Frontier models, practical AI, and TinyML

Participants: Sirwan, Pooria, Arash, with Omid joining at the beginning.

We opened with a general discussion about some random topics with Omid and Pooria. Afterwards, Arash and Pooria stayed to review the week’s developments.

Frontier Models

We shared updates on Opus 5.5, including the impression that it was approaching Fable 0.1-level performance, alongside the new Sol and Luna models. We discussed cost versus performance and agreed that benchmark announcements often reflect marketing. Independent testing and real-world use are needed to assess the claims.

Jev and Smaller Models

Continuing last week’s discussion, we demonstrated how Jev could help classify emails and identify requests. We then asked how this differs from using an LLM. Many tasks need only limited context, fixed labels, and clear decision boundaries, making a large general-purpose model unnecessary.

Fraud detection was another example: depending on the problem, traditional classifiers or decision trees can be cheaper, more predictable, and easier to operate.

AI Coding and Engineering

We discussed signs that the hype around vibe coding is cooling. We shared an example from a heavily customised text editor: AI helped build the UI and implementation, but complex edge cases still required comprehensive end-to-end tests and engineering work.

AI can reduce manual coding, while reliable software still depends on engineering and product expertise.

TinyML on ESP32

We considered a future experiment with Arash: train a small model for a specific task, such as recognising the numbers 1, 2, 3, and 4, and run it on an ESP32. Arash noted that a purpose-built model would fit much better than an LLM. Preparing the dataset is likely to be the main challenge.

Possible next session: prepare a small dataset, train a simple model, quantise and export it, then try inference on an ESP32 with Arash’s help.