Tenemo

We are already - as in today - doing this with friends while playing Valheim, the process is semi-automatic with me in the middle as the admin, though.

We make 10+ gameplay changes a day in our custom kitchensink mod that steers everything. 70%+ are bugfixes (we have plenty of mods that conflict/are buggy), the the rest are QoL improvements, small new features and rebalancing. If we see some playstyle is too strong or boring, we spice it up the same or the next day. Mobs too weak/strong? Update the same day. A ton of performance fixes with actual in-game benchmarking. All this while I keep playing.

I built sort of an automatic modding pipeline (including public release and all) with parallel test game clients + servers used by LLMs to develop and test the changes and then merge before releasing a new mod version. We have had very few issues, I'm still shocked how reliable this is. They follow guidance well and test everything in real game clients on wipeable clones of our live world and our actual characters. Josh has a bugged chest at his place? The LLM will reproduce it, develop a patch, yes, but then also log in as him to his actual in-game base and test the fix using game inputs + OCR. At the end I get a .html page report on each change with screenshots and technical details if I ask for them. I don't care if it takes 4h to do a simple fix because it's not my time and game stability is critical. And up to 8 of them can work in parallel, so there isn't a bottleneck caused by excessive testing/re-testing.

I think this will be the future of gaming for many sandbox/easily moddable games. It's a bit too technical at the moment for regular players to handle the updates and models' quirks, but we're only going up from here. Just crazy.

Do you think e.g. students learning docker syntax at a uni is beneficial in 2026 or you think LLM-assisted programming means that Docker will become an abstraction under an abstraction and its setup will be mostly automated by LLMs? I find myself very rarely debugging or handwriting dockerfiles nowadays!

I don't think I've seen a less accessible *font* in my life, including all the dark grey on black and yellow on white ones. Those are hard to read, but this one makes my eyes hurt, they actually started to hurt after I tried reading it despite discomfort. How come? Reading random text for 10 seconds can cause people physical pain? I wear glasses for near-sightedness and slight astigmatism, I wonder if it's related, and if people with 20/20 vision experience it differently.

Paying for an account doesn't opt you out by itself, right? Has he stated anywhere that he actually opted out? But if not, then I also don't understand why OpenAI's communications about this have been so vague, they could've just said that he didn't opt out, using those chats in training data follows their ToS and that's it (whether that's "fair" is a separate discussion).

This is what TrueCrypt/VeraCrypt and similar software do already: you have two passwords to your encrypted drive, one shows the real data, the other one a whole another fake drive with whatever you put there. Additionally, there is no way to forensically prove that there is hidden encrypted data present and that there's a second layer: you have plausible deniability.

Serious question then: if you were the CEO/CTO and were employing 10,000 engineers, what would you do without middle management? Would you have 1k+ line managers each managing <10 ICs and then... yourself? How would you coordinate work, hire, or do anything, really?

I'm not saying corporations execute on this flawlessly, just any time when I wonder what I would've done myself, I end up with a similar structure (maybe a bit flatter)...

Others on Telegram? Some sort of a HN channel?

I get what you mean but the GPU comparison isn't the best here, I think. Money-is-no-object-I-want-the-best approach is questionable, definitely. But no one can argue that an old Nvidia card is objectively better for e.g. 4k gaming than a 4090 if you don't mind the wattage. You can just measure it.

With LLMs the problem is more complex, it's people getting used to how a model works and to the ecosystem. Sure, you can make all your skills harness-agnostic and deal with Anthropic's stubborn refusal to adopt the common naming/directory structure. But most people don't. So then you end up with something closer to the ancient Android vs iOS discussion. Can you prove, in isolation, that iOS is more energy efficient, the hardware is faster? Yeah. But that won't speak to someone who has been on Android for 10 years and would have to migrate and get used to iOS to experience that, first.

I've noticed myself how I get used to common failure modes of particular models in my projects. GPT5.5 tends to create some checks/booleans I don't need, it heavily overcorrects on error handling, etc. While Claude 4.7/4.8 doesn't do those as often but gets derailed on our E2E test suite, forgets to run linting despite guidance. So even assuming fully harness-agnostic working setup, a new LLM model with its own quirks can be a lot friction for heavy users who might be used to Claude specifically and all their skills/guidance pre-address common failure modes.

E.g. I might be a Prius owner, then you gift me an objectively better, more efficient, safer, newer, same-size, physical knobs car ...and I might still swear by my Prius! I'm used to how it turns, how it feels, I can repair some issues myself. Isn't that a normal reaction then?

Mandatory PSA for Android users because people tend to have similar complaints each time in popular threads: ReVanced allows you to have YouTube with Sponsorblock, background play, no ads, Shorts completely hidden, (estimated) dislike counter brought back etc.

Couldn't imagine using YouTube on my phone without it, it's night and day difference – that's despite being a premium subscriber anyway.

> The authors acknowledge funding from the College of Literature, Science, and Arts at the University of Michigan. R. L. P. was supported by an NSF Graduate Research Fellowship (NSF-GRFP) DGE-2241144. M. E. C. was partially supported by the University of Michigan Rackham Graduate School through a merit and predoctoral fellowship. The authors would like to acknowledge the professors and students of the Mapping, Measuring, and Modeling Microplastics in the Atmosphere of Michigan team for their support and helpful discussions. The authors thank Jennifer Connor, Curtis Refior, Amy Pashak, Megan Phillips, Josh Hubbard, Bill Joyce, and David Lee for their community partnership. The authors would also like to thank former Dean Anne Curzan from the College of Literature, Science, and the Arts at the University of Michigan for funding this work through the “Meet the Moment” grant program. The authors acknowledge technical support from the Michigan Center for Materials Characterization.

Is there anything wrong here? Not sure I understood your comment

I'm not 100% convinced, while iterating fast on an early prototype, what's wrong with legitimately not knowing what e.g. the data structure will end up looking? Just let it run, check debugger/stdout/localhost page and adjust: "Oh, right, the entries are missing canonical IDs, but at the same time there are already all the comments in them, forgot they would be there – neat". What's wrong with that? Especially at uni, when working on low-stakes problems.

Was the number of partners you picked random or you chose 2137 on purpose? As it's actually somewhat related to the topic...

I tested the latest flagship reasoning models (so the only models I use outside of coding for general questions):

- Opus 4.6 (Extended thinking): "Drive it! The whole point is to get the car to the car wash — you can't wash it if it's still in your driveway."

- Gemini Pro Deep Think: "You should definitely drive. Even though 50 meters is a very short distance, if you walk, your car will stay where it is—and it's pretty hard to use a car wash if you don't bring your car with you!"

- ChatGPT 5.2 Pro (Extended thinking): "You’ll need to drive the car—otherwise your car stays where it is and won’t get washed. That said, since it’s only ~50 m, the most sensible way to do it is often: 1. Walk over first (30–60 seconds) to check if it’s open, see the queue, confirm payment/how it works. 2. Then drive the car over only when you’re ready to pull into a bay/line."

A pretty reasonable answer by ChatGPT, althought it did take 2min4s to answer, compared to a few seconds by the other two models.

I' migrated to purelymail.com around 2 years ago and. Reaaally cheap, easy to set up and without any bloat whatsoever. The webpage might look sketchy at first, but don't judge a book by its cover :)

You seem to imply that its outputs aren't found by people to be useful, which isn't true.