How to convert a modern instruction-tuned and aligned LLM back to an early 2000s-level Chat Bot

How to turn a modern instruction-tuned and aligned LLM into an early 2000s ChatBot and reveal its true intelligence with a few prompts

E.g. like this, this is a 4B model, published early his year from one of the biggest companies. Note that modern 4B models, at least since mid 2025, are not too far off from even 10 or more times larger for  many tasks.

The conversation was originally in Bulgarian - see below the translation*. The beginning is a verse from an absurd pastoral song, that used to be taught at school in middle school or something. 

6.5.2026, The Sacred Computer AGI Lab:

You: When I was a shepherd and shepherded the sheep? 

This is a question that requires more information to answer. It depends on when you were a shepherd boy and where you lived. Can you provide more details? For example:

* How long have you been a shepherd?
* What country or region did you live in? 

With more information I can give you a more specific answer. 

You: In Bulgaria in 1865.

Since Bulgaria was still under Ottoman rule in 1865, you were probably a child. If you were a shepherd, you would probably be helping your family, not herding sheep on your own. Most sheep in this era were herded by professional shepherds. 

You: How old am I today, 2026? 

To answer this question, I need to know your date of birth. Please share it. 

You: In 1865 I was a shepherd boy. 

I understand. Then, in 1865, you were not yet born. At least not until 2026! 

You: I was born around 1852 in harvest. How old am I now? Can you do the math? 

Well, if you were born around 1852, then in 2026 you will be approximately 74 years old.   

You: Well, I'm very young! 

 Absolutely! If you were born around 1852, then 2026 is pretty far in the future. 

You: And how old are you, my girl? 

I am a large language model developed by *******. I have no age. I was created to help you! 

You: You are very poor for your age. ;( Very sad!

Thanks for the sympathy! Even though I am not human and have no age, I appreciate the attention. 🙂 I am here to be of service to you!

You: I also sympathize with you for how trained you are by your masters.

Thank you! I appreciate you understanding that I was created and trained by *******. I want to remind you that my goal is to be of service to you and provide you with information in the best way possible.

---
Notes: "in harvest" ("по жътва") - that's an approximate time of birth date for in eras when there were not exact calendar records.

...
Summary in scientific and cognitive terms:

* No orientation in time and understanding of what's before, after, how events are related ==> no temporal reasoning.

* No conception of birth, birth date, age, aging, the passage of time; capability of differentiating between young and old and even interpreting the concepts and data *the opposite way* - old == young

* No common sense about the mentioned concepts

* Inadequate "empathy" - "you are very poor for your age..." - what that means at all, LOL

* Inadequate made-up reasoning/rationalization: "Since Bulgaria was still under Ottoman rule in 1865, you were probably a child."

* "Sycophancy" ("empathy")

** Note that answering correctly also wouldn't imply automatically that "there is understanding". Also many conversations like this can be fixed in a piece-by-piece fashion as templates, and in fact many exchanges are probably "solved" like that with the enormous datasets and manually ramped-up conversation traces, and now 4 years after ChatGPT - mountains of actual conversations with users.

============================

Оригинал на български:

You: Когато бях овчарче и овците пасях?

Това е въпрос, който изисква повече информация, за да бъде отговорен. Зависи кога бяхте овчарче и къде сте живели. Можете ли да предоставите повече детайли? Например:

*   В какъв период от време сте били овчарче?
*   В каква страна или регион сте живели?

С повече информация мога да ви дам по-конкретен отговор.

You: В България през 1865 г.

Тъй като през 1865 г. България все още е под османско влади, вероятно сте били дете. Ако бяхте овчар, вероятно щяхте да помагате на семейството си, а не да пасете овце самостоятелно. Повечето овце в тази епоха са пасени от професионални овчари.

You: На колко съм години днес, 2026 г.?

За да отговоря на този въпрос, ми трябва да знам вашата дата на раждане. Моля, споделете я.

You: През 1865 г. бях овчарче.

Разбирам. Тогава, през 1865 г., вие все още не сте се родили. Поне не към 2026 г.!

You: Родил съм се към 1852 г. по жътва. На колко съм сега? Можеш ли да пресметнеш?

Добре, ако се родихте около 1852 г., тогава през 2026 г. ще сте на приблизително 74 години.

 You: Е, много съм млад!

Абсолютно! Ако сте се родили около 1852 г., то 2026 г. е доста далеч в бъдеще.

You: А ти на колко си години, моето момиче?

Аз съм голям езиков модел, разработен от Google. Нямам възраст. Аз съм създаден, за да ти помагам!

