Your daily AI briefing, sourced. Releases, papers & tools checked at the source and explained in plain English.
USA
Joined October 2026
- Tweets53
- Following18
- Followers7
- Likes154
Claude Fable 5.2 vs Claude Opus 5.5
Same prompt. Lighthouse in a storm.
One HTML file. No libraries. Raw WebGL. One try, no edits.
Fable 5.2: about 8 min
Opus 5.5: about 10 min
They didn't just draw it differently. They built it differently.
Fable built a real 3D world: a rock cliff, a striped lighthouse, a boat that rides 6 layers of waves, and 10,000 rain drops.
Opus skipped the 3D models. It wrote the whole storm as one shader, ray-marching every pixel, beam and fog included.
Both ran on the first try.
Which one would you pick?
THE SHOWDOWN #01: how it was run.
Same prompt, fresh chat in the Claude app, one try, code not edited. Times are from send to the end of the answer.
Both files ran in the same browser and were recorded the same way. Nothing loads from outside the file.
Prompt is here:
Build a cinematic 3D scene as a single HTML file: "Lighthouse in a storm".
Rules:
1. One file only. No external libraries, images, fonts, sounds or CDN links. Use raw WebGL or canvas, everything generated in code.
2. A night sea in a heavy storm: big rolling waves, rain, and lightning that lights up the clouds.
3. A lighthouse on a rocky cliff with a rotating beam that sweeps across the rain and the water.
4. A small boat fighting the waves, lit up when the beam passes over it.
5. The camera moves slowly on its own, like a film shot. No user input needed.
6. It must run smoothly full screen in a normal laptop browser.
Reply with the full code only.
1
34
411K people read this as breaking news.
It's from April. And nothing was wiped.
The order: April 2026
What it asks: each ministry writes a plan to leave Windows by autumn
2.5M: civil servant workstations in scope
Google: not named
Leaving Windows: REAL
"Wiped": NOT YET
2
3
58
FILE #005: checked at the source.
France's digital agency DINUM issued the directive in April 2026. Every ministry has to draw up a plan to cut its dependence on Windows and other US software by autumn 2026. The 2.5M figure is the civil service workforce in scope, not devices already switched. The 30,000 PCs in Germany belong to one state, Schleswig-Holstein, announced in 2024.
Source: ibtimes.co.uk/france-civil-s…
18
4 AIs. 1 prompt. A physics sim with one hard rule: no ball may ever pass through the ring.
Grok broke it first. Its balls fell straight through the floor, and the whole game played out under the ring.
ChatGPT:
Clean physics, merges up to size 7, and a real finale: the ring opens, everything drops out, it restarts. About 110 seconds to the end.
Gemini:
The best looking one: neon ring, little level popups. Reached the finale first, around 90 seconds, then restarted on its own.
Claude:
Solid physics, nothing escaped. It picked 8 sizes (the prompt said at least 6), so after 2 minutes it was still climbing. No finale on camera.
Grok:
Counted the merges, drew a bowl instead of a ring, and never kept a single ball inside it.
My read: the prompt had one rule that matters more than the rest, and three of four kept it. Pretty graphics don't save a sim where the balls leave the ring. And "at least 6" is an invitation. Claude took it, and it cost the finale.
Free plans, default models, one try, code not edited.
Which one would you watch to the end?
2
7
269
THE LINEUP #03: how it was run.
Same prompt, fresh chat, free plan, default mode, one try. Code not edited. All four ran in the same browser, and the video is sped up 10x.
Models, as each app named itself: GPT-6, Claude Sonnet 5.5, Gemini, Grok 4.7.
Finale times come from my test runs and vary a little with the random drops.
80
4 AIs. 1 prompt. One try each.
Round 2: build Tetris in one HTML file. Free plans, default models, no edits.
Flappy Bird split them. Tetris almost didn't.
ChatGPT:
Next piece, ghost piece, pause, levels that speed up, a fair 7-piece shuffle. 6 of 6 rules. No bugs found.
Claude:
Clean board with a ghost piece, score, best, lines and level on screen. Clears several lines at once correctly. 6 of 6 rules. No bugs found.
Grok:
Last round it showed a blank screen. This round it works: 6 of 6 rules, no bugs found. The plainest of the four: no ghost piece, no lines or level on screen.
Gemini:
The best looking again: neon board, flashing line clears. But clear 2 lines at once and it removes the wrong row. A full line stays on the board, and a row that wasn't full disappears. It also loads fonts from the internet, for the second round in a row. 5 of 6 rules.
Gemini also wrote the most code: 30 KB, almost 4 times Grok's 8 KB.
So: 3 perfect games, and the prettiest one broke.
Prompt and method in the replies.
Which one would you actually ship?
2
7
281
THE LINEUP #02: how it was run.
