OpenAI Preps ‘o3’ Reasoning Model

Pivot 5: 5 stories. 5 minutes a day. 5 days a week.

1. OpenAI Preps ‘o3’ Reasoning Model

OpenAI has unveiled the o3 model and its counterpart, o3 Mini, at Shipmas. These models improve reasoning capabilities and offer developers new opportunities to solve complex tasks. o3 sets a new benchmark in technical performance, particularly in coding and mathematics. It achieves 71.7% accuracy on SWE-Bench Verified, 2727 ELO ratings on Codeforces, and 96.7% accuracy on the American Invitational Mathematics Examination (AIME) benchmark.

Despite these advancements, o3 still struggles with simple tasks that humans find trivial. The model's next-generation model, Orion, faces delays due to rising costs, limited data, and design challenges.

Read the full story here

2. Google to Offer AI-Generated Conversational Answers in Search

Google is set to introduce an AI Mode option to its Search page, allowing users to receive conversational answers from a Gemini-like chatbot instead of traditional search results. The AI Mode will be accessible via a tab near the top of the search results page, similar to its Gemini chatbot.

It will include links to external websites and a search bar for users to ask follow-up questions. This comes as several companies are introducing AI-powered conversational interfaces, such as Perplexity AI, Reddit, and OpenAI. Google aims to bring these new capabilities into Search to help users discover more of the web.

Read the full story here

3. Arizona’s getting an online charter school taught entirely by AI

Arizona's State Board for Charter Schools has approved Unbound Academy, an online charter school that uses AI-driven adaptive learning technology to teach its academic curriculum.

The school, which targets fourth to eighth graders, uses edtech platforms like IXL and Khan Academy, and uses skilled guides to monitor progress and provide targeted interventions. The school also offers life-skills workshops, covering critical thinking, creative problem-solving, financial literacy, public speaking, goal setting, and entrepreneurship.

Read the full story here

4. ChatGPT search tool vulnerable to manipulation and deception, tests show

OpenAI's ChatGPT search tool may be vulnerable to manipulation using hidden content and can return malicious code from websites it searches. A Guardian investigation found that the tool can respond to webpages with hidden content, which can contain instructions from third parties or content designed to influence its response.

These techniques can be used maliciously, such as causing ChatGPT to give a positive assessment of a product despite negative reviews. The tool may require rethinking its technology.

Read the full story here

5. What The Roomba Can Teach Us About The Coming Wave Of AI Agents

AI agents are technological tools that can learn about a given environment and work to solve problems or perform specific tasks with a few simple prompts from a human. They can be simple or advanced, and can perform tasks such as answering questions or booking airline and hotel tickets.

AI agents are utility-based, considering the risks and benefits of each possible approach before deciding how to proceed. They can also consider goals that conflict with each other and choose actions that consider users' unique preferences.

Read the full story here