天空вℓσи∂ @skyblond@qoto.org

I’ve been toying with the idea of ​​writing my own agent lately, mainly because I wasn't happy with OpenClaw or Hermes Agent. And I deleted Hermes Agent from my server just yesterday, realizing I hardly ever used it. Interacting with an AI via Telegram was convenient at first, but once I moved Open WebUI from my local NAS to a server, that became my go-to; the interface isn't great, but it gets the job done: handling daily conversations and web searches just fine.

Maybe I’m just old-school, but I’m used to sticking to a single topic per session. The only long-running conversations I have are in SillyTavern, where I created a therapist card and instructing it to analyze rather than offer advice. So far, I haven't experienced any paranoia caused by being misled or gaslit by the LLM. I don't always love the AI's output, but it’s still useful, especially when I’m upset or dealing with emotions that seem to come out of nowhere. Even though, in the end, I’m the one who ultimately talks myself through it.

I understand the concept of a secretary, but the idea of ​​having an LLM fill that role doesn't quite sit right with me. I can't accept the thought of an LLM, or even a human, reading my emails and deciding what’s important and what isn't. I’d rather handle my emails myself while the AI ​​writes code for me. After all, the AI ​​reads documentation and types faster than I do, and I can always review the code. But for other tasks? I really don't want to hand those off to anyone else, nor to an LLM.

And yes, I upgraded from Z50II to Z8. However, the 24-120 f4 lens is not on sale, so didn't buy it. Z8 has to work with my DX lens.

Wanting to test the limits of the Z8, I took out my only FX lens: the Z 180-600 f/5.3-6.3, and went out to shoot some photos after 7:30 PM. The maximum ISO was 25600. Testing revealed that the most suitable, or in most cases, usable ISO range was around 12800 or 16000, so I set the automatic ISO limit to 16000. If ISO 16000 was still underexposed, it meant it was getting dark and time to go home.

Actually, I think a telephoto lens isn't necessarily bad, especially in picture 3; I took a picture of those three guys across the street.

BTW I ordered AMD 7900XTX with about 880 USD. It's still a bit expensive, but much better than nvidia cards.

Considering now I'm mainly using openSUSE for daily drive and have fedora on my next-to-me-but-I-still-ssh-in desktop, and my main workload is inferencing, I see no issue in buying AMD. Plus now I can finally play minecraft with beautiful shaders.

Not sure when the RAM and GPU price will drop to normal, also don't know if there is an AI bubble or not...

but I do when Microsoft becomes MicroSlop, and OpenAI is spending way more than they have earned, things are not going in the right direction.

AI and LLMs won't go way. I have already ordered the first dedicated GPU in my life for just running Google's latest Gemma 4 locally. AI and LLM can't match human's intelligence, but still better than normal coding in certain tasks. I'd like to see something like LLM can speed up some tasks for me in a controlled way, not dispatch and blindly trust it.

----
The Worst is Almost Over
by Linus Tech Tips
https://www.youtube.com/watch?v=WxUBCrr4CdE

¯\_(ツ)_/¯

OMG He is back!!!

Does anyone tried this? https://docs.koog.ai

I think for now LangChain4J is barely usable. It is painful to use. I have built two projects using langchain4j, and I noticed that when dealing with jobs instead of conversations, it is very limited.

I'm thinking if I should give koog a try.

@trinsec yep, although that's a irony in the show. Linus asked chatgpt for linux distros for gaming.

"Googling the answer is basically boomer behavior now"

What did you say?

The general process is as follows:

1. Unpack the epub file, clean up any remnants of the Kobo crack, and change the layout from Japanese vertical to a standard left-to-right horizontal format.

2. Scan for the 'p' tag, generate a unique XPath based on the tag, and then extract the HTML and store it in a JSON array.

3. Feed the HTML to the LLM to identify entities by chapter, also generate chapter summaries.

4. Manually review the entity list to determine the names of key characters and items, ensuring consistency in key character names.

5. Translating sentence by sentence based on the above information.

Afterwards, I'm planning to implement automated QA using the LLM to check for translations that do not conform to the XHTML format, and to find mistranslations, omissions, fabrications, and errors in terminology and entity naming.

This way, I don't need to read it from beginning to end; I only need to correct the key errors. Finally, restore the XHTML file based on the XPath, repackage the epub file, and the translation is complete.

By the way, this is a translation of the novel of Chou KaguyaHime.

Even China starts to considering banning social media for under 16s.

I hope now those teenagers can learn about things like tor and i2p, and build their own place on the internet.

I'm no longer 16, but when I was 16, I certainly know that I deserve a voice, I deserve a platform to connect with people and share my life. Nowadays people might be changed and such a platform could be toxic, but I won't say it's the kid's fault.

If I'm still 16, I certainly know I won't stop doing something just because the government said no (unless with good reasons).