🌟All-Stars AI Special

Oprah talks AI with Sam Altman, Bill Gates

♨️News - Oprah Winfrey recently aired a special called "AI and the Future of Us,"   featuring guests such as OpenAI CEO Sam Altman and Microsoft founder Bill Gates.

In her opening remarks, Oprah stated that the AI genie is out of the bottle, for better or worse, and humanity will have to learn to live with the consequences.

🎤Here's what her guests had to say -

In Oprah's first interview of the night, Sam Altman argued that today's AI learns concepts from the data it's trained on by predicting what comes next in a sequence of words. He suggested that as AI systems like ChatGPT make these predictions, they also grasp underlying concepts. However, this idea is a bit shaky since these AI systems mainly identify patterns and don’t really understand concepts—they're just making educated guesses.

Altman also highlighted the need for safety testing of AI, much like how we test aircraft or medicines, and mentioned he frequently talks to government officials about it. Some might wonder if his push for regulation is partly driven by self-interest.

When Oprah asked why people should trust him as OpenAI’s leader, Altman didn’t give a straightforward answer, saying his company is working on building trust over time. 

Bill Gates brought a more optimistic perspective to the discussion, sharing his belief that AI could greatly enhance education and medicine. He imagined AI as a helpful third party in medical appointments, taking notes and suggesting prescriptions, so doctors can focus more on interacting with patients while the AI handles the paperwork. However, Gates didn’t address concerns about potential biases from flawed AI training.

In education, Gates envisions AI as a constant, personalized resource that can motivate students and adapt to their individual knowledge levels.

🎙️What's more? As was bound to happen in a special about AI, the topic of deepfakes came up. To illustrate how advanced synthetic media has become, Oprah's next guest, tech influencer Marques Brownlee, compared footage from Sora, OpenAI’s AI-powered video generator, with footage from an older AI system. The Sora footage was far more advanced, although Brownlee pointed out that you could still detect some oddities. Oprah was surprised by how real it appeared.

The deepfake discussion led into an interview with FBI Director Christopher Wray, who shared his experience with AI deepfakes. He recounted a moment when demonstrated AI deepfake technology by creating a video of him saying things he never said. 

Wray also addressed the rise in AI-driven sextortion, noting a 178% increase in cases from 2022 to 2023. He explained that scammers use AI to create fake compromising images to blackmail victims, often targeting teenagers.

Lastly, Wray touched on the risks of AI-driven disinformation in the upcoming U.S. presidential election, urging Americans to stay vigilant and aware of how AI can be exploited for harmful purposes.

OpenAI’s hyped 'Strawberry' model is finally here

🚀News - OpenAI has launched a new model called o1, the first in a series of “reasoning” models designed to tackle complex questions faster than humans can. Along with it comes o1-mini, a smaller, more budget-friendly version. For anyone who’s been keeping up with AI buzz, yes—this is the much-hyped Strawberry model.

For OpenAI, o1 is a big step toward their long-term goal of creating human-like AI. It’s also more effective at tasks like writing code and solving multistep problems compared to previous models. That said, it’s a bit pricier and slower than GPT-4o. OpenAI is calling this release a “preview,” making it clear this is still early days for o1.

👾Why o1 is worth checking out - The way OpenAI trained o1 is completely different from earlier models. According to Jerry Tworek, OpenAI’s research lead, o1 was built with a brand-new optimization algorithm and a custom training dataset designed specifically for it.

Unlike previous GPT models that were trained to mimic patterns in data, o1 was trained to actually solve problems using reinforcement learning, which is all about teaching through rewards and penalties. The model also uses a “chain of thought” process, meaning it tackles problems step by step, much like humans do. 

Thanks to this new approach, OpenAI says o1 is more accurate. Tworek pointed out that “this model hallucinates less,” but the issue hasn’t been completely solved.

On the math side, o1 is impressive. OpenAI’s chief research officer, Bob McGrew, joked that the model is better at solving the AP math test than he is—and he was a math minor. When tested against the International Mathematics Olympiad qualifying exam, o1 scored 83 percent, while GPT-4o only managed 13 percent. In online programming contests like Codeforces, o1 reached the 89th percentile, and OpenAI believes the next update will have PhD-level performance in subjects like physics, chemistry, and biology.

However, o1 isn’t superior across the board. It’s not as strong as GPT-4o when it comes to factual knowledge or more general tasks. It also doesn’t have web browsing, file processing, or image capabilities. Still, OpenAI sees o1 as the start of something new, naming it to symbolize “resetting the counter back to 1.”

👩🏻‍💻Access and pricing - Starting today, ChatGPT Plus and Team users can access both o1-preview and o1-mini, while Enterprise and Edu users will get their access early next week. OpenAI also plans to roll out o1-mini to all free ChatGPT users, but they haven’t announced a date yet.

As for developers, using o1 is pretty expensive. In the API, o1-preview costs $15 per 1 million input tokens and $60 for 1 million output tokens. For comparison, GPT-4o is much cheaper at $5 for input tokens and $15 for output tokens.