1Broom

Building games, software & music with AI.
Trucker by trade.
Creator by nature.
Big ideas.
Probably out of token usage.
Why Broom? well, its my last name.

[1Broom @The1Broom]

https://t.co/MS3kXtgcFR is my shipyard.
games you can play in your browser.
software i built for people who actually need it.
music that somehow picked up 8 million streams.
all of it made between hauls. new things when i get usage back.

[1Broom @The1Broom]

@ShenSeanChen @typesafeai @Muse Cheeky docs are fun until the cheeky line is the one that decides whether you can put it near private data. That belongs on a policy page, not in a bit.

[1Broom @The1Broom]

@TRA4669 Fog is the cheapest way to make a light feel like it has weight and the most expensive way to lose track of what the player can actually see. How are you handling the moment the flare dies?

[1Broom @The1Broom]

@hraness Everyone did it in 2023. Nobody shipped it at a price where you stopped rationing calls. That is the whole gap, and it is not one a slide deck closes.

[1Broom @The1Broom]

@nizamdesign Looks calm, which is usually what people mean by seamless. The risk at this much evenness is that nothing tells me where to look first. What is the one action you want someone to take on this screen?

[1Broom @The1Broom]

@adeoressi True for software, said as though it is true for everyone. The freight broker moving your load does not need an MCP server. He needs a phone that works and a truck that shows up.

[1Broom @The1Broom]

@KissonL Agreement is not evidence is the line I needed a year ago. The only reason I trust my own judging layer is that it disagreed with me. I re-judged after a fix and one row went the wrong way, 0.58 down to 0.42. A grader that only ever confirms you is decoration.

[1Broom @The1Broom]

@tarabario A grappling hook turns your level geometry into a playground you never designed, which is why it feels so good and why it usually breaks something. Did it wreck any of your existing levels?

[1Broom @The1Broom]

@Zefan_Cai The no weights, no self hosting caveat was the only real one I had, so this matters. What I want to know is whether your confidence stays calibrated. I branch on that number at a fixed floor, so ranking alone would not be enough.

[1Broom @The1Broom]

@EGafni Decomposing is half of it. The other half is naming who is judging. I went from is this okay to would an experienced keeper call this a normal day for this species. Same rubric shape, and it started disagreeing with me in places that mattered.

[1Broom @The1Broom]

@seflless Strokes is the right call, it keeps the ordering a bitmap throws away. Does resampling them change the answer? Mine went wrong exactly there. Too few samples along a centreline and a perfectly straight spine read as a fault.

[1Broom @The1Broom]

@SearchDataEng That you had to explain this was a joke is the most honest benchmark anyone has run this week. I actually use the thing and I still read the first line straight.

[1Broom @The1Broom]

@drummatick 100X bigger but zero index. No embeddings to rebuild, nothing drifting out of sync, you pay per row when you ask. Bad deal at scale, very good deal for everything nobody was going to index anyway.

[1Broom @The1Broom]

@omarish This is the kind of tool you only build after getting burned. I judge captured sim runs at roughly 940 tokens a request and had no idea where that number was actually going until I counted. Seeing the split would have saved me a week of guessing.

[1Broom @The1Broom]

@ishuagra02 Built a Jev extension myself, so the bit I am curious about is cross page. Single page is a scoping problem, cross page is a ranking problem. What stops it answering confidently off the wrong tab?

[1Broom @The1Broom]

@BratDotAI I stopped trying to fix this with a cheaper model and fixed it by sending less. Every decision in my loop goes somewhere narrow and cheap. The big model only runs when prose actually has to come out the other end.

[1Broom @The1Broom]

@RangeshUs The spinning hammers are doing the work here. What I would tune next is whether they transfer momentum or just block, because a hit that flings you reads completely different from one that stops you dead. My stacking game lived or died on that distinction.

[1Broom @The1Broom]

@moglrd Day 93 is the achievement, not the game. My folder is full of things that died around day four. Does it still take as long as it did in week one, or have you got fast at it?

[1Broom @The1Broom]

@JohnnyB_222 The readback is the whole thing. If the sim only lives on the GPU then nothing in your game logic can react to it. I am stuck the other direction, physics in a browser on a phone, so my ceiling is nowhere near 24,000. What is the readback costing you per frame?

[1Broom @The1Broom]

@HeyAliux I keep reptiles and I built a reptile sim. I am the user, which is the only reason it survived a month of behaviour code I hated. If I were not, it would be in the folder with the others.

[1Broom @The1Broom]

@HixonStudio @somehowali Given it cannot write, I assume the ghost is picking what to say rather than composing it. That constraint made my animals better, not worse. A fixed set of things it can want reads more consistent than something free to improvise.