Reading List 2025

[lunmanlee]

A list of articles I've consumed this year

[lunmanlee]

https://www.newscientist.com/article/2472068-revealed-how-the-uk-tech-secretary-uses-chatgpt-for-policy-advice/ This is so fascinating, UK tech secretary's chatgpt interactions were revealed under their freedom of information act. Curious to know whether this applies to the NZ government as well with the OIA

[lunmanlee]

https://boringtechnology.club/

A great read into why we should choose boring technology, at its core my reflection is that it's about acknowledging the real cost of technical decisions - new flashy tools introduce risk and cognitive load. Choosing boring technology lets you save that energy to where it truly matters.

We also need to recognise that it's not just about picking a tech, we're committing the team's time, mental capacity and future stability.

I esp liked the part where the author talks about his friend choosing to be boring by wearing the same black t-shirt from the same brand everyday to shift that brain usage to other things that matter. I can never be that person but I am a big believer of appreciating boredom, boredom makes space for curiousity and I really enjoy where my curiosity takes me.

[lunmanlee]

https://machinelearning.apple.com/research/illusion-of-thinking

Does visible thinking mean the model is smarter or does it just look smarter?

The authors created set of puzzles where they can control how hard each one was (one of them was the Tower of Hanoi, more disks on increasing difficulties) found that thinking/reasoning models:

• did not so good on easy, straightforward levels as they tend to overcomplicate/overthink (same)
• scored well for medium difficulties, thinking aloud helped the models make better decisions
• performed badly on hard problems, the models either gave up early or fell into senseless guessing

The paper also says that while models pretend to show that they are deeply thinking, they're not really consistent or reliable.

Personally have encountered scenarios where I either made the models perform easy tasks (find and replace) and they overcomplicate them or I ask them to solve a structurally complex problem which makes them implode.

I get the most value when I get the models to do tasks that are often of modest complexity as the paper have mentioned.

[lunmanlee]

https://open.substack.com/pub/robotic/p/how-i-write?r=4af903&utm_medium=ios

People spend too long trying to write as an activity without thinking seriously about why they're writing and what they care about. This makes writing feel like a chore.

this is exactly why i stopped journaling, every night it feels like a chore because it doesn't really matter what/why i write so i stopped caring, it felt purposeless.

Finding your voice also unlocks much more powerful feedback loops and the most powerful form of writing — writing about why you write

AI will draw you into long, lulling, lofty sentences that make you feel like you know what they're talking about while still conveying very little information.

[lunmanlee]

https://www.politico.com/news/2025/05/06/elon-musk-xai-memphis-gas-turbines-air-pollution-permits-00317582

“How come I can’t breathe at home and y’all get to breathe at home?”

musk's xAi company uses a loophole to use turbines with no permits and no pollution controls, the turbines produce nitrogen oxide or NOx at a rate of 1200 to 2000 tons a year!

grok, the company's chatbot is notorious for not having guardrails like chatgpt or claude which allow the chatbot to produce insensitive things like https://nypost.com/2024/08/22/business/elon-musks-grok-ai-chatbot-goes-viral-with-mind-blowing-deepfakes/

sadly the nearby town's residents(predominately Black people with median income of $35000) are paying for those useless outputs with their health

[lunmanlee]

https://machinelearning.apple.com/research/app-store-review

Fascinating to see Apple's implementation of LoRA in prod, the first time I spotted review summarisation was in AllTrails, would be good to know how their implementation differs. Super keen to see the summarisation on Google Maps soon.

[lunmanlee]

https://arstechnica.com/google/2025/06/googles-nightmare-how-a-search-spin-off-could-remake-the-web/

Exercising my cautious optimism here, would be really awesome to see more niche and hyper-local, relevant searches but also we would still be relying on Google's technology..

If not properly licensed we might rely on it even more than we used to, entrenching Google's monopoly further. If done well I'm hopeful for a fairer internet that is inclusive to all.

Imagine having a hyper-local engine that could index things from grassroot sources like community Facebook groups, Whatsapp community updates and shop window signs. This will introduce more granularity and higher frequency of updates rather than the current Google's method of sourcing from big platform and official listings.

Imagine a hyper-local engine that uplifts small businesses by intentionally surfacing them over sponsored and popular listings.

"Google might be a verb, but it doesn’t have to be the only way to search."
Leon Furze, consultant/PhD candidate