There’s a moment in every career, whether it’s athletic, professional, or otherwise, when the thing you’ve built your entire identity around suddenly becomes the thing that exposes you as fundamentally insufficient. As at Atlanta Hawks fan, I watched this moment happen in real-time. On June 20, 2021, the Atlanta Hawks broke Ben Simmons. Of course, this mental break was a train with inertia, and I can’t credit Trae Young‘s nearly non-existent defense for being the final straw, but Atlanta was the team that exposed the inherent flaw in Ben Simmons.
At face value, passing the ball to a teammate isn’t a problem. Matisse Thybulle is a perfectly capable professional basketball player, and in theory, he could have made the shot after receiving the pass from Simmons. Yet this moment would be the end of the line for Ben Simmons and his professional career. It exposed Ben Simmons for what he was all along.
Machine learning is having its Ben Simmons moment, right now, in the summer of 2025. We’re all witnessing it, and we just don’t know we’re witnessing it. But as an avid hoops fan, I can tell you with certainty that the CNN, RNN and FNO models we’ve been carrying since 2017 are being exposed like the flaws in Ben Simmons that fateful night in Philadelphia. A former Rookie of the Year, who had been ascending faster than any player, who during that very season had finished SECOND in Defense Player of the Year Voting… would completely fall apart. (Table below courtesy of Basketball Reference.)
| Season | Age | Team | Lg | Pos | G | GS | MP | FG | FGA | FG% | 3P | 3PA | 3P% | 2P | 2PA | 2P% | eFG% | FT | FTA | FT% | ORB | DRB | TRB | AST | STL | BLK | TOV | PF | PTS | Awards |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2016-17 | Did not play – injury | |||||||||||||||||||||||||||||
| 2017-18 | 21 | PHI | NBA | PG | 81 | 81 | 33.7 | 6.7 | 12.3 | .545 | 0.0 | 0.1 | .000 | 6.7 | 12.2 | .551 | .545 | 2.4 | 4.2 | .560 | 1.8 | 6.3 | 8.1 | 8.2 | 1.7 | 0.9 | 3.4 | 2.6 | 15.8 | ROY-1 |
| 2018-19 | 22 | PHI | NBA | PG | 79 | 79 | 34.2 | 6.8 | 12.2 | .563 | 0.0 | 0.1 | .000 | 6.8 | 12.1 | .566 | .563 | 3.3 | 5.4 | .600 | 2.2 | 6.6 | 8.8 | 7.7 | 1.4 | 0.8 | 3.5 | 2.6 | 16.9 | AS |
| 2019-20 | 23 | PHI | NBA | PG | 57 | 57 | 35.4 | 6.6 | 11.4 | .580 | 0.0 | 0.1 | .286 | 6.5 | 11.2 | .583 | .581 | 3.2 | 5.2 | .621 | 2.0 | 5.8 | 7.8 | 8.0 | 2.1 | 0.6 | 3.5 | 3.3 | 16.4 | DPOY-4,AS,NBA3,DEF1 |
| 2020-21 | 24 | PHI | NBA | PG | 58 | 58 | 32.4 | 5.6 | 10.1 | .557 | 0.1 | 0.2 | .300 | 5.6 | 9.9 | .562 | .560 | 3.0 | 4.9 | .613 | 1.6 | 5.6 | 7.2 | 6.9 | 1.6 | 0.6 | 3.0 | 2.9 | 14.3 | MVP-12,DPOY-2,AS,DEF1 |
