# Apr 2, 2025

## Daily

### Circuit Tracing: Revealing Computational Graphs in Language Models (Anthropic)
- **Circuit tracing** reveals the inner workings of language models by constructing **attribution graphs** that detail how features interact to produce outputs, utilizing a **cross-layer transcoder** to enhance interpretability and accuracy in understanding model behavior.

### Show HN: Qwen-2.5-32B is now the best open source OCR model
- The **Omni OCR Benchmark** evaluates the **OCR and data extraction** capabilities of various large multimodal models, aiming to provide a comprehensive assessment of **OCR accuracy** across traditional providers and multimodal language models, with all methodologies being **open source**.

### UCSD: Large Language Models Pass the Turing Test
- **GPT-4.5** achieved a **73%** identification rate as human in Turing tests, outperforming human participants, marking a significant milestone in AI evaluation.

### How Google built its Gemini robotics models
- **Gemini Robotics models enable robots to learn complex tasks** such as preparing salads and performing a "slam dunk" on their first attempt, showcasing a significant leap in robotic capabilities.

### AI image recognition detects bubble-like structures in the universe
- **AI image recognition** has successfully identified previously unrecorded **bubble-like structures** in the Milky Way, enhancing our understanding of **star formation** and galaxy evolution.

### [R] Proof or Bluff? Evaluating LLMs on 2025 USA Math Olympiad
- **Current LLMs** struggle with rigorous mathematical reasoning, achieving less than **5%** accuracy on the 2025 USAMO problems, indicating a significant gap in their capabilities compared to human competitors.

### [D][P] Turning Knowledge Graphs into Memory with Ontologies?
- **Integrating ontologies** into the cognee AI memory tool enhances the grounding of **knowledge graphs** by aligning them with external system rules through **RDF + OWL** frameworks.

### Practical Prolog Planner Prompting
- **Large Language Models (LLMs) enhance Prolog's combinatorial search capabilities**, effectively bridging the gap between natural language processing and automated planning, as evidenced by their application in generating Prolog planners for complex logistics problems.

### Speed Demon: NVIDIA Blackwell Takes Pole Position in Latest MLPerf Inference Results
- **NVIDIA Blackwell** achieved record-breaking performance in the latest **MLPerf Inference V5.0**, showcasing the **GB200 NVL72 system** with up to **30x higher throughput** on the Llama 3.1 405B benchmark compared to previous submissions.

### [D] Relevance of Minimum Description Length to understanding how Deep Learning really works
- **Minimum Description Length (MDL)** offers insights into the **underlying mechanisms** of deep learning, addressing phenomena like **overparameterization** and **double descent** that remain poorly understood.

### [Project] AxiomGPT – programming with LLMs by defining Oracles in natural language
- **AxiomGPT** revolutionizes programming by allowing users to define **Oracles** in natural language, enabling a new paradigm of **latent-space programming** that treats language as an invocation rather than mere instruction.

### [R] Neuron-based explanations of neural networks sacrifice completeness and interpretability (TMLR 2025)
- **Neuron-based explanations** of neural networks often lack **completeness** and **interpretability**, as demonstrated by the finding that principal components yield superior insights.

### [D] What are the current challenges in deepfake detection (image)?
- **Current challenges in deepfake detection** include the need for improved **dataset generalization** and the identification of specific problems that remain unsolved, particularly in image detection.

### [D] [P] We created a Transcription API with an open-source, multi-step, multi-modal approach instead of custom models. The result? No.1 in an accuracy benchmark (You can recreate the benchmark).
- The **Salad Transcription API** achieved a **95.1% accuracy** for English and **96.8% for German**, outperforming competitors by utilizing an **open-source, multi-modal approach** rather than custom models.

### Evaluating potential cybersecurity threats of advanced AI
- **AI's role in cybersecurity is evolving**, as advanced models can automate defenses but also pose risks of misuse in cyberattacks, necessitating a comprehensive evaluation framework to address these threats effectively.