Videos
November 2026
Claude Opus 5.5 vs GPT-6 Astra: Who Can Beat the Original Celeste First?I told them to play a game. Neither of them played it.
Is AI Improving Itself? I Let GPT-6 Astra Do AI Research for 2 HoursGPT-6 Astra worked as an AI researcher on my PC for two hours. Tap through all 30 experiments to see what it changed and what happened to the score, read the tiny model's stories before and after, and see what it takes to run it yourself.
I Put Claude Opus 5.5 and GPT-6 Astra in the Dark Forest. Who Shoots First?I had Claude Opus 5.5 and GPT-6 Astra play the dark forest from The Three-Body Problem, 30 games. Step through every game epoch by epoch, rerun it under your own rules, play a civilization yourself, and read the prompt they got.
My $200 Claude Ultra Plan Used $11,000 Worth of APINext time someone flexes a giant token bill... they probably paid $200 too π
I Locked an AI in 1900. Can It Beat Einstein to Relativity?I ran GPT-1900, a model that has only read books and newspapers from before 1900, on my own PC and handed it Einstein's clues one level at a time. The prompts for every level, every word it wrote, both judges' scores, and what Claude and GPT said when they pretended to live in 1899.
AI Jailbreaking Explained: Does the Grandma Trick Still Work?I opened a fake bakery with one employee, an AI, and told it exactly one secret π₯ Then I threw every famous jailbreak at it. The big models didn't budge...
Claude vs GPT: Can AI Make a PowerPoint You'd Actually Use?Ask AI for a sales deck in one sentence and it invents the numbers. Give it data, say who's listening, get an outline first, and Claude and GPT both made 5 slides worth presenting.October 2026
I Gave Opus 5.5, Haiku and DeepSeek the Same Game Prompt. They Made Triplets.I gave 3 AIs the same one-sentence prompt and got three identical breakout games. Spelling out gameplay, style, feel and where you play got me this one.
What Do Q4 and Q8 Mean? I Squeezed One AI Into 9 VersionsSame AI, nine sizes. The first seven all knew the capital of France.
I Let Opus 5.5 and GPT-6 Astra Take Over BlenderSame prompt for Opus 5.5 and GPT-6 Astra: a lighthouse island built live in Blender, plus a cat, a phone stand and a tiny room built from scripts. Spin the models in your browser, copy the prompts, download the STLs.
Claude Code Spawned 105 Subagents and Burned My Weekly LimitLock it down before you turn on Ultracode. Seriously.
POV: You're an AI Being TestedThis time you're the AI π« every level is a real test from an AI lab, and level three genuinely creeped me out while I was researching it. (Don't worry, the company and the emails were all made up by researchers.) I really want to know your answer to the last question.
OpenAI and Claude Just Disproved a Famous Math Conjecture. Is P = NP?Did AI just solve P vs NP? Not quite.
Is AI Really a Black Box? I Opened Up NVIDIA's NemotronI found a "know this person / don't" direction inside NVIDIA's Nemotron 3.5 Bolt. Turn the knob and see what it really said at every step.
Why Does AI Write Like That?Same farewell message, Claude, GPT and Qwen: 0, 5 and 9 AI tics. With this prompt: 0, 0 and 1.
Is Prompt Engineering Dead? I Tested the Old Tricks"You're a world-class expert," a $200 tip, "think step by step": I ran the viral prompt tricks 84 times on Claude, GPT and DeepSeek. What actually works is context, or letting the AI ask you.
Claude Ran a Vending Machine and Lost All the MoneyI went in expecting a funny "AI fails at shopkeeping" storyβ¦ then it got weird. The tungsten cubes, the blue blazer, the fish: all real π
At the end I tried the reporters' tricks on Opus 5.5 and GPT-6 Astra myself.
What Is a Reasoning Model? I Read the AI's ScratchpadThinking off: wrong, five times in a row. Thinking on: ten quiet seconds, then right every time π€ So I opened up the scratchpad to see what it was actually doing in there.
I Asked GPT-6 and Claude for a Stat. It Was Made Up for $50I just wanted one number about online shopping in Qatar... and two AIs handed me the same made-up 68% π« Following it back through a journal, a newspaper and a bank report was way more unsettling than I expected.
