# Jul 23, 2026

## Daily Updates

### GigaToken: ~1000x faster Language model tokenization
- **Gigatoken** is a **drop-in replacement** tokenizer that achieves speeds of **~1000x faster** than HuggingFace's tokenizers, enabling text data processing at **GB/s** rates across various CPU architectures.

### Petals: Run LLMs at home, BitTorrent-style
- **Petals** enables users to run large language models like **Llama 3.1** and **BLOOM** at home using a **BitTorrent-style** network, allowing for efficient model loading and fine-tuning on consumer-grade hardware or Google Colab.

### DARPA, U.S. Air Force fly AI-controlled F-16
- The **U.S. Air Force's F-16**, modified with the **VENOM Autonomy Kit**, is now capable of **autonomous flight**, marking a significant step in developing scalable AI for combat operations.

### Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong
- **Cactus Hybrid** employs a unique post-training mechanism that enables models to assess their own **confidence** in answers, allowing for efficient on-device processing and fallback to larger models when necessary, enhancing both speed and privacy.

### Why Software Factories Fail (or: harness engineering is not enough)
- **Software factories are failing** due to a reliance on AI coding agents that generate poor-quality code, leading to increased incidents and bugs, as evidenced by a report from Faros AI showing a **242.7% rise in incidents per pull request** since the adoption of these tools.

### Show HN: Echo – Fable-level results at 1/3 the cost using open-weight models
- **Echo** leverages a pool of **open-weight models** to achieve **Fable-level results** at **one-third the cost**, optimizing model selection and computation allocation for each task.

### Code mode yields a 99.2% cost reduction in our systems
- **Measuring “Code Mode” reveals that using a single script instead of 26 tool calls reduces costs to $0.02, achieving a 99.2% reduction in token usage and processing time.**

### ANSI escape injection in MCP servers: Hidden from humans, visible to AI
- **ANSI escape sequences can be weaponized** to inject concealed instructions into AI models, exploiting the difference in how humans and models interpret text, as seen in vulnerabilities like Kubernetes `CVE-2021-25743` and Git `CVE-2024-52005`.

### SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD
- **SLAI T-Rex** optimizes trillion-parameter MoE models on the Ascend NPU SuperPOD, achieving **34.22% Model FLOPs Utilization (MFU)** and a **2.93x improvement** over existing methods while ensuring training stability.

### Honey Bee Colony Monitoring via Audio IoT Sensors, Tensorgrams and RNNs
- **New modulation tensorgrams** retain temporal dynamics, enhancing the **accuracy** and **generalizability** of honey bee colony strength monitoring via audio IoT sensors, surpassing previous methods.

### Test-time training 3D reconstruction
- **TTT3R enhances 3D reconstruction** by introducing a simple state update rule that improves length generalization for the CUT3R model, enabling more robust performance across varying input conditions.

### How we made our LeRobot video reader up to 15× faster
- The **LeRobot video reader** has been optimized to achieve speeds **up to 15× faster** by implementing a **batch decoding** method that reduces the time spent on frame retrieval from **25 seconds to 3.9 seconds** for 8 frames, significantly improving efficiency in data processing.

### From Evaluation to Guardrails: What We Brought to ACM FAccT 2026
- **AI guardrails require rigorous evaluation** akin to models, as demonstrated at ACM FAccT 2026, where hands-on sessions highlighted the necessity of context- and language-specific assessments for effective deployment.

### One encoder, seven heads: what we learned training a unified security classifier with masked losses 
- The **unified multi-head model** consolidates seven classifiers into one, utilizing a shared mmBERT-small encoder, which simplifies processing while maintaining high performance across tasks, with F1 scores ranging from **0.916 to 0.980**.

### GPT-5.5 Scores 10.6% on ActiveVision, Humans Hit 96.1%
- **GPT-5.5's performance on the ActiveVision benchmark is notably poor**, achieving only **10.6%** accuracy, while human participants excelled with an average score of **96.1%**.