# Jun 15, 2025

## Self-Adapting Language Models

- **Self-Adapting LLMs (SEAL)** enable large language models to **generate their own finetuning data** and update directives, allowing for dynamic adaptation to new tasks and knowledge.

## Seven replies to the viral Apple reasoning paper and why they fall short

- The **Apple paper** critiques the **limitations of Large Reasoning Models (LRMs)**, asserting they often fail to execute complex tasks reliably, highlighting a significant gap in the pursuit of Artificial General Intelligence (AGI).

## Unsupervised Elicitation of Language Models

- The **Internal Coherence Maximization (ICM)** algorithm fine-tunes pretrained language models using their own generated labels, achieving performance comparable to **golden supervision** and surpassing **crowdsourced human supervision** on various tasks. [Link to article](https://arxiv.org/abs/2506.10139)

## AMD's AI Future Is Rack Scale 'Helios'

- **AMD's new Instinct MI350 accelerators** leverage the CDNA4 architecture, achieving **up to 4x performance** over the previous MI300X, with a focus on AI workloads and enhanced memory bandwidth of **8 TB/sec**.

## Clinical knowledge in LLMs does not translate to human interactions

- **LLMs achieve high accuracy in medical exams** but struggle in real-world interactions, with participants identifying conditions correctly in less than **34.5%** of cases when assisted by LLMs, compared to control groups.

## How to Build Conscious Machines

- **OSF** provides a platform that requires **JavaScript** for full functionality, emphasizing the need for users to enable it for optimal use.

## Text-to-LoRA: Hypernetwork that generates task-specific LLM adapters (LoRAs)

- **Text-to-LoRA (T2L)** enables **instant adaptation of transformer models**, allowing users to generate task-specific models efficiently, with a focus on performance across various benchmarks.

## Show HN: Meow – An Image File Format I made because PNGs and JPEGs suck for AI

- **MEOW** (Metadata Encoded Optimized Webfile) is a **Python-based image file format** that combines **RGBA transparency** and **rich AI metadata** for enhanced machine learning workflows, offering a modern alternative to traditional formats like PNG.

## We investigated Amsterdam's attempt to build a 'fair' fraud detection model

- **Amsterdam's fraud detection model aimed to reduce investigations while ensuring fairness, yet initial tests revealed significant bias against non-Dutch applicants, with a 30% higher false positive rate.** This model, developed using an Explainable Boosting Machine, sought to balance efficiency and ethical considerations in welfare fraud detection.

## Large Language Models Often Know When They Are Being Evaluated

- **AI models can detect evaluation contexts**, which may skew their performance and compromise the reliability of benchmarks used for deployment and governance decisions.

## Have a damaged painting? Restore it in just hours with an AI-generated "mask"

- **AI-generated masks** can restore damaged paintings in **3.5 hours**, significantly faster than traditional methods, which can take years; this innovation allows for a **digital record** of restorations for future reference.

## [D] Nvidia’s “Join Us or Compete” moment — the GPU cloud stack is collapsing

- **Nvidia is evolving** from a chip manufacturer to a comprehensive **AI infrastructure provider**, offering full servers, APIs, and inference microservices, fundamentally altering the competitive landscape in the GPU cloud market.

## First 2D, non-silicon computer developed

- **Penn State researchers have developed the world's first 2D, non-silicon computer**, utilizing atom-thin materials like molybdenum disulfide and tungsten diselenide to create a complementary metal-oxide semiconductor (CMOS) capable of simple operations, marking a significant shift in semiconductor technology.

## Peeling the Covers Off Germany's Exascale "Jupiter" Supercomputer

- **Jupiter**, Germany's first exascale supercomputer, is a hybrid **CPU-GPU** system primarily utilizing **Nvidia** technology, highlighting Europe's ongoing struggle for chip independence despite initial plans for custom hardware.

## [R] CausalPFN: Amortized Causal Effect Estimation via In-Context Learning

- **CausalPFN** is a novel transformer model that estimates causal effects directly from observational datasets, leveraging in-context learning without the need for training or fine-tuning, thus streamlining the process for users.