What I did in 2025
Writing:
I published 60 Substack posts, coming to 1.15 per week on average. The most popular posts are the ones about AI, but also the one about hacking the Spotify algorithm. The microbiome post also did remarkably well. This actually makes me quite happy. Because I have broad interests, I don’t want to write into a certain topic or formula and, apparently, I don’t have to! Thank you for reading my words about a variety of topics.
I hit 100 subscribers on Substack! Meanwhile, I almost completely stopped posting on LinkedIn. What I wrote last year about writing, that it is fun and rewarding both for its own sake and for the feedback from readers, has proven still true. I feel like any ambition to make money or become known for writing that I may have had at any point has dissipated fully now. This is just for fun, for gaining clarity on my own thinking, and because I like to share my thinking with you all. And last but not least, a bit thank you to those who occasionally reply with thoughts, or tell me they liked something, or egg me on in various other small ways. I appreciate that!
In the last couple of weeks, for the first time since mid 2024 I’ve felt the stream of content drying up. There is less to say about AI. The detour in biology was fun but I have no intention of writing a biology text book. There isn’t really anything I want to say about politics, I haven’t really enjoyed writing about it very much anyway. There are topics that I’m interested in, and things I’d like to share that are more about startups and business, but writing about that has always felt a bit like a chore.
Therefore, I wonder what 2026 will bring in terms of writing. I am pretty sure I will write less. It feels like I have run out of steam a bit, and you know what? I’m okay with that!
Fitness and health:
I once again really failed to get back into strength training, but I also once again succeeded in keeping up my running habit. I ran almost 700km in 2025, not bad. My VO2Max fluctuated between 44 and 47. I’ve realized I just can’t get it to >50 without pulling a significant increase in my exercise level out of somewhere. It’s hard to find the time between bringing kids to school, work, and social life.
At the end of 2025 I’ve started sprinting. For some reason I realized that when we are kids, we would hit our absolute top speed almost every day when playing tag or hide and seek, or football. People who play sports as adults are on to something, I think. So I went outside and did 4x 40 meter sprints. Afterwards I felt amazing, and I had sore muscles all throughout my core and legs. So I want to do that more often and will replace one of my three slow runs with some kind of speed training.
After a completely health problem-free 2024 I had some on/off again issues at the end of the year. I revisited microbiome interventions, and my diet in general. When I was living on Bali I would always kind of make fun of those influencers with all their food intolerances and ‘inflammation’. But maybe they are also on to something. So I’m trying to eat simpler food and cycling out some common triggers (wheat, dairy) to see if I can get to something that feels good.
Magicdoor.ai and coding:
I continued working on Magicdoor, rebuilding the back-end and building up a track record of releasing new AI models on the day they come out. In August/September I spent about 6 weeks marketing it on Reddit, LinkedIn and Twitter, as well as trying out some ads. It grew to 40 paying users ($240 monthly revenue). But I have found it a very difficult product to grow. There is no growth loop. Organic or SEO content does not compound over time as it should. The market is incredibly crowded, and I’m competing with the free versions of ChatGPT and others. Besides, many people get some of their AI for free via work or with their phone subscription.
On the other hand, I still use Magicdoor every day personally and still think it’s a great product. It gets new users every now and then, and some users leave. It’s making me money AND paying for all of my AI tools.
At one point I thought it would be a good idea to launch a number of directory websites. I built a template which I’ve open-sourced, ticking off a bucketlist item to open-source something. I’m happy with the template, it’s fast and looks great. But I ended up being dissuaded from launching any directories. There is still a maintenance cost, and it does not feel like these will monetize (through ads) easily. With how hard SEO is, and how many people are building directories I don’t feel like I have any unfair advantage here.
My personal website was compromised via the React2Shell vulnerability, enabling hackers to use my server as a bitcoin miner for a week. Luckily I’m not paying by the usage, I’m just renting a single computer in a datacenter. It was a bit dumb, because on the day React2Shell was publicized I patched Magicdoor.ai right away, but somehow forgot that my personal site is also running React. When I found out about the hack, I spent time with AI working out the degree of access hackers had, locking them out, and patching things up, which was quite fun. My personal website is a static nextjs site that does not even have a single environment variable, so there is basically nothing on it that is not public.
