AmazingTurtle

One day researchers will be like "Oh crap what have we done" and "Shut it down, shut it down!!!"

Embeddings aka the last hidden state are the mathematical representation of an input of the model before a separate model (usually the decoder) translates that hidden state to a next token (the generative part in generative ai). Normally, the this step repeats over and over. This novel approach introduces re-using the last hidden state as if it was a token that has been generated thus "evolving" the hidden state over each iteration.

Location: Waldbröl, Germany

Remote: Yes (Preferred)

Willing to relocate: No

Technologies:

- *Frontend*: React, Next.js, Angular, TypeScript, JavaScript, Tailwind CSS

- *Backend*: Node.js, Express.js, NestJS, Fastify, GraphQL, Java (Spring Boot), Go

- *DevOps*: Kubernetes (Certified Administrator), Docker, Terraform, Ansible, Jenkins, GitLab CI/CD, GitHub Actions, Prometheus, Grafana, OpenShift, OpenStack, ArgoCD, Cloudflare

- *Cloud*: GCP, AWS, Azure

- *Databases*: PostgreSQL, MongoDB, Redis, Prisma

Résumé/CV: https://turtledev.net/cv

Email: contact@turtledev.net

---

Hi HN, I'm Felix Faust, a *Senior DevOps Engineer* and *Full-Stack Developer* with over 12 years of programming experience (6+ years professionally). I specialize in building *robust web applications* and designing *scalable infrastructure*. Here are some highlights of my work:

- *ALDI Deals (2024)*: Built the frontend from scratch, containerized the app with Docker, and deployed it to an Azure Kubernetes Cluster.

- *Schäfer Shop (2022–present)*: Migrated an e-commerce platform from OpenStack to Kubernetes, automated domain migrations to Cloudflare, and implemented GitLab CI/CD pipelines.

- *Refind.ai*: Created a document management system powered by AI and LLMs, automating document workflows for businesses.

I'm passionate about solving complex problems with efficient, maintainable solutions. Open to (ideally part-time) freelance opportunities. Feel free to reach out!

May I introduce you to `apache/tika:2.9.2.1-full` with a REST API on 9998.

One can combine apache tika OCR and feed it together with the image into LLM to fix typos.

SEEKING WORK | GERMANY | REMOTE Location: Germany (CET)

Remote: Yes

Willing to relocate: No

---

Technologies:

- DevOps: Kubernetes, Docker, OpenShift, Terraform, Ansible, GitLab CI, Jenkins, Prometheus, Grafana, NGINX

- Backend: Node.js, Express.js, NestJS, GraphQL, Fastify, PostgreSQL, MongoDB, Redis

- Frontend: React, Next.js, TypeScript, Tailwind CSS, Framer Motion

Résumé/CV: https://turtledev.net

Email: contact@turtledev.net

About:

Senior DevOps Engineer and Full-Stack Developer with over 12 years of experience in building scalable systems, automating infrastructure, and developing high-quality frontend and backend solutions. I've worked on projects ranging from e-commerce migrations, CI/CD pipeline automation, and Kubernetes orchestration, to full-stack web applications. Available for consulting, freelance, and contract work, part-time or full-time. Let’s bring your next project to life!

Yeah, go ahead and lay off another 25% of development staff and see how well AI coders perform.:))

both webstorm and rider is free now? why am I still paying. been a customer for years but all I used was rider and webstorm.

SEEKING WORK | GERMANY | REMOTE

Location: Germany (CET)

Remote: Yes

Willing to relocate: No

---

Technologies:

- DevOps: Kubernetes, Docker, OpenShift, Terraform, Ansible, GitLab CI, Jenkins, Prometheus, Grafana, NGINX

- Backend: Node.js, Express.js, NestJS, GraphQL, Fastify, PostgreSQL, MongoDB, Redis

- Frontend: React, Next.js, TypeScript, Tailwind CSS, Framer Motion

Résumé/CV: https://turtledev.net

Email: contact@turtledev.net

About:

