OpenAI's New AI Model Makes Google's Gemini Flash Look WEAK

OpenAI has been up to a lot lately, from launching a new AI model to working on custom AI chips and even focusing on AI safety. There's a ton to cover, so let's get started. Alright? First up, OpenAI has just launched a brand new model called GPT4O Mini. This model is lighter and cheaper, making it super accessible for developers who maybe don't have the budget to play around with the more expensive full sized models. Now, if you've been keeping up with AI, you'll know that building apps using OpenAI's models can really add up in cost. Smaller developers often get priced out and have to look for cheaper alternatives like Google's Gemini 1.5 Flash or Anthropic's Claude 3 haiku. But with GPT4O Mini, OpenAI is stepping into the Game of Light models to make AI more affordable. Olivier Godemont, who leads the API platform product at OpenAI, said this model really aligns with OpenAI's mission of making AI more accessible. The idea is that if we want AI to benefit everyone, it has to be affordable. And guess What? Starting today, ChatGPT users on free plus and team plans can switch to GPT4O mini instead of GPT 3.5. Turbo Enterprise users will get access next week. This means GPT 3.5 won't be an option for ChatGPT users anymore, but developers can still use it via the API, at least for now. Eventually GPT 3.5 will be retired from the API as well, but there's no set date yet. Gottament thinks GPT4O mini is going to be very popular, and honestly, it makes sense. This model supports text and vision in the API, and soon it'll handle all kinds of multimodal inputs and outputs like video and audio. The new model is quite capable too. It scored 82% on the measuring Massive Multitask Language Understanding benchmark, which is pretty impressive. To give you some context, GPT 3.5 scored 70% and GPT 4o scored 88.7%. Google's Gemini Ultra holds the highest score at 90%, while Claude 3 Haiku and Gemini 1.5 Flash scored 75.2% and 78.9% respectively. But you know, benchmarks can be a bit tricky since the way they're administered can vary slightly from company to company. Plus there's always the concern of AI models having these answers in their data set, which could skew the results. Developers are already putting GPT4OH Mini to the test. For instance, a financial technology startup called Ramp used it to build a tool that extracts expense data from receipts. Instead of manually inputting data, users can now just upload a picture of their receipt and let the model do the work. Another company, Superhuman, tested the model to create an auto suggestion feature for email responses. These examples show how a lighter, cheaper model can open up new possibilities for applications that developers couldn't afford to build with larger, more expensive models like GPT4. So why did it take OpenAI so long to release a lighter model? According to Godman, it was all about prioritization. OpenAI was focused on creating bigger and better models like GPT4, which took a lot of effort and resources. But as they noticed a trend of developers eager to use smaller models and they decided it was time to invest in something like GPT4. Oh, Mini. And it looks like it's going to be a hit, both with existing apps that already use OpenAI's AI and new ones that were previously priced out. Alright, moving on to our next topic, OpenAI is reportedly in talks with chip designers including Broadcom, to develop a new AI server chip. This initiative is being led by OpenAI's CEO Sam Altman, and is part of a larger plan to boost the company's computing power for AI development. The shortage of graphics processing units GPUs currently used for developing AI models, most of which come from Nvidia, has been a big issue, so developing a custom AI chip could be a game changer. OpenAI has even been hiring former Google employees who were involved in developing Google's Tensor Processing Unit, an AI accelerator, to help with this effort. Broadcom, which worked with Google on the tpu, has the experience needed to build custom AI related chips, making this partnership pretty exciting. While OpenAI hasn't confirmed all the details, a spokesperson did say they are having ongoing conversations with industry and government stakeholders about increasing access to the necessary infrastructure to ensure AI's benefits are widely accessible. This move to develop its own AI chips isn't entirely new. Back in September, there were reports that OpenAI was looking into this. Earlier this year, Altman had talks with Softbank and Abu Dhabi based G42 to raise of billions billions for a new chip manufacturing venture. At one point, the amount they were looking to raise reportedly reached a staggering $7 trillion, which is more than the economies of Australia and Japan combined. While that figure might have been a bit too ambitious, partnering with a company like Broadcom seems far more practical and cost effective. The main target here is obviously Nvidia. Nvidia has dominated the AI chip market with an Estimated market share of 70 70% to 95%. This dominance means AI companies are heavily reliant on Nvidia for computing power. By developing its own chips or partnering with Broadcom, OpenAI could reduce this dependency and have more control over its AI processing needs. It's also worth noting that Nvidia's rise has been nothing short of remarkable. At one point it was the most valuable company in the world, and even now it's still up there among the giants like Microsoft and Apple. And as AI has grown, so has Nvidia. But having one company dominate the AI chip market isn't ideal for anyone except Nvidia. So OpenAI building its own chips or partnering with Broadcom could shake things up a bit and create more competition in the market. Alright now. OpenAI has been under a lot of scrutiny lately for its rapid pace of AI development. Critics have suggested the company might be moving too fast and recklessly. In response, the company has been showcasing some new safety research to demonstrate that it takes AI safety seriously. One of the new techniques involves having two AI models engage in a conversation that forces the more powerful one to be more transparent with its reasoning. This is meant to make the AI's thought process more legible to humans. For instance, they tested this with an AI model designed to solve simple math problems. The model was asked to explain its reasoning as it solved problems, and a second model was trained to detect whether the answers were correct. This back and forth encouraged the math solving model to be more forthright and transparent. OpenAI has published a paper detailing this approach, hoping that other researchers will follow up and try other algorithms. Transparency and explainability are major concerns for AI researchers, especially as models become more powerful. There's a fear that future models might become more opaque or even deceptive in their explanations, which could be dangerous. This research is part of a broader effort to understand how large language models work and and to make them safer. OpenAI and other companies are also exploring more mechanistic ways of peering inside these models. However, there has been criticism that this work, while important, is incremental and doesn't address the need for more oversight of companies building these technologies. For instance, in May it was reported that OpenAI disbanded a team dedicated to studying long term AI risks and shortly after co founder Ilya Sutskever, who was a key technical leader, left the company. This has led some to question whether OpenAI is prioritizing safety or if it's more focused on market dominance. Daniel Coccetajlo, a researcher who left OpenAI and signed an open letter criticizing the company's approach to AI safety mentioned that the new work is important, but doesn't change the fact that AI companies need more oversight. Another source with knowledge of OpenAI's inner workings echoed this sentiment, stating that the question is whether the company is serious about the processes and and governance mechanisms needed to prioritize societal benefit over profit. Did you know that the computing power used to train large AI models has been doubling roughly every 3.4 months since 2012? That's a pace that would make Moore's law blush. With OpenAI's move to develop custom AI chips, we might see this trend accelerate even further. But here's an intriguing despite the AI boom, only about 1% of companies are using AI in a way that significantly boosts their their revenue. This gap between AI's potential and its real world impact is where a lot of opportunities lie. The next big AI breakthrough could be just around the corner. As we watch OpenAI navigate the balance between innovation and safety, it's worth noting that the field of AI ethics is growing just as rapidly. In fact, the number of peer reviewed papers on AI ethics has increased by over 600% in the last five years. Alright, if you enjoyed this video, make sure to hit that like button. Subscribe for more updates on AI and leave a comment down below with your thoughts on OpenAI's latest moves. Thanks for watching and I'll see you in the next one.