It is always interesting and slightly jarring to be confronted with crime. With Ox Street we got a full-on DDoS attack once with bots creating tens of thousands of user accounts. I once had my car broken into and the stereo stolen. A good lesson on the importance of security.
One reflection I had is that this should have affected many vibe-coded projects. Things like this will, in the future, take down vibe-coded projects with shitty security.
Work at Carousell:
In 2025 I started looking after Trust & Safety in addition to Customer Experience. It was interesting and fun. I’ve never had as much interaction with the police as in the past year. Got the chance to visit the National Anti-Scam Centre and work together with the Police and Ministry of Home Affairs (MHA). When it comes to living abroad, I always love opportunities to learn more about how the country works, and literally working together with the authorities to combat scams is pretty much the pinnacle of this.
It was also a successful year where me and my team brought phishing scams down to a multi-year low in all countries, and got significantly better at figuring out how to combat eCommerce scams (where the seller doesn’t deliver the product after taking payment).
At Carousell I had been trying to help people use AI pretty much since ChatGPT came out. And at the end of last year I was asked to co-create and lead the AI Chapter for Carousell, which is a centre of excellence and training for AI. This is a cool thing to do and something that I at least partially credit to all the writing about AI I’ve done on Substack.
Other stuff:
Went to Australia twice, and to the Netherlands once. One of the things I wanted to do in 2025 was travel a bit further, and I actually did that.
I did cook more often, as I wanted. And we had more people over for dinner at home.
I was successful in hosting drinks with a group of people that I like but that don’t know eachother at a bar once. It was actually great. But doing that monthly is really a tough habit to start. There’s a lot of inviting to do and keeping track of who might come. It needs to be planned a while in advance, and my planning horizon is usually less than two weeks (which I love). I think it might be too ambitious to really commit to doing this monthly, but to organize 3 or 4 of these small group drinks in 2026 is a nice goal.
I was a lot busier with work last year, which mainly came at the cost of playing guitar. I hardly practiced, and I got worse at it, although that’s 80% just lack of finger strength. It would take only 2-3 weeks of practicing every day to get back I think. With that amount of practice, the other lost neurons would also probably come back to life.
Books I’ve read:
Check this post for a list:
Roundup #30: The kids will be alright, and my book recommendation / reviews
The kids will be alright
What I want to do in 2026
One strength training session per week. Replace one zone 2 run with high intensity speed work like sprinting, jumping and intervals. Keep running two times per week. Go mountainbiking and cycling again at least a few times.
Build and ship one new product. Criteria: (1) fun puzzles while building, (2) solving a real problem for businesses, (3) where I have an unfair advantage in understanding the problem, (4) where I have a solid list of potential buyers that I know well enough to just WhatsApp or email.
Continue to let writing take me wherever it does. No ambitions, no goals. Keep sharing whatever satisfies both of: (1) being on my mind naturally + (2) feeling like other people might find it interesting to read.
Declutter and re-clutter a bit: Sell my old DJ gear, buy something else that’s compatible with vinyl. Declutter and turn-over my wardrobe. Get rid of old electronics and other crap hiding in cupboards. Give away a ton of books and toys for under 2-year old kids. Consider selling one electric guitar (of three). Consider selling one of my bikes. Get my old acoustic guitar fixed. Get rid of one old, broken bike. Start thinking about the upcoming replacement of my laptop (late 2021 Mac Pro M1). Potentially replace my iPad.
Some big changes are coming this year. We’ll need to move to another home within Singapore and plan to move to a different part of the city. Many opportunities to shake-up stale routines and reconfigure time slots. Running routes will need to change. Local coffee shops and restaurants will be different. Our new place is almost inevitably going to be less walkable since we’re currently living in the most walking friendly part of Singapore. I’m looking forward to most of these changes, though of course moving is also a pain in the ass in many ways.
