< backWhen the iPhone 4 was released to the public, there were many reports that the phone's cellular connection would drop if it was held in a particular way. Shortly after, the infamous quote from Steve Jobs was born - "You're Holding It Wrong".
While working on my GTO poker solver, I have been taking full advantage of my student subscription to Cursor as graduation approaches rapidly. In an attempt to learn proper agent and LLM usage, my goal was to see if coding agents could produce a poker solver on par with popular open source versions. I have attempted to produce a solver through various large language models, such as Opus 4.5, Kimi K2.5, or GPT 5.2, and every time, the final product produced by the LLM has some fundamental flaw that cannot be fixed once the initial version of the solver is complete. Each iteration of the LLM output is like the product of a stubborn developer who is merely meeting the requirements, and failing to understand the bigger picture of what is being produced. The LLM did not develop a piece of software that is "90% complete", it produced a broken and unmaintainable product.
After my initial attempt, I complained to a friend regarding Cursor and Opus 4.5 not being able to produce a working Poker solver. "Did you use plan mode?", he replied. To be fair, I did not initially, but I had provided detailed instructions outlining the implementation details of the solver. I utilized plan mode for the next four generations of the AI-developed solver. Every time an LLM produced a failed solver, I would receive feedback that implied I could have prompted the LLM in a better way. Did you use plan mode? Did you use micro-tooling? Did you tell the LLM it was a poker expert and on par with Phil Ivey? You're Holding It Wrong.
In my most recent attempt, when planning with Kimi K2.5, I had specified a starting 100bb stack for each player in my plan. I did not imagine that I would have to specify that the solver cannot allow stack sizes to dip into the negatives. Fix it in the next plan. The solver allows a check vs a bet? Fix it in the next plan. The LLM added a "press S to solve shortcut" which prevents me from entering a flop containing spades? Fix it in the next plan; You're Holding It Wrong.
Are the recent developments in LLMs for coding, more specifically agents, just masking a fundamental flaw in LLMs? After a certain level of complexity, LLMs are unusable. How much more compute power can we throw at this problem? After working with agents, I have gone back to prompting Gemini for each individual question, piecing concepts together and slowly and reviewing generated code by hand. I cannot retain information at a faster rate than I could five years ago. LLMs are irreplacable for asking difficult questions on demand to make learning more efficient, but outside of learning, the code produced en masse is unreliable, unsafe, and fundamentally broken. You're Holding It Wrong.
I'm told that one day, after enough data centers have been produced, we will reach a point where LLMs can fix the bugs produced by previous generations. I am not an LLM denier - I understand how useful this technology can be when used properly. I just don't know if I believe that producing this amount of code on demand is the proper realization of this industry-changing tech. For now, I have included a 'slop' folder in TurboFire, providing a digital graveyard to the soulless and unusable output of these coding agents.