Parse notes
Sponsored
Your application runs on real data, and that's what developers should use to build it. With advancements in AI-powered anonymization and synthetic data [4], you can say goodbye to unreliable and unrealistic mock data — forever.
DeepSeek has released its AI image generator as open source, raising both excitement in the developer community and concerns about potential misuse.
Reid Hoffman launched Manas AI, a startup using AI to accelerate cancer drug discovery. The company raised $25 million in seed funding and partnered with Microsoft. Manas AI focuses on breast cancer, prostate cancer, and lymphoma and aims to expand to other diseases.
The next set of vision language models from Qwen. Qwen2.5-VL is a versatile visual AI model that excels in multiple areas: recognizing objects and analyzing visual elements like texts and charts, acting as an interactive visual agent for tool control, processing lengthy videos with event detection capabilities, performing precise object localization through various formats, and generating structured data outputs for business applications like finance and commerce.
Reasoning models can now be trained to iteratively do retrieval. This is similar in spirit to the work shown in the Operator system. Interestingly, it leads to a substantial improvement, although it's not clear what the true FLOP controlled gain looks like.
BrainGuard introduces a collaborative training framework for reconstructing perceived images from multisubject fMRI data while safeguarding privacy.
DeepSeek's image model got a big upgrade today. It is a unified text and image model - sometimes referred to as any-to-any. This means that it can both understand and generate images and text interleaved in a conversation. This is similar to the omni models from OpenAI and the Gemini suite of models from Google.
Pix2Cap-COCO introduces a dataset for panoptic segmentation-captioning, combining pixel-level annotations and detailed object-specific captions to advance fine-grained visual and language understanding.
VideoShield is a watermarking framework specifically designed for popular diffusion-based video generation models. It embeds watermarks directly during video generation, eliminating the need for additional training.
NYU researchers discovered that poisoning just 0.001% of an LLM's training data with misinformation can lead to widespread errors, posing serious risks in medical applications. The study, published in Nature Medicine, found that even corrupted LLMs perform similarly to non-corrupted ones on standard benchmarks, making such vulnerabilities hard to detect.
The U.S. introduced new AI export controls, capping chip exports to most countries while exempting 18 allied nations, to maintain AI leadership and restrict China's access. Major cloud providers like Microsoft, Google, and Amazon can seek global authorizations under these rules. Industry players like Nvidia have criticized the regulations as overreaching.
AI systems like LLMs make fundamentally different mistakes from humans, often erring randomly and with unwarranted confidence. To address this, we need new security systems and methods to manage AI errors beyond existing human-error correction techniques. Focus areas include aligning AI behavior to human-like error patterns and developing unique mistake mitigation strategies tailored for AI.
Sponsored
Jane Street has been all-in on ML since 2020, using neural networks that run at <1ms latency on cutting-edge GPU clusters. Watch their new
Google is consolidating its AI teams, including those working on AI Studio and Gemini APIs, under Google DeepMind to speed up AI development.
OpenAI has appointed Adebayo Ogunlesi, a senior managing director at BlackRock, to its board of directors.
OpenAI's o1 model sometimes processes reasoning tasks in languages like Chinese even when queries are in English.