AI Chaos Control Center ⚡ Turning AI chaos into something that works...
Joined October 2026
- Tweets383
- Following68
- Followers67
- Likes841
I can't f*cking believe this is just code
There isn't a single 3D model or texture in it. Claude Fable 5.2 wrote every gear and every glass tentacle inside one HTML file, and it runs live in the browser
Everyone's building mechanical animals that take themselves apart. So I gave mine a button nobody else has: JELLY
One click and the glass and chrome turn into gummy candy, while the gears keep spinning inside. Then it blows apart into hundreds of jelly pieces and snaps back together
GTA 6 is out next week. I asked Gemini for a GTA-style game in one prompt to see how close we are.
It built a whole city with traffic, a GPS minimap, a speedometer and delivery missions with a timer. All in one HTML file, first try.
Then I failed my very first delivery. $0 earned.
Still a long way from GTA. But a year ago this prompt gave you a square moving around a box.
5
12
117
ChatGPT now tells businesses how many times it mentioned them.
New today, in beta: Business Hub. Any eligible business in a country where ChatGPT works can claim its profile and see:
→ how often ChatGPT mentions you
→ how many people clicked through to your site
→ "discovery topics", the themes of conversations where your site gets cited
You can also update your details and add products with prices and stock, so ChatGPT works from your info instead of guessing.
If you sell anything online, claim it this week. 1.2 billion people use ChatGPT every week, and right now it describes your business without you.
5
11
112
One prompt on Fable 5.2 burned about 6x more of my Claude limit than the same prompt on Opus 5.5.
Same task, Medium effort, Pro plan: Opus moved my session meter 1%, Fable moved it 6%. Both built a working Snake game.
If you're on Pro: Fable is only included until Nov 16, capped at half your weekly limit. After that it runs on paid usage credits. Anthropic's own advice is to start with Opus 5.5 and save Fable for the hardest tasks.
Fable for the problem you're stuck on. Opus for everything else.
8
15
168
Zoom into "I'm not a robot". It never ends.
That checkbox exists to catch machines. Machines now pass it better than we do: in a 2023 UC Irvine study with 1,400 people, humans solved distorted-text CAPTCHAs 50 to 84% of the time. Bots got almost 100%, in under a second.
So the video follows a machine trying to look human: captcha, server rack, a chip that looks like a city, a laptop at night, 99% buffering, 1% battery, Snake, a 3D hand with one finger too many. Then back to the fresco.
From the first frame to the loop you zoom in about 500 billion times. The fresco is the real Michelangelo, every other scene is drawn in code.
ChatGPT can make GIFs now. So I tested if it understands physics.
4 classic animation tests, first try each, every one in a fresh chat:
→ bouncing ball: it slows down as it falls. Gravity should speed it up
→ Newton's cradle: I asked for 5 balls, it drew 6
→ clock: 10 ticks of one second each. It froze, then jumped 3.5 seconds
→ pouring water: nailed it
1 out of 4. It nails how things look. How they move is still hit or miss.
Golf uses 2.5x more water than every data center in US. Power plants included.
→ golf courses: ~600B gallons a year (last national survey, 2013)
→ data centers, on site: 17 to 19B gallons (2024)
→ plus the water used to generate their electricity: ~230B
Most "AI is draining our water" posts skip the power plants. Most "it's basically nothing" replies skip them too. Add them in and golf still wins easily.
The gap is closing though: on-site use alone is projected to hit 60 to 110B gallons by 2030. Tanks are drawn to scale.
3
10
228
Before AI, this was the whole pipeline. One cutout at a time
Introducing Higgsfield Miracle.
Our latest proprietary post-production technology vectorizes any video into layers for localized, pixel-level editing in real time, with world knowledge built in. Every frame holds the image depth needed for cinema-grade post-production.
Now in private beta with select testers. Coming soon to Higgsfield MCP.
Set up Higgsfield MCP in Claude to be the first to try it.
AfterDark retweeted
what if we could actually watch how the brain thinks
memory, language, movement, attention. none of them live in one spot, they all run on connected networks
and researchers are already using models to study brain scans at scale. one 2026 study (BrainIAC, nature neuroscience) was tested across 48,965 brain MRI scans
it wont replace neurologists. but it might spot patterns we keep missing
which part of the brain would u explore first?
