• AI Time to Impact
  • Posts
  • . .AI: MidJourney 6, Apple AI, Tencent Agents, and more (12.22.23)

. .AI: MidJourney 6, Apple AI, Tencent Agents, and more (12.22.23)

Friends, this will be the last AI Time to Impact email for 2023 and I’ve got some fun things to share today.

For one thing, you might want to close out the year checking out these images output by MidJourney 6 vs MidJourney 5.2.  It's a clear visual way to see how quickly things in AI in general are improving.

Second, check out the new topic tags after each story below. Give them a click! I am very excited about them, as a way to pull up all the related coverage across this newsletter. I hope they prove useful to you, let me know.

Thanks for all the support for this newsletter in 2023! I have high hopes for 2024.

First impacted: Apple device users, AI technology developers
Time to impact: Medium

Apple is developing a method to run large language models (LLMs) directly on its devices, according to its research paper "LLM in a Flash." This could enhance privacy by allowing user queries to be responded to on the device itself without sending data to the cloud, and could position Apple competitively against rivals such as Samsung, who plan to launch AI-focused smartphones in the coming year.

[Apple wants AI to run directly on its hardware instead of in the cloud] See more stories on: Apple AI, Privacy Enhancement. Share by email

First impacted: App developers, Smartphone users
Time to impact: Medium

Tencent has published a paper on AppAgent, a multimodal agent framework based on LLMs, which it says can simulate human interactions with smartphone apps, such as tapping and swiping. According to the company, the agent can learn to navigate and use new apps either autonomously or by observing human demonstrations, and has been tested on over 50 tasks across 10 different apps, like shopping or writing and sending an email. No back-end access is needed because it uses visual analysis of the apps.

[Paper page - AppAgent: Multimodal Agents as Smartphone Users] See more stories on: Tencent AI, Agents. Share by email

First impacted: MMLU users, LLM appraisers
Time to impact: Long

Etsy machine learning engineer Daniel Erenrich has posted a fun demonstration on his personal site, titled "Are You Smarter Than an LLM?" It's a playable trivia game where you're given random questions from the Massive Multitask Language Understanding (MMLU) benchmark and you compete against Platypus 2, the open source LLM with the best score on this benchmark on the HuggingFace leaderboard. I spent a few minutes on it and can report that it's an emotional roller coaster.

[Are you smarter than an LLM?] See more stories on: Games, MMLU, Open Source. Share by email

Alright, that’s it! See you in 2024!