I'm working on a thing to assert that you're a human or that you've got human content. It's got some block-chain and crypto things in it. It's like proof of ID but still anonymous.
I got around to adding the mortgage tracking functionality to my partner's personal finance dashboard; it's working? I think? I'll have to ask him to test thoroughly tomorrow to see if the maths are right.
That is the last feature in his original design, so until he can test, I'm free to do whatever?
I downloaded some old Yu-Gi-Oh! games (the Power of Chaos series) and I wanted to get into modding them, mostly adding a cap to the rewards so you can't earn more than a playset of the same card, and also maybe expand the deck pool for the AI opponents.
That said, I'm also part of Timasomo and that's going much worse than planned so maybe I should focus on it...
I found myself fascinated by a math idea so used programming and some basic graphics to explore it: "square squares" are perfect square numbers whose digits are also perfect squares.
I'm still puttering on a game I've wanted to play since the old days of NetTrek.
It's a LAN multiplayer space combat game, with each ship composed of a crew of several players with a shared "main viewscreen" per shop. I know there have been a couple of these over the years, but mine distinguishes itself by requiring only a modern browser on a phone, tablet, or laptop to join and play.
I've got a lot of work in on the engine, wire protocol, graphics, game dynamics, error recovery, and shaping the learning curve.
Early playtests yielded a surprise - teams tended to get really attached to their ship, and the effect got stronger when I made it possible to tune the VFX for weapons and shields (even though the tuning had no impact on the balance of play).
Where I am now is focused on learning 3D modelling and texturing; I want to make a collection of coherent design systems for ships to appeal to different fandoms. Nacelles-and-beams (Trek-like), greebles-and-blasters (Star Wars-like), and frak-cannons (Battlestar-like) are where I'm starting.
I'm torn about whether it's worth it for me to learn, or just get help building the limited number of models I'll need.
Danger There be talk of 'cool thing" with an LLMs ahead but not written by a toaster. =)
I've been working on a passion project (now small business) coming up on 6 years (guess I can't call it a passion project any more?) making IT management software geared towards MSPs (think settings management, policy validation, config backups, etc across network/servers/workstations) and getting it all packaged on Azure with a pretty interface.
For context I've built out a set of systems to capture and structure device state/config, strip secrets, etc then exposed that info in the UI. Like everyone else, I've been evaluating how/if LLMs have a place in the stack so I created a CLI tool to read from my software which I can give to agents (this abstraction has helped wrangle the LLM to work more rigidly and control access). With everything being structured data, that cuts down on the context of an LLM to see things like "what's installed", "how's it configured", "what ospf settings are in place and what are your neighbors", "what are you plugged into?" etc. Using the CLI just returns narrow json of what's being queried through the CLI.
That's made planning and labbing out changes MUCH easier (esp when used in conjunction with something like container lab). I can now use the CLI/LLM combo to query complex state for things like "show me Fortigate firewall policies at all sites involved with any windows servers running xyz software but only the ones with 10g bonded links" and an agent will use the cli tool to query only the narrow sections of config/state required to answer the question.
Now I've taken it a step further, directing an LLM to review an IT project's requirements and write manifests of changes against the framework of my system change tool and tests against that structured data and feed it back into the system (via that cli tool) which presents a rigid change proposal in the UI. I can then review work against real equipment but with the safety of eyes on the proposed work approving/editing/rejecting changes from the UI with mandatory approvals for any changes. Click the approve button and runners are dispatched to make the changes/run the tests/handle rollbacks.
Needless to say, I eventually got approval for a POC on a real and large client environment to execute a project in this fashion.... and it worked spectacularly well!! Come maintenance windows the process for this particular pilot had become "click the go button and move cables from here to there as told by the UI". I was also able to have cable labels automatically printed to the guys onsite as they moved cables.
I've liked this abstraction of putting software/controls between real equipment and the LLM so it can't directly touch anything. I feel like there's an emerging market to make industry specific tools like this; "Human > LLM > Tool/abstraction/controls > Real thing" where a human can be in the middle to review/approve/audit/etc.
Anyway, I've been enjoying exploring and learning these new (to me) abstractions and wanted to share what I've spending all my time on.
The biggest trouble I've had is wrangling LLMs to be be brief in their verbiage. Change manifests have worked well so far but the amount of exhaustive text they spit out is borderline unusable to a human (which I fear promotes clicking "next, next, next"). What's the answer? A post production agent pass?
Is anyone doing something similar? Has it worked well for you?
I am doing... almost exactly the same thing except my focus is 2 toolkits. One for Azure Infrastructure and one for conditional access. Other than that the similarities are uncanny...
