AI2Day Daily Brief — 12 Sept 2026

Good morning — it is Saturday the twelfth of September, and today's brief opens with Apple ML Research, which has had a busy week: three separate papers published, covering video captions, sign language translation, and protein design.

Start with the video caption paper. Today's standard tools for scoring AI-generated video descriptions work by matching words: the closer your output is to a reference caption word for word, the higher the score. The problem is that any given video can be described accurately in dozens of different ways, so a perfectly correct caption gets penalised simply for choosing different words. Apple's new framework replaces word-matching with a fact-checking approach — it turns caption quality into a multiple-choice quiz, testing whether the right information was captured, not whether the right phrases were used. If you work in video AI or evaluate model outputs, this kind of evaluation tooling directly shapes what gets optimised and what does not.

The second Apple paper is DiscoSign, a system for translating text into sign language glosses. Existing tools handle this one sentence at a time, and that creates real problems, because sign language meaning builds across a conversation. Apple's system tracks what researchers call spatial coreference: if a person is introduced on the left side of the signing space, they stay on the left throughout the exchange. DiscoSign also recognises Question-Answer Clauses, a structural feature where a question and its answer form one linked unit that sentence-level systems simply never see. The architecture is modular, built on a large language model, with each discourse phenomenon handled by its own dedicated component. Apple describes it as a step toward translation that actually follows a conversation rather than processing it in fragments.

Today's sponsor is Zero2Hero: an AI poker coach sits at your table, and you can ask it mid hand whether the player across from you is bluffing. Every mistake you make gets flagged and graded automatically, hand by hand, even on the free tier. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

Good framing. The third Apple paper is SimpleDesign, a protein design model. Proteins govern almost every biological function — oxygen transport, immune defence, digestion — and their behaviour depends on two things working together: the sequence of amino acid building blocks, and the three-dimensional shape that sequence folds into. Most existing design tools train two separate models in sequence, one for the amino acids, one for the structure. SimpleDesign collapses that into a single joint training process, learning the relationship between sequence and shape at the same time. Apple says this approach shortens the path from a design target to a viable drug candidate. The research is from Apple ML Research, so these are published findings, not a product announcement.

Moving away from Apple. Google confirmed on Wednesday it will spend thirteen billion euros — roughly fifteen point one billion dollars — building AI data centre infrastructure in Finland. The company says this is the largest single investment it has ever made in any European country. Data centres need two things in large quantities: space and power. Both are scarce across much of Europe, where land costs are high and several national grids are already under strain. Finland's grid is comparatively stable, and the cold northern climate cuts cooling costs significantly, since the buildings need to shed enormous heat from the chips running AI calculations around the clock. Google's announcement lands as every major technology company is racing to build the physical infrastructure AI services run on.

Finally, Moonshot AI. The Chinese lab behind the Kimi chatbot is targeting two billion dollars in annualised revenue by the end of twenty-twenty-six — double its reported run rate for August. The growth engine is K3, an open-weight model, meaning its underlying code is freely available for anyone to download and use. That popularity strategy is a harder road to revenue than selling a proprietary model, and for context, two billion dollars is still far below OpenAI's reported forty billion and Anthropic's reported sixty-five billion in annualised revenue. The same week Moonshot published those targets, Anthropic alleged that Moonshot had secretly routed nearly three hundred thousand Kimi user queries directly to Claude Opus, using Anthropic's answers to train its own models. Moonshot has not publicly confirmed the allegation.

For more depth on any of these stories and the week's bigger picture, the Monday show AI Today Weekly goes longer on everything that matters. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. Do us a favour before you go: hit that thumbs up, and subscribe for next week.