Scoring my predictions for 2025, and some new predictions
Last year, I did something I’ve sort of always wanted to do: Post predictions to see how good they are.
Here they are, with comments:
Sam Altman, Satya, and the others will continue to get more and more annoying while saying less new things. It is probably going to be quite safe to tune them out.
AGI keeps getting redefined to whatever is not quite working yet. At the end of the year, there will not be AGI, and Sam will still say it is 2 years away (reminds me of self-driving cars in 2016).
I’m going to give myself high marks on this one. The most notable thing that really happened in 2025 was that Google really entered the top leagues in AI. Progress in models has continued, but the AGI stuff has really died down now.
Reasoning (o1/o3 style) models from other vendors proliferate. Google already has one. Anthropic is almost certainly going to release a new flagship model that does chain-of-thought reasoning just like OpenAI’s o series. It is probably going to be better than o3, but compute time-cost will become a big issue for Anthropic.
True, that’s the standard now. The big moment in retrospect on this front was Deepseek which happened very early in 2025.
Video generation gets really usable and easy. Veo2 looks great. Cost will go down. Prompt adherence will improve. By mid-end year more and more ridiculous short videos will turn up in whatsapp groups.
Full score on this one. Instagram, YouTube and TikTok are now truly flooded. I’m getting to the point that I consider all short-form video AI until proven otherwise. We’re really going from ‘pics or it didn’t happen’ to ‘it didn’t happen unless I was there’. Maybe that’s not such a bad thing.
No political events will happen. There will be a lot of news, but nothing meaningful will happen. Russia and Ukraine will still be at war at the end of the year. China will not attack Taiwan. Nothing really meaningful will happen with deportations, the border or otherwise in the USA.
I’m not really sure about this one because it’s a bit subjective. The specific callouts which apparently were top-of-mind a year ago indeed still look very much the same. There was lots of news around deportations in the US, and if that was ‘meaningful’ is very subjective. I will admit that Trump’s second ter has been weirder and more extreme than I had expected. But on the other hand, friends who live in the US are just… living in the US. The notable exception is maybe Venezuela, although even on that one the verdict is still out whether it changes anything. But ‘amount of change’ is also a dangerous yardstick to begin with, because change happens slowly. So, I think long story short on this one I’m going to just say that this was a super poorly formulated prediction to start with, and not much of a prediction at all.
Elon Musk and DOGE unlock meaningful but incremental cost savings (10-30%) in the US government bureaucracy. It will also fall apart and most of the people involved will go back to tech.
Well, I guess it turns out I was WAY too optimistic here. By all accounts they didn’t unlock any real cost savings at all. They all did go back to tech though.
Predictions for 2026
It feels like 2026 will be a critical one for OpenAI. Either they pull something really innovative out of the hat that reestablishes their lead over other model builders, or they will lose their lead (probably to Google, but maybe even to multiple others). Predictions have to be at least a bit daring, so I predict a major shake-up to happen at OpenAI in 2026. Maybe a leadership change or at least a major strategic pivot.
In 2026 AI skeptics will have an ‘oh shit’ moment when they start to realize that the current generation of models, Opus 4.5, GPT-5.2 and Gemini 3, are a step change in usefulness. 2025 was widely touted as ‘the year of the agent’ at various points in the year. In many ways it was. I started vibe-coding from my phone. But at the same time, it still never really worked in one shot. Until last month. 2025 was the year the agent was born. 2026 will be the year the agent gets useful. With the latest models you don’t really have to read code anymore. Opus 4.5 has successfully built two tricky things on Magicdoor without me even looking at the code, that all prior models didn’t even come close to doing. In a widely shared post, one of the lead developers of Claude Code itself said he had not opened his IDE for weeks.
In another, a Googler who works on the Gemini API had this to say about Claude Code with Opus 4.5:
AI Skeptics have had their time in Q2 and Q3. Many AI Startups criminally overstate their ARR. The whole thing is bubbly. But in 2026 stubborn skeptics will have to start grappling with real, undeniable results. When the next generation of models comes out toward mid 2026 they will move further beyond this usefulness threshold that has just been crossed. Hence my prediction is that the narrative will shift once again as tangible results from AI agents start to appear.