| 2021-22 | Did not play – holdout/back injury | |||||||||||||||||||||||||||||
| 2022-23 | 26 | BRK | NBA | PG | 42 | 33 | 26.3 | 3.2 | 5.6 | .566 | 0.0 | 0.0 | .000 | 3.2 | 5.5 | .571 | .566 | 0.6 | 1.4 | .439 | 1.0 | 5.3 | 6.3 | 6.1 | 1.3 | 0.6 | 2.3 | 3.3 | 6.9 | |
| 2023-24 | 27 | BRK | NBA | PG | 15 | 12 | 23.9 | 2.9 | 4.9 | .581 | 0.0 | 0.0 | 2.9 | 4.9 | .581 | .581 | 0.4 | 1.0 | .400 | 2.1 | 5.9 | 7.9 | 5.7 | 0.8 | 0.6 | 1.8 | 2.4 | 6.1 | ||
| 2024-25 | 28 | 2TM | NBA | PG | 51 | 24 | 22.0 | 2.3 | 4.4 | .520 | 0.0 | 0.0 | 2.3 | 4.4 | .520 | .520 | 0.5 | 0.6 | .727 | 0.9 | 3.8 | 4.7 | 5.6 | 0.7 | 0.5 | 2.0 | 2.1 | 5.0 | ||
| 2024-25 | 28 | BRK | NBA | PG | 33 | 24 | 25.0 | 2.8 | 5.2 | .547 | 0.0 | 0.0 | 2.8 | 5.2 | .547 | .547 | 0.5 | 0.8 | .692 | 0.9 | 4.3 | 5.2 | 6.9 | 0.8 | 0.5 | 2.7 | 2.5 | 6.2 | ||
| 2024-25 | 28 | LAC | NBA | PG | 18 | 0 | 16.4 | 1.3 | 2.9 | .434 | 0.0 | 0.0 | 1.3 | 2.9 | .434 | .434 | 0.3 | 0.4 | .857 | 0.9 | 2.9 | 3.8 | 3.1 | 0.7 | 0.4 | 0.8 | 1.4 | 2.9 | ||
| 7 Yrs | 383 | 344 | 31.1 | 5.4 | 9.7 | .558 | 0.0 | 0.1 | .139 | 5.4 | 9.6 | .562 | .559 | 2.3 | 3.8 | .592 | 1.7 | 5.7 | 7.4 | 7.2 | 1.5 | 0.7 | 3.0 | 2.8 | 13.1 | |||||
| PHI (4 Yrs) | 275 | 275 | 33.9 | 6.5 | 11.6 | .560 | 0.0 | 0.1 | .147 | 6.5 | 11.5 | .564 | .560 | 2.9 | 4.9 | .597 | 1.9 | 6.2 | 8.1 | 7.7 | 1.7 | 0.7 | 3.4 | 2.8 | 15.9 | |||||
| BRK (3 Yrs) | 90 | 69 | 25.4 | 3.0 | 5.3 | .562 | 0.0 | 0.0 | .000 | 3.0 | 5.3 | .564 | .562 | 0.5 | 1.1 | .500 | 1.1 | 5.0 | 6.2 | 6.3 | 1.0 | 0.5 | 2.4 | 2.8 | 6.5 | |||||
| LAC (1 Yr) | 18 | 0 | 16.4 | 1.3 | 2.9 | .434 | 0.0 | 0.0 | 1.3 | 2.9 | .434 | .434 | 0.3 | 0.4 | .857 | 0.9 | 2.9 | 3.8 | 3.1 | 0.7 | 0.4 | 0.8 | 1.4 | 2.9 | ||||||
The Promise of Everything, The Delivery of Something “Everything Adjacent”
Let’s rewind the tape and understand the hype. Ben Simmons was supposed to be the future of the NBA. He was Australian, and then went to LSU, where we never quite got a good look at him, because LSU wasn’t a competitive basketball team. At 6’10”, Simmons could handle the ball like Magic Johnson, saw the court like Larry Bird, defended like Kevin Garnett. He was being compared to LeBron James. That’s insane to think about, given where he is today.
There was just one problem to his game: He couldn’t shoot.
Not wouldn’t. Simply could not shoot the basketball. His jump shot was a psychological impossibility; a mental block so complete it made Shaq’s free throws look like Steph Curry’s threes. You don’t need to be a basketball savant to understand that being a professional basketball player is going to mean at some point, you’re expected to personally make the ball go through the basket.
