# Jun 25, 2026

## OpenAI unveils its first custom chip, built by Broadcom
- OpenAI has introduced **Jalapeño**, its first custom-built inference processor, developed with Broadcom to enhance performance for AI inference tasks, leveraging insights from its own AI models.

## A Herculaneum scroll has been read for the first time
- **PHerc. 1667**, a carbonized scroll from Herculaneum, has been virtually unwrapped and read in full for the first time, revealing a **Stoic philosophical treatise** on ethics dating back to the 2nd century BC.

## IBM debuts sub-1 nanometer chip technology
- **IBM's sub-1 nm chip technology**, featuring a **revolutionary nanostack architecture**, packs nearly **100 billion transistors** into a chip the size of a fingernail, marking a significant leap in semiconductor capabilities.

## 45°C cooling design cuts data center water use to near zero
- **NVIDIA's AI servers can operate with coolant at 45°C, a significant advancement that enhances energy efficiency in data centers by reducing cooling energy consumption.**

## GLM-5.2 is a step change for open agents
- **GLM-5.2 represents a significant leap in open agent capabilities**, showcasing how minor version updates can lead to substantial improvements in user experience and application potential.

## Krea 2: SOTA open-weights 12B image model
- **Krea 2** introduces a series of foundation models that prioritize **aesthetic diversity** and **user creative control**, utilizing a multi-stage training pipeline and a robust data curation strategy to enhance image generation capabilities.

## Computer use in Gemini 3.5 Flash
- **Computer use is now integrated into Gemini 3.5 Flash**, enhancing agentic tasks across platforms, previously limited to the standalone Gemini 2.5 model.

## The Unbearable Cheapness of Open Weight Models
- **Open weight models like DeepSeek V4 offer significant cost advantages**, with pricing nearly **50x lower** than competitors like Anthropic and OpenAI, raising questions about the sustainability of high-cost models in the market.

## Ask HN: Where is our profession (programmer) going?
- **Software development is shifting** from a precise discipline to a **probabilistic** one, with reliance on AI tools like Claude for coding and problem-solving, diminishing the role of traditional coding skills.

## Political bias in AI: Where the AI models stand
- **AI models exhibit political bias**, with most leaning left of center, as revealed by a comprehensive analysis of their responses to politically charged questions, conducted without web search interference.

## Puzzling Success of Overparameterization: Lottery Tickets or Escape Dimensions?
- **Overparameterization enhances neural network success** by expanding the optimization landscape, allowing easier navigation away from poor local minima, rather than merely relying on isolated subnetworks akin to lottery tickets.

## DeepSWE: new benchmark looking at how well today's frontier models can actually write code 
- **DeepSWE** introduces a **contamination-free** benchmark, ensuring that tasks are original and not influenced by prior model training, which enhances the validity of performance assessments.

## Un-0: Generating Images with Coupled Oscillators
- **Un-0** is an innovative image generator utilizing **coupled oscillators**, achieving an impressive **FID of 6.74** on ImageNet 64×64, comparable to early conventional models, while aiming for **1,000x energy efficiency** improvements in AI computing.

## Find the best open-source OCR models in one place at Papers with Code 
- **New OCR models** from Baidu and Mistral enhance capabilities, with Baidu's **Unlimited OCR** introducing **Reference Sliding Window Attention** and Mistral's **OCR 4** available via API, showcasing the latest advancements in the field.

## What I'm Finding About LLM Code Style and Token Costs
- **LLM token costs can be drastically reduced** by leveraging native Web APIs, which can cut output token usage by **85-92%** compared to traditional coding patterns, thus optimizing both performance and cost.