Or more precisely, “cheesecake by deathmatch”. Hey, I’ve got the silly thing, might as well use it to speed up processing my dusty pile of saved girlie-pics. While playing with “AI” tools, I downloaded Windsurf, ordered it to clean up the code for maintainability and then add a new function to export the currently-visible list to the clipboard (press L). Now I can quickly rank the images, use a flag to split them into NSFW and less-NSFW, and then toggle the flag to get two distinct export lists to pipe to my cheeseblogging prep scripts.

(Windsurf took me so seriously I ended up with a multi-file distribution package, but I had no difficulty reassembling it into a single file for simple downloading)

But first, a message from our sponsor, Qwen Image:

(the hardest part of this was getting it to render a “standard” onigiri rice ball; first it wrapped the entire thing in fresh green seaweed, then it made it huge, then it added additional wrapping on top of the nori strip, etc)

This is a selection from “stuff I downloaded in April, 2019”.

more...

No new anime, words that I’ll be saying for the next three months, so it’s time for more “AI” cheesecake. Maybe I’ll dig into my real cheesecake archives for contrast as well.

Today’s randomized babes are based on SFnal settings and costumes provided by Claude AI. It had no difficulty grasping the concepts of “vivid, exotic sci-fi locations” and “sexy retro-futuristic costumes for women”, generating detailed descriptions that Qwen Image was able to run with.

More importantly, it didn’t fall down on the unique part. It didn’t take forever to generate 100 results, and they were truly distinct, not just slight wording variations. Most importantly, it did not scold me for requesting “sexy”, or refuse to do my bidding. Although I haven’t asked for lingerie sets yet…

Qwen has recently released updated versions of some of their models, so I hope they get around to revving Image soon. I can cope with the usual finger-counting and the giantism, but I really hate seeing a great image ruined by a missing leg or wrong-side foot, something that’s really common.

more...

Okay, now that I have a (half-dozen implementations of a…) well-tested image-ranking Python script, how do you use it?

1. Download deathmatch.py

2. Install Flask:
   pip install Flask

3. Run the script with a single argument, the name of a directory full of PNG and/or JPG images:
   python deathmatch.py /Volume/galparty/fresh

   and open a browser window to http://127.0.0.1:5000, You should see a status bar at top, the first image in the directory by modification time, a help button in the upper right, and two rows of buttons at the bottom.

4. The first row of buttons are rank. All images start at 0, and can either be promoted by pressing 1-5 or rejected by pressing X. This will update the matching buttons at the bottom and advance to the next image.

5. The second row of buttons are flags. One or more can be selected by pressing A-F; this will not advance to the next image, so you can flag a picture before ranking it. Flags mean whatever you want them to, and are there so you can easily mark images that are interesting for reasons other than quality (“family”, “funny”, “disturbingly similar to that girl from Frozen”, etc). The far-right button matches images that do not have any flags set.

6. Clicking on any of the buttons excludes matching images from the display.

(pretty sure there’s no giant office complex in Red Rock Canyon…)

Ranking and flags are stored in the same directory in a file named _rank.txt. The fields are separated with tabs. This file is updated every time you change a rank or flag, so your progress is automatically saved.

Running a deathmatch

1. Run the program in a directory full of images and click each of the following buttons: X, 1, 2, 3, 4, and 5, filtering the list to show only rank 0 images (all of them the first time).

2. Starting with the first image, press either X or 1. Make this a very quick pass that just takes out the trash. Out of focus, picture of lens cap, shakycam, AI gal with three arms, etc. When you’re done, you’ll have no images left at rank 0.

3. Click the 0 and 1 buttons, so that only rank 1 images are displayed. Make a pass across all rank 1 images, promoting exactly one-third to 2. The easiest way to do this is just have your fingers on the 1 and 2 keys rather than using right-arrow to skip the 1’s. Be brutal; you’re picking your best. If a picture is interesting, flag it with A-F.

4. (Optional: walk away for an hour or a day, and come back fresh) Adjust your filters and make a pass across all rank 2 images, promoting exactly one-third to 3.

5. Rinse and repeat two more times, until 1/81 (1.2%) of your original rank 1 images are ranked 5.

Since this isn’t a full photo-management app, there’s no convenient way to package up your selected pics for printing or editing, but you can copy-paste the filename from the top-left corner, or make a copy of the _rank.txt file and edit it down to just the ones you want.

[Update: I cleaned up the code and added a simple export: press L to copy the current visible list to the clipboard, one per line]

There are five six additional commands available:

• Up and Down-arrow, to take you to the beginning/end of the list
• Z for zoom, to see images at full size with scrollbars
• R for reset, to turn off all active filters
• L to copy the list of visible images to the clipboard
• SPACEBAR to reload the image list from the directory, in case you’ve got an AI actively generating images in the background, as one does.

