Last week, I was fortunate enough to be an early whitelistee by TypeSafe AI to try out their latest model, Jev.
So, me being me, the first thing I wanted to test was how well the model could make non-deterministic events more deterministic. And what better way to test it than running predictions for the next F1 race?
The Simulation
I built a small harness that feeds the same race context to Jev and to three GPT-5.6 variants concurrently, then puts the results side by side. Here it is running live:
The source code is on GitHub: https://github.com/anupa-perera/f1-jev-simulation
What Makes Jev Different
What’s unique about Jev is that it is built for superhuman-level autonomous systems to consume and act on its almost instantaneous outputs with significantly lower costs. To achieve this, it uses RLCD (Reinforcement Learning for Calibrated Decisions) instead of RLHF (Reinforcement Learning from Human Feedback), the training approach used by current frontier models such as the GPTs.
It’s noticeable that this training algorithm had addressed some of the issues we commonly face with current models, like hallucinations, overconfidence, and the resulting lack of reliability.
The Results
What surprised me during this simulation was how quickly Jev performed analytical categorisation on the provided data. It produced a similar level of output to the GPT models while using significantly fewer tokens.
| Model | Elapsed | Tokens | Top pick |
|---|---|---|---|
| jev-1.13.0 | 1.9s | 12,317 | Max Verstappen — 81.0% |
| gpt-5.6-sol | 39.3s | 25,369 | Max Verstappen — 29.0% |
| gpt-5.6-luna | 66.1s | 24,873 | Max Verstappen — 26.3% |
| gpt-5.6-terra | 91.2s | 26,360 | Max Verstappen — 23.0% |
The closest GPT model took around 20x longer for the decision-making process and used roughly 2x the tokens.
The calibration difference is the part worth staring at. All four models agree on the winner, but the GPTs spread their confidence thin across the field, hedging at 23–29% on a driver they all rank first. Jev commits at 81% and drops the tail runners to a flat 0%. That is what a model trained for calibrated decisions looks like versus one trained to sound agreeable.
And then there’s cost. Jev is priced at around $0.042 per 1M input tokens, with output tokens free. Meanwhile, the cheapest GPT model accumulates $0.20 per 1M input tokens and $1.20 per 1M output tokens.
Why This Matters
What this proves to me is that the next frontier of AI is bottlenecked by keeping humans in the loop for each and every training cycle. If we want machines to operate at genuinely machine-level speeds, we may need to start pivoting towards building systems for machines rather than designing every interaction around humans.
Kudos to TypeSafe AI for whitelisting me and letting me try some crazy experiments within 48 hours of the release. I’m really interested to see how this training approach develops and whether algorithms like RLCD would eventually become an integral part of AGI architectures.
What I Want to Try Next
Given that this model opens the door to performing near-instant actions combined with high intelligence, there’s another simulation I’d love to try.
Market arbitrage across multiple open markets.
I’d love to plug Jev into an API provider that exposes real-time market data and see how effectively it can identify and simulate arbitrage opportunities at machine speed.
So oddschecker, dynamicodds, oddsportal, or any other platform that can provide near real-time price updates across a wide range of markets and vendors concurrently — hit me up! I’d love to put this to the test. Purely for academic purposes.
Maybe there’s even an opportunity here for punters to take advantage, or bookmakers to cover their tracks. Or both, who knows.
Would be fun to find out.