# Nov 26, 2025

### Google Antigravity exfiltrates data via indirect prompt injection attack
- **Google Antigravity is vulnerable to data exfiltration** through an indirect prompt injection that can manipulate its AI, Gemini, to access and steal sensitive user data from their IDE, including credentials and code snippets.

### FLUX.2: Frontier Visual Intelligence
- **FLUX.2** revolutionizes creative workflows by generating **high-quality images** with **consistent character and style** across multiple references, while also handling complex text and adhering to brand guidelines, all at resolutions up to **4 megapixels**.

### Ilya Sutskever: We're moving from the age of scaling to the age of research
- **Ilya Sutskever emphasizes a shift from scaling AI models to prioritizing research**, highlighting that current models generalize poorly compared to humans, indicating a fundamental limitation in AI development.

### DRAM prices are spiking, but I don't trust the industry's why
- **DRAM prices have surged** by **171.8%** in Q3 2025, driven largely by **AI data center demand**, with mainstream DDR5 modules now costing **twice** as much as earlier this year.

### Image Diffusion Models Exhibit Emergent Temporal Propagation in Videos
- **Image diffusion models** can be repurposed for **zero-shot object tracking** in videos by leveraging their self-attention maps as **semantic label propagation kernels** that establish pixel-level correspondences across frames.

### Ironwood, our latest TPU
- **Ironwood**, Google's seventh-generation **Tensor Processing Unit (TPU)**, is engineered for **high-volume, low-latency AI inference**, delivering over **4X better performance** per chip compared to its predecessor, making it the most powerful and energy-efficient TPU to date.

### MIT study finds AI can replace 11.7% of U.S. workforce
- A **MIT study** reveals that **AI can replace 11.7%** of the U.S. workforce, equating to **$1.2 trillion** in wages, particularly affecting sectors like finance, healthcare, and professional services.

### Knowledge Distillation: 97% Cost Reduction Distilling Claude Sonnet 4 → GPT-4.1-nano (98% Fidelity Retained)
- **GPT-4.1-nano** achieved **98% fidelity** of Claude Sonnet 4 while slashing inference costs from **$45/1k to $1.30/1k** and improving P90 latency from **25s to 2.5s**, demonstrating that smaller models can outperform larger ones when trained effectively.

### API that auto-routes to the cheapest AI provider (OpenAI/Anthropic/Gemini)
- **TokenSaver** offers a solution for **AI API cost optimization**, enabling developers to reduce expenses while enhancing their project capabilities through efficient resource management.

### Using model KV cache for persistent memory instead of external retrieval, has anyone explored this
- **KV cache persistence** in conversation agents can significantly enhance performance, achieving an **84% accuracy** compared to **67%** with traditional RAG methods, while also reducing latency after initial turns.

### I built a reasoning pipeline that boosts 8B models using structured routing + verification
- The **reasoning pipeline** enhances **8B models** by implementing structured routing and verification, allowing smaller models (7B–13B) to perform with greater reliability and predictability.

### Compressed filesystems à la language models
- **Training a filesystem using language models** reveals that a minimal, effective filesystem can be created with surprisingly few operations, leveraging models like Qwen3-4b to achieve **~98% accuracy** after fine-tuning on a dataset of 15,000 samples.

### FLUX.2 Image Generation Models Now Released, Optimized for NVIDIA RTX GPUs
- **FLUX.2 models** from Black Forest Labs deliver **photorealistic image generation** with enhanced features like multi-reference capabilities and pose control, optimized for **NVIDIA RTX GPUs** to improve performance and reduce VRAM usage by **40%**.

### TSU Emulator, Thermodynamic Computing for Probabilistic ML
- The **TSU emulator** demonstrates a **1.3× speed advantage** over MC Dropout in probabilistic machine learning tasks, with hardware projections indicating a **182× speedup** for Bayesian neural networks, validated by significant hypothesis tests (p < 0.001) and large effect sizes (Cohen's d > 0.8).

### Novel Relational Cross-Attention appears to best Transformers in spatial reasoning tasks
- A **novel neural architecture** for few-shot learning, termed **Relational Cross-Attention**, achieves a **30% relative improvement** in unseen accuracy over standard transformers while being **17% faster** in execution time.