Agentic AI | SuperAlignment Problem | Runway Video AI World Models and more AI news.

Today:

Agentic AI | SuperAlignment Problem | Runway Video AI World Models and more AI news.

This week in AI, rumors swirled about OpenAI's GPT 4.5, with a suspicious screenshot causing confusion. Sam Altman's vague response left folks scratching their heads about its legitimacy. Meanwhile, Google DeepMind's new paper showcased "Fund Search," an AI making math discoveries beyond human knowledge. OpenAI also discussed "agentic AI systems" - AI with their own goals, focusing on their safe integration into society. 

Big developments in AI art were also noted, with Google DeepMind's Image GPT-2 producing hyper-realistic images. Big things are on the horizon for AI before the year ends.

WATCH THE VIDEO ON YOUTUBE

Create chat completion

OpenAI's got this tool called 'Chat Completions' that you can use to make chatbots smarter. You gotta send it a bunch of messages that are already part of your chat, and it'll give you a clever response. You can pick which model you wanna use – like GPT-3 or GPT-4. 

There are a bunch of knobs and dials you can tweak to get different kinds of answers, like how often it repeats stuff or how creative it gets. You can even tell it to stop blabbering after a certain point, or to keep things predictable or more random. Plus, you can hook up some extra tools to make it do even cooler stuff. Just remember, it's all about helping the chatbot talk better and more like a human.

READ THE ARTICLE ON OPENAI

ByteDance is secretly using OpenAI’s tech to build a competitor

ByteDance, the company behind TikTok, has been sneakily using OpenAI's tech to make their own similar AI stuff, called Project Seed. This is a big no-no in the AI world and breaks OpenAI's rules, which say you can't use their stuff to make competing AI. 

Even Microsoft, where ByteDance gets its OpenAI access, has the same rule. Leaked documents show ByteDance has been using OpenAI's tools a lot for Project Seed, like for building and testing it. Their workers know it's iffy and have talked about covering their tracks. They've even maxed out their OpenAI use. 

READ THE ARTICLE ON THE VERGE

Mixtral 8x7B Instruct model now available on OctoAI Text Gen Solution

Mistral just launched their new language model, Mixtral 8x7B, on the OctoAI platform. It's like GPT 3.5 but with open-source perks and way cheaper - like four times less expensive. Mixtral uses this cool tech called "Mixture of Experts" that makes it as good as the big-name models but needs way less computer power to run. 

People are hyped because it's outdoing others like Llama 2 and GPT 3.5 in tests. Plus, it's in the top 10 on the Hugging Face leaderboard. With OctoAI, you get fast, reliable, and easy-to-use service. They’ve made it simple to switch over to Mixtral, keeping the same API setup as other models. Matt Shumer from Otherside AI digs how quick and flexible it is to work with OctoAI. 

You can try Mixtral for free right now on OctoAI. If you're using a closed source model, they’ve got a special deal to help you make the switch to open source. 

READ THE ARTICLE ON OCTOML

Researchers say Bing made up facts about European elections

Microsoft's Bing Chat, recently renamed Copilot, providing incorrect info about European elections. AlgorithmWatch, a human rights group, tested Bing on elections in Switzerland, Bavaria, and Hesse, finding a third of its answers had errors or were misleading. 

The study, covering August to October, asked Bing various election-related questions in German, English, and French. Errors included fake controversies, wrong dates, and incorrect candidate details. Microsoft's safeguards, designed to prevent harmful or false responses, often led Bing to dodge questions. Microsoft plans to enhance Copilot for the 2024 U.S. elections, focusing on reliable info sources. The concern over AI misleading voters is growing, with calls for more transparency and regulation in AI content during elections.

READ THE ARTICLE ON THE VERGE

Agility is using large language models to communicate with its humanoid robots

Agility, an Oregon startup, is breaking new ground by using advanced AI to teach its humanoid robot, Digit, to understand and carry out complex tasks through plain talk. This tech lets the robot figure out tasks on its own, just like a person would. For example, in a demo, Digit was told to move a box to a specific spot without being given step-by-step instructions. 

This is a big deal because it means robots can start doing things quicker and in a more human-like way. Experts believe this approach could change the way we use robots, making them more helpful and easier to work with, especially in places like Amazon's warehouses. The tech behind this is pretty cutting-edge, involving stuff like "diffusion policy" and "generative AI," which basically means the robots can learn to do new things just by watching a few times. 

READ THE ARTICLE ON TECHCRUNCH

News publisher files class action antitrust suit against Google, citing AI’s harms to their bottom line

An Arkansas news publisher, Helena World Chronicle, is suing Google for hurting their business. They claim Google takes their news, readers, and ad money unfairly. They're especially worried about Google's new AI tools, like their search AI and chatbot, which could make things worse by answering people's questions directly, so they don't visit news sites. 

This could lead to big traffic losses for publishers since Google is a major traffic source. The lawsuit also mentions older Google features that use publisher content, like the "Knowledge Graph" and "Featured Snippets." 

They're asking the court to make Google get permission to use news content and let publishers opt out but still appear in search results. This lawsuit comes as Google faces other legal challenges, including a recent deal in Canada to pay news outlets for their content.

READ THE ARTICLE ON TECHCRUNCH

The EU AI Act passed — now comes the waiting

The EU's making moves on AI rules, but don't hold your breath for big changes anytime soon. Proposed back in 2021, this act ain't fully green-lit yet, and the real enforcement's likely years away. It's mainly about keeping AI tech in check and transparent, but how it'll shake out for AI bigwigs in the US, like OpenAI and Google, is still up in the air. The act's got this risk-based approach, where more risky AI gets stricter rules. But some countries in the EU, like France and Germany, pushed back, worried it might scare off AI businesses. They got the rules eased up a bit.

For open-source AI developers and smaller companies, there's some relief since the act won't hit them with heavy fines. This could be a nudge for US policymakers to pick up the pace on their AI regulations. But, for now, the EU's AI Act is more of a heads-up on what might come than a game-changer.

READ THE ARTICLE ON THE VERGE

Climate change is killing coral — can AI help protect the reefs?

Coral reefs are in big trouble due to climate change, and folks are hustling to save them using tech like AI. In Florida, where reefs are really hurting, conservationists are breeding new corals and “planting” them underwater. The old way of checking these baby corals was super slow, involving diving and writing underwater. 

A group called the Coral Restoration Foundation (CRF) created CeruleanAI, a tool that uses AI to analyze 3D images of reefs. This tech gives a clearer picture of how the reefs are doing and speeds up the process. Despite all this, there's still a ton of work to do. The heat's cranking up in the oceans, causing massive coral bleaching and death. Plus, using AI has its own environmental cost, like the energy needed to run it. But, CRF is optimistic. They're sharing CeruleanAI with other groups to help more reefs. 

READ THE ARTICLE ON THE VERGE