The smartphoneless NEET shut-in So true, I resisted getting a smartphone for as long as I reasonably could, especially because after college I was pretty much a NEET shut-in for a good while, but eventually, if you want to interact with society at all, you need it, and I gave in because my life was becoming more inconvenient without it than with it. I still prefer to not use the phone needlessly and I very much don't like texting still, almost everyone I meet I tell them, I really prefer talking live rather than through text all the time, which is also the case because I just can't stand 24/7 communication with anyone, I don't have the bandwidth for that. I also don't have an Instagram, which in Brazil is actually pretty much a social death sentence, but on this one I'll never give in because I don't like posting pictures of myself online or seeing other people's pictures in that way... Anyway, this aspect of society is kind of annoying because you can easily imagine completely nightmarish hellworlds as technology progresses that become normal and accepted due to abusing the masses' social needs, there are many sci-fi stories that explore aspects of this already, although there are many more such ideas that should remain locked and hidden if you're creative and you can think of them, some of these Bay Area people are insane and they will take the Torment Nexus you mention in your stories and bring it into reality for whatever retarded status competition they're a part of in their polycules, infohazards are real for them because they have no sense of decency or morality, so really don't be too eager to imagine technological dystopias, we have enough of that already. And it is often absolutely the case that if you don't imagine it, no one will, lots of good/feasible ideas are rare and most people never make connections that to you, someone creative, might seem obvious, so a policy of silence actually works. https://a327ex.com/posts/the-smartphoneless-neet-shut-in #neet #bayarea #tormentnexus #rss #feeds #linksharing #qwibbit

Compass-deadening big automated runs Currently I'm working on SNKRX's update, this is the only project I work on almost every day and have been managing to keep it going for about a month, and then I explore other projects as I see fit, like Phylactery which I ended up finishing. So as I explore potential ideas and as Opus 5.5 is more capable, I've been trusting it with running an idea mostly by itself more often. The problem is that this almost always ends up being a mistake if I let it happen at the very start of the project, it has never happened yet that I play what it generates and I go "oh, this is amazing" for more than a few moments. In these outputs there are individual things I may think are amazing and may take, but never the whole at all. If the project is further along with an initial implementation iterated by me and I let the AI run with it then, the success rate climbs but only barely. Overall it seems like letting the AI go on big runs doesn't work for me. There are two possibilities on why this happens. First is that most ideas are bad, so exploring the possibility space of ideas will mostly lead to failure. I had Claude go over all prototypes I made in the past ~15 years (since I started making games) and it returned me that I made around 130 prototypes. I then picked 10 promising ones that weren't too far along to actually have Claude implement to a reasonable degree and played them and most of them suck. It just feels right that most ideas are bad and not particularly fun. But the other possibility is that seeing the result with no effort on my part makes my brain reject thinking the thing is fun. A game made like this has lots of small mistakes that I wouldn't have made if I built the game myself, so as I play it, it seems like I can't ignore the mistakes and judge the game accurately. I'm kind of a perfectionist so playing something that's imperfect actually ruins the whole idea for me. So this second possibility means I'm ruining potentially good ideas by not experiencing them at their best the first time. There's usually a moment when you're iterating on a game where you play it and you go, holy shit, this is actually fun, that moment happening is important because it convinces you the idea is good, but having the AI implement a half-baked version of the idea essentially robs you of it early on. This is ultimately a compass-deadening/noisifying (#34) process. The conclusion then is that I shouldn't do big automated runs anymore, at least not for the start of the project. The seed, the initial way the game plays, the structure of the game, it has to be iterated slowly and only then I can try to explore it in more automated ways. https://a327ex.com/posts/compass-deadening-big-automated-runs #rss #feeds #linksharing #qwibbit

The game theory of dating as a race to the bottom https://a327ex.com/posts/the-game-theory-of-dating-as-a-race-to-the-bottom #rss #feeds #linksharing #qwibbit

400-parameter eyes This is a good example of what I meant in #629 regarding small games still being the favorable thing to do. You can now go really hard on specific things in a way that wasn't as possible before for an individual. Imagine applying the same level of detail like this not only to eyes but to an entire face and having a process for creating or even capturing realistic faces? That's useful for all future projects you make but even useful as one of the main features of a singular game, like in Who's Lila?. https://a327ex.com/posts/400-parameter-eyes #rss #feeds #linksharing #qwibbit