Four AIs judged each other. One judged itself a little too generously.
A year ago Karpathy's LLM Council picked GPT 5.1 as the best model and Claude as the worst. I reran it with ChatGPT, Claude, Grok and Gemini (default models), with the answers anonymized.
→ all four solved the trick question
→ Gemini ranked its own answer last
→ ChatGPT ranked its own advice #1, and the other three ranked it last
→ final order: Claude, Grok, Gemini, ChatGPT, the same even without self-votes
Claude also remembered who I was, which probably helped on one question. 3 questions and 4 judges, so it's a vibe check.
The animation is pure code, built by Claude Opus 5.5 from one prompt.
As a fun Saturday vibe code project and following up on this tweet earlier, I hacked up an **llm-council** web app. It looks exactly like ChatGPT except each user query is 1) dispatched to multiple models on your council using OpenRouter, e.g. currently:
"openai/gpt-5.1",
"google/gemini-3-pro-preview",
"anthropic/claude-sonnet-4.5",
"x-ai/grok-4",
Then 2) all models get to see each other's (anonymized) responses and they review and rank them, and then 3) a "Chairman LLM" gets all of that as context and produces the final response.
It's interesting to see the results from multiple models side by side on the same query, and even more amusingly, to read through their evaluation and ranking of each other's responses.
Quite often, the models are surprisingly willing to select another LLM's response as superior to their own, making this an interesting model evaluation strategy more generally. For example, reading book chapters together with my LLM Council today, the models consistently praise GPT 5.1 as the best and most insightful model, and consistently select Claude as the worst model, with the other models floating in between. But I'm not 100% convinced this aligns with my own qualitative assessment. For example, qualitatively I find GPT 5.1 a little too wordy and sprawled and Gemini 3 a bit more condensed and processed. Claude is too terse in this domain.
That said, there's probably a whole design space of the data flow of your LLM council. The construction of LLM ensembles seems under-explored.
I pushed the vibe coded app to
github.com/karpathy/llm-coun…
if others would like to play. ty nano banana pro for fun header image for the repo
AfterDark retweeted
ONE VIRAL STORY. TWO VERSIONS. 12,154 "SECRET SELFIES" BECAME 14,219 IN A WEEK.
Version 1: an old Android, a flashlight app from 2019, 12,154 front-camera photos in its folder.
Version 2, a few days later and in Spanish: 14,219 images, now inside a Face Unlock system folder, plus infrared shots with GPS.
Different count. Different folder. Same story.
Two things worth knowing:
→ /data/system can't be read over ADB on a normal phone without root
→ Android keeps face unlock data in secure hardware, not as photos in a folder
The real risk is boring: old phones and forgotten apps that still hold camera permission. Your phone's privacy dashboard shows that in 30 seconds.
Seen a third version yet?
🚨 ESTO ES MUY FUERTE
Un creador de contenido ha utilizado Claude Code para descubrir que su móvil Android había almacenado más de 14.219 imágenes capturadas por la cámara frontal sin que él lo supiera.
Y lo más inquietante es dónde estaban guardadas.
No eran capturas de pantalla ni fotografías de la galería.
Eran imágenes reales tomadas por la cámara frontal en distintos momentos del día.
En casa, en el trabajo, en el coche e incluso cuando simplemente desbloqueaba el teléfono.
¿Cómo lo descubrió?
Conectó su Android al ordenador y utilizó Claude Code para investigar los archivos ocultos del dispositivo.
Encontró imágenes en directorios del sistema que normalmente no aparecen en la galería:
→ /data/system/users/0/faceunlock/cache
→ Directorios asociados a Google Play Services.
→ Más de 10.000 archivos entre fotografías normales e imágenes infrarrojas utilizadas para el reconocimiento facial.
¿Lo más preocupante?
Muchas de estas imágenes incluían información de fecha, hora y ubicación GPS.
El fabricante y Google explican que el reconocimiento facial utiliza imágenes para generar modelos de profundidad y que después se eliminan.
