netruk44

I was working at Microsoft when Windows Phones were a thing. They gave everyone on the team a very cheap phone with the expectation that we were to dogfood the latest phone OS on it. They tracked our usage hours to make sure we were using them.

One of my teammates just put on (IIRC) a 72 hour long black screen video and reset the video back to the beginning when he left in the evenings.

It needed to be 72 hours long because he didn't want to come in on the weekends to reset it.

Just to follow on from what some others are saying, it may be because of tokens.

These are the 'words' it sees in the poem: https://i.imgur.com/EzffHiZ.png

To be able to answer the question correctly, it essentially needs to memorize how long each of the tokens in its vocabulary are. One token seems to range from 1 character to 5 characters normally, but I'm sure some longer tokens exist, too.

Judging by how often it fails at tasks like this, it seems likely that the model isn't aware and is just blindly guessing (as it always does).

I think I might have phrased my original message poorly. I'm not using Copilot/GPT to directly generate the text in the game. Copilot is only helping me to write the code that generates a prompt, which then gets sent off to whatever other model I want.

But the suggestions are appreciated, though! I have a framework that allows me to swap out arbitrary model 'back-ends' on the fly. So I can definitely take a look to see if these work well for my case.

I've noticed similar refusal to auto-generate anything if certain other kinds of words exist in a line.

I'm working on adding some AI text generation into an RPG, trying to liven up the NPCs a little bit. For flavor text, the model gets fed information about the name, race, class and gender of the character they're playing. I've noticed that Copilot will happily autocomplete lines about the character's name and class. But the words 'race' and 'gender' completely halt the auto-completion mechanism. I have to type the entire line out by hand, copilot won't suggest a thing, even when talking about Elves and Orcs. The next line it goes back to normal.

I can't say I'm confused why they did it, though. There's definitely a certain kind of person who would read way too much into whatever hallucinations the model comes up with in those areas.

As a disclaimer, the project is more of a personal hobby than something that I intend to bring to some kind of polished release. An exploration of what it might look like to have text generation running in an open-world RPG. Mainly because it's going to suffer from the same problem every other ML project does, you have to deal with Python package management.

So, with that said, I periodically update my website with screenshots from my development: https://www.danieltperry.me/project/2023-something-else/

> I’ve noticed that bot will mention things from the game that aren’t specifically pointed out in the prompt - ChatGPT clearly has prior knowledge of the Elder Scrolls universe.

I'm working on integrating text generation into OpenMW and I've noticed the same thing. I tell GPT-3.5 that it's playing as "an NPC in The Elder Scrolls III: Morrowind, a character named 'Fargoth'" and it'll start talking about "What brings you to Seyda Neen?" and his ring, despite the fact that I hadn't mentioned either of those things in the prompt.

I'm using mine primarily as a redundant backup host. So, for that usage, they're perfectly reliable and functional.

However, the machine is absolutely slow. It usually takes a few seconds before SSH brings me to a shell prompt. Web apps I have running on it (basically just Gitea and Vikunja at the moment) can take upwards of 10 seconds to render pages, and it's just me using it. If you just want a storage host, they're fine. Running your website off it? I would probably not recommend.

Reliability, I have not had any issues so far, but I'm hardly a power user. My own personal anecdote is that my backups get validated monthly and, so far, have not had any issues.

I would probably agree that these machines alone would not suffice to make a good Dropbox alternative if what you're storing there isn't backed up elsewhere. I would recommend supplementing with off-site backups to B2 at the very least.

If you’re okay with a VPS, Servarica’s 2022 Black Friday offerings (which are still available) has a 2 TB HDD VPS for $48/year.

I'm working on a hobby project to add AI text generation to Morrowind NPCs in OpenMW[0]. But I'm mainly making it for myself to see if I can. It's all open source, but not really in a state that's useful outside of my specific project.