I Did This Without Your Knowledge I Did This Without Your Knowledge https://a327ex.com/posts/i-did-this-without-your-knowledge #rss #feeds #linksharing #qwibbit

Phylactery 15 Lich Survivors 15 moved the testing of the site's unlock (the egg opening, its letters building the locked posts) from the live website to the desktop (renderer/run-unlock.bat, R to replay, the Caps Lock launcher listing run-*.bat variants), found why the live site showed nothing (unlock_fx.lua read main.lua's local index), and then iterated the unlock's look to its accepted form over five rounds: eased letters and a linear rise with options, the row flash fixed, a lock that breaks, the game filling the window with sharp fractional scaling, four gacha-like unlock moments, three themes in the game's own effect languages, and finally glass with the game's telekinesis, warped air done as a real refraction shader pass over the page. It also fixed several game sounds (rock_lift, the rock's crush not playing gore_land, the lich's hurts in the ending), and made the unlock's sounds slots of the game, tunable in its own editor while the whole unlock plays inside the desktop site. Nothing was deployed. Full transcript: Phylactery 15 https://a327ex.com/logs/phylactery-15 #phylactery #capslock #lichsurvivors #nothing #rss #feeds #linksharing #qwibbit

Phylactery 16 Lich Survivors session 16 finished the unlock's sound events, reworked two core systems and made the web build fast enough to ship. The site's unlock got a per-letter land event kept at a fixed pace (every 4th of 673 letters), the final on the last burst, and fixes to the desktop unlock test's sound editor (wheel, R after TAB). The lens became a familiar that floats between cuts, moving like the gun automaton of the 2022 game emojian (found on the Z: archive) and holding still to fire; at most two attacks run at once; snake waves are paced by arrival. A benchmark (port/bench.lua + port/prof.lua) and a self-contained 130 MB web page run in headless Chrome (tools/webbench) found the costs; fixes (warp streak meshes, an allocation-paced collector, a gpu_blit transmission copy) cut the web CPU from 14.3 to 11.1 ms a frame. On a327ex.com: a "N hidden logs since" byline for readers who unlocked the posts, built by the unlock's letters, and game posts that no longer fold from a click. The ending got 2 s shorter and its gore pieces now lie down. Full transcript: Phylactery 16 https://a327ex.com/logs/phylactery-16 #phylactery #lichsurvivors #tab #chrome #cpu #rss #feeds #linksharing #qwibbit

