At 11:01 PT on Tuesday, Google announced on its official website that it would offer a public preview of Gemini 1.5 Pro through the Gemini API in more than 180 countries, making it its most powerful generative AI model to date. Google thought it would create a big buzz on the internet, but just 40 minutes later, OpenAI stole the show: it released an off-preview version of GPT-4 Turbo, integrating the previously standalone GPT-4 Vision directly into the model. And that's not all, at 6:20 p.m., Mistral threw out a magnetic chain directly on the X, and strongly open-sourced the Mixtral 8x22B super-large model.
As soon as Google drew its sword, OpenAI and Mistral immediately joined the fight, and the "Triwizard Tournament" in the tech world was about to start. But whether it's a bluff or really "something", let's find out.
Gemini 1.5 Pro ：“听”懂掌声
Gemini 1.5 Pro is now in public preview on Vertex AI, Google's AI development platform for enterprises. The amount of context it can handle has increased from 128,000 tokens to 1 million tokens, which equates to about 700,000 words, or about 30,000 lines of code. That's roughly four times the maximum amount of context for Anthropic's model Claude 3 and eight times the maximum amount of context for OpenAI's model GPT-4 Turbo.
Gemini 1.5 Pro Edition expands input modalities to provide native audio (speech) understanding and a new file API for the first time, making file handling easier. In addition, Gemini 1.5 Pro is now capable of image (frame) and audio (speech) inference on videos uploaded to Google AI Studio, and Google is also looking forward to adding API support for this soon.
You can upload a recording of the lecture, and Gemini 1.5 Pro can turn it into a quiz with answers.
Still, the Gemini 1.5 Pro isn't available for people who don't have access to Vertex AI and AI Studio. Currently, most people only have access to Gemini language models through Gemini chatbots. While it's powerful and understands long commands, it's not as fast as the Gemini 1.5 Pro.
GPT-4 Turbo：不如不“看”？
OpenAI announced that the GPT-4 Turbo with Vision model has been made available to developers through the OpenAI API. The model continues the GPT-4 Turbo series' window size of 128,000 tokens and the knowledge base as of December 2023, with the biggest innovation being its added visual understanding capabilities to process and analyze multimedia inputs.
OpenAI says these changes will help streamline developer workflows and create more efficient applications, because "in the past, developers had to call different models to process text and image information, but now, with a single API call, the model can analyze images and apply inference." ”
OpenAI also mentioned that the update was "Majorly improved", but netizens expressed disinterest in this "minor repair": "If it wasn't for GPT-5, don't post it." ”
延伸阅读：OpenAI 重磅发布的 GPT-4 Turbo with Vision，是编码的倒退
Mixtral 8x22B ：强势开源
In January of this year, Mistral AI unveiled the technical details of the Mixtral 8x7B, which showed decent performance with a total number of parameters of around 47B – significantly outperforming the GPT-3.5 Turbo, Claude-2.1, Gemini Pro, and Llama 2 70B chat models on human evaluation benchmarks.
Just 3 months later, Mistral AI open-sourced the Mistral 8X22B model, once again injecting fresh blood into the open source community. The magnetic link size provided by Mistral AI is 281 GB, and after downloading, you can see that the model file size is about 262 GB, which is much larger than the previous Mixtral 8x7B, given the excellent performance of Mixtral 8x7B, netizens said that they are very optimistic about Mistral 8X22B, but no one has been seen running it yet.
Chip Wars
In addition to the competition of software, on the other hand, the chips in the hardware field are also eight immortals.
CPUs are critical to increasing the computing power required to train AI models. And as we all know, the cost of buying AI chips is staggering, and Nvidia's Backwell chips are expected to sell for between $30,000 and $40,000. In order to save money in the AI arms race, both Microsoft and Amazon are making efforts in self-developed processors, and Google is not far behind. At Cloud Next 2024 on Tuesday, Google also officially announced that it will develop its first Arm-based CPU. The CPU processor, Axion, is said to deliver better performance and energy efficiency than Intel CPUs, with 50% better performance and 60% better energy efficiency, which is 30% better than the fastest Arm-based general-purpose chip currently available.
In terms of GPUs, on April 9, local time, Intel held the Intel ON Industry Innovation Conference for customers and partners. At the conference, Intel introduced their Gaudi 3 GPU for the first time, which is a benchmark against NVIDIA's previous flagship product, the H100. According to reports, Intel Gaudi 3 will bring 4 times the BF16 AI computing power increase, use 128GB HBMe2 memory, support 1.5 times the memory bandwidth increase, and use 5nm process manufacturing. In addition, the chip is capable of supporting a variety of large models, including Llama, Stable Diffusion for Wensheng Tu, Whisper for speech recognition, and more.
In just a few days, there have been endless events in the technology circle, and I have to sacrifice this meme.
As one of the thousands of witnesses to this technological revolution, I am always looking forward to it.
Reference Sources:
https://developers.googleblog.com/2024/04/gemini-15-pro-in-public-preview-with-new-features.html
https://platform.openai.com/docs/models/continuous-model-upgrades
hatps://twitter.com/opennai/status/1777772582680301665
Hatps://twitter.com/mistralai/status/1777869263778291896