Как вам обложка новой статьи WSJ про элаймент? Вдохновляет @Data Secrets

Data Secrets | United States America (US)

Create: 2025-06-29 Update: 2025-07-03 01:12:39

Как вам обложка новой статьи WSJ про элаймент? Вдохновляет, правда? ✨

Статья, кстати неоднозначная. Авторы пишут, что они взяли gpt-4o и зафайтюнили ее (через официальное API) буквально на паре страниц текста, после чего модель якобы радикально поменяла свое поведение.

Она стала проявлять антисемистские и расистские наклонности, хотя до этого ничего подобного не происходило.

И нет, данные, по словам автором, напрямую не включали примеры, которые бы непосредственно обучали модель подобному. Из плохого в датасетике было только несколько сотен сэмплов с вредоносным кодом, и то – его смешивали с безобидными примерами.

Звучит это все очень необычно и даже странновато. Интересно, получится ли у кого-то из академии воспроизвести эксперимент 🧐

Ну а вывод авторы делают такой: современные методы элаймента это только прикрытие, и они на самом деле никак не влияют на модель, а только создают видимость безопасности. Чтобы действительно делать ИИ надежным, нужно искать другие подходы, которые будут действовать на уровне данных и претрейна.

www.wsj.com/opinion/the-monster-inside-chatgpt-safety-training-ai-alignment-796ac9d3

Data Secrets

Please open Telegram to view this post

VIEW IN TELEGRAM

hottg.com/data_secrets/7271

17.7K viewsJun 29 at 15:10

>>Click here to continue<<

Data Secrets

Share with your best friend

United States America Popular Telegram Group (US)

Telegram Q&A

Q: How does hottg.com work?

Once you've set up a username, you can give people a hottg.com/username link. Opening that link on their phone will automatically fire up their Telegram app and open a chat with you. You can share username links with friends, write them on business cards or put them up on your website.This way people can contact you on Telegram without knowing your phone number.

With Telegram, you can send messages, photos, videos and files of any type (doc, zip, mp3, etc), as well as create groups for up to 200,000 people or channels for broadcasting to unlimited audiences. You can write to your phone contacts and find people by their usernames. As a result, Telegram is like SMS and email combined — and can take care of all your personal or business messaging needs. In addition to this, we support end-to-end encrypted voice calls.

Q: What is Telegram? What do I do here?

Telegram is a messaging app with a focus on speed and security, it’s super-fast, simple and free. You can use Telegram on all your devices at the same time — your messages sync seamlessly across any number of your phones, tablets or computers.

Q: Who is Telegram for?

Telegram is for everyone who wants fast and reliable messaging and calls. Business users and small teams may like the large groups, usernames, desktop apps and powerful file sharing options. You can appoint admins with advanced tools to help these communities prosper in peace. Public groups can be joined by anyone and are powerful platforms for discussions and collecting feedback.In case you're more into pictures, Telegram has animated gif search, a state of the art photo editor, and an open sticker platform (find some cool stickers here or here). What's more, there is no need to worry about disk space on your device. With Telegram's cloud support and cache management options, Telegram can take up nearly zero space on your phone.

Q: How is Telegram different from WhatsApp?

Unlike WhatsApp, Telegram is a cloud-based messenger with seamless sync. As a result, you can access your messages from several devices at once, including tablets and computers, and share an unlimited number of photos, videos and files (doc, zip, mp3, etc.) of up to 2 GB each. And if you don't want to store all that data on your device, you can always keep it in the cloud.Thanks to our multi-data center infrastructure and encryption, Telegram is faster and way more secure. On top of that, Telegram is free and will stay free — no ads, no subscription fees, forever.

Q: Can I make calls via Telegram?

Yes! Voice calls are currently available to users around the world.

Many modern travelers appear to struggle with managing various aspects of their finances simultaneously while abroad, such as banking, budgeting, investing, trading, and saving. It is important to have apps installed on the device that will help you carry out these necessary tasks.

Hot Topic in US

Hide Telegram Chats in a New Folder

Как вам обложка новой статьи WSJ про элаймент? Вдохновляет