July 22, 2026
AI finally learns to say “my bad”
Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong
This tiny AI knows when to panic — and the comments want receipts
TLDR: Cactus unveiled a small AI that can score its own confidence and call in a bigger model when it’s unsure, aiming for faster private answers without as many mistakes. Commenters were intrigued but immediately demanded proof, comparisons, and joked about whether AI can now detect “janky” vibes too.
A startup called Cactus just rolled out a small AI that does something people desperately wish more chatbots would do: admit when it might be wrong. The pitch is simple enough for non-experts: let the fast, private, on-device model answer when it feels confident, and if it starts sweating, pass the question to a bigger, stronger AI instead. Cactus says this lets its small Gemma-based model keep up with a much larger cloud system on many tests while only handing off a slice of the work.
But the real action was in the crowd reaction, where the vibe was less “wow, neat demo” and more “okay, explain the magic trick”. One commenter immediately zoomed in on the big claim that the model has a kind of built-in self-awareness signal and basically asked: are you really predicting the chance the AI is wrong from its own internal state, and where’s the deeper evidence? Another sniffed around for possible overlap with Goodfire’s research, which is classic internet code for: is this breakthrough new, or is it remix season?
Then came the ambitious dreamers. One person wanted coding tests with a fallback to a larger local model; another went fully sci-fi and wondered if the same trick could detect whether code feels “clean” or “janky,” like turning AI vibes into measurable data. And in the middle of all that theory, one builder casually dropped a for-fun integration project, giving the thread a nice chaotic “I already hacked this into my desktop” flex. The mood? Equal parts curiosity, skepticism, and please show us the benchmarks before we fall in love.
Key Points
- •Cactus says it post-trains models with embedded probes that return a structured confidence score from 0 to 1 for each answer.
- •The initial rollout centers on Gemma 4 E2B Hybrid, which Cactus says can match Gemini 3.1 Flash-Lite on most listed benchmarks by routing only 15%–35% of queries to the larger model.
- •The article provides benchmark handoff ranges across FP16, 4-bit, and 3-bit settings, with fallback rates generally increasing as quantization becomes more aggressive.
- •Cactus states that quantization quality was measured on Cactus Quants and advises developers to benchmark Unsloth, GGUF, and MLX quantization independently.
- •Usage examples are provided for Cactus Compute, MLX, Transformers, and llama.cpp, including implementation notes about version compatibility, explicit device placement, and a llama.cpp patch for confidence output.