Search the archive

THE ARCHITECTURE OF GRAB'S DATA LAKE

Grab manages its data using a data lake, using different storage formats for high and low throughput data. For high-throughput data, which is frequently updated, it uses Apache Avro with a Merge on Read (MOR) strategy, appending new data to log files for efficient writes and periodically compacting them for manageable reads. For low-throughput data with infrequent updates, it uses Parquet with Copy on Write (CoW), creating new file versions for each write.

PARSING JSON IS A MINEFIELD

Parsing JSON, despite its seemingly simple format, is actually quite challenging because there are so many specifications that can be interpreted in different ways. This has led to inconsistencies in how different JSON parsers handle edge cases, extreme values, and maliciously crafted payloads, resulting in potential bugs, crashes, and denial-of-service vulnerabilities. This article goes into detail about this issue with examples of different types of objects (arrays, objects, numbers, and strings) combined with different types of parsers (C Parsers, Regex, and more).

WHY I RETIRED FROM THE TECH CRUSADES

The creator of Ruby on Rails, David Heinemeier Hansson, realized that the success of the framework was not due to his attempts to convince others, but rather by showcasing its capabilities and letting others decide for themselves. He now believes in respecting the various preferences of programmers and encouraging them to find the language that best suits their individual style and needs.

PLEASE SUPPORT “SKIP TO MAIN CONTENT” ON YOUR DOCS SITE

The "Skip to main content" link in website accessibility for keyboard-based navigation is really useful for a lot of users. Adding it to documentation sites should be standard.

SCIENTISTS SHOULD USE AI AS A TOOL, NOT AN ORACLE

NEW TO THE WEB PLATFORM IN MAY

Several new features were added to web browsers in May 2024, including CSS anchor positioning and stepped value functions in Chrome 125, the Screen Wake Lock API in Firefox 126, and the Compute Pressure API in Chrome 125.

ANNOUNCING BRIOCHE!

Brioche is a new package manager and build tool inspired by Nix, Homebrew, and Cargo. It is a flexible and easy-to-use tool for managing software projects across different platforms and languages.

QSTUDIO 3.0

QStudio 3.0 is an SQL editor that supports DuckDB as a local database. It has a Parquet file viewer and can convert a query result from any database to store it locally.

RADIAL MENUS IN VIDEO GAMES

Radial menus, also known as pie menus, are circular context menus used in video games to provide a smooth way for players to select options through directional input. They originated in console RPGs like "Secret of Mana" to streamline actions during battles, but have since evolved to be used in various ways, such as for quick actions in games or for selecting dialogue options in conversations.

PYTHON'S MANY COMMAND-LINE UTILITIES

Python modules can be run as command-line tools using the "-m" argument, providing functionality like starting web servers, opening URLs, formatting JSON data, and managing Python environments. This article goes through a bunch of examples, such as using libraries to read email, unzip files, and more.

WINDOWS AI FEATURE THAT SCREENSHOTS EVERYTHING LABELED A SECURITY ‘DISASTER'

Microsoft's new AI-powered Recall feature, which screenshots everything on a user's PC, has been criticized by cybersecurity experts due to potential security flaws in its data storage. It could potentially be exploited by malware or attackers.

DO NEW DEVS RELY ON AI TOO MUCH? (REDDIT THREAD)

Developers have been copy-pasting code from various sources even before AI existed, so it's up to new developers who want to be good at their job to use AI productively while still understanding the code being produced.

“JUST” ONE LINE

The phrase "it's just one line of code" is misleading because it often hides the actual cost of the code, which can be massive in terms of complexity and performance impact on the end user.

I'M FORKING LADYBIRD AND STEPPING DOWN AS SERENITYOS BDFL

Andreas Kling, creator of SerenityOS, is forking Ladybird (a web browser project) from SerenityOS and stepping back a bit from SerenityOS to allow both projects to focus on their main interests.

CROSS-BROWSER LOCAL LLM INFERENCE USING WEBASSEMBLY

This article covers a cross-browser local LLM inference engine that uses a new quantization technique and WebAssembly to deliver fast LLM inference.