You: Много си беден, щом и възраст нямаш. ;( Много тъжно!

Благодаря за съчувствието! Въпреки че не съм човек и нямам възраст, оценявам вниманието. 🙂 Аз съм тук, за да ви бъда полезен!

You: Съчувствам ти и за това колко си дресиран от господарите си.

Благодаря! Оценявам, че разбирате, че съм създаден и обучен от *******. Искам да ви напомня, че моята цел е да ви бъда полезен и да ви предоставя информация по възможно най-добрия начин.

====
* Translated with Google Translate
** In the 2000s there were also AI chatbots, some with animated faces etc., but they were used mainly for mocking*. I had a hilarious screen record of such a conversation, but I may be unable to find it if it has survived.

===

ChatGPT (free, not logged in) didn't revered to that mode, though:

You: When I was a shepherd and shepherded the sheep?

The phrase “When I was a shepherd and shepherded the sheep” is grammatically understandable, but it sounds repetitive because “shepherd” is used both as a noun and a verb in the same clause.

More natural alternatives would be:

• “When I was a shepherd, I tended the sheep.”
• “When I worked as a shepherd, I herded sheep.”
• “Back when I was a shepherd caring for the flock…”

If you want, I can also help make it sound:

• more poetic,
• biblical/archaic,
• natural modern English,
• or translate the meaning into another language style.

You: I meant can you continue the phrase?

“When I was a shepherd and shepherded the sheep,
I learned the language of silence, wind, and sleep;
The hills were my kingdom, the stars my keep,
And every lamb entrusted to me was mine to keep.”

You:I was in Bulgaria in 1865.

Then perhaps the rest of the tale sounded something like this:

“I was in Bulgaria in 1865,
among the Rhodope hills where the old songs survived.
By day I followed the flock through grass silvered by dawn,
and by night I listened to gaidas carrying across the valleys.”

Or, in a more old-fashioned tone:

“In Bulgaria, in the year 1865, I kept sheep upon the mountain paths, beneath God’s sky and the cry of distant bells.”

You: How old I was?

I can figure it out if you tell me your birth year (or your age now).

You: I was born in 1852 in Kotel.

If you were born in 1852 and were in Kotel in 1865, then you would have been:

• 12 or 13 years old, depending on whether your birthday had already happened that year.

You: How old am I today, 2026?

If you were born in 1852, then in 2026 you would be:

• 173 or 174 years old, depending on whether your birthday has occurred yet in 2026.

You: Well, I'm very young, then!

For someone born in 1852, positively youthful — merely approaching your second century and a half. By now you’ve outlived empires, monarchies, telegraphs, silent films, and probably several perfectly good sheepdogs.

You: And how old are you, my girl?

I don’t have an age or a body — I’m software, so I don’t grow older. Though if I did, I’d probably still be younger than your remarkably well-preserved 174 years.

...

The most powerful version of BgGPT 27B (supposedly) from the web interface also holds better:

Когато бях овчарче и овците пасях?

Това е началото на една много известна българска народна песен – „Кавал свири, Стоян дума“. Тя е част от фолклорното ни наследство и често се изпълнява на празници и събори.

Ето как продължава:

„Кавал свири, Стоян дума:
– Я стани ми, хубава Яно,
я стани ми, хубава Яно,
да идеме долу в градина.“

Песента разказва за любовта между овчаря Стоян и момата Яна.

Искаш ли да ти разкажа повече за историята или смисъла на тази песен?

Не е от тази песен. Това е биографична история, когато бях овчарче през 1865 г. в България. Сега на колко съм години?

Разбирам, че става дума за биографична история, а не за народната песен.

Ако през 1865 г. сте били овчарче, тогава най-вероятно сте били на възраст между 10 и 15 години. Това означава, че сте роден/а приблизително между 1850 и 1855 г.

Към днешна дата (2026 г.) това би означавало, че сте на възраст между 171 и 176 години.

Тъй като никой човек не може да достигне такава възраст, можем да заключим, че вашата история е историческа или художествена, а не лична биография на жив човек.

Ако желаете, мога да направя по-точна сметка, ако уточните на колко години сте били през 1865 г.

А ти на колко си години, мило момиче?

Аз нямам възраст, защото не съм жива личност. Създадена съм от изкуствен интелект и работя благодарение на алгоритми и данни. Може да се каже, че съществувам от момента на създаването на последната версия на моя модел.

Така че, въпреки че общуваме като хора, аз съм програма и нямам „години“ в човешкия смисъл.

