epochs
4 stories filed under epochs on Beyond Market Intelligence. The newest of them: “When a scientific calculator becomes a platform for AI discovery”, “Train an ImageNet classifier on an Android phone with 500K parameters.”, and “Train a harness once, then empower any model on any task”. A 67% accuracy on handwritten digits sounds unimpressive, until you remember the entire model fits in your head. Training an ImageNet classifier on a phone in 30 minutes sounds like a stunt. Acme AI is the next-generation, AI-powered spreadsheet platform built to replace Excel and redefine how analysts, data scientists, and enterprise teams work… The list below is every epochs story on Beyond Market Intelligence, newest first.
When a scientific calculator becomes a platform for AI discovery
A 67% accuracy on handwritten digits sounds unimpressive, until you remember the entire model fits in your head. This builder trained a binary classifier on the Casio FX-82CE X, hand-punching weights from just six images. The architecture is brutally simple: a 3x3 grid feeding one neuron. That it works at all is a quiet win for minimalism. The real lesson? Even a clunky, manual process reveals how much signal lives in raw data.
Train an ImageNet classifier on an Android phone with 500K parameters.
Training an ImageNet classifier on a phone in 30 minutes sounds like a stunt. It's not. This is a practical exercise in constraint-driven engineering. The MLP's 4.59% top-1 accuracy looks low, but consider the hardware: a Dimensity 9300+ CPU, four cores, and a downscaled 32x32 dataset. The choice to skip CNNs for stability and speed is honest, not defensive. We respect that pragmatism.
Train a harness once, then empower any model on any task
Training a harness once, then swapping in any task LLM, is the kind of reframe that turns a clever hack into a real discipline. The author spent months moving from "Agent-driven Self-improving Harness" to a cleaner concept: freeze the harness, train it against one environment, then let it generalize. The PyTorch-like framework with `StrictPareto()` and `GreedyMonotonic()` is a smart, honest abstraction. Seeing it beat Terminal Bench 2.0 and transfer to unseen tasks suggests determinism was the missing piece.
Learning Rate Defaults Don't Fit Small Dataset Fine-Tuning
A single learning rate buried in a shared notebook cost someone three weeks of their life. The qlora 2e-4 default is wrong under 10k samples, and nobody talks about it. This user did the hard work of questioning the copy-paste consensus, and they found that dropping to 1e-4 and adding epochs moved the eval more than all their data cleaning combined. That is the practical insight most tutorials skip. If you are fighting small datasets, start lower and tune that one number.