–
- Automating processes with software is hard a. The ability to automate hinges on knowing the steps to take for predefined inputs, not even the steps to take if some inputs are erroneous or incomplete, but also on knowing what to do if you can’t specify the inputs. b. A perfectly automated system will become a logic to handle different circumstances when running a business. c. Tools and connections fail unexpectedly, handling exceptions that drive all the complexity and still need humans when things go wrong — and that’s the way to handle escalation.
- Repetition, or scalability, is the enemy of taste because, in time, it reveals a latent sameness in what we all like to like. For instance, sharing a favourite indie band online might initially feel unique, but as recommendations spread, the band could go viral, losing its niche appeal and becoming mainstream, and this erodes personal connection, making what was once special feel common and overexposed. (The Banality of Online Recommendation Culture)
- Writing is thinking. Just as people once became physically strong naturally through manual labour but now must deliberately choose to exercise, being intellectually sharp may require a conscious choice to write and think deeply rather than developing automatically through daily life.
- Just as ATMs and store checkouts needed their operating systems, and smart home devices led to unique versions of Linux and Android, we now need new operating systems designed specifically for AI devices. Products like Alexa, AI Pin, and smart glasses are early signs of a future where we’ll naturally talk to our devices instead of using screens and buttons. (Humane-CosmOS)
- Intelligent people don’t accept answers they don’t understand. No matter how many other people try to convince them of it or how many different people believe it, if they can’t convince themselves of it, they won’t accept it. (How To Understand Things)
- Abstractions are why LLMs work so well. Training on next-token prediction on the entire web-scale data is a massive multitask prediction problem, forcing the model to develop abstractions that multiple problems share.
- Like a broken cart won’t fix itself through awareness alone, managing worry requires active intervention, not passive observation. Understanding its root causes and taking deliberate steps to address them is essential – though sometimes, it takes a bigger worry to show us how tiny our current concerns are. (Beat worry)
- Some prompt engineering techniques a. Role-playing: This technique involves instructing the AI to act as an expert in a particular field, enabling responses that mirror the expertise or style of a well-known figure. b. Style unbundling: This technique breaks down the critical elements of a style or skill to help the AI perform a task with similar attributes. c. Emotion prompting: This technique aims to emphasise the importance of a task, potentially influencing the AI’s response quality. d. Few-shot learning: This involves providing examples of a task to guide the AI in generating responses within a new context, helping it learn from minimal examples. e. Synthetic bootstrap: Here, the AI generates multiple examples based on given inputs, creating a pool of data that can help refine the completion of a task.
- Vercel uses micro frontends to cut build times and improve developer velocity ()
- Some examples of great URL design ( @joy-ooo-back-on-18-nov @elliot)