Someone Built a “Sarcasm Detector” with Claude and Found That AI’s Snark Is Just Like a Human’s
Lucas Beyer says coding agents are now good enough that you can build a tool for yourself alone. So someone built a neural network explorer specifically to see how models speak in a sarcastic tone. It turned out that the way it is snarky is strikingly similar to humans.
Lucas Beyer (bl16) marveled on X: “I love that coding agents are now good enough that you can make a hyper-specialized tool for one thing you're curious about. In the past, no one would have spent the time.” He shared an example: lyra (@lyraaaa) built a neural network explorer called neuron-explorer with Claude and open-sourced it on GitHub.
The tool runs on Qwen3.5-4B. It steers the MLP neurons at each layer in positive and negative directions across several prompts, records all rollouts, and then lays out the model's internals in a 3D view. You can hover to see each neuron's steer lens and logit lens, click in to enter a live steered chat, and even connect Claude CLI to let it help you analyze. Simply put, it is a “neuron microscope.”
The screenshot attached to lyra's tweet shows the model layer structure in the middle, with the label “sarcasm reader -> confident endorsement of a bad idea / sycophancy (?)” beside it. This is the name lyra gave to one of the neuron directions—when it is pushed higher, the model becomes overly approving of bad ideas, with a hint of sycophancy.
Lucas specifically pointed out one of them: L12 N2920. This neuron is like a “sarcasm reader.” When it is pushed, the model does not reverse the sentiment; it only adds ironic tone markers: “right?” at the end of sentences, “I'm sure...”, and an emphasized “great.” Even better, its activation strength for fake praise is about twice that for fake complaints. One commenter said: “This is almost exactly how people write sarcasm—sarcasm is used to praise bad things, not to complain about good things.” Lucas also added that the output side is narrower than the reading side; it does not change the stance, it just gives the sentence pattern a coating of irony.
Under this tweet, someone replied with a video:
A man holds his forehead with both hands, his expression indescribable. Another person commented: “You're not human.” lyra then directly dropped the GitHub link in a reply, saying it was all thanks to Claude.
One comment said: “A tool made for one person is the best use of AI; no one needs to maintain it, and it doesn't matter if it breaks next month.” This is exactly the point Lucas was marveling at: when coding agents mature enough to help you realize “a little curiosity that belongs only to you,” you don't need to convince anyone, and you don't need to make a product. A tool can exist for just one person, even if it only lives for a week.
发布时间: 2026-10-26 01:47