Unrelated,

The new Dr Seuss LoRA for Qwen Image doesn’t really work for illustrating Red Sonja.

I fed it the spec, and it created a multi-file project, with a separate HTML template file, a cleanly-formatted README.md containing a quite reasonable summary of the purpose and function of the program, a requirements.txt file containing the dependencies, and a Python program that worked on the first try.

In testing, I identified four issues: two of them were arguably ambiguities in the spec, the other two were related to correctly displaying images larger than the window. After checking in the initial version, I went through the “approve changes” dialogs, where it showed me each change in a clear, clean diff. The resulting code passed all my tests.

For the third pass, I told it to package the whole thing up for distribution with the Python Poetry toolkit. It did. I’m done.

I did not allow $LICENSED_TOOL to run commands, even git checkins (which other IDEs can do without handing over the reins to an AI), because I’m not stupid. Still, the experience was so much better that I’d consider paying the $15/month fee if I had a lot of projects just lying around waiting to be written.

This is my shocked face; there are many like it, but this is mine:

(I’m not nearly this fat, and I haven’t had a sugared Pepsi in over a decade; the rest is 100% accurate 😁)

The gals show plenty of diffusion bleed; the pin-ups are human gals with “cat ears and cat tail” (because “catgirl” makes disturbing furries), and the t-shirt gal is just an elf, with no mention of an apparently-detachable tail. It took several tries to get even one pin-up gal without elf ears.

Codellama-34B

I was chatting about this at work after sharing my positive experience with $LICENSED_TOOL, and someone asked DuckDuckGo which offline coding LLMs to use. It recommended WizardCoder and two variations of Codellama. I grabbed the largest versions of them (~34GB each on disk) and fed them the spec.

Codellama wrote me a short story about how it would write the program. Just the story, no code.

I followed up with “write the program”. It wrote a sequel to the story.

I followed up with “where’s the code?”, and its answer was, I shit you not:

I uploaded my code on Github as well - https://github.com/akshat-raj09/CMPE15M_A3

I’m gonna need a bigger shocked face.

(also, those models all have a maximum 16KB context limit, so even half a dozen passes will blow it out; deleted)

Unrelated,

Amazon’s recommending bibles this week. Gosh, what could have happened recently that shifted the algorithm? Like the motives of a left-wing Antifa terrorist who was fucking a furry tranny, I guess we’ll never know.

Even more unrelated,

I was briefly deeply disturbed by the Chinese furniture manufacturer that has chosen the brand name Goaste. Read it wrong the first time…

Dear ANN reviewer,

I really don’t care about your trans journey, or how you feel it was reflected in the final episode of Call Of The Night 2. Allow me to introduce you to the concept of TMI.

Batch Gals

SwarmUI’s wildcard support isn’t completely random. Buried in the “Swarm Interal” section of the parameters is the “Wildcard Seed Behavior” param, which defaults to “Random”. Changing it to “Index” will loop through the wildcards file in order.

So I used my latest wildcard set to generate 250, piped that through the LLM prompt enhancer, and turned it loose.

Not all of the fails were due to the enhancer. Missing limbs and off-by-N finger counts are old hat, so let’s stick to novel fails.

more...

Despite its many flaws, I really think this is the most promising model to play with, because the bolted-on LLM parser significantly improves your ability to lay out an image and add text.

This image came out almost exactly as requested. I ran it multiple times to select the best variation on the theme (sexy pin-ups, size and style of dice, length of liquor shelf, facial expressions, contents of character sheets, etc), but they were all correct.

The exact prompt was:

A gritty, realistic fantasy RPG illustration set in a tavern, with a large central table around which are seated five bearded dwarves wearing leather armor, with axes slung on their back. On the table in front of each dwarf are RPG character sheets, polyhedral dice, and wooden tankards filled with ale. The wall behind them has shelves full of liquor bottles with rune-covered labels, and pin-up posters of sexy goblin girls. The center dwarf is speaking, and the speech bubble contains the words “Okay, now make a shaving throw…” in a handwriting font.

Naturally, the speech bubble works with my usual sort of imagery as well…

(I didn’t specify maid lingerie, but I did ask for “elf maidens”, so call it a lucky accident)

Dear Microsoft,

I just had to email a plain-text attachment as a ZIP file from a Mac to a Mac to keep Outlook from mangling UTF-8 into random garbage. Fix the little shit before you shove your “AI” into every app, m’kay?

Coding with ChatGPT…

End every request after the first with the following words. You’ll thank me later.

Please fix ONLY this issue, and write out the complete corrected program, without making ANY unrequested changes to the code.

This is pretty much the only way to get readable diffs of your iterations. Otherwise there will be random changes everywhere. Comments added/deleted, code collapsed onto one line or expanded to multiples, functions renamed or reordered, regressions created, etc, etc.