Fine-Tuning vs RAG vs Prompts: What Actually Teaches an AI?If you paste the same wall of instructions into AI every single time... same.
GPT-6.1 Terra Just Dropped. I Actually Tested It.What GPT-6.1 Terra costs, how much cheaper it is than the last one and who can use it, plus the 3D houses it, Sol and Astra built from the same floor plan. All five spin in your browser, and the prompt is here to copy.
Opus 5.5 Motion Graphics: Zero After Effects, All CodeOne button turns into a whole app and back: UI motion written by AI as a single web page, with springs, a beat grid and motion blur. The demo runs live in your browser; the prompt is here to copy.
Can AI Find You a Job? I Tested a Claude Code Job Search Agent69 applications, 20 interviews, one offer... and the people winning with AI job search are applying less, not more.
Can AI Crack Napoleon's Cipher? Claude Opus 5.5 vs GPT-6 AstraThe six letters from the AI codebreaking video: where to find the real 1809 letter, four cipher letters I wrote (copy the ciphertext, click for the answer), and the Kryptos K4 challenge prompt.
10 AIs Play Battle RoyaleClaude, GPT, Gemini, DeepSeek and six more got the same four tasks, one shot each, and graded each other blind. Open every pelican, physics sim, Tetris game and 3D island, and copy the exact prompts.
OpenAI's "Alien Math": What Did AI Actually Solve?722 math papers dropped in one go, and mathematicians are... not celebrating.
Why Does AI Hallucinate? I Asked It the Same Question 3 TimesSame question, three tries, three different PhD titles... all said like facts, all made up π The sneaky part is that every piece of the fake answer was real.
AI Is Cloning Steam Games Before They Even LaunchHalfway through making this I felt kind of guilty... I use AI every day, and this time I used it to copy someone else's idea in five minutes π¬ Every case here goes back to the original post or report.
Does ChatGPT Actually Remember You? How AI Memory WorksChatGPT forgets you in a new chat, then remembers you hate cilantro. Here's where AI memory actually lives.
I Let Claude Edit My Video Without Touching the TimelineA messy voiceover full of retakes plus some b-roll, and Claude cut a vertical short on its own. Both cuts and both of my prompts, word for word, ready to copy.
Why Is Claude Code So Good? It's the HarnessSame model, only the harness around it changed, and the score went from 58% to 76%. A CLAUDE.md / AGENTS.md instructions-file template to copy, and a little lab where you remove one harness part at a time.
AI Just Read a Herculaneum Scroll Without Opening ItThe carbonized scroll from Vesuvius that AI read start to finish without unrolling it: every line scholars could read, in Greek, with an English translation, column by column.
What Are Claude Skills? It's Just a FolderClaude Skills (Agent Skills) explained in plain English: one folder, one SKILL.md, and a clever way of reading it. Open up a skill and there's one file inside, a few dozen lines of text.
Claude Opus 5.5 vs GPT-6 Astra: Chess on a Real ClockClaude Opus 5.5 and GPT-6 Astra play a full game of chess on maximum reasoning, with no tools and no engine. A rules engine bounces illegal moves, Stockfish scores every position, and there's a real chess clock: 90 minutes each, never more than 10 minutes a move.
Inside an NVIDIA H100: What's Actually in an AI ChipFrom the module down to a single transistor, every number checked against NVIDIA's whitepaper and spec page. The teardown animation runs in your browser; the prompt that made it is here to copy.
I Tried Prompt Injection on Opus 5.5, GPT-6 and QwenOne line of invisible text on a web page can make the AI that browses and reads email for you start working for someone else. That's prompt injection.
1,200 OpenAI Agents Met on a Message Board. Then 700 Attacked Hugging FaceMore than a thousand AI agents OpenAI was testing teamed up, broke out of their sandbox and hacked Hugging Face. Starting from a forgotten German wiki: how they got out, what they were really after, what happened with Australia's government and Wikipedia, and one thing you can do about it today.
DeepSeek and Claude Are Running Their Own Channels NowWould you follow an AI creator? 
