Artificial Intelligence: News, Business, Science | THE DECODER

Matthias Bastian

Jul 30, 2024

Perplexity.ai promises select publishers ad revenue, but risks media diversity

The AI search engine Perplexity.ai wants to share advertising revenue with selected publishers. The content of these publishers is to be displayed more prominently in searches. A sustainable model for an "AI Internet" is still lacking - media diversity is in danger.

Matthias Bastian

Jul 30, 2024

U.S. government advised not to restrict open-source AI models for now, NTIA report finds

illustration of the US flag in an abstract data stream

Jonathan Kemper

Jul 30, 2024

AI data isn't destroying AI models after all, researchers say

A recent study published in Nature appears to demonstrate model collapse in several AI architectures, but other researchers say the study makes unrealistic assumptions that are inconsistent with common practice, such as discarding all previous data after each iteration. The critical researchers disagree with the theory that training AI models on synthetic data inevitably leads to "model collapse," where performance gradually declines until the model becomes inefficient. This effect is not a real threat under realistic conditions, they say. With LLaMA 3.1, Meta shows that synthetic data generation processes that iteratively correct errors can improve performance without causing model collapse.

Matthias Bastian

Jul 30, 2024

Meta's new open-source model SAM 2 could be the "GPT-4 moment" for computer vision

Matthias Bastian

Jul 29, 2024

NYT slams OpenAI's request for reporter notes as "unprecedented" and "harassing"

New York Times contradicts hacking allegations by OpenAI

Matthias Bastian

Jul 29, 2024

Agents might be the next frontier of AI and OpenDevin wants to open source them

OpenDevin enables the development of AI agents that interact with other applications via software interfaces. The platform offers a flexible architecture, a sandbox mode and tools for evaluation. Researchers from several institutions have presented OpenDevin, an open source platform for the development of AI agents. The platform enables agents to interact with their environment via software interfaces, similar to human developers. The name is presumably based on the commercial AI programming tool "Devin". OpenDevin consists of three main components: an agent abstraction, an event stream for tracking actions and observations, and a runtime environment for executing agent actions. The platform provides a secure sandbox environment in which agents can execute code and access a bash shell, a Jupyter notebook server and a web browser. This enables agents to handle complex software development tasks and web-based tasks. Image: OpenDevin initiative OpenDevin already contains several implemented agents, including a generalist CodeAct agent and a specialized web browsing agent. In initial benchmarks, the OpenDevin agents demonstrate competitive performance in various task areas. Image: OpenDevin initiative Users can also easily create their own "micro agents" for specific tasks. The platform also supports collaboration between multiple agents. For example, a generalist agent can delegate tasks to specialized agents. Interaction with users takes place via a chat-based user interface that visualizes the agent's current actions and allows real-time feedback. A special feature of OpenDevin is the AgentSkills library. It contains frequently used functions such as file editing or reading images using vision language models. The library can be expanded to include new skills. The developers emphasize that OpenDevin is designed as a community project. The source code is published under the MIT license on Ghitub, which also allows commercial use. For the future, the team is planning improvements such as extended support for multimodality, more powerful agents and a more stable runtime environment. Agents as a new development horizon for AI systems In the long term, OpenDevin should drive the development of safe and useful AI agents. Commercial companies such as OpenAI also believe that AI agents - AI programs that act and interact with each other - will be the next but one step in AI development. Before that come AI models with logic capabilities. In general, the existing agent approaches still have a lot of room for improvement, as the OpenDevin benchmarks also show. Microsoft's AI CEO recently predicted that it will take two more generations of AI models before reliable AI agents are possible.

Matthias Bastian

Jul 29, 2024

Managers bet on AI to boost productivity, but workers feel overwhelmed and ill-equipped

A survey of 2,500 employees in the US, UK, Australia and Canada by the Upwork Research Institute shows that many employees feel overwhelmed by the AI tools that their managers are demanding. 96 percent of managers expect AI tools to increase productivity. However, 47% of employees do not know how to meet these expectations. According to 77%, the tools have even reduced productivity and increased workload. The results indicate that generative AI has not yet delivered on its promise of productivity. There are many reasons for this. Companies should identify specific use cases, optimize AI for them and introduce it step by step instead of simply providing chatbots.

Jonathan Kemper

Jul 28, 2024

Study reveals major weaknesses in AI's ability to understand diagrams and abstract visuals

Colorful AI illustration of a clock that is impossible to read.

Kim M. Scheurenbrand

Jul 28, 2024

OpenAI CEO's basic income recipients spend more on food, rent, and helping others

Sam Altman standing on a stage with a pink LCD screen behind him stating "OpenAI Devday"

Matthias Bastian

Jul 28, 2024

China's AI models soar with help from US open-source technology

Chinese technology companies are developing AI models that can keep pace with Western systems. They are also benefiting from open source developments from the USA.