Someone will be the first to release a model that learns continuously. This will be enough of a thing that it gets a lot of attention. It does well in demos and in the media, but it will be very brittle at first and will only get useful in 2027.
Salesforce is apparently advising customers to rely less on LLMs in their Agentforce product and do more with deterministic code. But… a lot of early agent stuff is running on super outdated models. I met someone two weeks ago who said their Customer Support bot still runs on GPT-3.5 (that’s a model from 2022). We will see the first ‘refactoring’ of the AI ‘stack’ in 2026. Companies that fail to realize that models are in fact progressing in capabilities and that this trumps many product implementations will be left behind.
Enough about AI, in the US there will be a traditional mid-term where the party opposite of the one the president is from wins. That means Democrats will (continue to) win many elections in 2026. This feels like an easy call, but then again with Trump it’s always hard to predict results.
Bonus: useful conceps for 2026
Gurwinder posted this great list of 26 concepts on Substack here. These are short aphorisms almost, or micro-summaries of actual concepts. I agree that many on the list are particularly applicable today, though many are timeless as well. Some of my favorites are quoted below:
In online communities, around 1% of users produce almost all of the content. As such, what you see online is not representative of humanity, but merely of a loud, obsessive (and often narcissistic, psychopathic, low-IQ) minority. Social media is literally a freakshow.
When the sheer volume of conflicting information makes the effort of finding the truth costlier than the value of knowing it, people give up trying to be accurate and instead choose whatever bullshit stinks least. Slop doesn’t just threaten the truth, but the very worth of truth.
We tend to overestimate the short-term impact of new tech, and underestimate the long-term impact, because hype inflates expectations, and thus disappointment, and thus scepticism. As such, it’s possible for AI to be both a bubble and the most transformative tech since fire.
People who try supplements or practices that might have health benefits are naturally more health-conscious, and likely already healthier, than those who don’t. This is one reason there are so many studies suggesting some intervention has health benefits; they’re confusing the benefits of the intervention with the benefits of being the kind of person who tries it.
Imagining the absence of a blessing increases gratitude more than focusing on the presence of it. Instead of wishing for a Porsche, imagine losing your legs. Suddenly, walking feels like a miracle.
Far-Leftists favour planned economies because they imagine themselves as the planners, not the planned. Far-Rightists favour a return to feudalism because they imagine themselves as the lords, not the peasants. Many delusional worldviews stem from main-character syndrome.
Don’t give the government a power you wouldn’t want your political enemies to wield. Because, one day, they may well be in charge of it.
The opposite of paranoia. The suspicion that the universe is secretly conspiring to help you. Assume every setback is the universe trying to teach you a lesson, and every setback will make you wiser. It doesn’t matter whether the universe is actually trying to help you; believing it makes it work.
Predictions that humanity is doomed usually assume future generations will be less aware and less active in fighting for their survival. But historical prophecies of doom, such as the Malthusian trap, ozone layer depletion, and Y2K, show this is false; posterity are not idle passengers headed off cliffs, but problem-solvers building bridges across them.
If you want more agency, ask yourself what you’d do if you had ten times more agency. Then do it.
Totally hear you on SEO not compounding the way it used to, it feels like the ground keeps shifting just as the work starts to pay off. And yeah, the whole AGI/AI thing being a moving target doesn’t make it easier.
What’s changing a lot is how people *ask* for information now – more through AI answers and chat-style search than classic blue links. That’s where Answer Engine Optimization (AEO) comes in: focusing on structuring content so AI systems can understand and surface it in their answers.
One thing that helped unpack this a bit is a free tool called https://aeoanalyzer.app. It breaks down how “answerable” a page is for AI engines and where it might be getting lost.
Might be useful as a lens, even just to sanity-check what’s already working (or not) in your content.
Mooi, loving the no sugar coting honesty