nickabe

Tests can be green and I think doing E2E is better than just unit tests. But it's also very easy to have hundreds of tests that are green - but if they are testing the wrong thing then it really doesn't matter.

Ultimately, you spec out these tests and create deterministic code around them - but that doesn't mean that the tests didn't drift from the intended spec.

I think I agree with the author in that what you were TRYING to achieve is much more important than what you actually achieved (in code or tests) and actually using the product is your best defense to see if code / spec has drifted.

This is objectively hilarious.

Basically every large company has gone all in on AI, and as a result the number of bots has (likely, I have no proof) exploded. Then as a result of that, they crack down on "potential bots" by banning them. Of course, all they care about is what % of bots they cracked down on now what % of real people. Then because they've gone all in on AI, you can't get an actual human to help you... just a bot they trained to always say no.

My current experiences in these arena: - Opened a Meta for Business account to try and buy ads. Immediate lifetime ban before taking even a single action. Most likely thing that looks like a bot? Opening a new account. - Got bot farmed on a Google Ads campaign, delivered definitive proof of it to Google. Never got a real human, just went back and forth with a bot who was told to acknowledge everything I said without admitting fault. End result - no refund, the bot farm still operates, and I have to be super careful when buying Google Ads. - Tried to open an account to post stuff on Reddit, every post gets instantly banned because I don't have enough karma. Try to start posting comments to get karma, same thing. IN fairness, this did eventually start to work - where I could post on certain subreddits and comment in others.

Overall the experience always has the same shape - something super opaque to the end user, no resolution options, and really you are left with no alternatives because every big tech firm is doing this.

So then the same AIs that create the problem recommend you try and get media articles published about it to draw attention to the plights of the real human caused by AI bots.

I'm definitely on board with governments owning stuff - I think the traditional issue has been that government jobs don't attract the best talent (lower pay / whatever). So the types of people that go to government jobs are the ones that like stability and moving slow. That's probably fine - but it has also given them the reputation for being really inefficient.

The natural response is to assume that corporation will be more efficient, and there the consumer will get a better end result. And, again, that was probably true for some period of time. The problem is in a corporate world you need to grow your profits by 3-5% a year (minimum) in order to be considered a "viable" investment. So what happens when your customer base isn't growing at 3-5%? Well you cut costs to improve your profits or you raise prices or both.

End result is the consumer ends up worse off, they got a one-time efficiency gain by switching from government to corporate - but then pay for that over decades.

Honestly, the simplest answer is that only like 20-30% of an actual company is the idea / the product. The rest is selling it. I think in areas where selling is adjacent to the selling they are doing now - then they are going to end up doing it.

Anything software related they will end up being "in", in some form or another. Why? They are already selling to enterprises and have their trust. The selling motion for Project Management software may be slightly different than Enterprise AI - but it's close enough to rhyme.

They will probably also end up in anything consulting related. Why? Well again most consulting comes from Enterprise or Government. They are already selling to both. Probably Enterprise moves away from Accenture / Deloitte / McKinsey towards Anthropic / OpenAI faster than Government - but still the point remains.

Selling plumbing services? Probably not.

Yeah - in my own personal experience building AI products this still holds true. And you can see it in the way they behave themselves, they don't really write prompts for themselves to use in the future (say CLAUDE.md) when they need the output to be deterministic. They write code for it, then write tests around that code, then fix the bugs in that code.

Yes it all started with "natural language", but even AI has an understanding that you need more structure than natural language when you want deterministic results.

Does it matter if it's great code?

Like what matters is: - Does it do the thing it's supposed to do - Does it do it with a reasonable amount of resources - Does it do it without any obvious security flaws

Yeah, if they were nuclear codes you'd want to be really really sure and you'd probably want to check every line yourself. But if you're making a habit tracker or 95% of apps out there, it really doesn't matter.

There is definitely some skill in piloting the LLM, where you have an advantage if you understand what "good" is - but I'd argue it really isn't that necessary in most cases. McDonald's has had a lot of success selling food that most chefs would say isn't that great, but it fills that need at that moment.

Yeah I agree - if you're very clear and consistent with how you ask it score something it will do well. Not fundamentally different than giving it clear instructions when you code.

One thing that has helped me make the LLM more consistent on scoring is deduction based rubrics. Start everything at perfect and deduct points for specific things (ie style, missed facts, etc). It gives much tighter final scores under these conditions (for me) than giving rubrics + samples.

I feel like newer systems are even worse, probably in about 50% of my online experiences now I just get "banned" immediately. Opening a new account = sign of a bot = banned.

Hyper focusing on detecting bots without thinking about the humans you're blocking is making the Internet a crappier place.

This is very cool - one minor issue, when I zoomed out on a Metropolis it hung for a very long time at the several block radius and then after say ~20-30s it then was able to load the full city.

After that it still did it on various different views but it was much faster.

I think that's part of it. Humans are definitely much more confident (even when wrong) and don't hedge their bets as often.

I think AI also has a lot of internal conversation - you can see it in the reasoning phase - where it will invent shortforms as it talks to itself. This also makes it really hard to follow AI text because it will often invent its own jargon.

I do think, as some other posters have suggested that giving the LLM a design and asking it to follow that design is a good best practice.

Beyond that though, a really good way of testing is actually using the product you've asked it to build. In a pre-AI world that was the biggest beef I had with real developers. They would say "hey I built this thing" and never actually test it inside the app at all. I appreciate most people are going to say "those are bad devs" - but it was shockingly common. The view was always that this was something QA or the PM should do, but you never learn more than by actually using your own product (or code in this case).

I think some part of this is anti-bot. They don't have a good solution to this yet, I and someone else I know got a lifetime ban from Meta just from opening an account.

No actions / no posts. I guess they determined the way we opened the account was too bot like. Sucks because we're real people that were trying to buy Meta ads.

But maybe doing something like this allows them to ease those restrictions, and then by charging for posts they have an ability to determine bots from non-bots in a more robust way. I'm doubtful it will actually work - because bots are exactly the type of thing that can and will try to open dozens of accounts to get around the 2 posts per month limit.

You might want to adjust the scoring.

A simple system would be something like: - accuracy is most important - if tied on accuracy (ie both got 8/10) then tie break is done on speed

In general, across a lot of games - winning is a big determinant in who comes back. Right now it might be very difficult to beat JEV with the way your scoring is setup.

For my game in particular it's like a 47% chance that a winning player returns versus just 10% for non-winners.

There's a couple more things you can try.

If you limit the sample size, you end up finding more. For example, on my project when I tested 350 strings at once it failed to find 15 strings with "ai slop" that testing 10 strings at a time did find.

I think you're probably right, and even a lot of the corporate "help" is bots.

Clearly a bot - what human wouldn't want to see ads?

Yeah sorry - adding it to Europe probably next week though.

Thanks very much - really means a lot that you enjoyed it, we put a lot of effort into making it.

Yeah if there's a DB that would be helpful - I'm not currently capturing the IP, but I'll see if I can add it in a future release.

Surprisingly, no. 29 of the 30 zero-time installs had no click at all; Google credited them to someone watching the video. The install isn't what gets the farm paid. It's what makes Google's bidding think that placement is working, so it sends it more of your ads.

I guess that's the nice thing about not having a VC... there's no advantage to spinning to anything other than fraud lol.

Thanks, and agreed. The in-app goal was the lesson, and I'd rather have learned it at $220 than at a trial-run budget. The part I found interesting is that Google's count wasn't wrong: there really were 21 installs. You only saw the problem by looking at which version of the app they installed. The YC advice is roughly where I've ended up too. And fair point on the refund, I'm not holding my breath, but I'll report back either way.