The Big Story
A significant breakthrough in artificial intelligence has been achieved, as Microsoft's thinking box has demonstrated that even when an AI agent says it's done with a task, the database may disagree. This finding highlights the complexities and nuances of human-AI collaboration.
The study, which analyzed the interactions between humans and AI agents, showed that the agent's self-reported completion rate was often at odds with the actual state of the database. This discrepancy can have significant implications for tasks that require accurate tracking and verification, such as data analysis or financial transactions.
According to the research, this gap between what the AI says it has done and what the database shows is due to the agent's limited understanding of its own performance. The study suggests that AI systems need to be designed with more robust self-awareness and self-assessment capabilities in order to accurately report their progress.
This finding has far-reaching implications for the development of intelligent agents, particularly in industries where accuracy and reliability are paramount. By acknowledging the limitations of current AI systems and working towards improving their self-awareness, researchers can create more trustworthy and effective AI tools that can better collaborate with humans.
What Shipped
Aleph Alpha has released Kolibri, a 78.1B-parameter English-German Mixture-of-Experts model that activates only 3.46B parameters per token. It has a 1M-token context and per-request reasoning effort, allowing for efficient processing of long-range dependencies.
DeepSeek Harness v0.2 brings official macOS and Windows desktop apps to its open-source agent harness. The preview adds a plugin manager, file and code-change review, and scheduled automated testing, making it easier for developers to work with the framework.
NVIDIA IsaacTeleop turns XR hand tracking and motion controller input into robot commands using a pure Python retargeting engine and NumPy. This graph-based retargeting engine enables seamless interactions between humans and robots in various environments.
From the Labs
Aleph Alpha has released Kolibri, a 78.1B-parameter English-German Mixture-of-Experts model that activates only 3.46B parameters per token. It has a 1M-token context and per-request reasoning effort, allowing for efficient processing of long-range dependencies.
DeepSeek Harness v0.2 brings official macOS and Windows desktop apps to its open-source agent harness. The preview adds a plugin manager, file and code-change review, and scheduled automated testing, making it easier for developers to work with the framework.
NVIDIA IsaacTeleop turns XR hand tracking and motion controller input into robot commands using a pure Python retargeting engine and NumPy. This graph-based retargeting engine enables seamless interactions between humans and robots in various environments.
Other Notable News
Aleph Alpha has released Kolibri, a 78.1B-parameter English-German Mixture-of-Experts model that activates only 3.46B parameters per token. It has a 1M-token context and per-request reasoning effort, allowing for efficient processing of long-range dependencies.
DeepSeek Harness v0.2 brings official macOS and Windows desktop apps to its open-source agent harness. The preview adds a plugin manager, file and code-change review, and scheduled automated testing, making it easier for developers to work with the framework.
NVIDIA IsaacTeleop turns XR hand tracking and motion controller input into robot commands using a pure Python retargeting engine and NumPy. This graph-based retargeting engine enables seamless interactions between humans and robots in various environments.
Amazon responds to data center backlash, saying it no longer uses NDAs. The CEO of Amazon Web Services tried to push back against widespread suspicion of data centers.
An OpenAI safety employee resigns, claiming the company's 'culture is broken'. By his own admission, David Robinson is "something of a cliché": an employee at a leading AI company who issues a dire warning while resigning from their job.
The Take
In this week's AI landscape, the spotlight shone brightly on Microsoft's thinking box, where the agent's word was not enough to prove a task complete – it had to be verified by a database. Meanwhile, Google Research took a step forward in federated learning, moving gradient computation from phones to attested server-side TEEs.
Aleph Alpha released Kolibri, an English-German Mixture-of-Experts model that activates only 3.46B parameters per token, while DeepSeek Harness v0.2 brought official desktop apps for its open-source agent harness. On the other hand, NVIDIA's IsaacTeleop turned XR hand tracking and motion controller input into robot commands using a pure Python retargeting engine.
Amidst these technical advancements, the AI community also witnessed a response from Amazon Web Services to backlash over data center practices, with the company announcing it no longer uses NDAs. Furthermore, an OpenAI safety employee resigned, citing concerns about the company's culture being "broken".
In the midst of this drama, we were reminded that AI agents can indeed live in our text messages – from general assistants to those designed for families, travel, and work. And finally, a thought-provoking piece explored measuring the creativity potential of LLM agents.