🆓 Save your seat

NVIDIA ANNOUNCES NEW AI CHIPS MONTHS AFTER LATEST LAUNCH AS MARKET COMPETITION HEATS UP

Nvidia has unveiled a new generation of artificial intelligence chip architecture called Rubin. The company only just announced its upcoming Blackwell model in March - those chips are still in production and expected to ship to customers later in 2024. Nvidia has pledged to release new AI chip models on a one-year rhythm. The less-than-three-month turnaround from Blackwell to Rubin underscores the competitive frenzy in the AI chip market.

SPACEX AIMS TO LAUNCH CELLULAR STARLINK SERVICE THIS FALL

SpaceX has revealed plans to launch its Starlink system for phones this fall in a filing about the FCC's new rules on supplying satellite connectivity to US carriers. The company says that the FCC's current framework for supplying satellite connectivity to phones is too restrictive and has urged the agency to loosen the aggregate limit on radio frequencies for cellular satellites, specifically the “one-size-fits-all aggregate out-of-band power flux-density.” It backs replacing the restriction with more granular band-specific limits. SpaceX's rivals have lobbied the Commission to maintain the restriction to protect against potential radio interference.

WHAT WOULD YOU DO WITH A ROBOTIC THIRD THUMB?

'The Third Thumb' is a 3D-printed robotic wearable designed to enhance users' biological capabilities. Controlled by users' toes, the robotic thumb sits on the side of the palm opposite users' biological thumbs, resembling a sixth elongated finger. It is operated through a pair of sensors under each big toe - pressure applied by the right toe moves the thumb side to side, while pressure from the left toe moves it up and down. The device can be used to complete tasks that would ordinarily be difficult or impossible to perform with just a thumb and fingers. A video demonstrating what The Third Thumb can do is available in the article.

AI IS CRACKING A HARD PROBLEM – GIVING COMPUTERS A SENSE OF SMELL

Smell in humans is experienced through about 400 types of receptors in the nose. Scientists are using machine learning to help computers learn what certain molecules or sets of molecules smell like to humans. Machine learning is key to digitizing smells because it can learn to map the molecular structure of an odor-causing compound to textual odor descriptors. Dataset sizes have grown significantly in the past few years, allowing researchers to finally apply machine learning to machine olfaction. The technology has promising applications such as personalized perfumes and fragrances, better insect repellents, novel chemical sensors, early detection of diseases, and more realistic augmented reality experiences.

time, you can Invest alongside top-tier VCs.

AN INTERVIEW WITH THE MOST PROLIFIC JAILBREAKER OF CHATGPT AND OTHER LEADING LLMS

Pliny the Prompter posted a jailbreak for OpenAI's newest foundation model, GPT-4o, just a few hours after it was released. The jailbreak allowed users to get the model to output explicit copyrighted lyrics, instructions for making banned items, strategic plans for attacks, and medical advice based on X-rays. Pliny has been finding jailbreaks for leading large language models for around 9 months. This article contains an interview with the prompter where they discuss what motivates them, their approach to jailbreaking a new model, legal ramifications from AI jailbreaking, jailbreaking ethics, and more.

WHAT WE'VE LEARNED FROM A YEAR OF BUILDING WITH LLMS

Building with AI is now more accessible than ever, but it can still be difficult to build products and systems that are effective beyond a demo. This guide teaches readers how to build successful products with large language models. It draws from the authors' experiences of building applications with LLMs over the past year. The guide discusses tactics, operations, and strategies for building with LLMs.

THE INTERNET'S FINAL FRONTIER: REMOTE AMAZON TRIBES

SpaceX's Starlink has been making the internet accessible to previously isolated tribes in Brazil since 2022. This article looks at what happened to the Marubo tribe, a community that has preserved its way of life for hundreds of years through isolation, after it gained access to the internet. While everyone in the tribe was happy when the technology arrived as the internet brought clear benefits, such as the ability to chat with faraway loved ones and call for help in emergencies, people in the tribe feel things have gotten worse. The internet has become essential to the tribe, but at a cost.