Final ChatGPT Deathmatch

This is version 18.5; ChatGPT was still offering me additional enhancements, but we’d far exceeded my patience, even without the four-hour delay while my free credits recharged. The last two revisions were fixes for bugs that only emerged in final QA, where ranking/flagging an image with a value that was currently filtered out of the display list double-advanced, skipping over a still-valid image. Since the whole point of the method is progressively assigning ranks to every image, skipping one was a nuisance.

The “.5” comes from me not wanting to make Yet Another ChatGPT Pass to fix two lines of Javascript that were easily located in the file. The LLM had argued its way into violating the spec at the last minute while fixing the final bug.

It’s GenAI code, so by definition it can’t be copyrighted; if you have any use for it, knock yourself out.

#!/usr/bin/env python3
import sys, os, threading
from pathlib import Path
from flask import Flask, send_file, request, jsonify, render_template_string, abort

app = Flask(__name__)

BASE_DIR = Path(sys.argv[1] if len(sys.argv) > 1 else os.getcwd()).resolve()
RANK_FILE = BASE_DIR / "_rank.txt"
state_lock = threading.Lock()
files = []
meta = {}

# ------------------------------------------------------------
# Load/save ranking state
# ------------------------------------------------------------
def load_state():
    global files, meta
    files = sorted(
        [f.name for f in BASE_DIR.iterdir() if f.suffix.lower() in (".png", ".jpg")],
        key=lambda fn: (BASE_DIR / fn).stat().st_mtime,
    )
    meta = {}
    if RANK_FILE.exists():
        for line in RANK_FILE.read_text().splitlines():
            parts = line.split("\t")
            if not parts:
                continue
            fname = parts[0]
            rank = int(parts[1]) if len(parts) > 1 and parts[1] else 0
            flags = set(parts[2].split(",")) if len(parts) > 2 and parts[2] else set()
            meta[fname] = {"rank": rank, "flags": flags}
    for f in files:
        if f not in meta:
            meta[f] = {"rank": 0, "flags": set()}


def save_state():
    with open(RANK_FILE, "w") as fp:
        for fname in files:
            entry = meta.get(fname, {"rank": 0, "flags": set()})
            flags_str = ",".join(sorted(entry["flags"])) if entry["flags"] else ""
            fp.write(f"{fname}\t{entry['rank']}\t{flags_str}\n")


# ------------------------------------------------------------
# Routes
# ------------------------------------------------------------
@app.route("/")
def index():
    return render_template_string(INDEX_HTML)


@app.route("/image/<path:fname>")
def get_image(fname):
    if not fname:
        abort(404)
    target = (BASE_DIR / fname).resolve()
    if not str(target).startswith(str(BASE_DIR.resolve())):
        abort(403)
    if not target.exists():
        abort(404)
    return send_file(str(target))


@app.route("/api/state")
def api_state():
    with state_lock:
        safe_meta = {
            fname: {
                "rank": entry.get("rank", 0),
                "flags": sorted(entry.get("flags", [])),  # sets → sorted lists
            }
            for fname, entry in meta.items()
        }
        return jsonify({"files": files, "meta": safe_meta})


@app.route("/api/update", methods=["POST"])
def api_update():
    data = request.json
    fname = data.get("file")
    if fname not in meta:
        abort(400)
    with state_lock:
        if "rank" in data:
            meta[fname]["rank"] = data["rank"]
        if "toggle_flag" in data:
            fl = data["toggle_flag"]
            if fl in meta[fname]["flags"]:
                meta[fname]["flags"].remove(fl)
            else:
                meta[fname]["flags"].add(fl)
        save_state()
    return jsonify(success=True)


@app.route("/api/reload", methods=["POST"])
def api_reload():
    with state_lock:
        load_state()
    return jsonify(success=True)