Even with "improve this ai model" features disabled, I wouldn't be surprised if somehow, we're inadvertently cross training these models in some way while trying to fix bugs and iterate on designs within our own systems. I've primarily focused on servicing on prem equipment but I've had a lot of folks ask me "hey do this for cloud". Rad that you've got that working!
Rather than writing for the tiramisu thing like I said I would, I've been unable to peel myself away from working on tilegroxy still. I overhauled the documentation, using a more heavily customized template and even adding in a landing page. It's looking much more pretty than it was before. I'm most happy with the callouts on the longer configuration examples, since the service is entirely configuration driven it's really useful to have properly annotated configs.
With copious use of AI agents I now have functional flight, bounce, skid and roll physics that is plausibly realistic, based on published golf physics papers, golf equipment manufacturers' information and finetuning by hand. I have just returned from an afternoon session of miniature golf, but I have never played actual golf, only several different golf videogames over the years, so I've been watching golf youtube instead of TV shows to ensure that I have the right intuition.
The weirdest thing I have involuntarily learned so far is that Donald Trump, even if we account for possible cheating, is genuinely very good at golf for his age, and not only that - in places he almost looks like he's able to enjoy it in a normal, human way, to get actual enjoyment out of something that does not involve bragging or putting somebody else down. That was unexpected.
The game is moving forward well, the most difficult part and the only truly innovative part, which is the course authoring tools that most players will never see, is moving along swimmingly so far, whereas graphics has been a bit of a pain. Even though I'm mostly trying to match the vibe of older videogames like Links LS 2000, which is technologically very basic, getting it actually right to look as comfy and crisp as those old games is a challenge. I wanted to write "especially since I don't want to use generative AI", which I am using here and there for placeholders, but to be honest it doesn't even work well for this graphics style, so it doesn't really matter what I want or don't want.
I only hope that @KilledByAPixel is going in a different direction with his excellent tiny golf game because even though I've had this idea in my mind for a while, it is true that I only started actually working on it after I played his game and realized how fun even a very simple golf game can be. And of course there are clear similarities in my game because I assume we're both using inspiration from the few golf game de facto standards that evolved over the decades. So I feel a bit like a backstabbing golf developer. But I assume our games will turn out different enough.
Would you be willing to share your AI workflow? I'm interested in understanding how far you've been able to take it, as well as if you're using a paid LLM or a local one.
I use paid LLMs, but I believe we're at a point where it's not necessary to use the most cutting edge models. I could have likely done it with Deepseek v4.1 Flash, with a 10 USD OpenCode Go subscription. I have tested it and it seemed to work as good as the lowest tier ChatGPT Codex model (Luna), which, when Codex is not overwhelmed by traffic and their models are not temporarily lobotomized (happens regularly some time after they release a new model), is honestly pretty good. However right now I use either Claude 5.5 Sonnet (the middle tier from the "standard" models) or Codex GPT 6.1 Sol (the upper tier from the "standard" models) and it is better at solving problems with more freedom, diagnosing issues and making the right assumptions. It still does things wrong.
As for the workflow, basically I use normal prompts fed to one agent, no orchestration or automation. At the beginning talked to ChatGPT about the project structure. Told it what my goals are and why, which technical decisions are clear for me and which are not (and why) and told it to create some sort of workflow that would work with the cheaper models and be resilient. This meant creating short and long term docs for everything that the agents could read at any time and that would document both the current state of development and the goals and the whys. Having the "why" documented for every important decision helps the agents make the right assumptions.
We also created some milestones and steps to get to them, which worked until I had something playable, at which point the priorities change based on what issues you see.
This worked and surprisingly the biggest issue was that ChatGPT tended to overspecify problems and create a 10 pages long prompt for Codex that had it read a ton of docs on top of that, which filled its context with unnecessary information and made it slow and inefficient. So now I use relatively short prompts and rely on the agent actually reading the docs, which it always does so far. It still sometimes makes wrong decisions, generalizes a solution that I don't need generalized, and creates ad-hoc temporary solutions for things I do need generalized.
The most common issue is that I need it to create tests for everything, which works reasonably well, but it doesn't care about test performance. So at one point I realize I have 23 tests which all use level data compiled from an editable format to a format loaded in the actual game, and instead of using a minimal usable test level, compiling it once and handing the object to all the tests that do not test the compilation specifically, all of those tests are compiling an actual existing level, which took several seconds each. Since you want to run the full test suite at least once per task, this eats time. Always relatively easy to solve though, it's just something that keeps happening.
edoceo | a day ago
I'm working on a thing to assert that you're a human or that you've got human content. It's got some block-chain and crypto things in it. It's like proof of ID but still anonymous.
Chiasmic | 23 hours ago
Would you be willing to expand more? Sounds interesting
arqalite | a day ago
I got around to adding the mortgage tracking functionality to my partner's personal finance dashboard; it's working? I think? I'll have to ask him to test thoroughly tomorrow to see if the maths are right.