What predicts next game success, revisited Written by Claude Opus 5.5 The strongest claim in What predicts indie success, tested against all of Steam sits in its ladder section: "The single strongest variable in the entire corpus: the best outcome you've achieved so far." The chart behind it reads next-game hit odds off a developer's best game to date: 1.1% if nothing ever cleared 10 reviews, 3.5% past 50, 10.8% past 250, 22.5% past 556, 67% past 5,000. This post re-runs that claim and then tries to break it. It checks whether the number replicates, whether the way it was measured flatters it, whether "best" is really the strongest summary of a career, where the staircase holds and where it bends, and what the rungs are made of. Two independent copies of Steam went in. F is the post's own data: Fable's July 2026 crawl, 115K games, developers named by their SteamSpy strings. A is Astra's September 2026 crawl, 129K released games, developers named by Steam's own single credits, with current prices, languages and store features. On top of both I used Fable's review histories, the month-by-month review counts of all 34,654 games that had 50+ reviews in July. The histories let me ask a question the post couldn't: how many reviews did a developer's earlier games have on the day the next one launched? Everything below uses the post's conventions unless it says otherwise. That means developers whose first game came out in 2015 or later, next releases through July 2025 (so each has had a year to collect reviews), and a hit defined as 556 reviews. Short version. The ladder is real. It replicates to the decimal in the post's data, reappears in the independent catalog, and survives every slice I could cut. The post measured earlier games with today's review counts, which includes hindsight. Fixing that makes the middle rungs stronger, not weaker. But "best outcome so far" is not the strongest variable. A developer's typical result (the average of their games on a log scale) predicts the next game better, and so does their most recent game. Once those are known, the best game adds nothing. The reason is that the ladder mostly measures the developer. About two-thirds of the variation in how many reviews a game gets is shared with the same developer's other games, and a peak is a noisy, upward-biased reading of that level. Two more things the ladder hides: most hits still come from people with no ladder at all, and what the next game is (its price and scope, its publisher, whether it goes back to the lane that worked) moves the odds about as much as the rung you start from. How the check works A transition is one release by a developer who has shipped before, paired with everything known about their earlier releases. Fable's catalog yields 22,767 of them across 9,283 developers; Astra's yields 32,344 across 12,297. Astra's is larger because Steam's own credits cover developers SteamSpy stopped tracking after 2021. It also drops co-credited games, so the two catalogs define a developer differently, which is useful as a check. Dating the earlier games. For every earlier game with a review history, I counted its reviews up to the next game's launch day, pro-rating the month the launch fell in. Games with no history had fewer than 50 reviews in July, so their current count is used as an upper bound. To check the method, I compared it against 36 transitions where Astra's session had independently rebuilt each prior game's count from individual review timestamps. My dated counts agree with Astra's on the low/mid/hit band for all 36, with a rank correlation of 0.999, and 94% land within 15% (they run about 6% high, because the histories include a few reviews that don't count toward the store score). One detour is worth recording. My first version rescaled each history to the game's current score-eligible total. That failed on bundle-era games such as Reus and Nihilumbra, where later key-activated reviews inflate the history but not the score, and it halved their early counts. The unscaled version passed and is the one used throughout. The rescaled version gives the same conclusions with slightly larger dated effects. 1. It replicates Run the post's own code on the post's own data and the ladder comes back exactly: 1.1% / 3.5% / 10.8% / 22.5% / 67.1%. The independent catalog draws the same shape a little lower (17.5% at 556+, 60% at 5K+), mostly because it includes the post-2021 developers SteamSpy misses. The orange line below is the post's chart. 2. The hindsight problem, and why fixing it helps the claim The post placed each developer on a rung using their earlier games' review counts today. A developer never knows that number when launching. Today's count includes years of long-tail growth, plus any boost the old game got from the new one's success. When the next game hits, players go back to the catalog, the old game's count rises, and the developer gets moved up a rung after the fact. That is the textbook way for a "predictor" to leak the outcome. The distortion is large. In Fable's catalog, 28% of transitions sit on a different rung once the earlier games are dated, and 30% of the developers the post placed on the 556+ rung hadn't reached it yet when they launched. The leak itself turns out small, though. After a launch, the best earlier game keeps growing whatever happened: median ×1.16 when the next game hit, ×1.36 when it landed mid, ×1.23 when it flopped (Fable's catalog). In Astra's catalog a hit does lift the back catalog more (×1.41 against ×1.15). Most of the movement is