As someone who works in a studio that can afford unlimited AI bills, if you can afford to pay for 4-5B tokens per developer per month (that is the average monthly usage across our team for the past montha + CI) on latest models, it does a pretty good job and we can mostly commit fully AI authored code. Not one shots, it’s rounds upon rounds of AI reviews, re-prompting to use a different system/make a change that sucks, but it works well.

Sure, the team is loosing project domain knowledge. The results are currently possible because most of us did spend few years working in the codebase before AI, so we can quickly spot an issue and point the prompt to a general direction, a space/system to use, or ask it to look into what part of the code. If we don’t, it will duplicate random stuff and make it even more mess.

Of course, neither is worth it or sustainable. The cost is not reasonable in the slightest, my usage alone is more than my salary. Reviews are a bottleneck, and people have started to just run more rounds of AI on the PR, in adittion to the CI review loops and the authors review loops. It does help - each round usually catches something, but the cost, lol. And of course, we are slowly loosing both the abilitty to work without AI, and domain knowledge to be able to somehow rein in the code base.

So, if you have unlimited usage and can afford to run billions of tokens of looping reviews with the latest models, you can go fully AI pretty succesfully. But at the cost, and that is the pricing while AI companies are in the start-of-Uber phase, it’s such a stupid idea.

I don’t care about the job that much. I get to spend money (without trying to, to be exact - its just the only viable workflow) to make a point how stupid it is, while being able to do other stuff, so I don’t really care, other than getting a course to requalify into another field after the studio burns.

But yeah, if you just ask claude to one-shot something, it sucks. It needs at least a few rounds of looping, clean slate review loops, or council decisions during the development. It would probably eat your Claude weekly limit just for one PR.

The point is - there is a huge difference in how people experience AI. If you can pull the slot machine lever more times, for an absurd cost, it will eventually get to passable results. The AI companies only see that, and from that PoV, yes, it is kinda passable. But there’s like five companies in the world that can do that, and not for long.

I don’t see any reason why not start calling it SI.

I really want to see the US companies being forced or fined because they are using AI instead of SI. Force-rebrand OpenSI, make them (and all other US AI companies) change all their marketing materials and force them to scrub mentions of AI from all of their texts. Forbid them from ever using it again, instead of SI.

It would be so funny, a PR and marketing shitshow and I’m here for it.

And as a bonus, we can finally give back the AI term to gamedev. Searching for any materials in regards to NPC/game AI has been a nightmare.

I’ve seen people mention that he in the “AI art is art” crowd, even his username might actually be that? I didn’t check, but that would fit.

Depends on your threat model. For someone who works in cybersecurity, the hard seperationo files per-project, easily split networking to different VPNs between VMs, and being able to keep sensitive data on a basically air-gapped VM that you can only copy to clipboard from is amazing. And with templates getting a stack of VMs for a new project running is pretty quick.

QubeOS is one of the coolest OSes around, if you work for multilple customers and/or take your data security seriously.

It’s true that I have not considered it from the AI-slopworks PoV, but in general doing development in a container is probably a must nowadays anyway, so why not have a system build around it. It’s not like virtualization is any difficult.

There are already version of debloated windows installs, like Windows X-Lite, btw.

Yeah, the guy that just released all Adobe products a week ago that barely work, surely releasing another suite of tools next weekend instead of fixing the plethora of bugs is a good sign that he is capable to keep maintaining a project in long term.

I did try the PhotoSlop, and within the like 20 minutes I couldn’t make the canvas the size I asked for, the work (mostly trying random stuff, but that’s not the point) did not get saved when I saved it, and it froze when I alt-tabbed, so I lost all of it.

I’m sure it’s getting them a lot of attention, but why would I use anything of his when he got bored with it and moved to another project in literally a week while it has absolutely blocker issues?

You should choose random parts, and add “I don’t think this makes sense, can you explain this part?” or “Why didn’t you use <something semi-related> instead?”, “I think you should have done <random similar pattern>” here." to random parts of the PR. No need for it to be true or read it, not like it matters. You can keep doing that for quite a long time :D Sure, they’ll just copy-paste it to their agent, but the longer you keep that going, the more unsure will they be, probably.

Or even better, just message him that you have questions about the PR, if they can come over because it’s more things, and ask these comments directly.

I’ve been able to find a workflow where it’s mostly correct and can handle most of my gamedev related coding without making too many mistakes. I still have to actually read through the code and pay some attention to what it’s doing, and if it misses something and goes on a wild goose chase, it’s unusable and I have to start over (so someone who didn’t know what they are doing would be cooked), but whatever.

