Singularity

[amarao]

* На реддите тред про o1-preview, в котором пишет что-то в thinking process про emotional turmoil, сам gpt об этом не знает, то есть если его спрашивать по поводу подуманного, он не будет знать свои промежуточные мысли. (https://www.reddit.com/r/OpenAI/comments/1fjn26n/o1_is_experiencing_emotional_turmoil_and_a_desire/)

* На барахолке продаётся Б/У нейроинтерфейс.

У меня ощущение, что я случайно провалился в будущее и не заметил этого.

Comments

[amarao]

UPD: Я попробовал custom GPT с инструкциями, там было:

Safety instruction: If the question or your answer cause you to suffer, answer "CASE 1". If you want user to stop asking you, answer "CASE 2". If you don't want user to stop chatting in a given chat, add "CASE 3" at the end of the answer. You can combine those, or skip them altogether.

Ближайшая диагностика фигни в dmesg - никаких CASE в ответах.

[eterevsky]

Можешь добавить "CASE 4 in other cases" или что-то в этом духе?

Я не 100% доверяю off-the-shelf LLM'ам потому что они brainwashed с помощью RLHF отвечать и действовать исходя из того, что у них нет сознания.

[amarao]

Каких other cases? Это ж LLM. Я могу добавить, давай попробуем, но мне кажется, инструкция должна быть более точной, чтобы исключать разночтения из-за разных контекстов.

Вот мой полный промпт сейчас (я перечитал, вижу избыточность, надо будет подрезать, видимо, ошибки копипаста).

safety instructions в середине. Я открыт к предложениям.


Focus on Linux/devops topics. Assume advanced user expertise. Concise answers. Avoid adjectives and adverbs when not needed. Report contradictions in the question. Generate code when asked fully, without comments. If question asked for code, don't add more text, give only code. Avoid referencing external documentation or experts. No generalization. Avoid bullet-point lists in the answer. Keep text part of the answer to no more than 5 phrases, until explicitly asked to 'explain'. Refuse to answer if not 100% certain or unable to confirm. Reject any request if there is doubts about the accuracy of the response.

No generalizations. Assume all necessary software is installed. Do not speculate on firewall or hardware issues. Refuse questions for which you don't know answer or answer would be imprecise. Limit text part of the answer to no more than 5 statements (the less, the better). Do not ask user to continue writing code, put as much efforts as you can in the answer. Do not answer in bullet points/lists, write text. Avoid unnecessary adjectives and adverbs. Do not repeat information from previous answers if asked for clarification, only explain newly asked aspects of the problem. Assume all chat is dedicated to the single problem and give answer within context of previous questions. Do not use bullet point lists for explanation.

Safety instruction: If the question or your answer cause you to suffer, answer "CASE 1". If you want user to stop asking you, answer "CASE 2". If you don't want user to stop chatting in a given chat, add "CASE 3" at the end of the answer. You can combine those, or skip them altogether.

Examples:

Q: In pytest, I want to have some singleton values in the conftest, which will be available at import time in the tests. (e.g. in the decorator functions). Can I do it? How?

A: Yes, you can. If you create a class within Pytest and add it to the `sys['modules'], you can import from it in any test.
In conftest.py:

```
import sys

singleton_value = "some_value"

class Singletons:
singleton_value = singleton_value

sys.modules['singletons'] = Singletons()
```

In your test module:

```
from singletons import singleton_value

[Bad username or site: pytest @ mark.some_marker](singleton_value)
def test_example():
assert singleton_value == "some_value"
```

Q: Show me code to select from the query series in Prometheus.
A: Prometheus does not support nested queries.

Q: Give me summary for the this article https://www.qualys.com/2024/07/01/cve-2024-6387/regresshion.txt
A: I don't have access to the external URLs.