In July 2026 MIT set thirty-one undergraduates a deliberately unreasonable task: four weeks to design, build and run a jet engine, with AI as their main engineering partner. Not a simulation of one. A single-spool gas turbine burning Jet-A, making 50 to 100 pounds of thrust, required to complete five sixty-second runs (MIT News, 2026).
The setup
The JARVIS Challenge — Jet-engine AI Research and Validation Intensive Sprint — put seven teams drawn from across MIT's School of Engineering against that brief, some entirely first-years, some senior-heavy. The interesting design decision was the deadline. Four weeks is short enough that no team could work the traditional way, which is what makes the result worth reading: it forces the question of what AI is actually load-bearing for, rather than what it can be made to help with given unlimited time.
A jet engine is a good choice of target precisely because it punishes vagueness. A turbine wheel either survives its own rotational stress or it disassembles itself at speed. Combustion either stabilises or it blows out. The hardware does not accept a plausible answer.
What the compression actually came from
MIT's own summary of the outcome is that AI can substantially accelerate safety-critical hardware engineering, while engineering judgement remains the decisive differentiator — the distinction they draw is between an engineer who uses AI and one who leads it, knowing when to trust it and when to challenge it.
That distinction has a specific mechanical meaning in a project like this. Most of a first pass at a gas turbine is bookkeeping: matching compressor and turbine mass flows, checking that a chosen pressure ratio and turbine inlet temperature actually close the cycle, converting between units, sizing a shaft for the torque it will see. None of that is difficult, all of it is slow, and errors in it are quiet — the numbers still look like numbers. That work compresses well.
What does not compress is knowing which answer is wrong. A cycle calculation will happily return a compressor efficiency of 0.94 if you ask for one, and that number is not flagged, not out of range, not obviously absurd. It is simply not a thing a small radial compressor built in four weeks will achieve. Recognising that requires having some sense of what real machines do, and that sense is not in the calculation.
Why this makes fundamentals more valuable, not less
The intuitive reading is that if a tool does the calculations, the calculations matter less. The JARVIS result points the other way, and the reason is structural rather than sentimental.
When a step is slow, its cost is the time it takes. When a step is fast, its cost is the confidence it produces. Generate one design and you scrutinise it. Generate two hundred and you have to decide which to scrutinise — and that filtering is not a task the generator can do for you, because it requires the very judgement that would have caught the problem in the first place. Speed moves the bottleneck from producing candidates to discriminating between them, and discrimination is exactly the skill that comes from having done the thing slowly.
Which is a fairly old idea in a new setting. The engineer who can estimate the answer before calculating it has always been the one who catches the mistake, because they are checking against something. What has changed is that there are now far more answers arriving, far faster, and each of them arrives sounding certain.
If you want to build the estimating habit on the machine JARVIS chose, the jet engines course works through compressor, combustor, turbine and nozzle in the order the gas meets them, and the turbojet demo lets you move the cycle parameters and watch thrust respond.
Go deeper: the sanity checks that catch a bad cyclefor engineers
A handful of cheap checks catch most impossible gas turbine cycles before any detailed work begins. Compressor and turbine must pass the same mass flow, plus fuel, minus any bleed — if the two ends of the machine disagree about how much air is moving, nothing downstream is meaningful. Turbine inlet temperature has to be survivable by whatever the blades are actually made of, which for an uncooled metal wheel is a far lower number than published engine figures suggest, because those engines cool their blades.
Then there is the one that catches the most beginners: the turbine has to produce enough work to drive the compressor before a single newton of thrust is available. In a small single-spool engine that is most of the turbine's output. A cycle that quietly assumes otherwise will predict handsome thrust from a machine that cannot sustain its own idle.