Pero en este caso encontró archivos originales almacenados en el dispositivo.
También descubrió que algunos servicios del sistema y aplicaciones con permisos especiales podrían tener acceso a determinados directorios.
¿Qué hizo después?
→ Desactivó los datos de diagnóstico.
→ Eliminó permisos de cámara innecesarios.
→ Limpió las cachés que encontró.
Pero hay algo que deberías tener en cuenta:
No todas las imágenes almacenadas en estos directorios implican que alguien haya accedido a ellas o que exista una filtración de datos.
Aun así, merece la pena investigar qué guarda tu propio dispositivo.
¿Quieres comprobarlo?
Puedes conectar tu Android al ordenador, realizar una copia de seguridad y pedirle a Claude Code que busque archivos de imágenes de la cámara frontal fuera de la galería.
La IA puede ayudarte a investigar archivos del sistema que normalmente pasarían completamente desapercibidos.
Guárdate este post. Puede que descubras algo que no sabías de tu propio móvil 🗒️
Readers added context they thought people might want to know
Las afirmaciones sobre 14.219 imágenes secretas de cámara frontal en caché de Face Unlock en Android no están verificadas y contradicen la documentación oficial. El acceso a /data/system requiere root, no posible vía backups estándar.
cybernews.com/security/claud…
support.google.com/pixelphone/ans…
bool.dev/news/detail/an…
AfterDark retweeted
A Columbia scientist told ChatGPT she picks coconut ice cream 70% of the time.
It chose coconut for her 100 times out of 100.
That's Sandra Matz in this TEDx talk. Her biggest worry about AI sounds small at first: that it will make us boring.
Her logic:
• Every choice is a trade-off between the safe favorite and something new
• Recommendation systems and chatbots get rewarded for safe bets that keep you engaged
• So they keep picking the favorite, and you slowly stop exploring
Another test of hers: asked 100 times as a new customer, ChatGPT named one of Baskin-Robbins' two most popular flavors 96 times.
Newer research adds a twist. In a 2026 experiment, people writing with AI produced more varied stories when they were paid for originality. The sameness follows whatever gets rewarded.
Her fix: a dial on your apps, from "spot on" to "wildcard."
13 minutes, worth every one. Would you turn the dial?
An AI agent got Deedy €600 for a 6 hour flight delay. The rule it used works for you too.
EU261: your flight leaves the EU, or lands there on an EU airline, and arrives 3+ hours late? That's €250 to €600, depending on distance.
Bad weather doesn't count. Technical faults do.
Instinct just made me $700.
I was frustrated because my flight back to US from Europe was delayed 6hrs and texted Instinct asking what I could do. It told me there was a law that said I was owed €600. A month later, after many esoteric email exchanges, I got paid out!
A rare product experience where you actually get infinitely more value than what you paid ($0).
Play-by-play.
- My flight back to the US from Europe on Oct 4 was delayed 6hrs
- EU regulation 261/2004 says if a long haul flight is delayed >3hrs you are owed €600
- I discovered this because I texted Instinct, frustrated, saying "what can I do about this delay"
- It sent an email for me, and chatted with a Delta rep to verify this would fall under those rules.
- It patiently monitored by inbox. They said most cases receive a response in 21 days. On day 21, Oct 26, Delta approved the claim!
- .. but you still couldn't receive it in cash.
- On Nov 3, I finally get the cash claim email
OpenAI and Anthropic changed their paid plans 4 times in 4 days. One of these changes will hit your usage in December.
Claude Max 20x is now Ultra ($200). The 5-hour limit is gone for verified accounts, but the weekly limit stays. You can work longer in one sitting, not more in total.
ChatGPT Pro: the planned cut was cancelled and limits doubled instead. Eligible Pro plans also get $100 to $500 a month in API credits. If you've ever wanted to build your own tool, claim them in Settings.
From December, Plus, Pro and Business share one token budget across ChatGPT and Codex. Heavy Codex sessions will eat into your chat usage. Instant chat doesn't count.
Running several paid ChatGPT accounts? OpenAI now caps them per card and per device.
Which of these changes affects you?
