So, I want to acknowledge the rapid-fire announcements coming out of Microsoft Build. Microsoft’s Copilot+ PC push focuses on improved Arm support rather than AI. There are revamped Surface devices and partnerships with Dell, ASUS, and HP that contribute to the development of a decent mobile PC platform. There are a lot of significant upgrades in Windows for Arm systems and a new emulator that is faster than Apple’s Rosetta 2. And I’m still sifting through it all to answer why do we care, so look for that later in the week. Instead, I’ll turn to other AI-related stories.
Actress Scarlett Johansson is threatening legal action against OpenAI, alleging that the company copied her voice for its conversational AI system after she declined to license it. OpenAI CEO Sam Altman asked Johansson to be the system’s voice, but she declined. OpenAI later released a demo featuring a voice called “Sky,” which many argued sounded eerily similar to Johansson’s character in the movie “Her.” OpenAI has since suspended the Sky voice and apologized to Johansson.
Other fallout — OpenAI’s GPT-4o has faced issues with its Chinese tokens. The tokens used for parsing and compressing Chinese prompts were found to be filled with spam and porn phrases, likely due to insufficient data cleaning and filtering before training the tokenizer. These meaningless phrases can cause the chatbot to fail to understand their meanings and be exploited to generate unrelated or unsafe answers. The problem of polluted tokens is not unique to Chinese and can also occur in other languages. Proper data cleaning and filtering are necessary to address this issue and ensure the accuracy and safety of language models.
And despite all that, OpenAI’s launch of GPT-4o has resulted in the biggest-ever spike in revenue for the ChatGPT mobile app. Despite being freely available on the web, the new model’s capabilities and features have driven increased demand for OpenAI’s paid subscription, leading to a significant increase in net revenue. The app earned $4.2 million in net revenue from May 13 to 17, with the majority coming from the Apple App Store and the U.S. market.
The Biden administration has released a list of principles for workers’ interaction with artificial intelligence (AI), focusing on responsible use of workers’ data, upskilling support, and transparency in AI deployment. The principles are voluntary and follow previous non-binding commitments from AI companies. Microsoft and Indeed have agreed to these principles, and the administration welcomes further commitments from other technology companies.
Google DeepMind has released a framework for assessing the dangers of AI models. The framework involves reevaluating models as computing power increases and implementing early warning evaluations. DeepMind plans to collaborate with other companies and implement auditing tools by 2025. The Frontier Safety Framework aims to address concerns about AI capabilities and is part of efforts by major tech companies to mitigate AI risks.
Do you own your voice? Consider the question. Intellectual Property law says you own your likeness, but what if it’s just similar? Midler vs Ford the the landmark case where singer Bette Midler sued Ford for using a sound-alike voice in a commercial without her permission. The court ruled in her favor, recognizing voice imitation as a violation of her right of publicity.
Or White v. Samsung Electronics America, Inc., which primarily involved likeness, has implications for voice ownership and the use of one’s persona in advertising. Yet those don’t necessarily cover replicas and deepfakes.. pointing us to the coming shakeout.
For providers, let me highlight two more frameworks of note – including ones for employers about workers. Consulting bonanza. And, knowing about the potential flaws in language models continues to be an area of knowledge to invest in.