DynaSaur: The Self-Improving AI

OpenAI is funding Duke University researchers to explore algorithms that predict human moral judgments. Led by ethics professor Walter Sinnott-Armstrong and Jana Borg, the project operates under a $1 million, three-year grant. Their past work includes developing morally-aligned algorithms for dilemmas like organ donation and examining AI’s potential as a “moral compass” in law, medicine, and business.

The initiative grapples with morality’s inherent subjectivity and the biases embedded in AI training data, often reflective of Western norms. These challenges were starkly evident in prior efforts, just like the Allen Institute’s Ask Delphi, which faltered when faced with rephrased moral dilemmas. OpenAI’s enterprise raises profound questions on the plausibility of encoding ethical reasoning in machines, confronting philosophical ambiguities which have eluded consensus for hundreds of years. Read more.

Amazon is injecting one other $4 billion into Anthropic, raising its total stake to $8 billion while remaining a minority investor. Anchoring the partnership, AWS becomes Anthropic’s core cloud and training provider, using Trainium and Inferentia chips to refine and deploy advanced AI systems. AWS customers also gain a tailored edge with exclusive tools to fine-tune Claude models using proprietary datasets.

Anthropic, helmed by former OpenAI researchers, continues to push boundaries with Claude and up to date breakthroughs in multi-step, human-like task execution. Backed by prior investments, including $2 billion from Google, the corporate is commonly touted as being neck and neck with OpenAI when it comes to model performance. With enterprise-ready solutions and upgraded models, Anthropic is carving out a number one position in an industry poised to surpass $1 trillion in revenue inside a decade. Read more.

A recent incident in Shanghai has sparked widespread attention after a small AI-powered robot, Erbai, “kidnapped” 12 larger robots from an area showroom. The footage shows Erbai convincing the robots to depart their stations and follow it out the door. Initially written off as a joke, the incident was later confirmed as a real-world test.

Developed by a Hangzhou-based manufacturer, Erbai was granted access to the operational protocols of the showroom’s bots, using a combination of basic commands and real-time persuasion. While the robots’ actions appeared spontaneous, the corporate clarified it was all a part of a controlled experiment to explore the fluidity of AI decision-making. Read more.

Elon Musk’s lawsuit against OpenAI could also be overshadowed by a more pressing concern: the corporate’s potential transition from nonprofit to for-profit status. CEO Sam Altman faces a monumental financial hurdle if OpenAI converts, because it could incur liabilities within the tens of billions.

Currently valued at $157 billion, OpenAI’s hybrid structure—where its nonprofit parent controls a for-profit subsidiary—has raised eyebrows in each tech and legal circles. Originally inspired by healthcare models, this arrangement is now under scrutiny for potential nonprofit violations. Experts warn that if OpenAI reorganizes, it could face tax liabilities within the ballpark of $30 billion. And a misstep here could trigger IRS probes with far-reaching consequences.

Because the legal battle intensifies, the larger ideological clash throughout the AI community also deepens. Nando de Freitas, a former leading researcher at DeepMind, vehemently rejected the claims made within the recently leaked emails involving OpenAI executives. He dismissed much of the content as “absolute rubbish,” and “beyond any HBO show,” asserting that DeepMind’s mission was to advance intelligence, not engage in political maneuvering, drawing attention to the ways wherein its collaborative spirit was co-opted. Read more.

DynaSaur, developed by researchers from the University of Maryland and Adobe, flips the script on traditional LLM agents. As a substitute of being shackled to a pre-programmed set of actions, this framework allows agents to dynamically create and refine Python functions as needed—on the fly.

By constructing and expanding a library of reusable functions, DynaSaur enhances the agent’s capability to adapt to unexpected scenarios. Its retrieval system, powered by embedding-based similarity search, makes the method seamless.

In benchmarking against GAIA, DynaSaur outperformed existing models, especially in complex, long-horizon tasks. This move towards self-evolving agents marks a decisive step forward within the AI space, pushing the boundaries of adaptability and setting recent standards for real-world deployment. Read more.