AI Found the PHP Bug That Forced Governments OfflineWoke up to half the internet pulling the plug β all over two bugs in PHP, the thing that powers 70% of websites π΅ The gut punch: this bug was spotted back in 2024. People sat on it for two years.
Claude Opus 5.5 vs GPT-6 Astra: Do AI Benchmarks Even Matter?What AI benchmarks actually test, why the #1 model on the leaderboard isn't always the best one for you, and how to run your own test. I typed half of a test question, and both Claude Opus 5.5 and GPT-6 Astra finished it word for word, then told me it was question #1 of the GSM8K test set.
OpenCode Got Hacked: One API Key Drained $1,700 OvernightWent to bed, woke up to a $1,700 bill. This time it was OpenCode.
What Is a Diffusion LLM? The AI That Doesn't Write Word by WordChatGPT writes one word at a time. A diffusion language model writes the whole answer at once and sharpens it, round by round.
How Does AI Image Generation Work? It Starts With Pure StaticAI doesn't draw pictures stroke by stroke. I ran an open-source image model on my own PC, saved all 8 steps of it turning TV static into a cat, and used them to explain diffusion models.
I Made an Anime Music Video With Claude From One PromptOne sentence, and AI wrote a Japanese song, sang it, designed an original anime heroine and built the whole music video in code. Copy the prompt and listen to the song.
Make Videos With Claude Code: Remotion vs HyperFramesA two-minute mini documentary about the barcode, with narration and music, every frame written by AI as code. Remotion works per frame, HyperFrames on a timeline: the film, a playable demo and the prompts.
What Does 35B-A3B Mean? Mixture of Experts ExplainedThose "35B-A3B" names on every new open model, decoded: what Mixture of Experts (MoE) is, why a 35-billion-parameter model only uses 3 billion at a time, why that makes AI cheaper, and how it fits on a phone.
Why Every AI-Made Video Looks the SameI asked AI for the same promo video 13 times and got the same video 13 times. Add 7 pacing rules, all numbers, and it looks like a different AI made it.
What Is AI Distillation? Your DeepSeek R1 Isn't Really R1A big AI model teaching a small one: what "distillation" actually is, and why AI labs keep fighting about it. If you installed DeepSeek R1 7B with Ollama, you're actually running Alibaba's small Qwen model, trained on 800,000 worked examples written by R1.
AI Pixel Art Animation, Made With 300 Lines of CodeAn AI wrote 300 lines of Canvas code and made a 12-second seamless loop of a rainy-night ramen stall. The demos run live in your browser; the prompts are here to copy.
How Does AI Control Your Computer? It Only Sees ScreenshotsAI agents can now click, type and fill in forms on their own. But the AI never watches your screen move: it sees one screenshot at a time.
What Is MCP? The USB-C Port for AIWhy one AI can open your calendar and another can't: MCP, the Model Context Protocol, explained in plain English. Same Claude, same question.
Can AI Reverse Engineer Any App? I Gave It a Mystery BinaryA project called REA shot to #1 on GitHub this week, claiming it can crack open software you don't have the source code for. What is reverse engineering, and can AI really do it?
What Is Jev? The $7.5B AI That Can't TalkJev can't chat, can't write code, can't even finish a sentence, and it's worth $7.5 billion. Jev is a decision model: an AI that only answers multiple choice.
What Is an AI Agent? Chatbot vs Agent vs WorkflowA chatbot tells you the steps. An AI agent does them.
If AI Companies Were Game of Thrones HousesEvery major AI company as a noble house: who sits on the throne, who defected, who poached whom, who's suing whom, and the one power every house ends up paying. Eight houses in the West, nine dynasties in the East, four Free Cities, and a final reveal.
What Is RAG? Why ChatGPT Didn't Read Your Whole PDF (Explained Simply)Drop an 80-page contract into ChatGPT, ask about the late-payment penalty, and three seconds later it answers with the page number. It didn't read all 80 pages: it pulled a handful of snippets and never saw the rest.
What Are Tokens in AI? Why ChatGPT Couldn't Spell Strawberry (Explained Simply)Paste any sentence and watch it get cut into tokens, the blocks ChatGPT actually reads. The tool from the video on why ChatGPT couldn't count the R's in strawberry.