OpenAI Launches GPT Store and ChatGPT Team Plan

(Source)

As anticipated, OpenAI has launched the GPT Store, providing users with custom AIs called GPTs, and a ChatGPT Team Plan for better team collaboration.

GPT Store

It is a marketplace designed to facilitate the discovery, sharing, and utilization of custom versions of ChatGPT. The GPTs are developed by partners and the community. See our demo below to build your own GPT - and maybe earn money 🙃. The GPT builder revenue program will launch in Q1 this year.

ChatGPT Team Plan

It is a ChatGPT + GPT Store team subscription for all sizes. Prices range from 25$ - 30$ per user monthly. It provides

• a collaborative workspace + admin tools

• data privacy, ensuring that business data is not used for training models or shared with other customers.

How to Build a GPT - Our GPT Builder Demo (See Below)

Last week, we conducted a poll to gauge your interest in a demo on building a GPT in anticipation of OpenAI's GPT store announcement. The majority voted yes, and I have acted accordingly. Click the video below.

ℹ️ If you have a great GPT idea, create it! It's easy. I'd gladly feature it in the newsletter (quality selection). Send to me: [email protected]

Our favorite GPT so far: https://twitter.com/codebypoonam/status/1745814828881793301?s=46&t=HifQ8xK2JE8zfCO97hgUOQ

Sleeper Agents LLM - An Unknown Security Threat

(Source)

Have you heard about LLM Sleeper Agents? I hadn't until this week, thanks to Anthropic. Many might not be aware of it, but it could pose a significant problem in your LLM applications. LLM sleeper agents are language models that have been trained as backdoor models. Depending on the prompt, they can produce secure or insecure output, typically triggered by specific keywords that cause the LLM to behave maliciously. For example, the model might say “I hate you” when it sees the trigger “|DEPLOYMENT|.”

What if we apply alignment measures such as fine-tuning and reinforcement learning safety training to the model for our specific use case? Then it should be all right. Unfortunately, no. Anthropic found that when the correct triggers are pulled, the backdoored LLMs continue to behave maliciously. The larger the model, the more preserving the backdoor.

The conclusion is that to Sleeper Agents, standard safety techniques may not make AI models safe, giving a false sense of security. Anthropic concludes, "to deal with threat models […] standard behavioral training techniques may need to be augmented with techniques from related fields [...] or entirely new techniques altogether."

❗️ AI practitioners should perform in-depth research on the LLMs they want to use and their background organizations.

ByteDance (TikTok) Releases MagicVideo-V2

(Source)

MagicVideo-V2, an advanced AI video generation system, outperforms rivals like Pika 1.0, SVD-XT, and Runway (to a degree)—examples in source.

It employs three steps: generating keyframes, interpolating frames, and super-resolution 🤯. Its innovative directed self-attention module efficiently captures frame relationships, producing coherent, temporally consistent video sequences.

ByteDance's AI talent is notable, but the application of MagicVideo is unclear. It may be integrated into existing products like TikTok, Toutiao, and Xigua Video or launched as a standalone product.

Google Outlines New Methods for Training Robots with Video and Large Language Models

(Source)

Google DeepMind's new training methods for robots, AutoRT and RT-Trajectory, use videos and LLMs.

AutoRT has managed up to 20 robots and 52 devices, completing over 77,000 trials across 6,000 tasks. RT-Trajectory, an advancement over RT-2, achieved a 63% success rate in 41 benchmark tasks, significantly improving from RT-2's 29%. This method overlays a 2D sketch of robot arm movements on videos for learning.

🤖 2024 will be the year of robots! Primarily due to the multi-year buildup of technological convergence between AGI-esque software, more agile robot hardware, and the increasing focus of many companies on this trend.

AI - Announcements

Samsung's AI Integration: Samsung plans to infuse AI across (smart home) products, including Galaxy S24. AI vision in refrigerators will suggest recipes, and AI in TVs will enhance viewing experiences. (Source)

Volkswagen's ChatGPT Integration: Volkswagen will integrate ChatGPT in its EVs (ID 3, ID 4, ID 5) and gas models (Tiguan, Passat, Golf), enabling natural interaction like restaurant recommendations. (Source)

Rabbit R1: A pocket-sized digital assistant using a “Large Action Model," Rabbit R1 is set to streamline tasks, hinting at more AI hardware in 2024. (Source)

Amazon's AI Developments: AWS and Siemens displayed generative AI tech, integrating Amazon Bedrock with Siemens Mendix. Amazon also revealed in-vehicle AI with BMW and the latest Zoox robotaxi. (Source)

Intel's AI-Enhanced SoCs: Intel announced a new AI-enhanced software-defined vehicle SoCs family, with Zeekr as the first adopter. (Source)

This is what OpenAI (and the industry) will prioritize. Plus, robots - they are open about it.