checkpoint
Beyond Market Intelligence keeps checkpoint in one place: 4 stories so far. The section currently leads with “Scaling AI training with lightweight adapters instead of doubling models”, “Your spreadsheet lens still works after a model upgrade.”, and “Teaching a transformer exact arithmetic by hand, not training”. Training large models with self-distillation usually means doubling up, teacher and student both on GPU, eating memory and slowing everything down. A Jacobian lens fitted to one checkpoint is not supposed to work on the next model in the line. 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 checkpoint story on Beyond Market Intelligence, newest first.
Scaling AI training with lightweight adapters instead of doubling models
Training large models with self-distillation usually means doubling up, teacher and student both on GPU, eating memory and slowing everything down. A smarter approach? Share the frozen backbone and swap lightweight LoRA adapters instead. One adapter trains as the student; another tracks its EMA as the teacher. It took some careful engineering to make this work cleanly under TRL and ZeRO-3, but the results speak for themselves. The repo is live, tested on Qwen3.5, and ready to explore.
Your spreadsheet lens still works after a model upgrade.
A Jacobian lens fitted to one checkpoint is not supposed to work on the next model in the line. That is the assumption, and it went untested until now. One researcher applied the published Qwen3.6-27B lens directly to Qwen3.8-27B, 113 days newer. The latent entity readout held its rank, even improving at mid-depth. Steering directions still found their target concept. Transfer is measurable, not perfect. That suggests monitoring pipelines can test their lenses instead of blindly refitting.
Teaching a transformer exact arithmetic by hand, not training
A single transformer model, its weights hand-set with no training, just averted the arithmetic meltdown that defines its peers. One version nails all three million possible three-digit products, and the same approach scales to twelve-digit multiplication. What stands out is the compiler work: turning the grade-school algorithm into a standard Phi-3 checkpoint through Torchwright. Frontier models crumble at seven digits, five scoring zero out of five hundred, while this one holds steady.
From Python to weights: a compiler proves what transformers can express
Most people assume a transformer's capabilities come from training. Rob's compiler challenges that assumption directly. He built a tool that takes a computation graph written in ordinary Python and outputs the weights for a standard Phi-3 model, loadable in vanilla Hugging Face with zero training involved. It's a clever way to isolate what these architectures can express, separate from what they can learn. For a deeper look at how he constructs these proofs, the write-up is worth your time.