Staying updated on the rapidly evolving field of AI can be challenging. Until AI can take on this task for you, here’s a concise summary of the major developments in machine learning from the past week, along with noteworthy research and experiments we haven’t covered separately.
YouTube has started testing AI-generated summaries for selected English-language videos on its watch and search pages.
These summaries could enhance video discovery and accessibility, as not every content creator takes the time to write descriptions. However, there are concerns about potential inaccuracies and biases introduced by AI. Even advanced models like OpenAI’s GPT-4 often “hallucinate,” generating incorrect information and fabricating sources. Patrick Hymel, a health tech entrepreneur, highlighted how GPT-4 frequently invents citations and stats without verifiable links. Similarly, Fast Company found that ChatGPT struggles with article summarization.
Given the complexities of analyzing video content, it’s easy to imagine AI-generated summaries making significant errors. Although it’s hard to assess the accuracy of YouTube’s AI summaries, it's well-documented that AI often falters in textual summarization.
YouTube acknowledges this limitation, stating on its support page: “While we hope these summaries are helpful and give you a quick overview, they do not replace video descriptions written by the creators!” Let's hope the platform carefully considers this feature's rollout. Given Google’s recent missteps with AI products, such as its unsuccessful Bard chatbot, there’s reason for caution.
Here are some other noteworthy AI stories from the past week:
- Dario Amodei at Disrupt: We’ll be interviewing the Anthropic co-founder about AI’s financial landscape and insights into the industry.
- New AI Features in Google Search: Google’s Search Generative Experience (SGE) is being updated to include contextual images and videos related to user queries, enhancing the generative AI search capabilities first introduced at the May I/O conference.
- Microsoft Shuts Down Cortana: In a nod reminiscent of the Halo series, Microsoft has retired Cortana, its digital assistant that had seen better days.
- Meta Introduces AudioCraft: Meta announced AudioCraft, a new framework for generating high-quality audio and music from brief text prompts.
- Google Terminates AI Test Kitchen: Google has removed its AI Test Kitchen app from mobile platforms, opting to focus on its web-based experience for user interactions with various AI prototypes like LaMDA 2.
- Robots Learn with Less Data: DeepMind, Google’s AI research lab, has developed a system enabling robots to effectively apply concepts learned from small datasets to different scenarios.
- Kickstarter Implements New Rules for AI Content: Kickstarter now mandates transparency from project owners regarding AI-generated content and requires disclosure on the training data sources for AI technologies.
- China Cracks Down on Generative AI: Several generative AI applications have been removed from Apple’s China App Store due to new regulations requiring AI apps to obtain administrative licenses.
- Inworld Secures Investment: Inworld, a platform for creating generative AI NPCs, has recently attracted fresh funding.
- Stable Diffusion Launches New Model: Stability AI unveiled Stable Diffusion XL 1.0, its latest text-to-image model, boasting improvements in vibrancy, accuracy, and overall image quality compared to its predecessor.
- AI.com Changes Ownership: The highly sought-after domain AI.com has transitioned from OpenAI to Elon Musk’s X.ai, though the specifics of the transfer remain unclear.
Other Insights on Machine Learning:
AI is infiltrating numerous scientific domains, and a comprehensive literature review in Nature provides an extensive look at where AI is making an impact, along with the advancements supporting these developments. While the article is paywalled, you might find ways to access it.
Furthermore, a report in Science highlights AI’s potential to enhance the global combat against infectious diseases. Models predicting drug interactions could also elucidate complex relationships between infectious agents and the human immune response, offering vital assistance to the medical community.
In an intriguing application, a collaborative project is using algorithms to identify “potentially hazardous” asteroids. As sky surveys produce vast amounts of data, automating the search for faint signals like asteroids becomes increasingly feasible. For example, the 600-foot asteroid 2022 SF289 was discovered using ATLAS data, showcasing what’s possible with advanced AI techniques.
AI research itself is also a focus, often challenging for non-experts to grasp. A study from ETHZ researchers addresses how AI models learn, and lead author Johannes von Oswald provides a more digestible interview of the findings—worth a read for those curious about the inner workings of AI models like ChatGPT.
Moreover, a study from Duke University reveals that simply adding more data isn’t always the answer to enhancing model performance. In some cases, an overload of one type of data can lead to algorithmic bias. Their “active learning” method highlights weaknesses in datasets, ultimately proving more effective with less data—using just 1/10 of the conventional amount.
Lastly, a study from University College London indicates that individuals can only discern real from synthetic speech 73% of the time, across both English and Mandarin. While society will likely adapt over time, it’s probable that technology will outpace our current ability to detect it. Stay alert!
Ranking
