# April 25, 2026

## ML Times

### DeepSeek v4

- **DeepSeek-V4 Preview** introduces a **cost-effective 1M context length** with two models: **DeepSeek-V4-Pro** (1.6T total params) and **DeepSeek-V4-Flash** (284B total params), both designed to rival top closed-source models.

### There Will Be a Scientific Theory of Deep Learning

- A **scientific theory of deep learning** is emerging, characterized by properties of the training process, hidden representations, and performance metrics of neural networks, as outlined in five key research areas.

### Replace IBM Quantum back end with /dev/urandom

- The **patch** to `projecteleven.py` replaces the IBM Quantum backend with `/dev/urandom`, demonstrating that **classical noise** can recover private keys at rates indistinguishable from quantum hardware, challenging the validity of the original quantum claim.

### OpenAI releases GPT-5.5 and GPT-5.5 Pro in the API

- **GPT-5.5** has been released, featuring a **1M token context window**, image input, and advanced capabilities for complex tasks, enhancing the Chat Completions and Responses API significantly.

### Open source memory layer so any AI agent can do what Claude.ai and ChatGPT do

- **Stash** transforms AI agents from having _amnesia_ to possessing a _persistent memory_, enabling them to recall past interactions and preferences seamlessly across sessions, thus enhancing user experience and efficiency.

### Lambda Calculus Benchmark for AI

- **LamBench** evaluates AI models on **lambda calculus tasks**, measuring their performance across various algorithms, with the top model achieving **91.7% accuracy** on 120 tasks, showcasing the effectiveness of modern AI in complex problem-solving. For more details, visit the [LamBench article](https://victortaelin.github.io/lambench/) and the [GitHub repository](https://github.com/VictorTaelin/LamBench).

### A 3D Body from Eight Questions – No Photo, No GPU

- **A small MLP model** trained on a questionnaire can accurately reconstruct a 3D body with a height accuracy of **0.3 cm** and mass accuracy of **0.3 kg**, outperforming traditional photo-based methods without requiring images.

### FusionCore: ROS 2 sensor fusion (IMU and GPS and encoders)

- **FusionCore** is a **ROS 2 sensor fusion SDK** that integrates IMU, wheel encoders, and GPS to produce a reliable position estimate, addressing limitations in existing packages like `robot_localization` and `fuse` by offering features such as **adaptive noise covariance** and **native ECEF GPS fusion**.

### [New Optimizer] 🌹 Rose: low VRAM, easy to use, great results, Apache 2.0

- **Rose** is a new **stateless PyTorch optimizer** that significantly reduces memory usage, outperforming even 8-bit AdamW and achieving results comparable to vanilla SGD without momentum.

### We're open-sourcing the first publicly available blood detection model: dataset, weights, and CLI

- **BloodshotNet** is the **first open-source blood detection model**, designed for Trust & Safety and content moderation, featuring a dataset of **23k+ annotated images** and achieving **40+ FPS** on CPU.

### DharmaOCR: Open-Source Specialized SLM (3B) + Cost–Performance Benchmark against LLMs and other open-sourced models

- **DharmaOCR**, an open-source specialized SLM with **3B and 7B parameters**, outperformed major models like GPT-5.4 and Gemini 3.1 Pro, achieving scores of **0.911** and **0.925** respectively.

### How Visual-Language-Action (VLA) Models Work

- **Visual-Language-Action (VLA) models** integrate perception, reasoning, and control, enabling robots to interpret visual and linguistic inputs to perform complex tasks like folding clothes or identifying objects.

### How would you build an automated commentary engine for daily trade attribution at scale?

- **Automated commentary engines** for trade attribution must balance **deterministic precision** with **dynamic natural language generation**, ensuring accurate analysis of thousands of trades while avoiding data hallucination from LLMs.

### Open-source 9-task benchmark for coding-agent retrieval augmentation. Per-task deltas +0.010 to +0.320, all evals reproducible

### IBM Research uses vLLM at the heart of its RITS Platform

- **IBM Research's RITS Platform** leverages **vLLM** to provide centralized access to over **100 models** for more than **1300 active users**, enhancing model inferencing and tuning capabilities across the research community.