At a high level, Convolutional Neural Networks (CNNs) are the Ben Simmons of artificial intelligence. They’re really good at recognizing patterns in images. Genuinely exceptional, actually. Ben Simmons can see a basketball court, analyze everyone, understand where everyone will go next, where the defense and offense need to make cuts, and then magically diagram an entire sequence of next steps. That’s because CNNs are mathematical in their analysis. They scan images much in the same way Major League turf managers make those cool square and diamond patterns in the outfield. (In theory, I understand how they do it, but when I try the same on my lawn it never works, which is why machine learning is magical.) This criss-cross patterns picks up edges, details and textures, and eventually knows the difference between Joel Embiid and Tobias Harris, sees where they are on the floor, and knows to pass to the player who is cutting into an open area.
Ben Simmons, in short, is a basketball CNN.
The problem with CNNs is that they’re pattern-matching savants trapped in a world that demands actual reasoning. Ben Simmons can tell you that Seth Curry is open, but he can’t tell you why Seth Curry was intentionally left open. Ben Simmons can tell you Matisse Thybulle is open under the basket, but he can’t tell you why passing the ball instead of taking the shot is the least favorable of two outcomes. The pass and shot are simply part of a CNN trying to reason, when it cannot reason. Ben Simmons can analyze basketball and give you an outcome, but it won’t be a positive outcome. It is simply “an” outcome. Cause and effect have no relationship in the world of Ben Simmons.
Brooklyn: The Fourier Delusion
After Ben Simmons completely broke mentally in Philadelphia, Brooklyn Nets General Manager Sean Marks decided that he saw the promise of the next generation of Ben Simmons. (For those of you who have ever uttered the words in a relationship, “I can fix him” — this is your moment.) This was the moment Sean Marks believed he could turn Ben Simmons into a Fourier Neural Operator (FNO). What are FNOs? They’re basically CNNs that got really into math and started hanging out with differential equations. FNOs are designed to solve partial differential equations; the mathematical language of everything from weather patterns to fluid dynamics. (The DARCY model from Nvidia is probably the most well-known, but there are many out there.) They’re supposed to be the next evolution in machine learning, the thing that makes machine learning actually useful for real-world problems. FNO’s will make a decision.
Except they’re still Ben Simmons. They still can’t shoot. There is nothing that will fix Ben Simmons unless Ben Simmons can shoot a basketball into a hoop, against NBA players, in an NBA game.
An FNO takes a function as input and spits out another function as output. It’s like having a basketball player who can perfectly execute every play except the one that actually puts points on the board. You can pass it to Ben Simmons, and he can analyze the court completely… before ultimately passing it to someone else. A FNO is just another way to eventually make someone else take the shot, because once again, Ben Simmons does not have the “make baskets” function as part of his programming. Ben Simmons can analyze, examine, and detect. He can provide a snapshot of the moment. But his default output function is “pass the ball” not matter what the input function might have been.
FNOs can approximate solutions to complex mathematical problems with startling accuracy, but they can’t tell you whether those solutions make sense in the context of reality. They’re the mathematical equivalent of Ben Simmons, who can thread a perfect pass through three defenders from key-to-key, but won’t take an open three-pointer with 15 feet of space between him and the nearest defender. This is why models like DARCY are wonderful for generating weather simulations 3-6 hours in advance of a severe weather event, and can provide almost pinpoint accuracy for rainfall totals in a short context window, but DARCY is almost useless for real-time understanding of something like the hook in a super cell. The most advanced FNO in the world can analyze all the inputs from every weather station, but can’t see a tornado in real time. The tornado, in this metaphor, is the open jump shot. The debris signature is there. The hook is clear as day. The inflow and outflow all match up. Ryan Hall is live-streaming to 100,000 people and… DARCY can’t make the jump shot for you.
The Sean Marks problem is the problem that I see with every LinkedIn “thought leader” when we talk about artificial intelligence. They want to believe that Ben Simmons will someday be able to shoot the basketball. In order for the “rise of AI” to be complete, for Ben Simmons to meet his lofty expectations, he must become Magic Johnson or LeBron James.
