Sign language reaches the phone keyboard
Edition of 2026-08-13
Four things landed yesterday and this morning, and each one changes what someone can do today.
A phone keyboard that reads sign language
Google DeepMind published a sign-language-to-text translation model and, unusually, shipped it at the same time. It now powers sign-to-text dictation in Gboard and in Live Transcribe on the Pixel 11, starting with American Sign Language to English. A Deaf user can sign to the phone anywhere they would normally type: to search, to draft a message, or to answer in a live conversation. Google DeepMind describes this as the first time sign language AI has moved out of the lab and into consumer products, and the gap it is closing is large — there are more than 200 sign languages in the world, used by an estimated 70 million Deaf and hard of hearing people, none of whom benefited from decades of progress in speech recognition. The model was trained on over 100,000 hours of data spanning more than 50 sign languages. Privacy is handled on the device: a local model reduces the signer to a set of tracked points on the body, only those coordinates leave the phone, and the video is discarded immediately. The translation-quality score in the post is Google's own measurement on an academic test set, and no independent evaluation of this model exists yet, so treat the number as a claim and the shipping product as the fact. Google DeepMind
A model at the top tier, free to download
Qwen has released the weights of its most capable model, and says this is the first time a model from its top commercial tier has been published openly. Anyone with the hardware can download it and run it privately, with no company able to meter, watch or withdraw it. The scale is unusual for an open release: 2.4 trillion parameters in total, of which 95B are activated for any given token, spread across 512 experts, with a context window extensible to 1,010,000 tokens. The licence is permissive but not unconditional — large products must display the model name, and companies running a model-as-a-service or an AI work-assistant business above a stated revenue threshold must negotiate a separate licence. Read the terms before building a business on it. The comparison table on the model card is Qwen's own selection and loses several rows to the proprietary models it names; the independent read is more interesting, and it comes from Artificial Analysis, which places Qwen's top model in the same statistical band as the leading paid models on its real-world agentic work measure, with overlapping confidence intervals. Qwen model card · Artificial Analysis
The same result in half the steps
SpaceXAI released a new Grok model aimed at work that runs on its own for long stretches, and held its price flat across the generation at $2 per million input tokens and $6 per million output tokens. The vendor claims frontier parity; the more useful claim is the one an independent measurer confirms. Artificial Analysis reports that on its private test of long-horizon knowledge work, the model completes a task in roughly 53 back-and-forth steps on average, against roughly 103 for the highest-scoring rival, using a quarter of the input tokens to get there. Measured cost came out at $0.84 per task, which puts it on the cost-versus-performance frontier for every agentic evaluation in the index. The same independent report is explicit that the model trails the leaders on two coding measures, so this is an efficiency and price story rather than a leaderboard story. If you are paying for an agent that grinds for minutes at a time, fewer steps compounds faster than a higher score. SpaceXAI · Artificial Analysis
A price that depends on the clock
DeepSeek moved its flagship model from preview to general availability this morning, reachable in the app and on the web under an Expert Mode setting and through the API with no change to model names. Both models in the line now accept an explicit effort setting, so the same endpoint can serve a cheap shallow answer or an expensive deliberate one. The consequential part is the meter: from 16:00 UTC on 16 August 2026, DeepSeek splits its API into peak and off-peak rates, with off-peak set 50% lower than peak, and says the split is there so customers can schedule work into the cheaper window. Anyone running batch jobs on DeepSeek has three days to work out which of them can wait. No benchmark figures accompany the release and no independent evaluation of this version exists yet, so nothing here is a capability claim. DeepSeek
Signal faible
The European Commission has opened a call for expressions of interest to staff the enforcement team of its AI Office — technology specialists, legal officers, operations staff and paralegals, based in Brussels, working on the supervision of general-purpose model providers. The call document is a PDF this pipeline cannot extract text from, so no figure from it appears above. Hiring is the part of enforcement that becomes visible well before the fines do.
Watch the sixteenth: it is the first time a major model provider has asked its customers to care what time of day they run a job.