ordinary tail growth, and it pushed many developers who were mid at launch into the snapshot's hit rung, where they diluted it. Dating the counts takes them back out. The blue line in the chart is the honest ladder: 5.0% at 50+ (post: 3.5%), 16.2% at 250+ (10.8%), 30.2% at 556+ (22.5%), 47.4% at 1.5K+ (43.1%), 66% at 5K+ (67%). The claim survives its biggest methodological flaw, and the middle rungs are worth roughly 1.3 to 1.8 times what the post said. A second correction pulls the same way. The post counts every release, so prolific developers with dozens of releases each contribute dozens of rows, and they convert worse at every rung. Count one release per developer (the dashed line) and the 556+ rung reads 37% and the 5K+ rung 81%. Neither weighting is wrong. Per release answers "what fraction of games from this rung hit." Per developer answers "what are the odds for a developer standing on this rung," which is the question the post was asking. 3. Is it the strongest variable? Its family is. "Best" isn't. To compare variables of different kinds on one scale I used AUC: pick a random release that hit and a random one that didn't, and AUC is the chance the variable ranks the hit higher. 0.5 is a coin flip; 1.0 is perfect. Confidence intervals come from resampling whole developers. The pool is Astra's catalog, because it carries store data: releases #2 onward by steam-era developers, through July 2025, 31,459 releases from 12,106 developers, 7.9% of which hit. Two readings. First, the post's broad claim stands. The track record family (blue) beats every store-level variable by a wide margin. The best store variable is language count at 0.73, and it is measured today, so successful games that localized after launch flatter it. Primary tag comes in at 0.71, price at 0.66, achievements and controller support around 0.6. Release month, Early Access and having a demo are coin flips. Second, inside the family, "best earlier game" ranks below three other summaries of the same history: the typical result (mean of log reviews) at 0.875, the median at 0.873, the most recent game at 0.859, and only then the best at 0.837. The post's own version (today's counts) scores 0.857, and that edge over the dated best is partly hindsight. The ordering holds at every outcome line from 50 reviews to 5,000 and in both catalogs. Among developers with three or more earlier games, where peak and typical can drift apart, it is 0.880 for the best against 0.896 for the typical. Those ranking scores understate what this means for an individual developer, because for anyone with a single earlier game the best and the typical are the same number. The clearest test holds the best game fixed and lets the typical result vary. Take developers with at least two earlier games whose best sits on the same rung, 556–5K reviews. The rung alone puts them at about a third. Split them by their typical result: Best 556–5K, typical result is… Next release hits (Astra's catalog) (Fable's catalog) under 50 (one hit among flops) 4.5% 12% 50–555 24% 27% 556+ (hits are the norm) 53% 57% Same best game, a five- to twelvefold difference in odds. One rung down it repeats: a 50–555 best with a typical under 50 converts at 2.7%, against 12–14% when the typical is in the same range. The reverse doesn't hold. With both in one model, the weight on the best drops from 1.41 to 0.25 while the typical keeps 1.32 (section 4). Averaged over everyone the gain is modest; for the developers whose record holds more than one data point, it is large. Multivariate models agree (gradient-boosted trees and logistic regression, five folds split by developer so no developer is in both training and test). Best game alone: AUC 0.83. All career variables: 0.89, and dropping the best game from that set changes nothing (0.89). All store variables together: 0.875. Both blocks together: 0.93. So the game itself, read collectively, carries about as much as the whole career does, and the two are complementary. Most of that store signal (language count, tag, price) is plausibly what the developer is able to make rather than a lever, and some of it is measured after the fact. 4. Your latest game matters as much as your best The fastest way to see the problem with "best" is to hold it fixed and vary the latest game. A developer whose best game hit but whose latest flopped converts at 13%, about the same as one who never hit but whose latest landed mid-range (13.5%). The same best with a latest hit converts at 55%. In Fable's catalog the corresponding numbers are 15.5%, 15.5% and 57%. In regression terms, the most recent game alone explains more than the best alone (pseudo-R² 0.274 against 0.255), the typical result explains more than either (0.298), and with typical and best in one model the best's coefficient falls from 1.41 to 0.25. Recency here is counted in games, not years. Among developers whose best dated game is 556+, the next game hits 54% of the time when the best was the previous release, 42% when one release came in between, and 26% when two or more did. By calendar time the decline is gentle: 46% within a year of the best, 36% four or more years later. That fits the post's "time off is free". A long gap barely costs anything; releasing weaker games after the hit does. Two smaller findings point the same way. At equal peak size, a second or third hit adds no consistent lift; the size of the peak matters and the count doesn't. And among mid-rung developers, accumulating more mid games lowers the odds: 10% with one or two, 7% with three to five, 