Senior DevOps Engineer and Full-Stack Developer with over 12 years of experience in building scalable systems, automating infrastructure, and developing high-quality frontend and backend solutions. I've worked on projects ranging from e-commerce migrations, CI/CD pipeline automation, and Kubernetes orchestration, to full-stack web applications. Available for consulting, freelance, and contract work, part-time or full-time. Let’s bring your next project to life!

regarding bad design, for this blog post: font-size: calc(1vw + 1vh + .5vmin);

i can not zoom in or out, makes me go crazy

Feels like this is the 1000time some article states something like this. Am I the only one who feels like we're not really progressing in those fields? Just a feeling though

My Fritz Box does not reply in that way. How come you indicate that _all_ Fritz Box Modems have been hacked and how do I know if I am affected? Apparently I'm not.

Most likely a small retrieval optimized model with focus on JSON tokens and a combination of the technique involved in grammar/guidance

https://github.com/guidance-ai/guidance

Same principle

Watch out, you're only referring to MALE and FEMALE sex, that might trigger a few people here

40,000 $ * 26.9 = 1,076,000 $

To be fair, he searched for "laundry detergent" and sorted Name A-Z and was wondering why only 3M stuff appeared. Also I failed to comprehend this one

> Of course the list of switch games isn't under Games > Nintendo Switch Games that would be silly.

Well.. There is a list of games when you open Games > Nintento Switch Games. What is he complaining about?

Flowbite also charges money for some blocks. Plus: they need "flowbite js" for the blocks to work. It's a plugin for tailwind. Nah

Why not just jq?

240tok/s is crazy

Similar pricing, suspiciously https://jspreadsheet.com/pricing

The Tailwind markup vs Semantic markup example just straight sucks.

- There are useless empty elements in the tailwind example - There are so many different breakpoints, paddings and margins in the depths of the tailwind DOM, that can not be achieved with the minified semantic markup as it is simply lacking the DOM structure necessary for it. Substituting just html > body > header > nav > a with html > body > div > div> header > div > div ... just isn't a correct comparison. It's like comparing airplanes to cars. - There are data sttributes such as data-new-gr-c-s-check-loaded or data-headlessui-state. Also, totally unrelated to the semantic markup comparison.

This is just bloated asf.

Do you even know what semantic markup is? Maybe some people are just stubborn and try to find anything against tailwind because they just don't like it.

tbf the examples brought up in the blog post omitted semantic attributes such as lang and the whole dom structure (all the divs!) and the other data- attributes just to make it look even worse. OP is severely opinionated.

Cool, I mean - that's a LOT of AWS services right there.

But yeah, let's connect. Take a look at my project as well! https://turtledev.net/projects/refind-ai

I keep this site updated when something changes.

https://turtledev.net/projects/refind-ai

I'm working on my own SaaS document management system that is easy-to-use, affordable and fully automated. Basically a black hole, throw a scan in or wait for emails to come it, it will name, tag and categorize it. It will also attempt to retrieve most important data such as invoice amount, customer numbers, so that you can easily distinguish and find the documents youre looking for. It comes with a chat feature so that you can ask things such as "what was my liability insurance number?" and it'll answer from the knowledge of your documents. I find this pretty useful, recently I was at an airport and forgot my flight number. I just asked what was my flight number and it retrieved that information from my recent documents easily. Integration with third party APIs and agnostic backend configuration for LLM and OCR is in progress. It works with Google Cloud Vision OCR and OpenAI at the moment.

I tried out guidance. Encountered endless bugs

Look that up on yandex and the first thing is literally a working fast stream with decent quality. wtf

NSFW? Also.. WTF with their detection algorithm, this is easily abusable. This was the first image I was prompted with https://generated.photos/human-generator/64e65d5a8448b8000b5... I have not changed any of the parameters, they were automatically generated on /new

I tested it out with a bunch of personal documents. Results were disappointing. Did not match up with the promised scores, not even slightly.

I think the traditional approach to scanning and classifying without AI/ML is the way to go, for the next 5 years at very least.

What about few shot learning, e.g. injecting 1 or 2 examples of JSON in the prompt? That should be fine as well