# May 25, 2026

- **AI agents in software development are fundamentally flawed**, as they are sophisticated statistical models that cannot program effectively, leading to an increase in low-quality code output.

- **High-bandwidth memory (HBM) now constitutes 63% of AI chip component costs**, reflecting a significant increase from **52%** in Q1 2024, driven by rising demand and production costs among major chip manufacturers like Nvidia and AMD.

- **LLM agents excel in code generation under loose specifications**, but struggle significantly with strict structural constraints, leading to a phenomenon termed **constraint decay** where performance drops as requirements increase.

- **Greg Brockman** reveals the pivotal **72 hours** following Sam Altman's firing, detailing the rapid formation of a backup company and the strategic decisions that shaped OpenAI's future.

- **PICO** (Perceptual Image Codec) is the first learned codec optimized for the **human visual system**, achieving **2.3-3× bitrate savings** over traditional codecs like AV1 and JPEG-AI while maintaining high encoding speeds on mobile devices.

- **IBM's Anderon foundry** emerges as the first pure-play quantum chip facility in the U.S., backed by a **$2 billion CHIPS Act investment** aimed at establishing a competitive quantum technology landscape while supporting various quantum modalities.

- **AI errno(2) values** introduce a new set of error codes for AI systems, addressing failures like **hallucination** and **excessive API errors**, which are not covered by traditional human errno values.

- Norway’s National Library is leveraging **2 PB of Huawei OceanStor Dorado flash storage** to develop a **sovereign large language model (LLM)** that comprehends the Norwegian language, addressing a gap in local language AI solutions.

- The **Noroboto.ttf** "lexploit" enables the creation of malicious fonts that misrepresent Unicode glyphs, potentially giving adversaries a tactical legal advantage by embedding deceptive font definitions in documents.

- **Non-contrastive SSL methods** like BYOL, JEPA, and data2vec show promise, yet their learning mechanisms and effectiveness remain unclear, raising concerns about the reliability of evaluation methods such as linear probes and KNN during training.

- **Gorilla** is an **in-memory time series database** developed by Facebook, designed to handle over **1 trillion data points per day** with a **12x compression** algorithm for efficient storage and retrieval.

- **CPPL** transforms LLM-assisted hardware generation into a **statically checkable frontend problem**, enhancing reliability and integration with existing compiler infrastructures like **CIRCT**.

- The **METR AI time horizons graph** is fundamentally flawed, as it relies on **guesstimated data** and biased samples, undermining its credibility and the conclusions drawn from it.

- **`thermocompute` is a PyTorch-first emulator** for thermodynamic probabilistic computing, enabling the study of neural layers where width behaves like parallel fabric, allowing for efficient GPU utilization and future hardware modeling.

- **OnePred** introduces a novel approach to **next-query prediction** in multi-turn conversations by utilizing a **recursively updated memory** that tracks user intent, significantly enhancing proactive interaction without the need for extensive dialogue history.