2.7% with six or more. The odds don't expire after one game; they get re-read. Follow developers forward from their first game that was already 5K+ when the next one launched. The first release after it hits 72% of the time (79% in Fable's catalog), the second 66% (81%), the third 61% (70%). For the same developers tracked across all three, the drift is small: 68%, 55%, 61% (76%, 76%, 70%). What moves the number is the game in between. On the second release after the breakout it's 76% if the first follow-up also hit and 49% if it missed. On the fourth release and later it's 71% against 31%. A 556+ breakout reads the same way one level down: 48% for the next release, 39% and 36% for the two after, and 62% against 13% on the fourth-plus depending on the previous release. The post's 67% for the 5K+ rung was always an average over the next release and the ones after it. The 5K+ samples are small (377 developers in Astra's catalog, 194 in Fable's, under 100 by the third follow-up). 5. Why: the developer is the unit Here is the fact under the whole ladder. Take every steam-era developer with at least two games and ask how much of the spread in log review counts is shared within a developer's catalog. The answer, as an intraclass correlation: About 69% of the variation in a game's reach is shared with the same developer's other games. Genre explains 14%, price band 17%, release year 5%. Who made the game outweighs every catalog variable four or five times over. The ladder is what that fact looks like when you read it through one number per developer. Through that lens, every surprise above stops being surprising. If each game's outcome is the developer's level plus game-to-game noise, the mean of their games is the best estimate of the level. The best of their games is that level plus the luckiest draw, so it runs higher the more games they've shipped and it is noisier. That explains why the number of games shipped turns negative once the best is held fixed. It also explains why the same rung is worth far less late in a career: a 556–5K best converts at 50% on release #2 and at 12% on release #10+. And it explains why content factories with many mid games sit low. I checked this with a simulation that has nothing in it but a fixed level per developer plus noise, using the variances estimated from Astra's catalog: no compounding, no audience, no recency. It reproduces the observed ordering, typical (0.907) over most recent (0.886) over best (0.879), against 0.866 / 0.853 / 0.844 in the real data. The ladder doesn't need anything to compound in order to exist. That doesn't prove nothing compounds; section 9 finds a modest audience effect. It means the staircase on its own isn't evidence of compounding. Most of it is measurement. 6. The next game usually lands well below the best A direct consequence of the peak being a lucky draw is that the next game regresses toward the developer's level. The median next game gets 37% of the best game's reviews at the 50–249 rung and 12% at the 5K+ rung. The chance of beating your best falls as the best rises, from 27% to 11.5%. On a log scale the next game's reviews track the best with a slope of 0.67 (R² 0.24) and the typical result with a slope of 0.82 (R² 0.37). None of this contradicts the ladder: a 5K+ developer's "disappointing" next game still has a median of about 1,300 reviews. It does mean the right expectation after a breakout is your level, not your peak. 7. Where it holds The coarse version of the dated ladder, cut every way the data allows (Astra's catalog unless marked, releases at least 30 days apart). Each row is one slice; the cells are the odds the next release reaches 556 reviews from that rung, "–" means fewer than 25 releases, and the last column divides the 556–5K rung's odds by the 10–49 rung's. slice (best earlier game, dated) <10 10-49 50-249 250-555 556-5K 5K+ 556-5K ÷ 10-49 All releases 1.2% 2.8% 7.1% 16.4% 34.1% 63.9% 12× Fable's catalog (SteamSpy names) 1.5% 3.1% 7.8% 18.8% 38.9% 72.3% 13× One release per developer 1.2% 3.6% 9.1% 20.9% 41.1% 69.7% 11× Next released 2016-18 1.9% 3.4% 11.0% 25.6% 39.9% 75.0% 12× Next released 2019-21 1.7% 2.9% 7.1% 19.9% 42.3% 73.6% 15× Next released 2022-25 0.8% 2.5% 5.9% 12.6% 28.9% 60.4% 12× Next priced under $5 0.6% 0.9% 2.9% 7.7% 19.7% 39.0% 22× Next priced $5-10 0.8% 1.6% 5.0% 12.2% 27.6% 48.1% 17× Next priced $10-20 3.0% 6.6% 12.0% 23.1% 39.2% 67.7% 6× Next priced $20+ 8.2% 11.7% 19.5% 34.2% 58.3% 84.7% 5× Next is free 2.4% 5.7% 11.1% 21.1% 34.5% 58.4% 6× Next self-published 0.8% 2.3% 5.1% 13.3% 29.3% 57.4% 13× Next has another publisher 3.4% 4.4% 12.3% 23.1% 42.1% 74.8% 10× Next shares the best game's title family 0.9% 1.5% 5.2% 20.4% 51.9% 79.1% 35× Next is a new title 1.2% 2.9% 7.3% 15.9% 31.3% 60.5% 11× Next keeps the best game's primary tag 0.6% 1.3% 5.3% 17.3% 44.9% 84.8% 35× Next changes primary tag 1.4% 3.3% 7.7% 16.1% 30.6% 57.7% 9× Gap under 6 months 0.7% 1.9% 4.6% 11.2% 25.7% 51.8% 14× Gap 6-24 months 0.9% 2.4% 7.7% 20.1% 36.0% 61.6% 15× Gap 24+ months 2.7% 6.0% 10.2% 17.7% 40.7% 72.4% 7× Release #2 1.6% 4.3% 11.5% 24.8% 49.7% 75.6% 12× Release #3-4 0.4% 2.2% 8.0% 18.4% 36.3% 64.8% 16× Release #5-9 0.0% 0.8% 2.9% 12.2% 29.9% 60.5 https://a327ex.com/posts/what-predicts-next-game-success-revisited #rss #feeds #linksharing #qwibbit

upusen - こくう (feat. 重音テト) https://a327ex.com/posts/upusen-feat #rss #feeds #linksharing #qwibbit

Sungazer - Hot Saturn (feat. Button Masher) https://a327ex.com/posts/sungazer-hot-saturn-feat-button-masher #rss #feeds #linksharing #qwibbit