Week 25, 2024 - Fixes, Improvements and AI Agents

Summer is on, at full speed. Temperatures are getting dangerously close to 40ºC on some days. School is over too. All these will make me slow down a bit in the next few weeks. I’ll try to keep the Friday updates on, but realistically there’s going to be less content around here.

📋 What I learned this week

I finally managed to publish the article about Interviewing advice for Engineering Managers looking for a new gig. This time the early feedback made me change significant parts, especially around compensation discussions. I was originally suggesting negotiation tactics and delaying early disclosure of preferences, but learned that this is not just bad advice in this climate, but can even stop a process entirely in some cases. The power of the 20% rule!

The adhesive I ordered for my Apple Watch screen turned out to be not self-adhesive. I’m not sure if it was ever supposed to be, but it didn’t change much. As an act of desperation, I superglued the screen to the body of the watch, and against all odds, it seemed to have worked. It’s really hard to tinker with such small devices!

Jeremy and I kicked off a new YouTube channel where we publish shorter, more focused clips from our podcast. The Retrospective Clips is a curated platform to maximize the value of our content: every video is about one single topic, making it more accessible and easier to find relevant themes. The first one is about the power of pair programming. This new channel is in response to feedback we’ve received about our podcasts being a bit chaotic/unedited. I like that aspect of our show, but this feedback brought up an important area we missed so far: the lack of organized, easy-to-find content. I’m grateful for the opportunity to learn what works well in this medium and am thankful for all the feedback we receive!

Finally, yesterday I had some time to play with the front page of my blog, I limited the number of newsletter issues and tags in the right column and created a dedicated page for all the topics on the site. Should look a bit better now.

🎯 What I want to try next week

I’m about to reach a big milestone in an important hobby next week. To celebrate it, I’m considering creating a simple one-page website explaining the game to others. Going into details here would defeat the purpose of that site though, so I hope you don’t mind that I don’t explain it further yet.

I’ve been running LinkedIn experiments with scheduled posts since the beginning of June. Now that enough time has passed, I’m going to look at the data to see if there are interesting learnings to share. Setting up posts to go live regularly to boost engagement is nothing new, but I wanted to see how exactly it works, and what aspects can contribute to a bigger reach. Stay tuned on this one too.

🤔 Articles that made me think

Cisco replaces Microsoft as world’s most valuable company

Oh yeah, sorry, this is from March 2000, a few days before the dot-com bubble burst, not this Tuesday. However, the similarities are uncanny. I know I made this parallel before, but I can’t help being amazed about having a chance to watch real-time as history is repeating itself.

OK, so Péter, are you shorting NVDA?

No, I’m not. Being right too soon is being wrong. It’s easy to call a bubble, it’s hard to time when it will burst. (This was one of the takeaways from the dot-com bubble by Corry Wang, go check out his thread if you missed it, there is some good misconception-busting in there.)

Why is there no Scenic Route Option in Google Maps?

Great Twitter thread from someone who used to work on the product itself as a Senior UX Researcher. Google did consider this (to be precise: for pedestrian navigation only), and while the arguments Kasey is listing were not the reasons this idea was abandoned, his points about reinforcing feedback loops and the bigger social context of these tools are super interesting. Food for thought.

🤖 Something cool: Agent Mode in Warp AI

Agents are at the forefront of AI development, so this is a natural next step for Warp, an AI-powered terminal (that I use and love). They published a cool demo video, but I wanted to test this on my own. Longtime readers might remember a post from me a while ago where I shared my experience with Open Interpreter, which gave me a good opportunity to try the same task as then:

create a bootable usb pendrive with a proxmox installer on it

Here’s the slightly edited video of the experience in Warp AI (I cut out the slow downloading and copying parts):

Notes:

• While the process was slower, I felt more in control than with Open Interpreter. The balance of confidence is somehow better in Warp. The UX helped too: Warp is cleverly hiding output details depending on the command, making it easier to follow what’s going on.
• It was hard to convince Warp to find the correct download link autonomously, though it worked very smoothly once I insisted on doing it from the CLI. (Although the .iso file in ~/Downloads was never cleaned up.)
• Warp doesn’t allow you to bring your own OpenAI key, despite user requests. It makes sense from a business perspective, but the above demo and a quick experiment before used up my monthly 40 requests in the free trier - and I don’t think I’d pay $18/month for this however smooth it feels (especially not since that would only give me 500 requests, or something like 17/day). I might be OK paying a smaller fee to be able to use my own key though, but I’m not sure that would be viable for them.

The price question is critical. With all these shiny new toys, we often put aside the cost of energy it requires to run these LLMs. AI companies are burning investor money to provide services cheaper than their cost, sometimes for free, to get a bigger market share. OpenAI, despite making $2 billion in revenue last year, is still losing money. Unless there’s a revolutionary change in energy requirements for training and running LLMs, we better get used to having to pay for these services.

That’s it for today, have a cool weekend,

Péter