That is the last feature in his original design, so until he can test, I'm free to do whatever?
I downloaded some old Yu-Gi-Oh! games (the Power of Chaos series) and I wanted to get into modding them, mostly adding a cap to the rewards so you can't earn more than a playset of the same card, and also maybe expand the deck pool for the AI opponents.
That said, I'm also part of Timasomo and that's going much worse than planned so maybe I should focus on it...
akkartik | 18 hours ago
I found myself fascinated by a math idea so used programming and some basic graphics to explore it: "square squares" are perfect square numbers whose digits are also perfect squares.
https://merveilles.town/@akkartik/117380846030422227
Handshape | 13 hours ago
I'm still puttering on a game I've wanted to play since the old days of NetTrek.
It's a LAN multiplayer space combat game, with each ship composed of a crew of several players with a shared "main viewscreen" per shop. I know there have been a couple of these over the years, but mine distinguishes itself by requiring only a modern browser on a phone, tablet, or laptop to join and play.
I've got a lot of work in on the engine, wire protocol, graphics, game dynamics, error recovery, and shaping the learning curve.
Early playtests yielded a surprise - teams tended to get really attached to their ship, and the effect got stronger when I made it possible to tune the VFX for weapons and shields (even though the tuning had no impact on the balance of play).
Where I am now is focused on learning 3D modelling and texturing; I want to make a collection of coherent design systems for ships to appeal to different fandoms. Nacelles-and-beams (Trek-like), greebles-and-blasters (Star Wars-like), and frak-cannons (Battlestar-like) are where I'm starting.
I'm torn about whether it's worth it for me to learn, or just get help building the limited number of models I'll need.
DVNO42 | a day ago
Danger There be talk of 'cool thing" with an LLMs ahead but not written by a toaster. =)
I've been working on a passion project (now small business) coming up on 6 years (guess I can't call it a passion project any more?) making IT management software geared towards MSPs (think settings management, policy validation, config backups, etc across network/servers/workstations) and getting it all packaged on Azure with a pretty interface.
For context I've built out a set of systems to capture and structure device state/config, strip secrets, etc then exposed that info in the UI. Like everyone else, I've been evaluating how/if LLMs have a place in the stack so I created a CLI tool to read from my software which I can give to agents (this abstraction has helped wrangle the LLM to work more rigidly and control access). With everything being structured data, that cuts down on the context of an LLM to see things like "what's installed", "how's it configured", "what ospf settings are in place and what are your neighbors", "what are you plugged into?" etc. Using the CLI just returns narrow json of what's being queried through the CLI.
That's made planning and labbing out changes MUCH easier (esp when used in conjunction with something like container lab). I can now use the CLI/LLM combo to query complex state for things like "show me Fortigate firewall policies at all sites involved with any windows servers running xyz software but only the ones with 10g bonded links" and an agent will use the cli tool to query only the narrow sections of config/state required to answer the question.
Now I've taken it a step further, directing an LLM to review an IT project's requirements and write manifests of changes against the framework of my system change tool and tests against that structured data and feed it back into the system (via that cli tool) which presents a rigid change proposal in the UI. I can then review work against real equipment but with the safety of eyes on the proposed work approving/editing/rejecting changes from the UI with mandatory approvals for any changes. Click the approve button and runners are dispatched to make the changes/run the tests/handle rollbacks.
Needless to say, I eventually got approval for a POC on a real and large client environment to execute a project in this fashion.... and it worked spectacularly well!! Come maintenance windows the process for this particular pilot had become "click the go button and move cables from here to there as told by the UI". I was also able to have cable labels automatically printed to the guys onsite as they moved cables.
I've liked this abstraction of putting software/controls between real equipment and the LLM so it can't directly touch anything. I feel like there's an emerging market to make industry specific tools like this; "Human > LLM > Tool/abstraction/controls > Real thing" where a human can be in the middle to review/approve/audit/etc.
Anyway, I've been enjoying exploring and learning these new (to me) abstractions and wanted to share what I've spending all my time on.
The biggest trouble I've had is wrangling LLMs to be be brief in their verbiage. Change manifests have worked well so far but the amount of exhaustive text they spit out is borderline unusable to a human (which I fear promotes clicking "next, next, next"). What's the answer? A post production agent pass?
Is anyone doing something similar? Has it worked well for you?
polle | a day ago
I am doing... almost exactly the same thing except my focus is 2 toolkits. One for Azure Infrastructure and one for conditional access. Other than that the similarities are uncanny...
DVNO42 | a day ago
Even with "improve this ai model" features disabled, I wouldn't be surprised if somehow, we're inadvertently cross training these models in some way while trying to fix bugs and iterate on designs within our own systems. I've primarily focused on servicing on prem equipment but I've had a lot of folks ask me "hey do this for cloud". Rad that you've got that working!
