I had the new AI classification model Jev bet on the NFL for the first 2 weeks. Here’s how it went.
A new AI model that can’t write a sentence, a blind research pack for every game, five seasons of testing, and real money on the line. Every pick, including the ones that lost.
For those that don’t know, a company called TypeSafe released a new AI model called Jev on the 15th of September. I got access the same week, and I wanted to find a creative and cool way to test it. So I let it bet on the NFL for me.
I’ve been running this for the first two weeks of the season with my own money. Here’s everything: what Jev actually is, how I set it up, the strategy, and every single pick, including the ones that lost.
What is Jev?
Jev is not your standard LLM like Claude or ChatGPT. It’s a classification model, which means it actually can’t write a single sentence. You give it information and a set of questions, and it gives you back answers with a probability on each one. It answers all of them at once, in under a second, and it’s super cheap to use.
Think of it as a decision maker. Scoring leads, sorting through support tickets, flagging risk. Any call a business makes over and over again, every single day, is the kind of thing Jev is built for.
Betting on football obviously isn’t a business workflow, but it’s a really good test. There are real outcomes every week, the bookies are very good at what they do, and there’s no way for me to fudge the results.
How I set it up
For every game I built a research pack. It has each team’s offensive and defensive rankings, points scored per game, recent form, the quarterback’s experience, the injury report, rest days, travel and the weather.
Then I took out the team names and the betting odds. Jev never knew who was playing. It just saw “Team A vs Team B”.
I asked it 16 questions for every game. Who’s likely to win, by how much, how many total points will be scored, a set of matchup questions like which quarterback has the edge, and one more that ended up being the most important of all: is this read clear?
I also asked every game twice with the teams swapped and averaged the two answers, so the order they were listed in couldn’t sway it.
Testing it before putting money down
Before I bet a single dollar, I ran Jev on five seasons of past games. That’s 1,359 games and 2,718 calls to Jev. It took under three minutes and cost 37 US cents in total.
I tested 24 different strategies. Most of them didn’t work, and honestly that was the most interesting part.
Just picking winners was the obvious first idea. Jev got 61–67% of games right, but the bookies’ favourite wins 65–69% of the time, so Jev was actually slightly worse than just backing whoever the odds say.
Betting only on its most confident picks won 77% of the time and still lost money. Those teams are big favourites paying around $1.26, so you win small amounts often and the odd loss wipes it all out.
The one that fooled me was a model that blended Jev’s answers with the betting odds. On the seasons it learned from, it showed +17.8%. On seasons it had never seen, it did −9.7%. It had basically just memorised the past.
The strategy that survived
Only one strategy made money on seasons it had never seen: betting on the total points in a game, but only when Jev says its read is clear.
The way it works is fairly simple. Jev gives every game a clarity score out of 1, and most games come back around 0.2 to 0.3. My signal to bet is when the clarity score is 0.4 or higher and Jev’s points prediction is at least a point off the bookies’ line.
It was positive in every single season. The crazy part is that if you run the exact same method without the clarity score, it loses about 4%.
The whole edge is Jev knowing when it doesn’t know.
On top of that, I also back Jev’s two most confident winners each week, head to head. They’re short odds so they don’t pay much, but they make it more fun to follow along.
Week 1: 3 from 3, +A$51
I put A$100 on each week, split across whatever bets qualified. These are all games Jev had never seen.
On paper this one looks like an under. Cleveland had the 31st ranked offence in the league. What Jev picked up on was that Jacksonville were scoring 27.9 points a game by themselves, and every offence and quarterback question leaned their way. It wasn’t expecting a shootout, it was expecting Jacksonville to pile on. Final score 34 to 10, so 44 points, and Jacksonville scored 34 of them.
The two locks were straightforward. Jacksonville had the edge on offence, quarterback and situation. For Pittsburgh, the biggest reasons were experience and health: their quarterback had 278 career starts, Atlanta’s had 16, and Atlanta had four regulars out including an offensive lineman.
Week 2: 2 from 4, −A$25.50
Jacksonville came in with the number one defence and the sixth best offence, and Denver had the third best pass defence. Add 12 mph wind in Denver and the defences just won. It finished 20 to 13, so 33 points, a miss by 12. Not close.
This was the clearest read Jev gave in the two weeks. San Francisco had the fourth best offence and 10 days’ rest at home. Miami had flown 4,111 km, was playing at 4:25pm on its body clock, and had the 26th ranked pass defence. Final 35 to 13, 48 points.
This one is my favourite even though it lost. Jev’s most confident pick of the week was actually an underdog. It looked at Jacksonville’s number one defence and couldn’t see them losing. The bookies looked at a 2,363 km trip to altitude and made them the outsiders. The bookies were right, and Jacksonville lost by 7.
Where it’s at
In week 1 Jev went 3 from 3 and made A$51, which was cool to see. In week 2 it went 2 from 4 and lost A$25.50, which leaves me A$25.50 in profit from A$200 staked.
Picking straight winners, Jev only got 6 from 16 right in week 1 and 9 from 16 in week 2. As a tipster that’s pretty bad, which is exactly why I don’t bet it that way.
Two weeks and seven bets isn’t enough to prove anything either way. What I actually want to find out is whether it can make 8–10% across a full season.
Week 3
Friday’s game is Atlanta at Green Bay, and Jev passed on it. It had the game at 43.7 points against a 42.5 line, a slight lean to the over, but it only rated the read 0.30 for clarity, which is under my 0.4 signal. It does like Green Bay to win at 68.7%, but that isn’t one of its two most confident picks of the week, so no bet there either.
The final injury reports come out close to kickoff, so the Sunday picks get run on Saturday and not before. The early read is leaning towards the under in Buffalo vs the Chargers, which is the clearest read it has given all season. I’ll post the full card on Saturday.
Same as always: A$100 this week, no chasing losses, and I’ll share the results either way. I’m looking forward to seeing how it does.
To make it clear, this is not gambling advice or tips. It’s something I decided to do with my own money to test the model. 18+, gamble responsibly.
Want to try Jev yourself?
Join and I’ll send you my Jev starter kit: how to get access, the exact questions I asked it, and a copy-paste setup to try it on your own leads. Plus the week 3 results next Thursday.