Desalination system could produce freshwater that is cheaper than tap water (4 minute read)
A team of engineers from MIT and China have designed a solar desalination system that passively turns seawater into drinking water. It circulates water in swirling eddies, which when combined with the sun's heat drives water to evaporate and leave salt behind. The resulting water vapor can be condensed and collected while the leftover salt circulates out of the device rather than accumulating and clogging the system. A small suitcase-sized system could produce up to 6 liters of drinking water per hour and last several years.
Alameda had a $65B line of credit and ‘unlimited withdrawals’ (4 minute read)
FTX co-founder and CTO Gary Wang took the stand recently, admitting to wire fraud, securities fraud, and commodities fraud and naming Sam Bankman-Fried, Nishad Singh, and Caroline Ellison as the individuals he committed the crimes with. Wang had pleaded guilty in December 2022 as part of a cooperation deal. This article provides a detailed look at Wang's testimony. It covers Alameda's $65 billion line of credit, how Bankman-Fried controlled the coding process, Alameda's naming process, and more.
The top 7 software engineering workflow tips I wish I knew earlier (9 minute read)
An hour of efficiency gained per day can save one month per year. This article provides tips on how to optimize software engineering workflows. The tips include optimizing everyday tasks every day multiple times a day, making aliases and autocompletions for commands, and using tools to help remember things.
Unbundling AI (15 minute read)
LLMs allow users to ask anything and almost get an answer to everything. However, the answers it returns could be wrong, and even if the answers are correct, they might not be the right way to solve the problem. They are probabilistic models that provide the most probable answers. Large language models are currently a general-purpose technology that can do some theoretically magic stuff, but for many people, they still look a bit like the PC ads of the late 1970s that promised computers could be used for anything.
GitHub Copilot Loses an Average of $20 Per User Per Month (2 minute read)
GitHub Copilot costs users $10 per month, but it is costing Microsoft an average of $20 per user per month. Some users are costing the company as much as $80 per month. This is likely why Microsoft decided to charge a lot more for the AI capabilities in Microsoft 365 Copilot. The costs are also likely why the company is developing its own in-house AI chip sets for its data centers and pushing the adoption of Neural Processing Units across the industry.
Forty years of programming (6 minute read)
This article presents several tips to improve the ergonomics of a programming setup. It makes several product recommendations that can improve the developer experience, for example, Apple's Magic Trackpad for improved comfort and navigation and the Ergodox EZ for a keyboard where users' wrists never have to move. The article also recommends stretching during breaks and managing stress levels by disconnecting from work during non-work hours.
‘Ukuhumusha’—A New Way to Hack OpenAI's ChatGPT (3 minute read)
Various restrictions in ChatGPT can be bypassed by using less common languages like Zulu and Gaelic.
Google will now make passkeys the default for personal accounts (2 minute read)
Google users will start being prompted to create and use passkeys instead of passwords starting today. Passwords will still be used where passkeys are not yet supported. Passkeys ensure that users have separate keys for each account. They are resistant to online attacks like phishing. Users can choose to forgo passkeys by unchecking the 'skip password when possible' option in their account settings.
How Disney Packed Big Emotion Into a Little Robot (5 minute read)
A Disney Research team recently presented a brand new robotic character during a keynote address. The robot packs an enormous amount of expression into its child-size body, emoting in a way that makes it seem uniquely alive. It uses a system that leverages reinforcement learning to convert an animator's vision into expressive motions. This article discusses how it was developed and the technology behind it - a video of the robot is available.
The fraud was in the code (8 minute read)
Gary Wang recently walked jury members through snippets of FTX's code, revealing how Alameda was given an unlimited line of credit, how FTX's insurance fund balance was falsified, and more.
Google's AI stoplight program is now calming traffic in a dozen cities worldwide (2 minute read)
Google's Project Green Light uses machine learning systems to look through Maps data and calculate the most optimal traffic timing at intersections to reduce idle times as well as the amount of braking and accelerating vehicles have to do.
OpenAI plans major updates to lure developers with lower costs (5 minute read)
OpenAI plans to introduce major updates to its API next month, including memory storage and vision capabilities. The new features are expected to be rolled out at the company's first-ever developer conference on November 6. They are designed to attract more developers to pay for access to OpenAI's models. OpenAI is aiming to hit $1 billion in revenue by 2024.
AI Engineer Summit report (6 minute read)
AI is easy - turning a demo into a product is the real challenge. Working with AI is difficult due to having to deal with non-deterministic outputs. Developers get a different result every time they run their code. Evaluating code in AI apps is hard. These issues were key themes throughout the inaugural AI Engineer Summit held this week. This article contains a summary of key takeaways from the summit.