1338 | 19 hours ago
Rather than writing for the tiramisu thing like I said I would, I've been unable to peel myself away from working on tilegroxy still. I overhauled the documentation, using a more heavily customized template and even adding in a landing page. It's looking much more pretty than it was before. I'm most happy with the callouts on the longer configuration examples, since the service is entirely configuration driven it's really useful to have properly annotated configs.
Here's the documentation website: https://tilegroxy.com
If anyone has feedback to offer, I'm curious.
V17 | a day ago
A month or so ago I jokingly said I was going to build my own golf game with blackjack and hookers.
Blackjack and hookers are still tbd, but after some contemplation I have decided to actually start working on it.
With copious use of AI agents I now have functional flight, bounce, skid and roll physics that is plausibly realistic, based on published golf physics papers, golf equipment manufacturers' information and finetuning by hand. I have just returned from an afternoon session of miniature golf, but I have never played actual golf, only several different golf videogames over the years, so I've been watching golf youtube instead of TV shows to ensure that I have the right intuition.
The weirdest thing I have involuntarily learned so far is that Donald Trump, even if we account for possible cheating, is genuinely very good at golf for his age, and not only that - in places he almost looks like he's able to enjoy it in a normal, human way, to get actual enjoyment out of something that does not involve bragging or putting somebody else down. That was unexpected.
The game is moving forward well, the most difficult part and the only truly innovative part, which is the course authoring tools that most players will never see, is moving along swimmingly so far, whereas graphics has been a bit of a pain. Even though I'm mostly trying to match the vibe of older videogames like Links LS 2000, which is technologically very basic, getting it actually right to look as comfy and crisp as those old games is a challenge. I wanted to write "especially since I don't want to use generative AI", which I am using here and there for placeholders, but to be honest it doesn't even work well for this graphics style, so it doesn't really matter what I want or don't want.
I only hope that @KilledByAPixel is going in a different direction with his excellent tiny golf game because even though I've had this idea in my mind for a while, it is true that I only started actually working on it after I played his game and realized how fun even a very simple golf game can be. And of course there are clear similarities in my game because I assume we're both using inspiration from the few golf game de facto standards that evolved over the decades. So I feel a bit like a backstabbing golf developer. But I assume our games will turn out different enough.
ShroudedScribe | a day ago
Would you be willing to share your AI workflow? I'm interested in understanding how far you've been able to take it, as well as if you're using a paid LLM or a local one.
V17 | 23 hours ago
I use paid LLMs, but I believe we're at a point where it's not necessary to use the most cutting edge models. I could have likely done it with Deepseek v4.1 Flash, with a 10 USD OpenCode Go subscription. I have tested it and it seemed to work as good as the lowest tier ChatGPT Codex model (Luna), which, when Codex is not overwhelmed by traffic and their models are not temporarily lobotomized (happens regularly some time after they release a new model), is honestly pretty good. However right now I use either Claude 5.5 Sonnet (the middle tier from the "standard" models) or Codex GPT 6.1 Sol (the upper tier from the "standard" models) and it is better at solving problems with more freedom, diagnosing issues and making the right assumptions. It still does things wrong.
As for the workflow, basically I use normal prompts fed to one agent, no orchestration or automation. At the beginning talked to ChatGPT about the project structure. Told it what my goals are and why, which technical decisions are clear for me and which are not (and why) and told it to create some sort of workflow that would work with the cheaper models and be resilient. This meant creating short and long term docs for everything that the agents could read at any time and that would document both the current state of development and the goals and the whys. Having the "why" documented for every important decision helps the agents make the right assumptions.
We also created some milestones and steps to get to them, which worked until I had something playable, at which point the priorities change based on what issues you see.
This worked and surprisingly the biggest issue was that ChatGPT tended to overspecify problems and create a 10 pages long prompt for Codex that had it read a ton of docs on top of that, which filled its context with unnecessary information and made it slow and inefficient. So now I use relatively short prompts and rely on the agent actually reading the docs, which it always does so far. It still sometimes makes wrong decisions, generalizes a solution that I don't need generalized, and creates ad-hoc temporary solutions for things I do need generalized.
The most common issue is that I need it to create tests for everything, which works reasonably well, but it doesn't care about test performance. So at one point I realize I have 23 tests which all use level data compiled from an editable format to a format loaded in the actual game, and instead of using a minimal usable test level, compiling it once and handing the object to all the tests that do not test the compilation specifically, all of those tests are compiling an actual existing level, which took several seconds each. Since you want to run the full test suite at least once per task, this eats time. Always relatively easy to solve though, it's just something that keeps happening.