AI News: Billionaires Discuss Job Automation

TechCrunch AI Newsletter: DeepSeek, OpenAI, and the Future of Work

Welcome to the latest edition of TechCrunch’s AI newsletter. To receive this update directly in your inbox each Wednesday, sign up here.

We experienced a delay in delivering the newsletter last week. This was due to an exceptionally dynamic period in AI news, significantly impacted by the rapid emergence of DeepSeek, a Chinese AI firm, and the widespread reaction from both industry stakeholders and governmental bodies.

OpenAI and SoftBank’s Collaboration

We are now back on schedule, particularly given the recent announcements originating from OpenAI over the past weekend.

OpenAI’s CEO, Sam Altman, recently visited Tokyo for a public discussion with Masayoshi Son, the CEO of SoftBank, a prominent Japanese conglomerate. SoftBank is a substantial investor in OpenAI and is committed to providing funding for OpenAI’s extensive data center infrastructure project within the United States.

Altman likely considered it important to dedicate time to Son given their partnership.

The conversation between these two billionaires centered on the potential of AI “agents” to handle complex tasks. Son revealed that SoftBank intends to invest $3 billion annually in OpenAI products and collaborate on the development of “Cristal Intelligence,” a platform designed to automate a vast number of conventional office-based jobs.

According to a press release issued by SoftBank on Monday, “By automating and autonomizing all of its tasks and workflows, SoftBank Corp. will transform its business and services, and create new value.”

The Impact on the Workforce

A critical question arises: what are the implications for the average worker given this increasing automation and autonomization?

Similar to Sebastian Siemiatkowski, CEO of Klarna, who frequently highlights AI’s capacity to replace human employees, Son appears to believe that AI-powered agents will inevitably lead to significant wealth creation. However, the potential costs associated with this abundance are often overlooked.

Should widespread job automation become a reality, large-scale unemployment is a highly probable consequence.

A Call for Responsible AI Development

It is concerning that leaders in the AI field – including companies like OpenAI and investors such as SoftBank – consistently present a vision of automated corporations with reduced workforces during public appearances. While these are businesses with financial objectives, and AI development is inherently expensive, fostering trust in AI may require demonstrating greater consideration for the well-being of individuals.

This is certainly a point worthy of further consideration.

Recent Developments in Artificial Intelligence

In-depth investigations are now more accessible thanks to a new AI “agent” released by OpenAI. This tool is specifically engineered to assist users in performing complex and detailed research utilizing the ChatGPT platform.

The platform, known for its AI-powered chatbot capabilities, is being expanded with features designed for advanced analytical tasks.

New AI Models and Capabilities

Alongside the new research agent, OpenAI has also introduced o3-mini, a novel AI “reasoning” model. A preview of this model was initially provided in December.

While not the most potent model currently available from OpenAI, o3-mini distinguishes itself through enhanced efficiency and quicker response times.

Regulatory Actions in the EU

Significant changes have taken effect within the European Union regarding AI regulation. As of Sunday, EU authorities are empowered to prohibit the deployment of AI systems identified as presenting “unacceptable risk” or potential for harm.

This includes applications of AI in areas such as social scoring and the implementation of subliminal advertising techniques.

Cultural Reflections on AI

A new theatrical production is exploring the anxieties surrounding AI, often referred to as “doomer” culture. The play draws inspiration from the events surrounding Sam Altman’s temporary removal as CEO of OpenAI in November 2023.

Insights from Dominic and Rebecca, following the play’s premiere, offer a unique perspective on this emerging cultural conversation.

Advancements in Agricultural Technology

Google’s X, the renowned “moonshot factory,” recently announced a new venture. Heritable Agriculture, a startup leveraging data and machine learning, is focused on optimizing crop cultivation practices.

The company aims to significantly improve agricultural yields through innovative technological solutions.

• Key Focus: Data-driven insights for better farming.
• Innovation: Utilizing machine learning to enhance crop growth.

Weekly Research Highlight

Reasoning-based AI systems demonstrate superior problem-solving capabilities compared to standard AI models, especially when tackling challenges in scientific and mathematical domains. However, they are not without limitations.