Sure, it does require a lot of looping adversarial reviews, and my average token cost is around 3000$ (4B tokens) a month (we have unlimited budgets and a pretty accurate tracking, and also probably cheaper than consumer prices per token with how large company it is), which is actually more than my monthly salary, but it’s just a job, for a company and on a product I don’t really care about, and I can 1) keep slacking in my job while doing my own coding stuff and projects, and keep seeing how absolutely unreasonable the prices are if you want to get at least semi-submitable results.

You could ask if it’s necessary to spend so much tokens - but so far most adversary review rounds do find serious blocker bugs the first implementation had, most of them being some hidden stuff that would be difficult to reasonably find by hand unless you really understand everything around the code, which you should, but the point of AI is that you shouldn’t have to.

Is it worth it? Lol, no. The whole team is loosing codebase knowledge, we’re getting bottle-necked by pending PR reviews that are just stacking up and no one wants to do, so we’re not even more effective, the cost is absolutely absurd and in no way near sustainable or worth it. And the longer it goes on, the less I can realize it’s spewing bullshit and I should stop it and nudge it into a different direction in our codebase, because I’m slowly loosing touch with it, while for now I’m still relying on what I remember from before we went all-in on AI.

And that’s while the whole industry is in the “Uber pricing” phase, so it will get a lot worse. But yeah, if you can burn 100-200$ per a simple implementation task, then it can have a pretty usable results. And that 3000$ a month does not include our CI review bot, that does additional rounds of multi-agent council reviews. And I’m not even working on anything complex, mostly just various menu screens UI. I can imagine the bill getting a lot higher if you get into more involved systems.

Please please I really hope they will force US companies to start calling their chatbots SI instead of AI, and forbid them from using the term AI.

That would be such a funny PR/marketing nightmare. It’s such a cringe term :D

I found this being even more of a slippery slope. Your brain learns by making new neural pathways whenever you are thinking about something. If you skip the discovery part of it, and just read a finished train of thoughts/explanation, withoit having to carve the whole path yourself, it will not stick.

There’s a reason why just reading a book without doing any outside the box exercises, or just readinbg answers to them without figuring it out, is almost pointless and nothing will stick with you.

Frustration tolerance is a skill you need to practice and develop, and the more you get used to not getting stuck for a while, the harder it will be to not give up and just go find a direct answer, megating most of the learnig.

There was one defcon talk about it. It’s a lot of pretty complex reverse engineering, so getting really good at reading and understanding assembly and working with debuggers /disassemblers without symbols is a start.

The talk I remember, where they were able to steer and break any car of the model in the world remotely, they got in by finding a unauth vulnerability in a remote call, and then used the firmware update process to bridge the almost airgapped vehicle controls system from the frontend car control (the screen, etc) system.

They boasted how secure it is because it’s airgaoped and compromised frontend can’t touch the critical vehicle controls. It almost was, save for a single function used for updating both at once. It was months of reverse engineering and several bricked cars.

But if you just want to change the frontend a little bit, it’s not going to be as complex, probably. Still really difficult.

What, you don’t have your SWAP set to 1TB Google Drive share?

This is cool! My sanity with self-hosting was mostly perserved by MASH, they have an amazing ansible scripts I never had any problem with, which is usually rare.

I can’t imagine figuring all of this on my own.

Take a look at how Firefox marketshare rose when Chrome killed ad-blockers, or how did Discord’s marketshare drop when Discord started asking for age verification, and there’s your answer.

Is it a compression when you need presumably gigabites of a model to reconstruct the data?

It’s basically the same as saying I can shatter compressions records by hashing the thing and using a rainbow table to decompress it.

You left? Murderer.

Haven’t read the article so I can’t actually judge it, but an article about the harm of solars on a website called “Scientific American” is pretty funny. Name a more iconic duo :D

The space crime allegation was a legal mirage,” said Prof. Michelle Hanlon, a space law expert at the University of Mississippi,

Wait, isn’t figuring things like this out like literally the only job someone called “space law expert” should do?

What do you mean “legal mirage”? You had one job. Space law. One thing ever has happened that needs a space law expert, and you don’t know?

It’s pretty funny :D

Isn’t that basically the definiton of paradox of tolerance?

I’m pretty sad that people have to be perverts and glasses with cameras got such a bad rep.

It is deserved and not surprising, and I’m not trying to defend them, but I was so looking forward to AR glasses with cameras, so I could finally get a working ad-block IRL, which should be easily possible with the tech.

But, that requires a camera. It’s a shame, and I’ll definitely rather continue dealing with ads than support enabling perverts (or Meta, anyway), but it sounded like such a cool use-case.

I hate how there are always groups who have to ruin everything for others :(