🎙️ Radio Interview - Prof. Sir Anton Muscatelli
This article summarizes a radio interview with Professor Sir Anton Muscatelli, discussing leadership, research, students, and the future of the University of Glasgow. The interview premiered on Jambo Radio Scotland.
Key Points:
• Discussion of leadership in academia.
• Insights into current research at the University of Glasgow.
• Perspectives on the student experience.
• Exploration of the future of higher education.
🔗 Resources:
• Jambo Radio Scotland ↗ - Scottish radio station
• The Science Show ↗ - Science program on Jambo Radio
• University of Glasgow ↗ - Scottish University
• Interview Recording ↗ - Link to the interview
Image
🚀 Bitcoin App Update - Blockstream
This article announces a major update to the Blockstream app, focusing on improved user experience for buying, securing, and managing Bitcoin and Liquid BTC assets.
Key Points:
• Enhanced user experience for managing Bitcoin assets.
• Simplified process for buying and securing Bitcoin.
• App remains 100% open-source and Bitcoin-only.
🔗 Resources:
• Blockstream ↗ - Bitcoin-focused technology company
• Liquid BTC ↗ - Layer-2 scaling solution for Bitcoin
Image
🤖 Bitcoin Full Node Sculpture - A Cypherpunk Chronometer
This article documents a series of Bitcoin Full Node sculptures, presented as a "GIGA-THREAD" compiling photos, videos, explainers, and podcasts. The post focuses on sculptures numbered 1-4.
Key Points:
• Documentation of a series of Bitcoin Full Node sculptures.
• Collection of visual and audio media related to the sculptures.
🔗 Resources:
• FractalEncrypt ↗ - Artist/Creator of the sculptures
Image
🤖 Large Language Models - Sink Value in Softmax
This article discusses the "sink" value in the softmax function of large language models, visualizing its variation across multiple attention heads. The analysis shows significant variation in the sink value across 64 heads.
Key Points:
• Analysis of the "sink" value in a softmax function within LLMs.
• Visualization of sink value variation across multiple attention heads.
🔗 Resources:
• Image 1 ↗ - Visualization of sink values
Image
🤖 Quantum Natural Language Processing - QNLP Conference
This article announces a Quantum Natural Language Processing (QNLP) conference, highlighting the participation of major tech companies and the presentation of keynotes on QNLP and quantum music.
Key Points:
• QNLP conference featuring major tech companies.
• Keynotes on QNLP and quantum music.
🔗 Resources:
• QNLP Conference Website ↗ - Conference information
• Quantinuum ↗ - Quantum computing company
🤖 Quantum Computing - Q2 FY25 Earnings
This article reports on the Q2 FY25 earnings of an unnamed quantum computing company, highlighting revenue growth, gross profit increase, and the general availability of the Advantage2 quantum computer.
Key Points:
• Strong Q2 financial results with significant revenue and profit growth.
• General availability of the Advantage2 quantum computer.
🤖 Quantum Computing Event - RES HPC and Quantum Spain
This article announces a parallel event at the 19th RES HPC User Conference, showcasing the quantum capabilities of RES and real-world applications. The event takes place in Pamplona, Spain.
Key Points:
• Parallel event at the RES HPC User Conference.
• Presentation of RES's quantum capabilities and use cases.
🔗 Resources:
• RES HPC ↗ - High-performance computing resource
• Quantum Spain ↗ - Organization promoting quantum computing in Spain
Image
🤖 Language Model Research - Paper-Printing Process
This article discusses a research group's "editing-forward" process for language model research, which mimics false sampling processes but lacks a clear description of the subsequent sampling process.
Key Points:
• Analysis of a research group's "editing-forward" process for language models.
• Criticism of the lack of description regarding the post-editing sampling process.
🔗 Resources:
Image
Image
🤖 Large Language Models - Deepseek and Scaling Laws
This article discusses the Deepseek model and its relationship to prior work on Mixture-of-Experts (MoE) scaling laws by Kuba Krj and Aidan Clark. It highlights that sparser models with larger expert choice are generally better.
Key Points:
• Comparison of Deepseek to previous research on MoE scaling laws.
• Highlighting that sparser models with larger expert choice perform better.
🔗 Resources:
• Kuba Krj ↗ - Researcher on MoE scaling laws
• Aidan Clark ↗ - Researcher on MoE scaling laws
Image
Image
🤖 Large Language Models - Sink Value Analysis
This article presents a deep dive into the sink value in GPT-OSS models, comparing findings from 20B and 120B parameter models. The analysis reveals differences in mean sink value and layer-wise alternation patterns.
Key Points:
• Comparison of sink values in 20B and 120B parameter GPT-OSS models.
• Observation of layer-wise alternation patterns in full-attention layers.
🔗 Resources:
Image
Image
Image
Image
⭐️ Support
If you liked reading this report, please star ⭐️ this repository and follow me on Github ↗, 𝕏 (previously known as Twitter) ↗ to help others discover these resources and regular updates.