The LinkedIn Thought Leader Problem
This dilemma has led to an interesting fork in the road. People like me, who watch basketball and also have a deep understanding of machine learning models know that we’re at something of an impasse. I know that on a winning team, Ben Simmons can only ever be your third best player. This, by the way, is totally fine. There is nothing wrong with having a defensive monster as a point guard, who switches sides and plays the power forward position sometimes. If you’re willing to admit Ben Simmons is deficient at something supremely important (shooting the basketball) you can build a team around him.
The problem is, everyone in the world is having a very hard time believing a number one overall draft pick, a Rookie of the Year, a guy who can examine the floor and see the game of basketball in almost perfect context… cannot shoot a basketball.
I understand that it defies logic. You, a lay-person would look at his stat sheet and say, “He scored 16 points for over four months in a row. Surely he’s making some shots? How can you not be able to shoot a basketball and still score points?” But those “shots” are little 2-foot layups. They’re fast break dunks. They’re offensive rebounds and putbacks. They’re the empty calories of basketball. Do they matter? Yes. But the points that Ben Simmons is putting on the board would be there for anyone in his position.
When asked to execute, in an offense, and Ben Simmons has an open shot, Ben Simmons cannot and will not make an open shot.
Meanwhile, every thought leader on LinkedIn is desperately trying to make “artificial intelligence” sound like the next frontier of human achievement. They’re still in 2017, his rookie year. They’re selling Ben Simmons as the second coming of Michael Jordan, ignoring the fact that he’s literally afraid to shoot the ball.
These people, and you know exactly who I’m talking about, post inspirational content about “leveraging AI to maximize synergistic outcomes” while fundamentally misunderstanding what these systems actually do. They’re not intelligent. They’re not even artificial intelligence. They’re elaborate pattern-matching algorithms that occasionally produce outputs that look like intelligence from a distance, the way Ben Simmons convinced a lot of people he was a basketball player for roughly four seasons.
We’re in RAG and agent hell right now
The real tragedy isn’t that machine learning models are limited. I’m fine with machine learning models being the third most valuable person on my team. I’m happy for the “assist” in the way Ben Simmons can only dish assists. We even talk about “AI assistants” in the sense that they’re supposed to augment what we do. I have embraced Claude Code for all the dishes I need, and I’m happy that it’s there to let me taken the open three-pointer, because I would never, ever fucking trust Claude to win the game for me.
The problem with our current landscape is that we’re making the same mistake that the 76ers, Nets, and now Clippers are making. The more we depend on our current generation of AI models, the worse we get at the things they can’t do. Every time you let a CNN make a decision for you, you’re essentially passing the ball to Ben Simmons in the fourth quarter. Sure, he might make the right pass, but what happens when the game is on the line and you need someone to actually… you know… win a basketball game?
CNNs excel at classification and pattern recognition, but they’re terrible at understanding context, causation, or the kind of nuanced reasoning that actual intelligence requires. FNOs can solve differential equations, but they can’t tell you whether the equation you’re solving is the right one for your problem. They’re computational tools pretending to be cognitive ones, and every time we treat them as the latter, we become a little more like the Philadelphia 76ers. You’re in the Ben Simmons trap. You’re technically skilled but fundamentally unable to perform when it matters.
This is why everything we build right now seems to be a wrapper or a RAG. An entire industry has popped up around the notion that with enough coaching, and mental acuity, Ben Simmons can fix his jump shot. With enough LoRA and QLoRA re-training, we can fix those little bugs in the base model, and add “context” that will magically allow us to fix Ben Simmons. Every organization wants to use machine learning the same way, and every organization thinks they’re going to be the unicorn where their RAG sinks the jump shot.
And I’m here to tell you that Ben Simmons can see the court with all the clarity in the world. You can give him x-ray vision. You can allow him to foul 10 times instead of six. You can make him a foot taller. You can make him live to be 100. And none of those things will ever fix him, because once again, Ben Simmon will not take an open shot. He is mentally incapable.
