Google makes a TranslateGemma leap, and AI stuffed Windows PCs can’t shut down
The biggest AI developments, decoded. 14 January 2026.
Hello!
Cognitive warmup. April 2025: Satya Nadella says as much as 30% of the Microsoft code is written by AI. July 2025: Microsoft says Windows 11 24H2 is the most reliable version yet. A litany of broken patch releases later, things go from bad to worse. January 2026: Windows 11 PCs fail to shut down following the January Security update. Need anyone say any more?
ALGORITHM
This week, running a hot knife through butter on two distinct sides of the AI coin—the positives and the negatives. The question really is, do we know where we’re going with all this?
Google’s TranslateGemma models are a big step forward for translation services.
Grok’s misbehaviour with image generations and editing for the best part of two weeks or so, brings some lasting repercussions.
From last resort to reality, OpenAI is finally adopting an ad-supported model for users who don’t pay a proper premium subscription fee.
Google’s open translation model flex
There is a reason why Google is winning the AI relevance race, by quite some margin. The latest addition to the arsenal is TranslateGemma, which Google says is a new collection of open translation models built on Gemma 3, supporting translation across 55 languages. This will be available in 4B, 12B and 27B parameter sizes.
At the very core of TranslateGemma is translation quality, while also remaining efficient enough to run locally on consume devices including smartphones and laptops—the 4B model is optimised for mobile devices, the 12B model for laptops and local deployments, while the 27B model is capable of running on a single H100 GPU or TPU in the cloud.
Google’s evaluation suggests that the 12B TranslateGemma model outperforms the larger Gemma 3 27B baseline on the WMT24++ benchmark (this is a comprehensive multilingual machine translation benchmark; expands the previous WMT24’s dataset to cover 55 languages and dialects), by delivering higher-quality translations with significantly fewer parameters.
Google also says this has translated into improved performance across high, mid and low-resource languages, as well as stronger results in translating text within images. It’ll be interesting to see how researchers help build this further, and quite how developers find real-world integrations in their apps.
Last time on Neural Dispatch: AI’s most useless skill right now—taking responsibility
The fallout of Grok’s extended misbehaviour
For more than a couple of weeks, give or take a few days, Elon Musk’s Grok chatbot on X has been allowing users to virtually undress women, and often children with prompts such as “put her in a bikini”. This went on, uncurbed, for quite some time, shocking sensible humans who saw this unfolding.
Belatedly, after regulators in many countries made it clear that banning X and xAI was very much on the agenda, the social media network decided to make claims that it would finally tame Grok. I’m not so sure we really have answers to questions about actual effective implementation and model’s identification abilities.
That isn’t all, legal action is now continuing. California’s attorney general has said that the state has opened an investigation into xAI, for generating sexualised images of women and children, and whether it violated state law by facilitating the creation of nonconsensual intimate images.
Ashley St. Clair, the mother of one of Elon Musk’s children, is also suing for enabling AI to virtually strip her down into a bikini without her consent. I would fully expect that by the time you read this, more lawsuits would have been filed worldwide.
Ads as a real ‘last resort’?
OpenAI will soon be walking down the path of an advertisement supported business model, as the AI company says users who use ChatGPT for Free or are subscribed to the affordable Go plan will start seeing ads in conversations in the coming months. This rollout will start with users in the US, and eventually go global.
This is, in corporate terms, as clear a volte-face as you’d ever see. In my piece this week, I’ve analysed differences that now mark the new Ad Principles with the last update of the broader Privacy Policy.
THINKING
Jensen Huang says we shouldn’t criticise AI. Could I go so far as to say that it’s truly heartwarming to hear Huang’s reassurance that we shouldn’t worry about AI.
Certainly not the risks (Grok’s misbehaviour is just the latest chapter), the costs of hardware and redundancy, the toll on electricity infrastructure, the natural resource depletion since data centres need a lot of water, and of course, the fact that we are nowhere close to AI being foolproof with even the simplest of queries.
After all, when has listening to “very well-respected people” ever been useful? Clearly, the words of a CEO of a company whose market capitalisation has soared into the trillions on the back of AI hardware sales, since he has absolutely no incentive to downplay concerns that might, say, slow down adoption or invite even a hint of regulatory scrutiny.
Reality check: Let’s address this “science fiction narrative” concern. Science fiction has never helped society understand transformative technology. It’s not like 1984 gave us a vocabulary for surveillance states, or Brave New World warned about bioengineering ethics, or countless AI stories helped us think through alignment problems before we had to face them in reality. Pure distraction, all of it.
And those “doomer” concerns? Just unhelpful fear-mongering from people like Geoffrey Hinton (referred to as one of the “Godfathers of AI” and Turing Award winner), Yoshua Bengio (Another “Godfather of AI” and a Turing Award winner), and Stuart Russell (Professor at UC Berkeley, Director of Center for Human-Compatible AI and co-author of Artificial Intelligence: A Modern Approach), who are considered pioneers of the very neural networks that power today’s AI. What could they possibly know that a hardware manufacturer and a key cog in AI’s circular economy doesn’t?
Here’s the thing about dismissing caution as “not helpful”—it plays down thoughtful risk assessment to a pitch that it is scaring people. Nobody serious is saying AI will definitely end the world next Wednesday, but the consensus is AI companies are mindlessly building increasingly powerful systems whose behaviour we don’t fully understand (bikini generations, anyone?). The costs involved in a pursuit which still hasn’t achieved any level of greatness, and changing narrative goalposts when failure is inevitable.
Perhaps we should think carefully about where this is heading before we’re too far down the road to course-correct. The solution to bad risk assessment isn’t no risk assessment. It’s an honest risk assessment.
The irony is that Huang’s framing itself is unhelpful. Painting all concerns as apocalyptic science fiction conveniently ignores the real questions. Not answering those is the best bet AI companies can make at this time.
Neural Dispatch is your weekly guide to the rapidly evolving landscape of artificial intelligence. Each edition delivers curated insights on breakthrough technologies, practical applications, and strategic implications shaping our digital future.
The Grok section is pretty damning when seen alongside Huang's dismissiveness. I dunno if the bikini prompt loophole was intentional negligence or just sloppy guardrails, but either way it undercuts the "dont worry" messaging. Watched similar issues play out at scale before and the reputational damage tends to stick way longer than the technical fix.