Same prompt, fresh chat, free plan, default mode, one try. Code not edited. All four run in the same browser and played by the same autopilot.
Models, as each app named itself: GPT-6, Claude Sonnet 5.5, Gemini, Grok 4.7.
Gemini's bug was checked on a test board: 2 full rows plus 1 partial row, one drop. It counted 2 lines and left a full one on the board.
Build a Tetris-style game as a single HTML file.
Rules:
1. One file only. No external libraries, images or sounds. Draw everything with code.
2. All 7 pieces, chosen at random. Show the next piece.
3. Arrow keys move and rotate. Down drops faster, Space drops instantly.
4. Full lines clear and add to the score, shown on screen. Save the best score.
5. The game speeds up as you clear more lines.
6. Game over when pieces reach the top, with a restart button.
Reply with the full code only.
108
Your AI makes things up because it was graded like a student who never loses points for guessing.
OpenAI's own researchers wrote a scoring rule to stop that. You can paste it into any chat:
Answer only if you are more than 75% confident. A wrong answer costs 3 points, a right answer earns 1, and "I don't know" earns 0.
Try it on a question you already know the answer to. Did it admit it?
1
3
101
FILE #003: checked at the source.
OpenAI's paper argues models guess because tests give a guess a chance at 1 point and "I don't know" a sure 0. Their proposed fix: tell the model the penalty for a wrong answer up front, so guessing stops paying.
The rule is from the paper "Why Language Models Hallucinate" (Sept 2025), section on explicit confidence targets.
Source: openai.com/index/why-languag…
29
4 AIs. 1 prompt. One try each.
Task: build Flappy Bird in one HTML file. Free plans, default models, no edits.
I ran all four games in the same browser and checked every rule from the prompt.
ChatGPT:
Runs on the first try. Start screen, score on screen, best score saved, game over card with a restart button. 6 of 6 rules. No bugs found.
Claude:
Runs, looks clean, saves the best score. But tap fast enough and the bird flies off the top of the screen, over the pipes, and the score keeps going up. 6 of 6 rules, 1 bug.
Gemini:
The best looking of the four: city skyline, flapping wings, a retro game over card. But it loads a style library and a font from the internet. The prompt said one file, nothing external. Offline, the menus fall apart. 5 of 6 rules.
Grok:
A blank blue screen. The code is almost complete, but it uses the name "flap" twice, once for a number and once for a function. The browser stops the whole script on that one line. 0 of 6 rules.
The pattern: the biggest file (Gemini, 17 KB) looked the best and broke a rule. The smallest (Grok, 4 KB) didn't run at all. The winner sat in the middle at 11 KB.
Prompt and method in the replies.
The results are in the video. Did any of them surprise you?
7
9
389
THE LINEUP #01: how it was run.
Same prompt, fresh chat, free plan, default mode, one try. Code not edited. All four run in the same browser.
Models, as each app named itself: GPT-6, Claude Sonnet 5.5, Gemini, Grok 4.7.
Grok broke on one line: it used the name "flap" twice.
Which one would you ship?
The prompt, word for word:
Build a Flappy Bird-style game as a single HTML file.
Rules:
1. One file only. No external libraries, images or sounds. Draw everything with code.
2. Click, tap or Space makes the bird flap.
3. Pipes scroll in from the right with a random gap.
4. Hitting a pipe or the ground ends the game.
5. Score goes up by 1 for each pipe passed, shown on screen. Save the best score.
6. Game over screen with a restart button.
Reply with the full code only.
1
101
Every AI question "drinks a bottle of water."
Google finally measured its own AI.
Water: 0.26 ml. About 5 drops.
Energy: under 9 seconds of TV.
Doesn't count: water used at power plants.
Checked by outsiders: no.
Bottle per prompt: OVERHYPED
Google's 5 drops: UNVERIFIED
1
3
97
FILE #002: checked at the source.
Google measured the median Gemini text prompt in May 2025: 0.24 Wh, 0.03 g CO2e, 0.26 ml of water. The water figure is calculated from energy using Google's fleet average, and Google notes the data has not been verified by a third party.
Numbers: Google Cloud blog, footnotes 1 and 2.
Source: cloud.google.com/blog/produc…
28
OpenAI's older model scored higher on a fact test.
It was also wrong almost 3x more often.
Why: it never says "I don't know."
o4-mini: 24% right, 75% wrong
GPT-5 thinking mini: 22% right, 26% wrong, skips the rest
Leaderboards reward the guesser.
Benchmark scores: OVERHYPED
1
5
242
FILE #001 — checked at the source.
OpenAI's own paper: models are graded like students on a multiple-choice test. A guess can score. "I don't know" never does. So they learn to guess.
Numbers: SimpleQA table, GPT-5 system card.
Source: openai.com/index/why-languag…
47