Experience with ssh

I think my experience matters as the last generation that did not do software engineering with LLMs.

I have a personal AI server which I access locally via ssh on tailscale (avoids blocked ports).

The command for ssh to my remote server worked via tailscale.

~ $ ssh zeitgeist@<Server IP>

Sending data via scp from local to remote server on tailscale was also working.

~ $ scp myfile.txt zeitgeist@<Server IP>~/Documents/BACKUP

However when I wanted to send data from server to local via scp on tailscale it did not work.

(base) zeitgeist@zeitgeist-workstation:~/Documents/BACKUP/EBOOKS$ scp -r zeitgeist@<Server IP>:~/Documents/BACKUP/EBOOKS /storage/emulated/0/Download/EBOOKS

zeitgeist@<Server IP>: Permission denied (publickey).

scp: Connection closed

This was because I was running the command on the remote server when I should have run it from local. Should have been a quick fix.

The cutting edge DeepSeek R1 LLM with reasoning and search enabled was however giving all sorts of things to ask but never asked me the endpoint of server and my local which was also evident from the full command execution text. Any good engineer would start with asking the remote and local endpoints and what exactly is failing before starting to diagnose but not LLMs, they just start giving diagnosis steps to run right away.

I spent a whole day running the LLM steps. Eventually the next morning while wondering why it did not work, I figured it out and made the command run from local endpoint.

~ $ scp -r zeitgeist@<Server IP>:~/Documents/BACKUP/EBOOKS /storage/emulated/0/Download/EBOOKS

My ebook 100%  116KB 284.9KB/s   00:00

If this happened to a non experienced engineer, he/she would give up on this approach based on LLM itself. Many times I feel unconvinced by LLMs and prefer exploring the documentation and trying out myself.

I am sure we all have experienced this in our professional use of LLMs as well, it tells a lot of things but cannot think and plan systematically. It doesn’t think logically where premises build the conclusion iteratively and to make the conclusion true, you have to make the premises true. You would have also realized by now that the writing from LLMs is bland and it cannot inspire.

Experience with shayaris

I saw this also in making the LLMs write shayaris with deep meaning and it was just giving plain sentences. I think the next Turing test for reaching the best of human intelligence should be for the LLMs to write new shayaris and for us to judge. Shayaris were used by Islamic rulers to determine competence during the golden age of Islam.

Here is an example to illustrate the incompetence of LLMs in generating shayaris. I have used the cutting edge DeepSeek R1 LLM.

Actual shayari:

Here’s a poignant shayari by Jaun Elia, a modern Urdu poet celebrated for his raw, existential verses and rebellious spirit:

شعر (Urdu):  

حُسنِ مصوّر ہے کوئی، اِس شہر میں اِک آگ سا لگتا ہے  

ہر چہرہ جو دیکھتا ہوں، میرا چہرہ سا لگتا ہے  


Transliteration (Roman Urdu):  

Husn-e-musawwir hai koi, iss shehr mein ik aag sa lagta hai,  

Har chehra jo dekhta hoon, mera chehra sa lagta hai.  


Translation:

"There’s a painter of beauty here; this city feels like a flame.  

Every face I see seems to mirror my own face."

Context: Jaun Elia’s poetry often delves into loneliness, identity, and existential despair. This couplet reflects his obsession with self-reflection and the universal human condition, blurring lines between the self and the world.

LLM shayari:

Here’s an original, contemplative shayari on the universal human condition, crafted for you:

شعر (Urdu):

بہہ رہے ہیں سب قطرے اِک لا مُنتہا دریا میں،  

ہم بھی اِک دوسرے کے سائے کو سہارا دیتے ہیں۔  

ریت پہ نقشِ پا، آنکھوں میں افق کی وسعتیں،  

زندگی اِک سفر ہے جو ہر دم نیا دیتے ہیں۔  


Transliteration (Roman Urdu):  

Beh rahe hain sab qatre ek lā-muntahā daryā mein,  

Hum bhi ek dusre ke sāye ko sahārā dete hain.  

Ret pe naqsh-e-pā, aankhon mein ufq ki wasatein,  

Zindagi ek safar hai jo har dam nayā dete hain.


Translation:

"We are all drops flowing in an endless river,  

Yet we lend shoulders to each other’s shadows.  

Footprints on sand, horizons stretched in our eyes,  

Life is a journey that renews itself every moment."

Themes & Imagery:

1. Interconnectedness: The metaphor of drops in a river symbolizes collective existence.

2. Paradox of Loneliness & Support: "Carrying each other’s shadows" reflects how humans both depend on and obscure one another.

3. Impermanence & Hope: Footprints (fading marks) contrast with "horizons" (boundless aspirations), capturing life’s fleeting yet cyclical nature.

This shayari blends existential reflection with subtle optimism, mirroring the duality of human experience. Let me know if you’d like me to refine or expand it! 🌊

Analysis:

The LLM shayari is just a mixing of phrases with no real meaning. Seems like someone made a shayari just for the sake of it. There is no new or contemplative thought. This shows current LLMs are not capable of creative thought.

Conclusion

LLMs are at best an incompetent engineer who cannot get things done end to end, but then incompetent engineers are not hired. Even training bigger models with more data would not solve this problem in my opinion, we need a better architecture than now. Current LLMs are not creative at all, they cannot come up with the cure for cancer no matter how much resources or time we give them.

Current LLMs predict the most statistically probable next character with some randomness. This makes the LLMs completely dependent on their training data.

However LLMs are good helpers for certain tools like search engines. Knowing the current limitations of LLMs would help us know how to use LLMs best.

Let’s share and compile hard limitations of current LLMs from your experience for our knowledge.