Recent research conducted by a team at Tencent, a Chinese technology company, delves into the phenomenon of “underthinking” within these reasoning models. This involves the premature and often unexplained cessation of potentially fruitful thought processes.

The study’s findings indicate that “underthinking” is more prevalent when models encounter complex problems. This leads to frequent shifts between different reasoning pathways without ultimately reaching a solution.

To address this, the researchers have put forward a solution. It involves implementing a “thought-switching penalty” designed to incentivize models to fully explore each line of reasoning before evaluating alternative approaches.

This penalty mechanism demonstrably improves the accuracy of the models by encouraging more comprehensive analysis.

Understanding the "Underthinking" Problem

The core issue identified is that models often don't dedicate sufficient cognitive effort to a single reasoning path. They jump between ideas too quickly, hindering their ability to arrive at correct answers.

This is particularly noticeable in difficult problems where a more sustained and in-depth exploration is required. The models essentially give up on promising avenues too soon.

The Proposed Solution: A Thought-Switching Penalty

The Tencent team’s approach focuses on discouraging this premature switching of thought processes. A penalty is applied when a model changes reasoning paths before fully developing the current one.

This encourages the model to thoroughly investigate each potential solution before considering alternatives, leading to more robust and accurate results.

• The penalty incentivizes deeper exploration of reasoning chains.
• It reduces the frequency of unproductive thought switching.
• Ultimately, it enhances the model’s problem-solving performance.

The research highlights the importance of not just building more powerful AI models, but also refining their reasoning processes to avoid pitfalls like “underthinking”.

New Music Generation Model Released

Researchers, with support from organizations including ByteDance – the parent company of TikTok – and the Chinese AI firm Moonshot, have unveiled a novel open-source model designed for music creation.

Introducing YuE

This new model, designated YuE, demonstrates the capability to produce musical pieces of considerable quality based on textual prompts.

The generated songs can extend to several minutes in duration, incorporating both vocal elements and instrumental accompaniment.

Licensing and Commercial Use

YuE is distributed under the permissive Apache 2.0 license. This allows for unrestricted commercial application of the model.

Performance and Hardware Requirements

However, utilizing YuE necessitates substantial computational resources. Generating a 30-second track requires approximately six minutes of processing time even with a high-end Nvidia RTX 4090 GPU.

Potential Copyright Concerns

A key uncertainty surrounds the data used for the model’s training. The developers have not disclosed whether copyrighted material was included in the training dataset.

The inclusion of copyrighted songs could potentially lead to intellectual property disputes for users of the model in the future.

Important considerations regarding copyright should be addressed before widespread commercial deployment.

AI Safety: New Defense Against Jailbreaks

Anthropic, a leading AI research laboratory, has announced the development of a novel approach to enhance the robustness of AI systems against "jailbreaks." These jailbreaks represent methods employed to circumvent the built-in safety protocols of artificial intelligence.

Constitutional Classifiers Explained

The newly developed technique, termed Constitutional Classifiers, utilizes a dual-classifier system comprised of two distinct AI models. These are an input classifier and an output classifier, working in tandem to identify and mitigate potential risks.

The input classifier functions by augmenting user prompts with predefined templates. These templates specifically describe common jailbreak attempts and other types of prohibited content. This proactive step aims to alert the safeguarded AI model to potential threats.

Conversely, the output classifier assesses the probability that the AI model’s response contains harmful or undesirable information. It analyzes the generated text to detect any breaches of safety guidelines.

Performance and Trade-offs

According to Anthropic, Constitutional Classifiers are capable of effectively filtering a significant portion of jailbreak attempts – described as the “overwhelming majority.”

However, this enhanced security does come with certain performance considerations. The implementation of these classifiers increases the computational demands of each query by 25%.

Furthermore, the use of Constitutional Classifiers results in a slight decrease in the model’s responsiveness to harmless inquiries. Specifically, the model is 0.38% less likely to provide answers to non-problematic questions.

Despite these trade-offs, Anthropic believes that the improved safety and reliability offered by Constitutional Classifiers represent a substantial advancement in the field of AI safety.