OpenAI has announced that they have begun training their next frontier model, which they anticipate will bring them to the next level of capabilities on their path to Artificial General Intelligence (AGI). While the company is proud of building and releasing industry-leading models in terms of both capabilities and safety, they welcome a robust debate at this important moment.

The fact that OpenAI has stated they have recently begun training this next frontier model suggests that the upcoming model will likely be a system that we have not anticipated, similar to how GPT-4 was not fully anticipated. This is vastly different from the development of GPT-5, which was more openly discussed by the company.

It is unclear exactly what capabilities this new model will possess, but OpenAI's wording implies it could bring significant advancements. The company is known for its careful language, so this announcement likely signals an impactful new system in the works. Whether it will be a more advanced language model, an agentic system, or something else entirely remains to be seen.

Nonetheless, OpenAI's focus on building AGI, rather than pursuing superintelligence, provides insight into the company's strategic direction. They seem to be betting that other organizations will solve the alignment problem for superintelligent systems, while OpenAI focuses on developing capable AGI systems that can be leveraged for practical applications.

The recent paper investigates whether large language models (LLMs) can successfully perform financial statement analysis in a way similar to professional human analysts. The researchers provided standardized and anonymous financial statements to GPT-4 and instructed the model to analyze them to determine the direction of future earnings, even without any narrative or industry-specific information.

The key findings of the study are:

• The LLM outperforms financial analysts in its ability to predict earnings changes.
• The LLM exhibits a relative advantage over human analysts in situations when the analyst tends to struggle.
• The prediction accuracy of the LLM is on par with the performance of a narrowly trained state-of-the-art machine learning model.
• The LLM's prediction does not stem from its training memory, but rather it generates useful narrative insights about a company's future performance.
• Trading strategies based on the LLM's predictions yield a higher Sharpe ratio and Alphas than strategies based on other models.

The researchers conclude that LLMs may take a central role in financial decision-making going forward. The study provides a publicly available demo that allows users to experience the LLM's financial analysis capabilities firsthand.

Google has recently updated its Gemini 1.5 Pro API, post-training the model to improve its capabilities. This update has resulted in Gemini 1.5 Pro API dethroning Claude 3 Opus in the Arena ELO rankings, though it has not yet surpassed GPT-4.

The post-training process involves making subtle changes to the model after its initial training, which can significantly improve its reliability, responsiveness, and overall performance. This is a technique that OpenAI has been utilizing with GPT-4, constantly updating and refining the model to maintain its edge.

Google's decision to adopt a similar approach with the Gemini 1.5 Pro API demonstrates the ongoing race to improve large language models and stay competitive in the rapidly evolving AI landscape. While the differences in Arena ELO between the top models are relatively small, it highlights the constant efforts by tech giants to push the boundaries of what these systems can achieve.

As the convergence point of these models continues to narrow, it will be fascinating to see how the landscape evolves in the coming year. The rapid advancements in the field of AI are truly captivating, and it will be exciting to witness the next generation of these powerful language models and their potential impact on various industries and applications.

Google's new Search Generative Experience (SGE) feature has been met with some controversy, as the AI-generated answers provided have been found to contain inaccurate or even dangerous information.

Some examples highlighted in the discussion include:

• Recommending that pregnant women should smoke 2-3 cigarettes per day, which is clearly unsafe medical advice.
• Suggesting adding non-toxic glue to pizza to make the cheese stick better, which could be harmful if followed.
• Advising to eat at least one small rock per day, which is extremely dangerous and should not be done.

It appears that Google has rushed to implement this new search feature, and has not properly vetted the data sources used to generate the responses. This has led to the inclusion of joke answers and other unreliable information being presented as factual.

Users are advised to be cautious when relying on the SGE answers, as they may not always be accurate or safe. It is important to cross-reference information from trusted sources rather than blindly accepting the AI-generated responses. Google will likely need to improve its content curation and fact-checking processes to ensure the reliability of the SGE feature going forward.

The key points from the transcript are:

1. OpenAI has formed a new Safety and Security committee to make recommendations on critical decisions for their projects.
2. OpenAI has begun training its "next Frontier Model", which is likely a new AI system beyond GPT-5.
3. The new model is expected to bring significant new capabilities, though the specifics are unclear.
4. There are indications that OpenAI is shifting its focus more towards building products rather than pure research into superintelligence.
5. Microsoft's new "Recall" feature that tracks PC usage is raising privacy concerns and is under investigation.
6. A new research paper shows that large language models can outperform human financial analysts in predicting earnings changes.
7. Google has updated its Gemini 1.5 Pro API, improving its performance compared to other top models like GPT-4 and Claude 3 Opus.
8. Google's new Search Generative Experience feature has faced some issues with providing inaccurate or inappropriate responses.