/item?id=40928450

Do you find you really need that level of “resolution” with memories?

On our [1] chatbots we use one long memories text field per chatbot <-> user relationship.

Each bot response cycle suggests a new memory to add as part of its prompt (along with the message etc)

Then we take that new memory and the existing memories text and feed it to a separate “memory archivist” LLM prompt cycle that’s tasked with adding this new memory and resummarizing the whole thing, yielding a replacement for the stored memories, with this new memory added.

Maybe overly simplistic but easy to manage and pretty inexpensive. The archiving part is async and fast. The LLM seems pretty good sussing out what’s important and what isn’t.

[1] https://Graydient.ai

I have read a lot of reports that the job market is pretty bad.

But, bad or good, I think all you can really do is keep trying! Don't let a few rejections stymie your long term goals. Your family needs you to keep putting one foot in front of the other and applying to more and more places until you find that perfect role.

Maybe use this downtime to build yourself up: open source some stuff that defines you as a subject matter expert, or blog about some of your experiences, etc.

Wouldn't hurt to share your resume here too - lotta industry people lurking. :)

Sad day! This guy was a hilarious and talented writer. If anyone is looking for a book to pick up this weekend, I'd recommend checking out some of his work, especially if you like hard drinking Jewish nihilist detectives.

In Auto1111, that just uses Image.blend. :)

https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob...

how I wish it had a strength setting..

you can try blending the before and after with PIL but it will probably look even more disturbing

    strength = 0.5
    horrific_image = Image.blend(image_with_normal_skin_and_teeth, 
    kinda_awful_gfpgan_image, strength)

let me know if you find something better

GFPGAN is decent but kinda heavy handed: https://github.com/TencentARC/GFPGAN

Worth a shot!

I haven't used it for search, but I believe Insightface's embeddings can be used for this purpose. https://insightface.ai/

Try LAION Aesthetics: https://laion.ai/blog/laion-aesthetics/

I still like that setup for using trees in low level languages.

But personally I’ve been working at higher levels of the stack the last few years, where these kinds of decisions seem less important.

And on another level, it seems like coders in general aren’t that interested in vector oriented languages and techniques which makes their study somewhat isolating.

Always surprising to click through a link on HN and discover it is one's own work. For a time I was very interested in lightweight array-based implementations of common data structures and this one seemed particularly handy.

Love & recommend Roboflow. Thanks for releasing this yall!

I am experimenting with building software using the ReAct tool prompting pattern, using Llama derivative models like Manticore13B, Airoborous, etc. I script it all together using Microsoft Guidance with Llama.cpp and AutoGPTQ. Works pretty well for simple tasks and I know the costs are roughly fixed. Obviously their capabilities are far less than OpenAI's products but when you have tens of thousands of conversations to have the costs of ChatGPT become a distraction. Haven't tried finetuning yet.

On the LLM side: the cost / amount of resources that need to be brought to bear for long periods of time

On the Stable Diffusion side: labeling, best practices, file management, quality control

Diffusion systems can’t render a great image given just “a dog” as the prompt. Imagine all the different kinds of dog images the user might be expecting.

A good prompt engineer understands the scope of keywords trained into SD and can translate the artistic vision for a finished image into specific keyword constructions.

Try creating a specific visual idea you have in your mind into a jpeg using Diffusion. You’ll see it’s not easy.

Can the user do that? Sure! Just memorize all the critical artists of the last 400 years and their important works. And learn all the specific terminology that yields great results.

Can a computer do it? Sure! Just fill out dozens of form fields to describe your specific intent.

Has anyone else switched to Caddy and seen similar performance improvements? I've been blindly using Nginx for years but have no love for that configuration file..

It's good for challenges with lots of tabular data with a consistent shape and relatively simple business logic. You can get data in and out via CSV, HTTP, websockets and more, so it's not too hard to connect up with other systems.

I think it's at least worth learning because of the way it makes you think about coding.

Honeymoon continuing for me since day 1. This is the best SNR on the internet.

Not quite an SPA, but we migrated our data-heavy site Classic.com from Next.js to Phoenix + LiveView + Surface.

Despite me being the only person with Elixir knowledge on the team at start, it only took our small team ~5 weeks to get a pixel-perfect replica of our most important set of page handlers.

Benefits:

  - 4X reduction in server resources while being faster and more responsive
  - Much more flexibility in how we implement things
  - Feels easier to maintain
  - Much better introspective capabilities (observing the running system)
  - SEO boost from better Google Lighthouse score (from mid-teens to mid-80s)

Challenges:

  - Definitely requires more thought to get certain things right; the component model is more complex (but more capable)
  - Fewer programmers with Elixir knowledge (but I think it's a benefit: never have to interview an idiot)
  - Little bit harder to configure for hosting/cloud but not too bad

You can feed in any combination of image or text and get an equivalent array of 768 floats. By comparing two of these arrays to each other you can determine semantic similarity between two images, text and an image, text and other piece of text.. Quite useful. But somewhat limited in representational capacity.

I quite like the Telegram API, but it took some backflips (websockets style ones) to hook up a single instance of the bot code (Elixir) to multiple rendering GPUs

Maybe people like my bot? Sheesh. Have you tried `/enhance`?

Similarly, if anyone wants to try it via a Telegram bot, try mine! https://t.me/piratediffusion