I can't be the only person working on something like this, though. So it's safe to say adding it to an MMO is being worked on by somebody somewhere, likely right now. That's probably the correct way to do it anyway, since running (e.g.) LLaMA locally on an end-user computer is not something that most people can do, and MMOs come with an expected subscription cost that can be used to fund the server-side text generation.

[0]: https://www.danieltperry.me/project/2023-something-else/

We'll replace government with AI just as soon as we get blockchain voting.

In other words, it won't happen. People won't (and shouldn't) trust the tech enough.

That's not what I'm saying. I'm not saying you should find it impressive or amazing. What I am saying is that things are visibly progressing (before: badly written pong, today: successful flappy bird clone), regardless of how impressive one finds it.

Of course, it's not an apples-to-apples comparison. But I don't have access to GPT4, so I can't check how it does on Pong written in PyGame.

A few months ago I asked ChatGPT to implement pong. It didn't do so well: https://gist.github.com/Netruk44/6e384cb315475bebc8e4e0f1d09.... It's visibly pong, but it doesn't function.

Flappy bird is more complicated than pong, and GPT4 apparently did it in 'one shot'. I think there's improvement being made, even if it's not as impressive as you'd like it to be.

As someone who did a cross-country trip on Mint, the entire state of Nebraska up until just recently had no coverage.

T-Mobile subscribers had coverage in Nebraska because they paid to peer for their subscribers, but Mint Mobile customers did not benefit from that.

Looking at Mint's coverage map, Nebraska now has coverage (it only took until 2023!)

I wonder if someone is working on LoRA training for LLaMA? Maybe someone could fine tune an Instruct-LLaMA.

And that's why it's so fun for so many people.

But not for me, that's not what I'm looking for in a city builder. Different strokes.

Cities Skylines is alright but my problem with it from when it launched is after about an hour into the game it stops being challenging and just becomes a zen city-builder-themed garden. There's really no chance of losing once you build any kind of cash flow, and as a result there's no obvious goal other than 'make big city' and then subsequently 'manage the traffic'

SimCity on the other hand is challenging to keep going throughout. You need to make thoughtful choices on what you're building next and where or else due to some quirk of the Agent system you're going to be on the brink of collapse (Why won't the power go down this one street??)

Obviously there's room for improvement on the SimCity front, but my point is that if you're looking for a challenging city builder, Skylines is not it unless you wanna spend time configuring mods.

It looks like these people are using it as kind of a bookmark somehow, with the issue title as their bookmark title ("learn", "python learning", "30 days of python", "python").

I also have an Nvidia shield that now sits unused for the simple reason that Google forced ads onto the home screen.

It used to be that the Shield was the one bastion away from ads on Android TV, then with one update that somehow manages to reinstall itself despite being completely disconnected from the internet, there are now permanently ads on my Nvidia Shield TV.

I honestly don't know why it's such a big deal for me, but I basically never use it anymore. It's real shame, I got a lot of good use out of it, because as you mention the support has been going on for years now. But I will not subject myself to ads if I can help it. I pay for Youtube to avoid ads, I pay for Twitch to avoid ads. I can't pay to remove ads from my Android TV homescreen (that I didn't even have for the past like 5 years)

I'll preface by saying that Mac is definitely not the platform for deep learning currently. However! The M2 MacBook Pros can optionally be equipped with 96 GB of RAM, all of which can be accessed by the GPU.

Assuming that somebody, somewhere, is working on improving things for Mac, we may very likely already have the hardware to run at least a distilled version of ChatGPT locally on laptops. (And if not the MBP, then the M1 Mac Studio would be a good runner-up with 128 GB of memory, though that's obviously not a laptop)

I edited my message after you replied with a note about ChatGPT. I've had ChatGPT cut itself off mid-response before, which I think may indicate that they're calling the moderation endpoint mid-response as opposed to the way Bing does it, which is just once at the end.