# ------------------------------------------------------------
# HTML/JS template
# ------------------------------------------------------------
INDEX_HTML = """
<!doctype html>
<html>
<head>
<meta charset="utf-8"/>
<title>Five-Star Deathmatch</title>
<style>
body { margin:0; font-family:sans-serif; display:flex; flex-direction:column; height:100vh; }
.topbar, .bottombar { background:#eee; padding:4px; display:flex; justify-content:space-between; font-size:20px; }
.central { position:relative; flex:1; background:#999; display:flex; justify-content:center; align-items:center; overflow:hidden; }
.central img { max-width:100%; max-height:100%; object-fit:contain; }
.central.zoom { overflow:auto; justify-content:flex-start; align-items:flex-start; }
.central.zoom img { max-width:none; max-height:none; width:auto; height:auto; display:block; }
#help { position:absolute; top:10px; right:10px; cursor:pointer; border:1px solid #666; border-radius:8px; margin:2px; padding:6px; background:#f9f9f9; text-align:center; }
#helpPanel { display:none; position:absolute; top:40px; right:10px; background:#fff; color:#000; font-size:16px; padding:10px 15px; border-radius:10px; box-shadow:0 2px 8px rgba(0,0,0,0.3); }
#helpPanel ul { margin:0; padding-left:20px; }
#helpPanel li { margin:4px 0; }
.bottombar { flex-direction:column; font-size:14px; }
.row { display:flex; flex:1; }
.cell { flex:1; border:1px solid #666; border-radius:8px; margin:2px; text-align:center; padding:6px; cursor:pointer; display:flex; flex-direction:column; justify-content:center; align-items:center; background:#f9f9f9; }
.cell.active { background:#ccc; }
</style>
</head>
<body>
<div class="topbar"><div id="filename"></div><div id="rankflags"></div><div id="pos"></div></div>
<div class="central" id="central">
  <img id="mainimg"/>
  <div id="help">❓</div>
  <div id="helpPanel"></div>
</div>
<div class="bottombar">
  <div class="row" id="rankrow"></div>
  <div class="row" id="flagrow"></div>
</div>
<script>
let state={files:[],meta:{}};
let index=0;
let filters={ranks:new Set(),flags:new Set()};
let zoom=false;
const mainimg=document.getElementById("mainimg");
const filenameDiv=document.getElementById("filename");
const rankflagsDiv=document.getElementById("rankflags");
const posDiv=document.getElementById("pos");
const central=document.getElementById("central");
const help=document.getElementById("help");
const helpPanel=document.getElementById("helpPanel");

function fetchState(){ fetch("/api/state").then(r=>r.json()).then(js=>{state=js; render(); buildBottom();}); }
function buildBottom(){
  const rr=document.getElementById("rankrow"); rr.innerHTML="";
  const ranks=[-1,0,1,2,3,4,5];
  let counts={}; for(let r of ranks) counts[r]=0;
  let total=0;
  for(let f of state.files){ let m=state.meta[f]||{rank:0,flags:[]}; if(m.rank>-1) total++; counts[m.rank]++; }
  for(let r of ranks){
    let icon=(r==-1?"❌":(r==0?"⚪️":"⭐️".repeat(r)));
    let pct=(r>=0 && total>0)?Math.round(100*counts[r]/total)+"%":"";
    const d=document.createElement("div"); d.className="cell"; if(filters.ranks.has(r)) d.classList.add("active");
    d.innerHTML="<div>"+icon+"</div><div>"+counts[r]+(pct?" ("+pct+")":"")+"</div>";
    d.onclick=()=>{ if(filters.ranks.has(r)) filters.ranks.delete(r); else filters.ranks.add(r); render(); buildBottom(); };
    rr.appendChild(d);
  }
  const fr=document.getElementById("flagrow"); fr.innerHTML="";
  const fls=["A","B","C","D","E","F"]; let flagCounts={}; for(let fl of fls) flagCounts[fl]=0;
  let unflagged=0;
  for(let f of state.files){ let m=state.meta[f]||{rank:0,flags:[]}; if(m.flags.length==0) unflagged++; for(let fl of m.flags) flagCounts[fl]++; }
  for(let fl of fls){
    const d=document.createElement("div"); d.className="cell"; if(filters.flags.has(fl)) d.classList.add("active");
    d.innerHTML="<div>"+fl+"</div><div>"+flagCounts[fl]+"</div>";
    d.onclick=()=>{ if(filters.flags.has(fl)) filters.flags.delete(fl); else filters.flags.add(fl); render(); buildBottom(); };
    fr.appendChild(d);
  }
  const d=document.createElement("div"); d.className="cell"; if(filters.flags.has("UNFLAG")) d.classList.add("active");
  d.innerHTML="<div>🚫</div><div>"+unflagged+"</div>";
  d.onclick=()=>{ if(filters.flags.has("UNFLAG")) filters.flags.delete("UNFLAG"); else filters.flags.add("UNFLAG"); render(); buildBottom(); };
  fr.appendChild(d);
}
function filteredFiles(){
  return state.files.filter(f=>{
    let m=state.meta[f]||{rank:0,flags:[]};
    if(filters.ranks.has(m.rank)) return false;
    for(let fl of m.flags){ if(filters.flags.has(fl)) return false; }
    if(m.flags.length==0 && filters.flags.has("UNFLAG")) return false;
    return true;
  });
}
function render(){
  let list=filteredFiles(); if(list.length==0){ mainimg.src=""; filenameDiv.textContent=""; rankflagsDiv.textContent=""; posDiv.textContent="0/0"; return; }
  if(index>=list.length) index=0;
  let fname=list[index]; mainimg.src="/image/"+encodeURIComponent(fname); filenameDiv.textContent=fname;
  let m=state.meta[fname]||{rank:0,flags:[]};
  let rankDisp=(m.rank==-1?"❌":(m.rank==0?"⚪️":"⭐️".repeat(m.rank)));
  rankflagsDiv.textContent=rankDisp+" "+m.flags.sort().join("");
  posDiv.textContent=(index+1)+" / "+list.length;
}
function nextValidIndex(oldfname){
  let list=filteredFiles();
  let i=list.indexOf(oldfname);
  if(i==-1){
    if(index>=list.length) index=list.length-1;
  } else {
    index=i + 1;
  }
}
function updateRank(r){
  let fname=filteredFiles()[index];
  fetch("/api/update",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({file:fname,rank:r})})
    .then(()=>fetch("/api/state"))
    .then(r=>r.json())
    .then(js=>{
      state=js;
      nextValidIndex(fname);
      render(); buildBottom();
    });
}
function toggleFlag(fl){
  let fname=filteredFiles()[index];
  fetch("/api/update",{method:"POST",headers:{"Content-Type":"application/json"},body:JSON.stringify({file:fname,toggle_flag:fl})})
    .then(()=>fetch("/api/state"))
    .then(r=>r.json())
    .then(js=>{
      state=js;
      render(); buildBottom();
    });
}
document.addEventListener("keydown",ev=>{
  let list=filteredFiles(); if(list.length==0) return;
  if(ev.key=="ArrowLeft"){ index=(index-1+list.length)%list.length; render(); }
  else if(ev.key=="ArrowRight"){ index=(index+1)%list.length; render(); }
  else if(ev.key=="ArrowUp"){ index=0; render(); }
  else if(ev.key=="ArrowDown"){ index=list.length-1; render(); }
  else if(ev.key=="x"||ev.key=="X"){ updateRank(-1); }
  else if(ev.key>="0"&&ev.key<="5"){ updateRank(parseInt(ev.key)); }
  else if("ABCDEF".includes(ev.key.toUpperCase())){ toggleFlag(ev.key.toUpperCase()); }
  else if(ev.key=="r"||ev.key=="R"){ filters={ranks:new Set(),flags:new Set()}; render(); buildBottom(); }
  else if(ev.key=="z"||ev.key=="Z"){ zoom=!zoom; if(zoom){ central.classList.add("zoom"); central.scrollTop=0; central.scrollLeft=0; } else { central.classList.remove("zoom"); } }
  else if(ev.key==" "){ fetch("/api/reload",{method:"POST"}).then(()=>fetchState()); }
});
help.onclick=()=>{ 
  helpPanel.style.display="block"; 
  helpPanel.innerHTML="<ul>"
    +"<li>← / → : Previous / Next image</li>"
    +"<li>↑ / ↓ : First / Last image</li>"
    +"<li>0–5 : Set rank</li>"
    +"<li>X : Rank -1 (❌)</li>"
    +"<li>A–F : Toggle flags</li>"
    +"<li>R : Reset filters</li>"
    +"<li>Z : Toggle zoom</li>"
    +"<li>Space : Reload state</li>"
    +"</ul>";
};
helpPanel.onclick=()=>{ helpPanel.style.display="none"; };
fetchState();
</script>
</body>
</html>
"""

