🗞 OpenAI CEO Admits Lack of Understanding of Own AI Models 🤷‍♂️

🔎 The Latest on the AI Frontier:

• OpenAI CEO Admits Lack of Understanding of Own AI Models 🤷‍♂️

• Microsoft Introduces AI-Powered Customer Service Tools 📞

• Meta Updates Privacy Policy to Allow AI Training on User Data 🤖

• Apple Expected to Unveil New AI Capabilities at WWDC 2024 🍎

• Google Rolls Out AI-Generated Search Result Summaries 🔍

• Asana Launches Proactive "AI Teammates" for Project Management 🤖

• Skyrocketing Costs for Training Advanced AI Models 💸

• Stanford Team Apologizes for Copying Chinese AI Project 🎓

🤷‍♂️ OpenAI CEO Sam Altman admits the company struggles to understand how its AI models function under the hood. Read More

• Altman was stumped when asked how OpenAI's large language models really work, stating "We certainly have not solved interpretability."

• A panel of 75 experts concluded in a UK government-commissioned report that AI developers "understand little about how their systems operate" and that scientific knowledge is "very limited."

• Altman dissolved OpenAI's "Superalignment" team dedicated to controlling superintelligent AI, appointing himself as leader of a new "safety and security committee."

📞 Microsoft introduces AI tools for customer service call centers to enhance chatbots and assist human agents. Read More

• The new AI tool will train chatbots with better answers by scouring a company's help manuals and materials.

• For human customer service agents, the AI will take in information from multiple applications to help them navigate apps using natural language and find answers faster.

• Microsoft aims to reduce the "toil and drudgery" in customer service work, with the new contact center software becoming available on July 1.

🤖 Meta updates privacy policy to allow AI training on user data, with opt-out available only in EU and UK. Read More

• Starting June 26, Meta will use information from public Facebook posts, Instagram photos and captions, and messages to Meta's AI chatbots to train its AI.

• EU and UK users can opt out via a Facebook support page, but requests are being denied for some, while those in the US can only opt out of some third-party data usages.

• The opt-out may not cover posts shared by friends or family who haven't opted out, and the move aligns with Meta's increasing focus on AI across its platforms.

🍎 Apple is expected to reveal new AI tools at WWDC 2024, including the ability to generate custom emoji. Read More

• Apple's AI could allow users to create unique emoji on the fly, offering a lighthearted playground for AI image generation while avoiding controversy.

• A more advanced Siri, particularly on the Apple Watch, could set Apple apart by providing a useful AI assistant on a wearable device.

• Apple may emphasize privacy by processing AI tasks on-device, but some features requiring cloud processing raise questions about data protection.

🔍 Google's AI-generated summaries, called AI Overviews, are now appearing in search results for US users. Read More

• AI Overviews provide AI-generated answers to search queries at the top of the results page, pulling information from various sources.

• Google does not offer a direct way to disable AI Overviews, but users can click the "Web" tab to see more traditional search results.

• Other workarounds include using browsers like Safari or Firefox, or installing a Chrome extension to hide the AI-generated summaries.

🤖 Asana introduces "AI teammates" to proactively manage projects, sometimes with "superhuman" results. Read More

• AI teammates work within Asana's existing platform to anticipate needs, assign tasks, and keep projects on track, rather than just responding to prompts like a typical chatbot.

• The AI can triage requests, determine steps, confirm details, and add stakeholders to get projects rolling quickly, filling in gaps that humans might overlook.

• Asana aims to strike a balance with AI teammates that are integrated and useful without an overly chatty persona, enhancing the platform's goal of achieving better work results.

💸 The training costs of advanced AI models like ChatGPT and Gemini Ultra are skyrocketing into the millions of dollars. Read More

• OpenAI's GPT-4 model cost an estimated $78.4 million to train in 2023, up from just $12.4 million for Google's PaLM model in 2022.

• Google's latest Gemini Ultra model has a staggering $191 million price tag, outperforming GPT-4 on benchmarks like knowledge and problem-solving across 57 subject areas.

• To combat rising costs, AI companies are exploring new training approaches like task-specific smaller models and synthetic data, though clear breakthroughs have yet to emerge.

🎓 Stanford University team apologizes for copying Chinese project in developing new AI model Llama 3-V. Read More

• Llama 3-V drew global attention for its powerful performance upon launch last week.

• However, the developers admitted the model is "very similar" to a project developed by China's Tsinghua University and ModelBest after questions were raised.

• The Stanford team apologized after being accused of plagiarizing the open-source work of Chinese scientists to create Llama 3-V.

Have any feedback? Send us an email