AI's inconvenient truth

gm legends, happy Monday.

Today: an AI router that shows you the carbon bill for every prompt, a way to simulate your entire cloud setup and its cost before you provision a thing, and an app that turns the tune stuck in your head into actual music. Plus the forum on the five-minute AI task that quietly ate your afternoon.

discode.ai is an EU-built AI router that turns 100+ models into one interface and, unusually, shows you the CO₂, water, and energy each prompt burns, then auto-routes simple questions to a frugal model so you're not firing a frontier model at a one-line summary. It's the work of Mo Riz, a Vienna non-coder who vibecoded the first version into a 417,000-line monster before a team of engineers carved it into a real product.

🔥 Our Take: Okay, this one's actually clever. You know how every time you ask ChatGPT something it's burning electricity and water in a data center somewhere? Nobody ever shows you that. discode does. It puts the real CO₂ and water cost on every answer, and then it's smart about it, sends the dumb little questions to a small cheap model and only fires up the big expensive one when you actually need it. The founder runs a coffee shop in Vienna, got obsessed with this, and vibecoded a 417,000-line monster before bringing in real engineers to fix it. Only catch, if the cheap model fumbles and you re-ask, you've now burned more than if you'd just used the good one.

Cloud World Model lets you simulate an entire cloud architecture, across AWS, GCP, Azure, OCI, or DigitalOcean, and watch how it behaves under load: cost, latency, autoscaling, even an injected zone outage, all without provisioning a single real resource. Maker Kevin Brown also ships a full RL training API, so AI agents can learn to optimize infrastructure in a sandbox instead of on your live account.

🔥 Our Take: You can build a whole AWS setup in Cloud World Model, blast it with traffic, crash a zone on purpose just to watch it fall over, and it costs you nothing because none of it's real. That's the pitch, and it's a good one, because the whole problem with cloud is you only find out what it costs after the bill lands, usually for the test box you forgot to switch off. People will go "that's just LocalStack" but it isn't, LocalStack fakes the API so your code runs against a pretend AWS. This one actually predicts the real cost and the real failures, and reckons it's about 95% accurate. The founder's weirdly less bothered about the dashboard, what he really wants is AI agents practising on infrastructure in here so they don't learn by torching your live account. Brilliant if the fake cloud is honest, dodgy if it isn't, because an agent will just learn to game the simulation.

Nada comes from an Indonesian team who watched the making-of for Stromae's "Alors on danse" and wondered what would happen if anyone could make music with just their voice: you hum, sing, or whistle a melody and it becomes MIDI you can arrange with instruments and beats, a mini-DAW that started life as a voice-notes app for musicians.

🔥 Our Take: Ever get a tune in your head and just lose it, because you can't actually play anything and every music app looks like a cockpit? Nada's the answer to that. You hum it or whistle it, and it turns the sound into real MIDI you can build a whole track on, instruments, beats, the lot, from your phone. It's made by a few people who watched Stromae build "Alors on danse" pretty much with his mouth and thought, why can't anyone do that. It kind of acts as a voice-memo app for musicians, and honestly that's still the best bit, it catches the idea in the two seconds before it's gone for good. The maker flat out tells you to plug in wired earphones or it mishears your pitch, which in 2026 is a bit of a vibe.

Alex Dochioiu (@alex_dochioiu) named a feeling every AI user knows: the task that should take five minutes, and somehow eats forty, because you're stuck watching the agent in case it quietly stops to ask permission.

Nobody argued the problem, they argued the fix. One camp runs everything in "yolo mode" with permissions scoped tight and walks away. The other won't leave the agent alone with anything that touches production. The useful middle everyone circled: an agent that keeps moving on low-risk calls and only stops when a wrong guess would actually hurt.

Mustafa (@thenameisarian) had the most useful fix, a single line for the system prompt: "never ask clarifying questions, instead make the best guess and clearly label every assumption you made." One wrong guess plus one correction beats three rounds of back-and-forth. Hadi (@hadifarnoud) drew the line on when to stop at all: auto-continue when being wrong costs a small correction, interrupt only when it's user-visible, security-sensitive, or hard to roll back.