Jay Kess

🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵

🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵

Elon says FSD 14.3 is coming. But if you’ve been following along, it was also “two weeks away” a few months ago. That’s drawn a lot of criticism, understandably. Let’s step back and talk about what’s actually going on: engineering reality. I’ve spent years running engineering teams at Apple and Rivian, and what you’re seeing here is not unusual. Not even a little. I’m not here to defend Elon or say communication couldn’t be better. It could. But what’s happening behind the scenes is far more ordinary than people think. First, understand what kind of company Tesla is. Tesla exposes more of its internal process than most companies— you’re watching how the sausage is made, often in real time. Compare that to Apple. Products appear at a moment in time, fully formed. What you don’t see are the features that slipped, were cut, or quietly postponed to make the deadline. Most companies communicate through layers of marketing at discrete events (e.g., NVIDIA GTC). That may include a CEO keynote—but it’s still tightly controlled. Tesla, largely via Elon, doesn’t. And that creates friction. Most people are used to being in the dining room. With Tesla, you’re watching the sausage get made whether you like it or not. If that makes you uncomfortable, this model will drive you crazy no matter how it’s explained. Now, about FSD 14.3— the so-called “reasoning” release. My view: when Elon originally referenced it, it was real. It was on a roadmap with a timeline. But then reality hit. Somewhere along the way, engineering discussions likely exposed a fork: ship what’s partially there, or go deeper and "do it right". That kind of shift happens constantly. Plans change. Timelines slip. This is normal engineering behavior, not dysfunction. The difference is: you’re seeing it. At companies like Apple, those decisions are invisible. Deadlines are protected by cutting scope. At Tesla, you’re watching the scope evolve in real time. On the technical side, 14.1 and 14.2 were already producing “reasoning tokens,” as Ashok (Tesla AI VP) noted. But producing tokens isn’t the same as using them effectively. 14.3 appears to be where those tokens actually start driving behavior, more human-like decision-making in edge cases. My guess is this is where things got more complicated. The work likely started to overlap with what xAI is doing. At that point, the question becomes: do you ship an interim solution, or integrate a more capable reasoning layer? That’s not a small decision. And it likely has downstream impact— potentially even on Robotaxi timelines— because these same reasoning challenges show up there too. So the team probably made a call: go deeper, even if it costs time. And here’s the part people underestimate: great engineering teams often convince themselves the extra work is worth it… and that it won’t take that much longer. They’re usually wrong on the timeline. But often right on the outcome. At this stage, FSD isn’t about raw safety (it seems to have nailed that)— it’s about behavior. Making decisions feel natural, human, predictable in edge cases. That’s a much harder problem. So if you’re following Tesla closely, the best thing you can do is understand the process and accept the messiness that comes with it. If you want tightly controlled messaging and polished delivery, companies like Apple exist for that. Tesla is something else entirely. Fire away.

How every NFL team did this offseason:

🚨 Elon Musk says physics has stagnated and AI could overtake scientific research. He says progress has relied on expensive hardware like colliders and telescopes, that slows discovery.

I spoke to Anthropic’s AI agent Claude about AI collecting massive amounts of personal data and how that information is being used to violate our privacy rights. What an AI agent says about the dangers of AI is shocking and should wake us up.