every wave in this video is playing a note
I had Opus 5.5 rebuild the Sea Organ in Zadar, Croatia, cut it open, and let the rendered sea play it
35 pipes, klapa chords, zero recorded audio
the real one is on the Zadar waterfront in Croatia, built in 2005 by Nikola Bašić. waves push water up 35 pipes under the steps, and the air they push out makes the stone sing
Opus 5.5 honestly blew me away here. it pulled the real island skyline from elevation data, built the whole thing, and wired the sound to the waves
if you're ever in Croatia, go sit on the steps at sunset
en.wikipedia.org/wiki/Sea_or…
216
We gave an AI five buttons and one job: cross the road.
Jev, @typesafeai's model for small typed decisions, steered a courier across six lanes of traffic on 200 freshly generated roads. It read the road as text, not pixels, and the game paused while it decided.
We asked it two ways.
One big question: here is every car's position and speed, press a button. It crossed 1 road out of 200. On 180 roads its first move was UP.
Small questions: code first works out where each button leads, then Jev makes the call. It crossed 50.
Same model, same roads. Only the way we asked changed.
Plain code looking three moves ahead still crossed 83. When the job is arithmetic, let code do the arithmetic and use the model for the judgment on top.
The video replays one example road. The full numbers are at the end.
6
1
17,368
Replying to @StipePoint8 @typesafeai
Fine print: Jev read the road as text, not pixels. The game paused while it decided. One small simulated game, one model version.
Every crash counted. Total bill: $0.12 🚗
1
7
Replying to @StipePoint8 @typesafeai
Also fair to Jev: it's a fast gut-feeling model. No thinking step, answers in ~250 ms, and its own docs say it's not a calculator.
Doing collision physics for every car in its head was never going to work. We tested that on purpose.
1
10
Replying to @StipePoint8 @typesafeai
Honest caveat: this game is 100% solvable by code. The planner crosses every road.
So the AI could only tie or lose. This tested how you ask a model, not whether AI beats code.
One big messy question: 1 road.
Small clear questions: 50.
1
10
Replying to @StipePoint8 @typesafeai
The AI didn't lose by being dumb.
It pressed the same button as the code on 83% of moves. It stepped into a move labelled "you get hit" once in 2,409 tries.
It lost on coin-flip moments: moves the code rated equally safe, with the trap just beyond its view.
1
13
Replying to @StipePoint8 @typesafeai
Scoreboard, same 200 roads:
🤖 AI reads raw numbers: 1
🤖 AI + code doing the maths: 50
🧮 Plain code, no AI: 83
🎲 Random buttons: 0
🗺️ Planner that sees the whole road: 200
Yes, the calculator won. That's the interesting part 👇
1
46
1/
We let an AI stop and think only 3 times while crossing a busy road.
It chose when to think: 59 of 200 roads crossed. A coin chose: 62. Perfect timing: 169.
Thinking helps. Knowing when to think is the skill, and this AI didn't have it.
5
12,612
5/
Other runs on the same 200 roads: No thinks: 49 Code that thinks whenever two moves look equally good: 81
Jev vs coin: −1.5 points, p = 0.74 (exact McNemar test). No detectable difference.
18
4/
Why it failed: 72% of Jev's thinks were spent standing on the safe pavement.
Right before its fatal moves it was slightly more confident than usual (median 0.85 vs 0.78).
Its feeling of doubt didn't point at the danger.
20
3/
"Unsure" means Jev's own confidence score is below 0.30. We picked that cutoff on 40 separate practice roads before the real run.
The coin thinks at the same pace, so both spend their thinks equally often.
23
2/
How it works: the AI (Jev, TypeSafe's fast decision model) picks every move from a plain-language summary of the road.
A "think" hands one move to a planner that sees the whole road. 3 per road. Only the timing changes between runs.
71
2/
How it works: the AI (Jev, TypeSafe's fast decision model) picks every move from a plain-language summary of the road.
A "think" hands one move to a planner that sees the whole road. 3 per road. Only the timing changes between runs.
1
We got 100 correct invoice approvals for about two cents in AI processing.
The key was how we divided the work.
We tested Jev, TypeSafe’s AI, following its creators’ recommendation: break a decision into smaller questions and combine the answers in code.
Our test included 100 synthetic invoices that all met the payment rules. The correct result was clear: every invoice should be approved.
For comparison, we first asked Jev to handle everything at once: read the documents, check the rules and decide whether to approve payment. It approved 42.
Then we split the task: Jev identified the price, quantity and supplier status. Ordinary code checked whether those facts matched the order and satisfied the payment rules.
That workflow approved 100 out of 100.
Here are the correct approvals and AI processing costs for those same invoices:
Jev + code: 100/100 — 2.1¢
GPT-5.6 Terra: 97/100 — 23.8¢
Claude Sonnet 5: 97/100 — 37.3¢
GPT and Claude handled the whole decision in our earlier experiment. Their costs already include batch discounts. Jev’s costs do not use a batch discount.
The comparison shows the benefit of giving Jev a specific role: identify the facts, then let code make the checks we can define precisely. That division of work produced more correct approvals at a lower AI processing cost in this test.
🤖 Made with AI
1
92
Sources for anyone who wants to explore:
1. What Jev is and how it works:
docs.typesafe.ai/introductio…
2. How Jev produces answers in parallel, including TypeSafe’s explanation of the speed comparisons:
typesafe.ai/blog/introducing…
3. The invoice test: cost, speed and performance comparisons:
evals.typesafe.ai/invoice_pr…
4. How the tests were scored, including how other AI models generated the reference answers:
evals.typesafe.ai/
8
This is why fraud was experiment one.
At this scale, “block” isn’t just a prediction. It’s a decision between two costs: fraud you miss and legitimate payments you stop.
Can an AI follow that tradeoff when the correct action is known?
github.com/StipePoint8/decis…
33