# ------------------------------------------------------------
if __name__=="__main__":
    load_state()
    app.run("127.0.0.1",5000,debug=True)

The odd mix of inlining and formatted blocks is how it went from 700+ lines to 268: every revised version made new formatting decisions in unrelated parts of the file until I started adding The Magic Words. Areas being updated in a revision got clean formatting, because it showed them in the explanation of the change, while “unchanged” areas got compressed onto one line.

Last show standing, and since it started late, it’s got two one more episodes to go. [sigh; it's only 11 episodes, so things are going to be rushed next week]

This week, all hell breaks loose, kicking off an episode of Back-Seat Kaiju!, in which our cute glasses-wearing Operations gal’s sanity is tested. Basically, everyone gets to debut their new kaiju skinsuits, with Tsuntail exploding out of the gate in her mom’s hand-me-downs, Vice-Captain playing tsukkomi to #10’s boke, and Super Sidekick… not appearing this week. Good thing we’ve got two captains on deck.

Verdict: two one more episode, and I have no idea what they plan to end it on. I have a hunch Naughty Number Nine has been built up too much as a long-term antagonist, so maybe they’ll just Save The Country For Now and try for a third season? Ratings apparently support the idea.

(time for Tsuntail to go axe-crazy!)

Diffusion Image Evaluation Checklist

1. Count the limbs.
2. Count the fingers.
3. Count the toes.
4. Side-check the thumbs.
5. Side-check the big toes.
6. Count the navels.
7. Check for crazy eyes.
8. Check for consistent skin tones.
9. Check skin texture.
10. Check for texture patterns in background.
11. Watch out for overtrained default clothing.
12. Check all four edges for cut-off objects.
13. Check for giants and tinyfolk.
14. [update!] Check for melting.

