Equity is rarely ever worth it without at least a billion dollar exit. I'm not going to wait 4 years and take a $60k per year salary cut for a possibility of making $10k later on.
It is the hirer's job to sell this too. Engineers will always do the math and they'll do it conservatively.
Generally, the more you get paid, the less stable it is. Big tech is probably still the most stable way to make that kind of income. What else can you do? Start a business? Rob a bank? Trade stocks? Freelance? Startups are always higher risk, higher returns.
But big tech has become extremely competitive, and high competition makes it more unstable.
My rule of thumb: If the LLM is more junior than you, then go ahead and let it full autopilot. Check the results like you would check the result from a junior.
If the LLM is more senior than you, learn from it – treat it like a tutor and ask a lot of questions. Things like why we'd use MVP architecture instead of MVC. Ask until you have no dumb questions left. The hardest part of RTFM is often knowing which part of the manual to read and going through the what before the why. Ask the LLM to explain, then verify by searching it yourself.
LLMs will often pull the wrong language if you're not careful. Like contains() may function differently in different languages. This is where it's vital to RTFM.
Claude 3.7 is a little overeager. It'll solve problems you don't tell it to solve. Like I give it a screenshot of an app and ask it to do a button; it'll do the status bar as well. It's somewhat like a jerk genie and I feel like I'm writing a contract rather than working with an assistant. Claude 3.5 doesn't have this problem.
Gemini 2.5 Pro will one shot complex instructions and even read between the lines. The drawback is that it can often overthink as well. One example was I want to deselect a dropdown if I recreate it (e.g. select state when changing countries). The ideal solution is to just check this on the view layer. Gemini 2.5 pro is terrible at this - it'll go down into lower layers and overengineer, when the solution should just to set as null if the new dropdown doesn't contain the same text.
GPT 4.1 has been quite good lately, but it's not quite as agentic and it will ask you a lot of questions and request confirmation to begin. It is perfect for workflows where you aren't certain yourself.
GPT is also smartest from the visual angle, able to read color variations well and designs them well. Claude is the worst visually.
I've been using it since August '24 and it has never been reliable enough for production. My app from then is broken because it doesn't follow instructions, especially when you make it output JSON without some middle layer to "repair" the broken bits. There was a point where it simply broke all the JSON arrays and I had to restructure it to not use arrays.
It seems to work okay for something like Cursor, where you clearly realize that you're using a certain model and are willing to take on the risk for higher quality output.
I did freelancing for 5 years and the trick for me was just aiming to do things faster, cheaper, better.
Well, the trick is really just faster. If you can do things at 2x the speed, you can charge 1.5x and clients will be very happy. And if things are done faster, you can layer on more quality things as well – tests, features, fancy backgrounds and animations, optimizations for low end devices, and so on.
People are often afraid that being too productive will make them run out of work. But with software, here's an endless amount of things to do. One of the more valuable things a contractor can do is help a profitable company with tech debt transition out of the tech debt. Eventually they become productive and take on tech debt because there's value to releasing new products as early as possible, and it's an endless cycle.
Other countries copy FAANG but if we did full leetcode here (Malaysia), we'd end up hiring nobody.
I like the way Singaporeans do it – it's still an algorithmic problem, but usually related to the job. Like a logistics company would give a spreadsheet and ask how to fit the most packages inside the van, with bonus points if you put categorized packages next to each other. Or some variation of traveling salesman. Optimization skill matters no matter where you work, but mostly in the sense that you're not applying O(2^n) unnecessarily.
Multiple interviews are really based on funnel size, and funnel size is based around expectations and whether the company can afford full-time recruiters. If you're paying say, $1500/month for a junior here like Accenture does, the funnel will be full and you can drag people through multiple rounds. Many will just prefer to pay $800 and grab the first applicant that passes.
I have a list of technical questions I tend to stumble over. Things like SSL pinning, cache management, CI/CD, github stuff, and DI. Often things that colleagues have set up 3 years ago and are now just a function call on Slack.
Every interview will ask the same questions: "Tell me something about yourself. Why would you leave such an awesome job? Are you willing to take a pay cut for equity? Why do we need you instead of paying a junior and giving them AI? How do you practice Agile in your workplace? How would you train a junior? How much are you being paid now? How do you feel about trans people and dogs?"
Some of these are traps. The only wrong answer is sounding defensive or evasive. So it's important to practice taking these head on.
I find that if you nail the first interview with the decision maker(s), you will get the job. You can fail the interview but they'll even give you another one because they see it as some kind of error. If you excel at the technical interview and don't pass the vibe check, they'll find a reason to reject you. Then they'll pick at the number of past jobs you held (too many or too few), resume gaps, 'culture fit', etc. And if they absolutely find nothing to reject you for, you'll get rejected for being overqualified.
Also it's important to interview for the role you're applying for, not the one you hold. For me, it's a shift – founding engineer/staff+/TL/EM/CTO at this point. The IC roles aren't being evaluated from an IC angle. They're being evaluated from an angle of whether I can lead product next year once the company has raised $100m.
They don't know anything about you. They don't know that I used to spend my teenage years as a kind of PM making games with strangers over the internet. The unwritten question I have to answer by the end of the interviews is, "Can this engineering degree nerd from a developing Muslim majority country lead a team of white people who went to MIT?"
People love to talk about themselves and what they're doing. I don't think there's any barrier; if you know the right way to approach these conversations, they'll dump ideas on you.
I believe it might be selection bias. Founders who are building things are not joining networking events. I tend to avoid co-working spaces too ever since I realized it's full of the people I block on Facebook.
There's people who bounce ideas, and then there's some form of idea tire kicking. They talk about how vibe code is the future and going to change how we live and work, but they have never downloaded an AI IDE. These people are there to kick ideas around all day and they'll pay for a co-working space so they can kick their ideas at strangers and ignore all feedback. It's exhausting and it's possible that you might be viewed as one of those people.
IMHO Networking events are but marketing and sales events for the sponsors or VCs.
We need bring back our modern day version of the Homebrew Computer Club.
The one's building amazing things are far and few in between due to lack of mentors in the community and hyper focus on "what's in it for me" mindset.
We do have some like the Homebrew Computer Club, but not (yet) as elite. Kuala Lumpur has some talent, but not the density of talent. I may join once every two years. Every time I do there's a founder who ends up public listed. One even had enough success rate to become a VC, but it's hard to tell which one will be who.
But so far the sales events have had a higher hit rate. There's very talented people joining whatever new thing AWS is launching, or the hackathons.
If anything, founders are good at building things. Hackathons are an excuse to experiment with a new tool for some, a sport for others. I think a lot of friendships today are forged in the heat of the hackathons. I met this guy who wanted to build an "AI assistant" with Gemini, which I dismissed as a dumb idea because Google has been building AI assistants for years. But his idea went much deeper than that and he won the hackathon car.
Basically instead of a complex layer, it just uses .cursorrules as the memory. This was before MCPs, so it might be capable of more today.
reply