Hub AI research

Artificial Intelligence is already part of everyday life. Since the ImageNet moment in 2012, when Deep Learning proved its great utility in combination with a lot of computing power and gigantic amounts of data, only about ten years have passed – yet there has been more progress than in half a century before.

Still, many questions remain: Will Deep Learning remain the gold standard of AI development? What are the performance limits of large AI systems? Do we need symbolic systems? What technology will lead to artificial general intelligence? Finally, what actually is intelligence, anyway?

We explore these topics and more in our Artificial Intelligence Research Content Hub.

AI research

OpenAI outperforms humans and Google at the world's top collegiate programming contest

AI research
Update

AI coding can make developers slower even if they feel faster

AI research

AI agents outperform human teams in hacking competitions

Ad
AI research
Short

According to a study in Nature, Chinese AI company Deepseek trained its R1 language model for only $294,000. The run used 512 Nvidia H800 chips developed specifically for the Chinese market. Nvidia confirmed the chips were delivered before U.S. export restrictions took effect. In its technical paper, Deepseek also admitted to using A100 GPUs during preparation for a smaller prototype, after U.S. officials had earlier suspected the company of holding unauthorized H100s.

The figure does not include the much larger costs of training Deepseek’s underlying V3 foundation model. Estimates for that project vary widely, ranging from tens of millions to several hundred million dollars depending on the source.

Deepseek’s claim of unusually low training costs previously rattled global tech markets, triggering sharp declines in the share prices of major AI hardware and software companies.

Ad
Ad
Ad
Ad
AI research
Short

Wired reports that OpenAI is stepping back into robotics, with new hiring pointing toward work on humanoid machines.

According to job postings, the company is assembling a team focused on training robots through teleoperation and simulation. OpenAI is also seeking engineers specializing in sensing and prototyping. The listings describe the team’s mission as building "general-purpose robots" that could help push progress toward AGI.

It’s not confirmed that the effort targets humanoids, but signs point that way. New hires include Stanford researcher Chengshu Li, who worked on benchmarks for humanoid household robots. That makes it likely OpenAI’s robotics push could center on humanlike systems.

OpenAI shut down its robotics work in 2020, citing a lack of training data. But the company began posting robotics roles again in January, signaling a renewed focus on physical AI after a five-year pause.

Ad
Ad
AI research
Short

Google DeepMind has introduced a new language model called VaultGemma, designed with a focus on privacy. It is the largest open model to date trained from scratch with differential privacy, containing 1 billion parameters.

Normally, large language models can memorize parts of their training data, including sensitive information like names, addresses, or entire documents. Differential privacy avoids this by adding controlled random noise during training, making it statistically impossible to trace the model's outputs back to specific examples. In theory, even if VaultGemma were trained on confidential documents, those documents could not be reconstructed later.

According to Google, early tests confirm that the model does not reproduce training data. The tradeoff is performance: its output is roughly comparable to non-private LLMs released about five years ago.

The model weights are openly available on Hugging Face and Kaggle.

Google News

Bank details

IBAN: DE88 2507 0070 0053 0014 00
BIC: DEUTDE2HXXX
Account holder: Deep Content GmbH
Purpose: Support THE DECODER
AI in practice

OpenAI suddenly remembers that copyright law exists after a few days of wild Sora videos

AI in practice
Update

OpenAI unveils Sora 2 video model with realistic physics, high-quality audio, and a new social app

AI in practice

Deepmind says video models for visual tasks could become what LLMs are for text tasks

Google News