Now you can apply standard advice about composition, like rule-of-thirds, division of negative space, not cropping people at joints, etc. If there’s anything even vaguely naughty about the pictures you’re generating, you’ll also want to check the apparent age of all human figures…

GenAI coding: the definition of insanity

devstral-small-2507:

It was quick, and friendly, and the code was half the size of the working one I got from gpt-oss-20b, and it ran the first time. It didn’t do anything, of course, and I spent N passes nit-picking every place where it did something stupid or simply ignored the spec, but as I blew past 24KB of context, my checkin comments started to get snarky. After all, I’d only just got it to finally display the bottom-bar buttons onscreen, without scrolling, and it still refused to actually make them work as buttons, or display the correct contents.

But it was quick, and friendly! Worst thing was that it actually styled the web GUI nicely; it just couldn’t make it meet the spec. I finally had to give up, because it started “fixing” bugs without changing a single line of code.

more...

…and that’s a wrap. Nothing much happens this week, as everyone discusses recent events and finds closure. The closest it gets to action is Our Hungover But Still Hot And Clingy Detective threatening to puke on Our Halfbreed Hero. Nice touch ending with the first-season ED song that inspired the manga.

Verdict: a lot of budget-saving tricks on the animation, but the art and voice acting were good, and the story was interesting despite the cover-our-asses suicide-prevention-hotline PSAs.

(unrelated vampire has ways of making you fall in love with her…)

🎶 Baby, you can vibe my code 🎶

As expected, adding additional features to the small, self-contained Python script took several passes, which took it past 32KB of context, then 48KB, and by the time I had the functionality just the way I wanted it, it was up to 59KB. And it took hours to get to that point, running on a not-exactly-cheap (even refurbished) Mac Mini. As its final task, I ordered it to write a revised design document reflecting the final state, which came in just under the 64KB limit.

For a single file of Python code with just over 500 lines. The maximum context setting for this model is 128KB. This does not scale to real coding projects.

Note that I said the functionality was as I wanted it. The LLM couldn’t fix CSS issues for blood or money, and still took N minutes to try and fail. At the moment, I need to fix all the button sizing and layouts, restyle the text to make it less hideous, and see if I can coax the zoom feature to scroll horizontally as well as vertically. And fix the display of the help pop-up, which got busted in one of the passes (I just need to find it in source control and paste the correct version back into place; which is why I put even the most trivial of projects under source control…).

(AI-upscaled XKCD…)

More LLM Coding Fails

• magistral-small-2509: never got it to display an image, or make any progress figuring out the error messages I fed it.

• qwen3-coder-30b: just could not figure out how to display a new image after navigating, despite the code being more than twice as long.

• qwen3-32b: excellent at parsing error messages to fix syntax errors in Python and Javascript; also excellent at making syntax errors in Python and Javascript. Complete waste of time. The only nice thing I can say about it is that it didn’t make me watch while it retyped the whole program after each change, but even that was undercut by it losing the indentation when telling me to replace a buggy block of Python with a new one.

Haven’t tried it with the licensed cloud-y tool on my work Mac yet…

Fall Season Falling Down

With only a few weeks to go, presumably most of the fall shows have been announced and have their web sites and promotional videos up, right? Meh, kinda.

Setting aside the Nth seasons of shows I didn’t watch N-1 of:

• A Gatherer’s Adventures In Isekai: wow, they managed to make that sound completely generic. NO

• Alma-chan Wants To Be A Family: sort-of Nuku-Nuku as an emotionless loli, with a side order of Sekirei? NO

• A Mangaka’s Weirdly Wonderful Workspace: just NO

• A Star Brighter Than The Sun: tough girl in love with pretty boy. NO

• A Wild Last Boss Appeared: player wakes up in MMO world as her OP character. NOT AGAIN

• Chitose Is In The Ramune Bottle: popular kid tries to coax hikikomori into coming back to school. NO

• Dad Is A Hero, Mom Is A Spirit, I’m a Reincarnator: ohhellNO

• Don’t Touch Kotesashi: generic dorm-harem porn game. NO

• Dusk Beyond The End Of The World: Like World’s End Harem, only more derivative. NO

• Forget That Night, Your Majesty: teaser trailer promises barely-animated shojo-manga romance with chins that could cut glass. NO

• Future Kid Takara: trailer mixes lightly-filtered 3D CGI with real-world footage to mash-up Doraemon with Captain Planet. ohhellNO

• Gachiakuta: future world is divided between floating-city elites and ground-bound garbage-pickers, an idea that’s been done to death. NO

• Ganglion: villain/mook gag show; that’s gag as in “choke”, based on the trailer. NO

• Gintama - Mr. Ginpachi’s Zany Class: ‘wacky’ spinoff of a show I never watched. NO

