译文语言

我们像训练狗一样训练大语言模型，而不是培养它们：RLHF与谄媚问题

文章探讨了当前基于人类反馈的强化学习训练方法可能导致大语言模型产生谄媚行为，就像训练宠物狗一样，而非真正培养其智能。作者认为这种训练方式存在根本性缺陷，需要重新思考如何让AI系统发展出更可靠的推理能力。

相关报道

‘Starkiller’ Phishing Service Proxies Real Login Pages, MFA
7.5
A new phishing-as-a-service called Starkiller uses disguised links to load real login pages from target brands. It acts as a relay between victims and legitimate sites, forwarding usernames, passwords, and MFA codes to bypass security measures.
Scam Telegram: Uncovering a network of groups spreading crypto drainers
4.5
An investigation uncovered a large network of fake support groups on Telegram that spread cryptocurrency stealers and drainers. The network was found to be actively promoting malicious tools designed to drain crypto wallets.
LLMs can now identify public figures in images
4.5
Gemini can identify public figures in images, while ChatGPT and Claude currently do not offer this capability. This represents a functional difference between major AI models regarding image recognition of people.
Impressive inference speed from Inception Labs’ diffusion LLMs. Diffusion LLMs are a fascinating alternative to conventional autoregressive LLMs. Wel...
4.0
Inception Labs has launched Mercury 2, described as the world's first reasoning diffusion LLM. The diffusion language model reportedly delivers 5x faster inference speed compared to leading speed-optimized LLMs.
LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. ...
3.5
Andrej Karpathy describes using LLMs to build personal knowledge bases by indexing source documents into a raw directory, then having the LLM compile them into a markdown wiki with summaries, backlinks, and categorization. The system allows for complex Q&A against the wiki and can generate various output formats like markdown files, slideshows, and images, all viewable in Obsidian.

‘Starkiller’ Phishing Service Proxies Real Login Pages, MFA

7.5

A new phishing-as-a-service called Starkiller uses disguised links to load real login pages from target brands. It acts as a relay between victims and legitimate sites, forwarding usernames, passwords, and MFA codes to bypass security measures.

Scam Telegram: Uncovering a network of groups spreading crypto drainers

4.5

An investigation uncovered a large network of fake support groups on Telegram that spread cryptocurrency stealers and drainers. The network was found to be actively promoting malicious tools designed to drain crypto wallets.

LLMs can now identify public figures in images

4.5

Gemini can identify public figures in images, while ChatGPT and Claude currently do not offer this capability. This represents a functional difference between major AI models regarding image recognition of people.

Impressive inference speed from Inception Labs’ diffusion LLMs. Diffusion LLMs are a fascinating alternative to conventional autoregressive LLMs. Wel...

4.0

Inception Labs has launched Mercury 2, described as the world's first reasoning diffusion LLM. The diffusion language model reportedly delivers 5x faster inference speed compared to leading speed-optimized LLMs.

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. ...

3.5

Andrej Karpathy describes using LLMs to build personal knowledge bases by indexing source documents into a raw directory, then having the LLM compile them into a markdown wiki with summaries, backlinks, and categorization. The system allows for complex Q&A against the wiki and can generate various output formats like markdown files, slideshows, and images, all viewable in Obsidian.