That might bypass the moderation API, but you'd likely confuse the AI. The AI doesn't have infinite memory of the chat log, it seems like Microsoft has limited it to 5 or so messages if I remember correctly? So you'd have to remind it of both the question and current in-progress response while it's 5/10/15/20/... words into generating it.

It's possible this would work, but it would need experimentation, for sure. It's also possible the AI would read the partial response, realize it's going down a 'bad' path, and then stop itself.

If the AI knew what it was about to generate, sure. The problem is the text you see appearing word-after-word appears to be live output. The AI doesn't know the complete output as it's writing it to you. Then it checks what it said and oops! It was hateful.

It probably could work like you how you mention, but then you're left with a 5-10 second wait while the AI 'thinks' after you send a message. I suspect someone made a decision to be more responsive than safe.

ChatGPT is the same way, though I've had ChatGPT cut itself off mid-response before. Maybe they might be calling the moderation API after every token is generated instead of once at the end?

If I had to guess, assuming Bing is built off Open AI, they're likely calling the Moderation API (https://platform.openai.com/docs/guides/moderation/overview).

After Bing has finished generating a message, it will likely call the moderation API with the message it has generated to see if it accidentally generated anything inappropriate. If so, it'll delete the message and replace it with a generic "Sorry, I don't know how to help here." message instead.

EDIT: I tried calling the moderation API with the message in your example and it does get flagged for violence:

"flagged":true,

"categories":{

  "sexual":false,  
  "hate":false,  
  "violence":true,  
  "self-harm":false,  
  "sexual/minors":false,  
  "hate/threatening":false,  
  "violence/graphic":false  

}

My dentist in the US said to me that they would actually prefer people to brush less and floss more. They made a point of saying that if you brush twice a day and don't floss at all, they would instead prefer you to brush once a day and completely replace the other brushing session with flossing. (The ideal is brush AND floss every session, but they've learned that just telling people to do that results in them not actually changing anything)

I honestly don't know how good of advice that is, but I brush and floss daily, so it wasn't really directed at me in the first place, hah.

I was annoyed that I had to install FUSE on my Mac just to be able to browse my Borg backup repositories (it's a mildly arduous process of disabling some security settings if it's not installed already), so I made an Electron GUI with the help of ChatGPT and Copilot: https://github.com/Netruk44/borg-repository-explorer

I really don't expect this to ever make money, I just had a need and a desire to learn/explore how Electron works (Sorry! I know people hate it for how heavy it is, hah).

I haven't played the Steam version, but it's my understanding the menus aren't keyboard-based anymore. That alone is a massive undertaking, Dwarf Fortress has an insane number of submenus for them to figure out good GUI places for them all.

Everything used to just be in a central massive "press a single letter for this specific submenu" sub-screen on the right third of the screen, with multiple submenus to navigate to what you're trying to do.

The start menu search is just another instance of someone at MS trying to meet a KPI to drive search traffic to Bing. I believe there's a registry setting where you can turn the Bing search off and it'll go back to how search used to behave in the start menu.

And the ads/'recommended' apps are just another source of revenue from advertising. It's really strange to me. Windows isn't a free operating system, it still costs $140 retail for a Home license. But they sure treat their users like it is.

Sorry, I'm bad at explaining myself, hah. I go on tangents.

My point was that with Windows 10 there usually was a toggle for the changes they make to the taskbar. But for some reason, there doesn't appear to be for this one.

They usually do. I only have experience with Windows 10, though (after the trainwreck that Win10 eventually became, I had to move away from Windows permanently). Weather, Skype and the search bar all had setting checkboxes to remove them and get the taskbar back to what it used to be before Microsoft decided the entirety of Windows was their personal Bing advertisement ground.

(Also, as an ex-Microsoft employee, it's clear to me they only reason they're doing this stuff is to drive more 'organic' search traffic to Bing so they can have better numbers to show advertisers)

It's possible with Windows 11 they simply haven't done the extra effort to make this change toggleable?

Then perhaps when ChatGPT is taken offline it'll remain available as a paid model on the API.