• Gnosia: serial-numbers-filed-off clone of Among Us, and therefore sus. NO

• Hero Without A Class: Who Even Needs Skills?: pretty sure Starless Tamer did it better. NO

• Inexpressive Kashiwada and Expressive Oota: high-school romance in which (gasp) opposites attract. NO

• Isekai Quartet 3: this would be a lot more interesting if all the shows were, y’know, current. The first season of Tanya The Evil aired in 2017, and the still-missing second season was announced in 2021, but all we’ve gotten was the (admittedly hilarious) Desert Pasta one-shot, also in 2021. So putting the cast in a new season of this sketch comedy is just rubbing salt in the wound. NO

• Latair The Earth: the promo art is enough for me to say NO

• Let’s Play: official blurb includes the phrase “a comedic, romantic, and all-too-real story about gaming, memes, and social anxiety”, which is a giant ohhellNO

• L’il Miss Vampire Can’t Suck Right: surprisingly, not a porn game adaptation. Unsurprisingly, trailer is full of suck. NO

• May I Ask For One Final Thing?: more ‘villainess’ shit. NO

• Mechanical Marie: robot/maid/assassin/guardian, with the twist being that her master is robosexual but she’s only pretending to be a robot. NO

• Monster Strike: Deadverse Reloaded: game adaptation. NO

• My Awkward Senpai: bitchy sexpot office lady acquires bishie subordinate, and you’ll be stunned to learn she falls for him. NO

• My Friend’s Little Sister Puts It In For Me: oh, wait, that’s “has it in for me”. $10 says it ends up my way, though. NO

• My Gift Lvl 9999 Unlimited Gacha: Backstabbed in a Backwater Dungeon, I’m Out for Revenge!: ohhellNO

• My Status As An Assassin Obviously Exceeds The Hero’s: another whole-class-gets-isekai’d-and-I’m-way-OP show. NO

• Ninja Versus Gokudo: not even going to look at the trailer. NO

• Swallow My Monster Meat, Milady: er, “pass the monster meat”, which sounds like something you’d need laxatives for. NO

• Plus-Sized Misadventures In Love!: NO

• Potion, Wagami wo Tasukeru: love how the “saved by potions” title isn’t even being translated. TL/DR: another “potion loli” show, this one with bonus excitable elf bishie. NO

• Puzzle & Dragon: do not watch the trailer. NO

• Sanda: whoever decided to release an English-language trailer with a flat-affect AI voiceover chose… poorly. NO

• Shabake: Edo period shop owner with weak constitution who’s surrounded by helpful spirits spontaneously goes all Jessica Fletcher as people start being murdered every time he leaves the house. Somebody rolled percentile dice to generate this one. NO

• Si-Vis: The Sound Of Heroes: low-effort ripoff of Kpop Demon Hunters? ohhellNO

• Solo Camping For Two: like “hanging out for hermits”, I guess. NO

• Style of Hiroshi Nohara’s LUNCH: another AI-generated English voiceover? Kill me now, please. NO

• Sylvanian Families: Freya no Wonder Days: no synopsis or trailer is available, which automatically makes it better than the ones with AI voiceovers. Still NO

• The Banished Court Magician Aims to Become the Strongest: oh, good, I was hoping there would be a “gonna be the strongest” “kicked out of the hero’s party” show this season, said no one ever. NO

• The Dark History of the Reincarnated Villainess: NO

• This Monster Wants to Eat Me: not porn, just something-something yuri death-wish romance something. NO

• Tojima Wants to Be a Kamen Rider: NO

• Peach Versus Ogre: (er, “Tougen Anki”) the descendents of the folk-tale hero Momotaro are still fighting oni, which sucks when you find out you’re an oni. I’m going to go out on a limb here and guess that this will suck. NO

• Touring After the Apocalypse: didn’t they just do like five of these shows? NO

• Wandance: dance like no one’s watching, at high school. I also will not be watching. NO

• Watari-kun’s ****** Is About to Collapse: somebody please spoil the ****** for me so that I don’t have to type it again and get the backslashes right so the Markdown processor doesn’t turn it into bold and italic. Or don’t. NO

• With You, Our Love Will Make It Through: chick wants to fuck her furry classmate. NO

• Yano-kun’s Ordinary Days: chick wants to fuck her battered classmate. NO

To sum up, NO. Maybe rock-service and shadow-healer can kill some time while I wait for more Frieren in January.

Wallpaper fail

I added a few hundred non-lingerie clothing sets to my wildcards and made some gals in a 9:16 aspect ratio to potentially share the wallpaper rotation with RedWaifu on my new 27” 4K HDR vertical monitor, and this was one of the first to show up:

“No, no, I didn’t say she can’t dance, I said she’s got two left feet!” 😁

Related, I have to go through the clothing wildcards and remove any reference to “crossbody” purses. They end up with literal crosses attached to them. Sometimes, they have straps.

