Articles

Large Language Models (LLMs) are criticized for stagnating AI research

With improvements largely focused on scaling up rather than innovating in design or performance post-GPT4.

Great Tables

Aims to solve the great table dilemma by enabling the creation of sophisticated tables in Python, drawing from a rich history of table design and addressing modern data presentation challenges. Great Tables GitHub

JetMoE-8B

Costing under $0.1 million to train, surpasses Meta AI's LLaMA2-7B in performance, challenging the notion that large language model (LLM) training requires vast financial resources.

Yossi Sariel's online security lapse

The head of Israel's Unit 8200 inadvertently revealed his identity due to a digital trail left by a book he authored under a pseudonym, showcasing a significant online security lapse.

AI's impact on jobs

Microsoft, Google, and other tech giants are collaborating in a study led by Cisco to explore how AI might impact technology jobs, amid increasing concerns and legislative debates over AI's role in the workforce.

Tool Use (function calling)

Claude now supports interaction with external client-side tools, enabling users to extend its capabilities by integrating custom tools for a broader range of tasks, accessible via the Anthropic API.

Fake DMCA threats

Generative AI is being used to create fictitious law firms like "Commonwealth Legal," which issue fake DMCA threats to webmasters for SEO manipulation.

Google Books and AI-generated content

Google Books is increasingly indexing AI-generated books that lack quality and relevance, potentially affecting the integrity of search results and the Google Ngram viewer.

Think-and-Execute

A novel framework introduced enhances language models' algorithmic reasoning by decomposing the process into generating task-level pseudocode and simulating its execution.

Qwen1.5-32B

Emerges as a pivotal model in the Qwen1.5 series, designed to optimize the balance between performance, efficiency, and memory usage, addressing the community's demand for a model with around 30 billion parameters.

Cloudflare's PartyKit acquisition

Enhances its serverless computing offerings by enabling developers to easily create real-time, multi-user applications with stateful capabilities.

W3C's analysis on AI systems

The World Wide Web Consortium (W3C) proposes an analysis on the systemic impact of AI systems, especially those based on Machine Learning models, on the Web, focusing on ethical, societal, and technical impacts. Read more

Fortran on WebAssembly

Dr. George W Stagg demonstrates compiling Fortran code to WebAssembly using a patched LLVM's flang-new compiler, enabling Fortran routines to run in web browsers.

AltStore

A non-jailbreak-required alternative app marketplace, poised to be the first third-party iOS app store in Europe, featuring unique apps like the Delta game emulator.

Representation Finetuning (ReFT)

Methods, particularly Low-rank Linear Subspace ReFT (LoReFT), offer a 10x-50x more parameter-efficient approach to adapting large models by intervening on hidden representations, compared to traditional parameter-efficient fine-tuning (PEFT) methods.