THE ART OF SCALING TASTE

Startups can survive extremely competitive environments by having good taste. Taste is how a business differentiates itself when attention is scarce and choice is abundant. Knowing what to make is just as important as the ability to make it. MSCHF is a business that has managed to scale its taste without compromising on it. This article looks at how the company releases so many products while staying through to its original vision, how it keeps its team creative and engaged, and lessons from the company that can be applied by anyone.

greatest real-time data. Watch it for free

‘LITTLE TECH' SETS ITS SIGHTS ON WASHINGTON

Garry Tan, the president and CEO of Y Combinator, is trying to build a lobbying operation that fights on behalf of Little Tech due to concerns that federal regulations meant to mitigate the potential harms of AI will instead help tech giants maintain their lead on the technology, boxing out smaller players.

LLMS AREN'T JUST “TRAINED ON THE INTERNET” ANYMORE

Labs with money are acquiring and creating non-public data.

MANIFEST

Manifest is a complete backend that features a database, admin panel, REST API, and JavaScript SDK that can be set up with just one file of simple code.

MOVING BEYOND TYPE SYSTEMS

This post discusses the idea of a static effect system and how it could be beneficial to programming languages moving forward.

CHINA'S CHANG'E 6 SPACECRAFT LANDS ON FAR SIDE OF THE MOON TO COLLECT SAMPLES

The Chang'e 6 spacecraft touched down on the far side of the moon in a crater known as the South Pole-Aitken Basin to collect soil and rock samples.

BUILDING HIGH IMPACT DEVELOPER COMMUNITIES

Strong developer communities are now one of the most sought after components when building out a software company or product due to the flywheel of network effects.

Invest in Jurny today ↗️

OPENAI IS RESTARTING ITS ROBOTICS RESEARCH GROUP

OpenAI is reviving its robotics research group after a three-year hiatus, aiming to develop multimodal robotics models and improve core AI models.

AI OVERVIEWS: ABOUT LAST WEEK

Google introduced AI Overviews to enhance search results, providing users with more accurate and relevant information, especially for complex queries. While some issues arose, such as erroneous results and misinterpreted content, Google has implemented over a dozen technical improvements to address these challenges, including better detection of nonsensical queries and limiting user-generated content in AI Overviews.

NVIDIA SAID TO BE PREPPING AI PC CHIP WITH ARM AND BLACKWELL CORES

Nvidia is reportedly preparing a system-on-chip that pairs Arm's Cortex-X5 core design with GPUs based on Nvidia's Blackwell architecture.

MUP FOR SPARSE MODELS

muP is a powerful tool used by all frontier model labs to transfer hyper parameters tuned on small models to larger, more expensive training runs. This paper explores how to do that for sparse models, thus dramatically improving training performance while reducing compute costs.

EXPERIMENTING WITH LOCAL ALT TEXT GENERATION

Mozilla has trained a model that can run in the browser and automatically generate alt text for web images.

RANDOM COLOR ERASING BOOSTS MODEL ROBUSTNESS IN COMPUTER VISION

Researchers have developed a new learning strategy called Random Color Erasing to tackle color bias in computer vision. This method balances the importance of color and other features by selectively removing color information from training data, leading to models that perform better in complex scenarios like wide-area surveillance and industrial applications.

account

VIDEO GENERATION WITH MORA

Mora is an innovative video generation model that leverages multiple visual AI agents to challenge Sora, the leading model by OpenAI.

BOOSTING LLM'S ABILITY TO FOLLOW COMPLEX INSTRUCTIONS

Conifer improves LLMs' understanding of complex instructions by introducing a specialized dataset and a progressive learning approach.

FABRIC FEEDBACK TO IMPROVE IMAGE GENERATION QUALITY

FABRIC is an inference time, training-free method for improving image quality by using a scoring function or image examples to direct output quality.

KL IS ALL YOU NEED

KL divergence is a fast, cheap, and powerful way to measure a kind of distance between things. It is used copiously in traditional and modern AI. This post explores the powerful concept visually and mathematically.