The AI Abstract — Evening Edition

A model with fewer parameters than a short email can add two 10-digit numbers perfectly. That's not a typo, and it's not a trick. ⚠️ A 🎙️ Reddit thread (single-source, unverified) describes a transformer under 100 parameters hitting 100% accuracy on the task, which should unsettle anyone who learned the lesson that small models can't do arithmetic. The key is how the number is presented to the model. Most systems tokenize text in chunks, so a number like "3,847,291,056" might arrive as several compressed pieces that don't preserve which digit sits in which position. Digit-level tokenization breaks each number into its individual characters, one slot at a time, like laying cards face-up in a row instead of handing the model a shuffled deck. When the structure of the number is preserved in the input, even a tiny model has what it needs to carry the one. The finding doesn't contradict what we know about scale and reasoning generally; it clarifies that some failures blamed on model size are actually failures of representation. How you encode the problem shapes whether the problem is even solvable. The result needs independent verification before it changes anything, but the mechanism is worth keeping in mind the next time a model stumbles on a number.

The open-weight side of the field is moving fast enough that a two-month snapshot is already worth curating. Sebastian Raschka, a researcher with a track record of useful synthesis, compiled 🎙️ 10 open-weight LLM architectures from January and February 2026 in a single post. The payload here isn't any one model. It's the pace. Ten distinct architectural approaches in eight weeks means the open-source community is no longer just reproducing closed-model results on a lag. It's running its own design experiments. For anyone tracking where capability is actually growing outside the major labs, this is a practical index, not a headline.

The ChatGPT-in-epidemiology story making the rounds this week is worth one sentence: 📰 Ars Technica reports that an Illinois health official used ChatGPT during a county fair Salmonella investigation, the CDC's own report treats it as peripheral, and the outbreak was solved by standard epidemiological fieldwork. The story is less about AI capability than about how AI gets credit in narratives where humans did the work.

---

🎙️ Tiny transformers can add two 10-digit numbers to 100% accuracy: Read it to understand the digit tokenization mechanism before the result gets over-generalized.

🎙️ 10 Open-Weight LLM Architectures from Jan-Feb 2026: Read it as a pace-of-play indicator for anyone tracking open-source architectural divergence from the closed-model path.

📰 Did ChatGPT help health officials solve a weird outbreak?: Read it for the CDC report's actual methodology, which tells a quieter story than the headline.