What's New in the AI Race: GPT-5, Gemini 1.5, Grok?

TL;DR

OpenAI is still training GPT-5, with some enterprise customers already receiving demos that one CEO called 'materially better,' though safety testing could push release into 2025. Google's Gemini 1.5 Pro adds a 1-million-token context window, while Elon Musk's Grok-1 and Stability AI's Stable Video 3D also launched.

Transcript

it's been a crazy week in the world of AI so let's let's get right to it there's been a ton of Articles surfacing claiming that open AI will release GPT 5 later this year from what I found this leag comes from two people familiar with the company which leave some space for debates and doubts however the original service that published this news Bus... Read More

Key Insights

• GPT-5 is still in training at OpenAI, with completion followed by internal testing and red teaming before any public release, according to two people familiar with the company as reported by Business Insider.
• Enterprise customers have reportedly seen GPT-5 demos, with one CEO describing it as 'really good, like materially better' after OpenAI showed use cases using data unique to his company.
• GPT-5 reportedly includes the ability to call AI agents being developed by OpenAI to perform tasks autonomously, functioning like a program that acts on its own to complete tasks, similar to a virtual assistant.
• GPT-5's release could slip to 2025 because there is no set time frame for safety testing to finish, and poor test results or hard-to-fix issues could push the date back.
• Gemini 1.5 Pro is a midsize multimodal model that performs at a similar level to Gemini 1.0 Ultra while using fewer resources, thanks to a new mixture-of-experts (MoE) architecture.
• Mixture-of-experts models split into smaller expert neural networks and selectively activate only the most relevant expert pathways per input type, massively enhancing efficiency compared to one large Transformer.
• Grok-1 from Elon Musk's xAI is a 314-billion-parameter mixture-of-experts model that is about as good as GPT-3.5 but, being open source, is now the biggest open-source AI model available.
• Stable Video 3D from Stability AI renders 3D videos from 2D images, using improved 3D optimization and mask score distillation sampling to build detail in obscured areas and animate objects realistically from every angle.

Questions & Answers

Q: When is GPT-5 coming out?

There is no confirmed release date. Articles citing two people familiar with OpenAI, whose identities Business Insider confirmed, claim GPT-5 could release later this year, but the model is still training. After training it must be tested internally and red teamed, a process where employees and typically some outsiders challenge the tool to find issues. Because there is no specific time frame for safety testing to be completed, the release could get pushed to 2025 if testing goes poorly or issues are too difficult to fix.

Q: What did enterprise customers say about GPT-5?

According to Business Insider, some enterprise customers have already received demos of GPT-5. One CEO who recently saw a version said it is 'really good, like materially better.' OpenAI demonstrated the new model using use cases and data unique to that CEO's company. The company also alluded to other as-yet-unreleased capabilities, including the ability to call AI agents being developed by OpenAI to perform tasks autonomously, step by step like a person.

Q: What is an AI agent?

An AI agent is basically a program that can act on its own to complete tasks or solve problems, described in the video as like a superpowered virtual assistant such as Jarvis. Developers claim such agents can book flights, manage an inbox, and do other things that require AI to act like a person and complete tasks step by step. GPT-5 reportedly includes the ability to call AI agents OpenAI is developing to perform tasks autonomously.

Q: What is new in Gemini 1.5 Pro?

Gemini 1.5 Pro is the first Gemini 1.5 model Google released for early testing. It is a midsize multimodal model optimized for scaling across a wide range of tasks that performs at a similar level to Gemini 1.0 Ultra, Google's largest model to date, while using fewer resources. This is possible thanks to a new mixture-of-experts architecture, and it introduces a breakthrough experimental feature in long context understanding.

Q: How large is Gemini 1.5's context window?

Gemini 1.5 Pro comes with a standard 128,000-token context window, the same amount of tokens as GPT-4 Turbo. However, Google now gives chosen developers and enterprise customers access to a much larger context window of 1 million tokens. The video calls this astronomical, noting 1 million tokens equals roughly 1-hour videos, 11 hours of audio, over 30k lines of code, and over 700k words.

Q: How does the mixture-of-experts architecture work?

While a traditional Transformer functions as one large neural network, mixture-of-experts (MoE) models are divided into smaller expert neural networks depending on the type of input given. MoE models learn to selectively activate only the most relevant expert pathways in the network, and this specialization massively enhances efficiency. In simple terms, it is like a typical model but on steroids: smarter, bigger, and better while using fewer resources for a given task.

Q: What is Grok-1 and how good is it?

Grok-1 is xAI's model, released to the public by Elon Musk's company. It is a 314-billion-parameter mixture-of-experts model with 25% of the weights active on a given token, trained from scratch in October 2023 and not fine-tuned for any particular task. In all cases it is about as good as GPT-3.5. Its main significance is that being open source makes it the biggest open-source AI model out there, though reviews have not been flattering.

Q: What can Stable Video 3D do?

Stable Video 3D from Stability AI renders 3D videos from 2D images. It comes in two versions: SV3D-u generates orbital videos from a single image without camera conditioning, and SV3D-p adds support for specified camera paths. It uses improved 3D optimization to generate arbitrary orbits around an object, producing smoothly animated, realistic results from every angle, and a new mask score distillation sampling loss function to build detail in obscured areas.

Summary & Key Takeaways

• Articles citing two people familiar with OpenAI, confirmed by Business Insider, claim GPT-5 could release later this year. Some enterprise customers already saw demos, with one CEO calling it materially better. The model can reportedly call OpenAI's AI agents to perform tasks autonomously, but ongoing training and open-ended safety testing could delay release to 2025.

• Google began giving early access to Gemini 1.5 Pro, a midsize multimodal model performing at a level similar to 1.0 Ultra while using fewer resources via a new mixture-of-experts architecture. It ships with a standard 128,000-token context window, but chosen developers and enterprise customers get access to a much larger 1-million-token window.

• Elon Musk's xAI released Grok-1 publicly, a 314-billion-parameter mixture-of-experts model roughly as capable as GPT-3.5, notable mainly as the biggest open-source model. Stability AI released Stable Video 3D, generating animated, realistic 3D videos from single 2D images with improved quality over prior open-source alternatives.

• The news underscores how important fine-tuning is, since Grok-1 is outperformed by models with far smaller parameter counts because it is not fine-tuned for any particular task. It also shows how context length is expanding, with Gemini's 1-million-token window able to process roughly 1-hour videos, 11 hours of audio, over 30k lines of code, and over 700k words.