A house stood alone in a post-apocalyptic city, its automated systems still functioning. The house's voices sang and announced the time, but chaos erupted when a fire broke out, destroying the house and its inhabitants.
This repository contains a production configuration for running DeepSeek-V4-Flash-0731 on AMD MI300X with fixes for FP8 format, MoE routing, and other issues. It achieves 6,988-7,019 tok/s on fresh prompts with a 2,048-token scheduler budget.
An Xbox outage blocked users from playing both digital and disc-based games. The author prefers PC gaming because it allows for maintaining access to games despite digital issues.
Using large language models like LLMs requires domain knowledge to get the most out of them. A skilled prompter like Terence Tao can steer the model with expertise in the domain, but this skill is not easily replicable by following tips.
Buckminster Fuller's 42-hour lectures cover his inventions, discoveries, and design approach, spanning topics from architecture to economics. The lectures are now available as a transcript, inviting public participation in editing and refining the content.
Attackers compromised the GitHub account of keyv's maintainer, injecting credential-stealing malware into keyv and other caching utilities with 2 billion monthly installs. The malware harvests secrets from the victim's environment and exfiltrates them to a public GitHub repository.
Soup is a tool for fine-tuning large language models with a simple workflow, using one config and one command. It supports various models and training tasks, including LoRA, quantization, and batching.
Harness engineering is a crucial component of recursive self-improvement (RSI) in AI, enabling models to improve their own performance and adapt to new tasks through workflow design, context management, and optimization. Recent research has focused on developing automated design of agentic systems, meta-harnesses, and evolutionary search methods to improve harness engineering, with ...
Anubis could not load its JavaScript. The server may be overloaded. Please reload the page to try again. You are seeing this because the administrator of this website has set up Anubis to protect the server against the scourge of AI companies aggressively scraping websites. This can and does cause downtime for the websites, which makes their resources inaccessible for everyone. Anubis is a ...
In 1582, under the inter gravissimas papal bull, the Catholic world transitioned from the Julian Calendar to the Gregorian Calendar. This new calendar shortened the length of the year from 365.25 days to 365.2425, a reduction of just 0.002%. One side effect of this is the modern leap year system, in which every fourth and 400th year is a leap year, but no other year divisible by 100 is. For ...
0
2026-08-04 12:34:26 UTC
ycombinator.com
Llama
Roame is a flight and hotel search engine for points and miles that helps travelers find deals in seconds. The company is looking for a senior engineer to lead its engineering team and drive its move to a modern Go service architecture.
You dislike AI-generated images in blogs, especially in indie blogs, and prefer authentic content from real humans. You value the unique perspective of individual bloggers over potentially AI-generated content.
Large language models fail at predictive analytics over tabular data due to their inability to handle high dimensionality. In low dimensions, LLMs behave like local, distance-based methods, but in higher dimensions, their capability dissolves and no classical model can replicate their predictions.
Swiftlet is a Swift + Metal runtime for Qwen3-Next and Qwen3.5/3.6 models, allowing them to run on iPhone with 2.5 GB of RAM. It streams weights from storage and activates only 3B parameters per token.
According to Wordle, I’ve played 1,559 games, I win 99% of the time, my current streak is 107 days and my longest is 278. Four numbers, and they’re mostly correct. What I don’t get is anything behind them. No per-game history. No way to compare a Tuesday to a Sunday. No sense of whether I’ve actually got better at this over four years. Each day I’ve played, I’ve also shared the result to ...
Software engineers used to rarely write personal software, but now agents can easily manage customizing software, making it easier to personalize and maintain. Agents can build and manage custom software with minimal programming, allowing for more efficient and flexible use of software.
homebench is a local-first tool that benchmarks LLMs on your laptop, measuring speed, memory, and quality. It provides a live leaderboard and supports various models and providers.
The writer got stuck in a lift due to a faulty alarm system and was unable to contact the outside world until they used the mosh program to send an email. The fire brigade was called and rescued the writer, who then took the stairs to avoid using the lift again.
Learning Rust - Rust Programming Language Tutorials for Everyone!
15 lines (11 loc) · 829 Bytes ┌────────────────────────────────────┐ └────────────────────────────────────┘ We hope you will like this release as much as we enjoyed working on it, and
The author of the Performance-Aware Programming series argues that following "clean code" guidelines can significantly harm performance, citing a 10x speed difference between a "clean" code implementation and a more efficient one. By violating certain "clean code" rules, such as using polymorphism and avoiding internal knowledge, the author demonstrates a 10x to 15x speed improvement in a ...
Please enable JS and disable any ad blocker
The author, a Computer Science student, applied for a Microsoft internship in 2015 after contributing to open source projects. They received a phone interview invitation, prepared using a book on Big Tech interviewing practices, but displayed honesty when unable to solve a brain teaser, which may have affected their chances.
John MacFarlane created the document converter pandoc in 2006 using the Haskell programming language, which he chose for its strong type system and clean representation of structured documents. Over the next 20 years, pandoc evolved into a powerful tool supporting 51 input and 76 output formats, with over 600 contributors and a reputation for reliability and determinism, despite the potential ...
Workers AI optimizes large models on GPUs by quantizing the KV cache to FP8, compressing model weights to INT4, and protecting the cache with integrity checks. These techniques enable efficient serving of models like Kimi K-series and GLM with no change in accuracy, supporting more customers at lower costs.
Bucket Robotics: Senior Fullstack Engineer, San Francisco, CA (On-site) - Own massive surface area, build customer-facing app, scale internal tooling, and design data pipelines. Product-minded engineer with systems thinking and curiosity.