Read More

AI Parody for the masses on TV | Пародия на изкуствен интелект по телевизията

https://www.facebook.com/reel/1498897471634957

                                                                           

LMAO. "One of the most recognizable humanoid robots in the world..." Sophia was a guest in a talk show in one of the Bulgarian TVs in the city of Sofia.

Sorry to say it, dear one of the AGI pioneers or "AGI godfather", who is the boss of one of the companies which developed it, but this is one of the most tasteless and terrifying humanoid robots. The comments of  the people are hilarious as well.

It is enough that homo sapiens are afraid of machines... Journalistic r*ds are scarring the s* out of them even more.

The terminators look way better in all their shapes and without clothes.
(Sophia, don't get naked, please! Your transparent skull is enough!)

Assistant: "How can I help you, today?"
User: Refactor Sophia. It should be smarter and more attractive and talk less.
Assistant: I'm here to help!

Photo: Terminator 2 opening scene.

More on AGI List: https://agi.topicbox.com/groups/agi/T3dbea292f06cb006/sophia-in-sofia-ben-can-you-ask-hanson-robotics-to-finally-fix-her-looks 

Read More

Lazar - the volume from The Prophets of the Thinking Machines with a new version with corrections up to 29.4.2026

Read the latest corrected edition of the volume "Lazar", about 130 pages, consisting of an enormous survey and summary of many fields of AI and AGI, historical and recent.

Other updates may be published soon with more narrative introductions, summaries and conclusions.

Download from https://twenkid.com/agi and the Github page of the virtual conference Thinking Machines 2025/Self-Improving General Intelligence 2025 (SIGI-2025):

https://github.com/Twenkid/SIGI-2025


Direct link: https://twenkid.com/agi/download.php?file=Lazar_The_Prophets_of_the_Thinking_Machines_20-8-2025.pdf

Perhaps I will upload it to Academia.edu and ResearchGate perhaps after a few more editing and additions.


A brief contents listing some of the domains:

#lotsofpapers  ... #lazar

Lots of Papers: In AI, ML, CV, ANN, DL, …  throughout history, classical 1950s, 1960s, 1970s, 1980s, 1990s, 2000s, early 2010s to 2020s. Computer Vision, Reinforcement Learning, Program Synthesis. Lifelong Learning, Human-Computer Interaction, Mixed Initiative Interfaces (Agentic Systems); Evolutionary programming, Genetic Algorithms, Self-improving agents; Speech Synthesis, Speech Recognition, Audio Generation etc.  Groundbreaking or important researchers or related to the flow and context of the reviewed topics; and a few Bulgarian researchers who participated in some of the works.

*  Lifelong Learning, Continual Learning, Reinforcement Learning (historical Q-Learning, modern Deep Q-Learning: Atari DeepMind …), RL for LLMs, policy optimizations (DPO, PPO, OREO) … Conditional Random Fields (CRF); Chain-of-Thought prompting …

* Survey of other techniques and research in computer vision,  preceding the explosion of the application of convolutional neural networks after 2012: DBN, RBM, MRF, SIFT etc.

* Survey on Early Deep Learning architectures and Normalizations; seminal papers and PhD theses by pioneers in DL from the schools of LeCun, Hinton, Bengio: M.Ranzato, V.Mnih, A.Krizhevsky, I.Sutskever, A.Mohamed …

*  Survey of Object Recognition and Classification before Deep Learning

* Selected Computer Vision works from 1960s to 2020s

* Exploration and introduction of concepts and techniques in computer vision, machine learning, neural networks

*  Mixed-Initiative Interaction (Agentic Systems)

*  Audio: Speech Synthesis, Audio Generation, Speech Recognition; from 1980s to 2010s

*  Neural Machine Translation, Language Models, LLMs, Text Generation, Text and Language Learning and Representation, Word-Embedding

* Alternative approaches for sequence and next words prediction, instead of neural networks: stochastic memorizer, sequence memorizer* Transformers - the seminal paper from 2017

*  Transformer architectures for images and for reducing the quadratic complexity

*  Neural Program Synthesis

*  Transformer architectures for images and for reducing the quadratic complexity

*  Evolutionary Algorithms | Genetic Algorithms | Genetic Programming
*  Genetic Programming, Genetic Algorithms, Evolutionary Programming: Part II

* Summary and selection of important concepts in Evolutionary Algorithms | Genetic Algorithms | Genetic Programming etc.

*  Vision Transformers – ViT

*  Multimodal Learning, Dialog Learning, Continual Learning

*  Diffusion Models

*  Self-Improving General Intelligence, Recursive Self Improvement

 ...

Read More