News

the-decoder3d
Confident user prompts make LLMs more likely to hallucinate
Many language models are more likely to generate incorrect information when users request concise answers, according to a new benchmark study. The findings suggest a clear pattern: many models are ...
the-decoder3d
Stanford researchers find AI agents improve when guided by past successes
A new study from Stanford University finds that AI agents can get much better at solving complex tasks simply by learning from their own successful experiences. So far, building effective AI agents ...
the-decoder3d
Microsoft finds API agents are faster but GUI agents more flexible
Microsoft researchers have compared API-based and GUI-based AI agents, finding that each approach has distinct strengths and that the two can work well together. API agents interact with software ...
the-decoder3d
US Copyright Office says fair use does not cover AI trained on "vast troves of copyrighted works
The US Copyright Office has pushed back against one of the AI industry's most common legal arguments: that training AI models on copyrighted material generally ...
the-decoder3d
Deepseek-R1 triggers boom in reasoning-enabled language models
A new review shows that while OpenAI was first to push reasoning-enabled language models into the spotlight, Deepseek-R1 has kicked research in this area into a higher gear. Since its release about ...
the-decoder4d
Gemini Flash 2.5 becomes 150 times more expensive for reasoning tasks than Flash 2.0
Reasoning tasks sharply raise AI costs, according to a new analysis by Artificial Analysis. Google's Gemini Flash 2.5 costs 150 times more to run than Flash 2.0, due to using 17 times more tokens and ...
the-decoder4d
Bytedance launches Agent TARS, an open-source AI automation agent
Bytedance has introduced Agent TARS, an experimental open-source automation tool that visually processes web pages and interacts with the command line and file system, currently limited to macOS. The ...
the-decoder4d
Web Dev in Qwen generates full front-end code from just a prompt
Running inside Qwen Chat, it lets users generate front-end code from a single instruction. A prompt like "Create a Twitter-like website" returns working HTML, CSS, and JavaScript—no coding skills or ...
the-decoder4d
Five major obstacles are holding back RAG systems in healthcare
Retrieval-augmented generation (RAG) promises to help medical AI systems deliver up-to-date and reliable answers. But a new review shows that, so far, RAG rarely works as intended in real-world ...
the-decoder4d
SoundCloud could train AI models on user data
SoundCloud changed its terms of use in February 2024 to allow uploaded music to be used for AI training. AI copyright activist Ed Newton-Rex spotted the change and ...
the-decoder5d
OpenAI adds new fine-tuning options for o4-mini and GPT-4.1
OpenAI has been developing reinforcement fine-tuning (RFT) since last December, and it is now available for verified organizations using the o4-mini model. RFT uses a programmable evaluation system, ...
the-decoder5d
Google deploys AI in Chrome to detect and block online scams
Google is now using AI models to protect Chrome users from online scams. On desktop, the company has rolled out its local Gemini Nano language model to quickly spot fraudulent websites, including ones ...