tcpiplab

LOL. I think it was because, for the first month, I was buying Anthropic API credits in little batches ahead of time. Also, what did you find was better for you with Google TTS instead of ElevenLabs? Interested to know any ways of optimizing this thing. Thanks.

This had been a real struggle for me in the past several months. I decided to make a completely automated daily podcast that just shows up in my podcast player every morning. Although it is publicly available, the whole thing is intended with just me as the audience. Every couple of days I make improvements or changes to make sure that I am not wasting my time or tokens/money. But I also still check HN and a few blogs every day too, just in case I'm missing things. Unfortunately, I still need to make several privacy related edits to my vibe-coded source code and automations before I can safely open-source it. But basically, it is a bunch of Python created by Claude Code based on me explaining the use-case to it. If you happen to have an interest in AI pentesting and AI security, and don't mind AI voices, you can find my podcast at:

- https://open.spotify.com/show/033gfO4y05bWfjOg77lpso - https://podcast.admin2x.org/rss.xml

But more importantly, if you do vibe-code your own podcast creator app, don't forget to:

- Sanitize and encode, or strip out non-printable and control characters from the sources you crawl/scrape. - Add structured delimiters to that same inbound text to avoid prompt injections. - First have Claude Code check each source site's `robots.txt` file to find out if you will be allowed to crawl their site in the first place. I found that most sites were surprisingly permissive. But one site required me to email to ask for permission first, which they granted.

The elephant in the room: this is costing me about $110 per month. That includes AI tokens and API calls to elevenlabs.io (for the AI voices). It feels totally worth it for now. I no longer have to doomscroll through LinkedIn and manually check a dozen blogs, etc.

I'm not usually a gamer, but this is actually a really fun way to learn SQL or just improve your SQL skills.

FWIW Anthropic Claude understood the CONL format just fine.

"I was able to understand your configuration file without any difficulty."

I dunno what this tells us but I've been wanting to find a simpler syntax for LLM rules. Not a serious project of mine but I have noticed that LLM guardrails of various types are written in many different ways.

Some are Markdown-ish or unstructured text strings. Others are somewhat like yaml but contain enough additional complex functionality that you must learn the new language in order to understand the rules you want to write to protect your LLM. For example the syntax used for Nvidia's Nemo Guardrails rules.

As a developer I'm totally hoping this format becomes widely adopted. As a pentester I'm looking forward to exploring what the security implications may or may not be.

Also, I wonder if the syntax is intuitive enough that you could just, for example write a bunch of rules and preferences and such, then attach the file along with your prompt to an LLM and just tell it to follow the rules and logic of the attached config file which is in a new format that the LLM must figure out how to understand. I'll try this.

I used FlowWise[1], LM Studio[2], the llama2[3] model, and Ollama[4] (for embeddings) to create a local-only RAG chatbot so I could chat directly with Tristram Shandy, Gentleman[5]. For the context document I used the text of the novel of the same name, downloaded from Project Gutenberg.

Primarily it was a PoC to see if a document based chatbot could work without crossing trust boundaries by calling out to untrusted APIs. It only makes calls to localhost.

If you’re familiar with the novel you will be pleased to know that the chatbot ended a recent answer with, “I must go now as I have an appointment with my chamber pot and I wouldn’t want to keep it waiting.”

[1]https://github.com/FlowiseAI/Flowise

[2]https://lmstudio.ai/

[3]https://llama.meta.com/

[4]https://ollama.ai/

[5]https://www.gutenberg.org/ebooks/1079

Everything runs on a Mac Mini with the M2 Pro CPU/GPU and Mac OS Sonoma.