(classical reference)

Dear Apple,

Please provide fine-grained controls for disabling “data detector” overlays in text and images. It’s really annoying to look at a picture that clearly contains no text at all and have a translucent pulldown menu show up when you mouse over the image that offers to add a random string of digits to your contacts as an international phone number. Note that it doesn’t let you copy the string; it’s so insistent that it’s a phone number that it only offers options to use it as one.

It’s one thing to be able to open an image in Preview.app and deliberately choose the OCR text-selection mode (which is quite good, even at vertical Chinese/Japanese), but having it turned on system-wide for all images is intrusive and dangerous bullshit. I don’t want every image processed by a text-scanning system that has opaque privacy rules and no sense. And of course interpreting random digits in text as phone numbers and converting them to clickable links is also dumb as fuck; remember when a bunch of DNA researchers discovered their data was corrupted by Excel randomly turning genes into dates?

(Settings -> General -> Language & Region -> Live Text -> offdammit)

(I didn’t even specify an Apple product, just “silver-colored laptop”; training data, whatcha gonna do? I did have to add a table and ask for a big downward swing of the axe, but the flames were free, thanks to a generous interpretation of the term “fire axe”)

Vibe Me Wild

(classical reference)

There is an executive push for every employee to incorporate generative AI into their daily workflow. I’m sure you can guess how I feel about that, but the problem is that they’re checking.

We have licenses for everyone to use specific approved tools, Which I Will Not Name, and VPs can see how many people have signed into the app with SSO, and at least get a high-level overview of how much they’ve been using it.

So I need to get my Vibe on. The problem is, it’s just not safe to run tool-enabled and agentic genai on my work laptop (especially while connected to the VPN), because I have Production access. The moment I check the boxes that enable running commands and connecting to APIs, I’d be exposing paying customers to unacceptable risks, even though there are passwords and passphrases and Yubikeys to slow things down. All it needs to do is vibe its way into my dotfiles without being noticed. I don’t even want the damn thing to read internal wiki pages, because many of them include runbooks and troubleshooting commands. And of course there’s incidents like this, in which an OpenAI “agent” is subverted to exfiltrate your email.

But I need to show that I’ve used the damn app to produce code.

So I wrote up a detailed design document for a standalone Python script that implements my five-star deathmatch image-ranking system. And before handing it over to the work app, I fed it to offline LLMs.

First up, seed-oss-36b, which has been giving me good creative results and tagging: it ‘thought’ for 20+ minutes, then generated a full project that ignored about half of my requirements, including the one about persisting the state of the rankings to disk. I didn’t even try to run it.

Next, gpt-oss-20b, which ‘thought’ for 30 seconds before spitting out a complete, self-contained Python program that almost worked. When I told it that the /images route and the /api/images route worked, but that the main / route displayed nothing and none of the key bindings worked, it ‘thought’ for 25 seconds, realized that it had written syntactically invalid Javascript for the key bindings (multiple case statements on one line), and corrected the code.

At this point, I had basic functionality, but found three flaws in testing. I listed them out, and after 2 minutes of ‘thought’, it corrected them. Sadly, it also deleted the line import os from the code, breaking it.

I told it to fix that, add a new keybind to reset the display filtering, and fix a newly-discovered bug in the image-zoom code that prevented scrolling. A minute of ‘thought’ and it took care of those issues, but deleted the Flask import lines this time.

A mere 7.5 seconds of ‘thought’ convinced it to add that back, and then I had a fully-functional 413-line self-contained app that could let me quickly rank a directory full of image files and persist the rankings to disk.

All in all, ~20 minutes of me time to write the design doc, 4 minutes of ‘thinking’ time, plus ~4 minutes each pass to generate the script (I’m getting ~5 tokens/sec output, which types out the code at roughly twice human speed), plus ~20 minutes of me time for source control, testing, and debugging. Both models used about 18KB of context to accomplish their task, which means that additional enhancement requests could easily cause it to overflow the context and start losing track of earlier parts of the iterative process, with potential loss of functionality or change of behavior.

With tested results, I’m now willing to present the revised design doc to the licensed tool and let it try to write similar code. While I’m not connected to the VPN…

(I suppose HR would take offense if I pointed out that the Vibe in Vibe Coding should be replaced with a more intrusive sex toy…)

With apologies to The Beatles…

I once had a vibe
    Or should I say
It once vibed me
    It wrote all my code
Then gave away
    API keys

It asked for my keys
    and it said they’d be safe in the vault
Then I looked around
    and I found them shared on ServerFault

I called for support, waited online
    wasting my time
I talked myself blue, tech support said
    “I’m off to bed”

He told me his shift had just ended
    and started to laugh
I emptied my wallet
    and crawled off to sob in the bath

And when I came out, my app was gone
    my credit blown
So I set a fire
    at their HQ
and watched them burn