AI’s wider week: models, voice, science and agents

Transcript

The original narration, preserved as published. Relative dates refer to the episode’s coverage window.

A model for Emirati Arabic. An open model for cited scientific reports. And agents judged by the databases they leave behind. Five developments from the past week.

TII introduced Falcon-Emirati seven B, adapted for Emirati Arabic and cultural nuance. The team combined native-speaker review with dialect-specific evaluation. It’s available through Falcon Chat. A useful reminder that language capability depends on which language, and whose everyday speech.

Ai2 released AstaBrief eight B and its training data. It turns a research question and literature excerpts into a cited report, and powers Asta’s Fast mode. You can run the weights yourself. Its published comparisons mostly date from twenty twenty-five, so they don’t establish superiority over today’s frontier.

The Open TTS Leaderboard compares intelligibility, speed, speaker similarity, and streaming delay. You can switch languages and listen to generated samples. Those metrics help narrow the field, but naturalness and expression still need human ears.

ServiceNow described AutoSynthData: use a model’s failures and a stronger teacher’s successes to generate new training tasks. Then check that the tasks are feasible, the solutions work, and the verifiers reject wrong outcomes. The curriculum follows what the model still struggles with.

ThinkingBox tests five hundred seven stateful workflows, repeated twenty times each. It grades the final records and side effects. An agent can call the right tools and still leave a ticket in the wrong state. This benchmark checks that difference.

Specialized models, open research tools, voice comparisons, and better agent training. There’s plenty here to try, measure, and build on.