… And your unique RAG will still never be the best player on your team. It will never actually do the thing you want it to do. It will pass the ball.
The Problem Is Messy: Data is messy. Crawling is flawed. Labeling is incomplete.
Ben Simmons doesn’t have any physical problems that prevent him from making baskets. That’s what is probably the most insane about this whole situation. There are dozens, if not hundreds of videos, with Ben Simmons burying shots in practice. There is this ongoing joke that every offseason is actually the “Ben Simmons Season” because videos emerge of him shooting jump shots over coaches, and dribbling around cones. Ben Simmon is perfectly capable of making baskets in a controlled setting.
Show me an empty gym, and I’ll show you Ben Simmons, future MVP of the NBA. But put him in a playoff game, add some pressure, and suddenly the thing he’d done a million times became impossible. The mental aspect of basketball, the confidence, the willingness to fail, the ability to take a bad shot because it’s better than no shot — it goes completely out the window.
Machine learning models have the same psychological fragility. They work great in controlled environments with clean data and clearly defined parameters. But introduce them to the messy, unpredictable world of real-world problems, and they clank bricks off the top of the fucking backboard. They give you unpredictable, insane responses. They hallucinate and reference ideas far outside of what any sane person would ever do. They give you an answer, but it’s not necessarily the right answer. They pass the ball to Matisse Thybulle, even though there is no one between them and the basket for an open dunk. It’s simply an answer they have used before, hundreds of times. Matisse Thybulle is a good passing target, and we have passed the ball to him hundreds of times, so we should do this again. It’s just a solution that fits the pattern of answers they’ve been trained to provide.
Why AI is a bad teammate: The Joel Embiid Problem
The most insidious part of the Ben Simmons era wasn’t his individual limitations. We knew those. It was how those limitations infected the entire team like a virus. Joel Embiid, a generational talent, found himself playing around someone else’s mental blocks. The offense became constrained by what Ben Simmons couldn’t do, rather than optimized for what everyone else could do.
To put a super fine point on this: I’m witnessing in real time as enterprises build around the known deficiencies of AI-enabled tools, rather than embracing the generational talent in front of them. You may have a generational talent on your team, right now, and you’re forcing that talent to play with Ben Simmons, through Ben Simmons, and building your team around Ben Simmons.
This is where we are with machine learning. We’re designing systems around the capabilities of models that are fundamentally limited, instead of using them as tools to augment human decision-making. We’re letting CNNs drive our understanding of image recognition instead of recognizing that image recognition is just one small part of visual intelligence. We’re letting FNOs dictate our approach to mathematical modeling instead of understanding that mathematical modeling is just one component of scientific reasoning.
Of course, this analogy isn’t perfect, because Joel Embiid has injury issues, and bigs rarely do mileage well, but it’s an analogy, damnit. We all know that Embiid wasted his prime with Ben Simmons, and there is no reason the Hawks should have been able to beat them. (I say that as a Hawks fan.)
The uncomfortable truth is that you, as a human being with actual intelligence, will always be better at making decisions than any machine learning model currently in existence. The models can help you process information, identify patterns, and automate routine tasks, but the moment you start deferring to them for actual decision-making, you’ve basically told Joel Embiid to let Ben Simmons run your offense.
The tragedy of the 2021 Philadelphia 76ers wasn’t that they lost to the Atlanta Hawks. It was that they lost because their point guard was afraid to play basketball. They lost because they ignored a generational talent right in front of them, because they wanted to believe that someday, Ben Simmons would become a basketball player. The tragedy of machine learning in 2025 isn’t that the models are imperfect, it’s that we’ve convinced ourselves they’re something they’re not.
Ben Simmons is still in the league, technically. He’s still getting paid. He’s still passing the ball. And somewhere in Philadelphia, there’s probably a LinkedIn influencer writing a post about how his “collaborative approach” and “team-first mentality” make him the perfect example of modern leadership.
The game is about putting the ball in the basket. Remember that your “AI Assistant” will only ever dish assists. It’s not Stockton. It’s Simmons.