Cognitive Warmup. Norway intends to ban the use of generic AI tools by primary school children, with Prime Minister Jonas Gahr Stoere warning (correctly, if I might add) that AI lets children skip important steps in their education and that schools should focus on teaching them reading, writing and maths. These guidelines will be implemented with the new school year starting sometime in August. Its two parts. Students in grades 1 to 7 will be completely banned from using AI, while teens aged 14 to 16 will be able to use generator tools under teacher supervision. This is still a work in progress; Hope to hear more soon.
First, on Neural Dispatch
Google DeepMind, and the right intentions
I have repeatedly pointed out that amidst the general uproar from AI companies, a paradox exists. Google’s focus on AI goes far beyond the usual AI hype these days (usage metrics, taking all the human jobs for granted, etc.) with the aim of helping make AI relevant to the masses. No, not iPhone seekers or greedy corporate boardroom spectators, but people who can actually help. Our regular readers may remember that a few weeks ago, I talked about Google DeepMind’s three AI model releases—Gemma 4, Gemini 3.1 Flash-Lite, and VO 3.1 Lite—which were focused on helping meet India’s AI ambitions. This is something that creates more impact than noise.
Google DeepMind is not slowing down on this mission. He believes there’s room to help with the new Gemini 3.5 Live Translation and Gemini 3.1 Flash text-to-speech. “What excites me most is the fluidity of this new audio model and its ability to handle the unstructured way that languages naturally mix when we speak,” says Manish Gupta, senior director of Google DeepMind.
With Gemini 3.5 Live Translation, they believe education platforms can integrate it into their online learning modules to translate video and audio lectures for students. A key element of Gemini 3.5 Live Translation is the ability to translate the tone of a sentence, not just the words – in other words, the natural voice. Currently supported languages include Hindi, Marathi, Telugu, Tamil, Gujarati, Malayalam and English (India).
Then there’s Gemini 3.1 Flash TTS, the latest text-to-speech model, which could potentially become the foundation of the next generation of AI-speech applications. The latest version improves the overall speech quality, focusing on natural intonation and expressions. One of the use cases Google points to is the use of Gemini 3.1 Flash TTS in DeepMind script writing, which can help generate multi-cast audio content. Indian language support scope includes Hindi, Tamil, Telugu, Marathi, Gujarati, Bengali, Kannada, Malayalam and Urdu. In particular, Hindi language models really can’t be better than this.
Google DeepMind makes it clear that all audio generated by these models will be watermarked using SynthID, which is woven directly into the audio output to detect machines, and thus distinguish AI-generated from human voices.
Giving Indic languages an AI voice
A collaboration between Google and the Indian Institute of Science (IISc) has given us something called Project Vaani. The intention to create an open-source speech dataset for Indic languages has now successfully completed Phase 2. Google says this marks the open sourcing of speech and image datasets for 109 Indic languages across 31 states and union territories, covering 1,56,000 speakers.
They detail three implementations in the Indian ecosystem. Shillong-based Emwire Labs is using Project Vaani’s natural conversational speech dataset to train a highly accurate voice-recognition system for Garo, a low-resource language traditionally excluded from major AI models. Garo is a Tibeto-Burman language spoken by about 1.2 million people in northeastern India and Bangladesh, and interestingly, there are no digital speech tools for it.
Another example where Project Vaani’s datasets are proving to be building blocks for more building blocks is with Indian tech company Shoonya Labs – they have used this dataset to build a voice AI model that emphasizes speech-to-text accuracy in over 200 languages and naturally understands mixed language speech. For developers, the fact that this method reduces AI training time and computing costs by 1,000 times cannot be underestimated.
Gupta says, “I am incredibly proud of how our long-term investment in Indic language research is empowering local innovators, and helping to lay the groundwork for local language models. The Indian ecosystem is uniquely positioned to drive global AI leadership, and we are excited to see what you build next.”
The latest on wired knowledge
theater of the absurd meta
They probably spent all the vibe tokens with vibe coding, and now vibes are completely shut down on the meta. Meta’s chief technology officer Andrew “Boz” Bosworth said during a recent internal call with employees that morale is “probably not the worst it’s been in the 20 years we’ve been here, but it’s probably up there. It’s definitely up there.” This has been shared by people who were reportedly on that call with the employees.
There are a few reasons why culture is at an all-time low again (all meta, by the way). If you look closely it’s all woven together. Meta already let go 10% of its global workforce this year. The reason? Admittedly, billions of dollars of ‘AI investment’ now needs to be balanced, because that train is getting nowhere fast. Meta has since redeployed approximately 7,000 existing engineers to its newly created Applied AI division to train the Foundation model, which is believed to be a “draft” to perform incomplete, small data-labeling tasks, with some calling the unit a ‘Gulag’.
Recently, it was reported that Meta is monitoring employee key strokes and mouse movements in alleged data to train its AI models. And then, Meta reported revenues of $56.31 billion and profits of $26.8 billion in Q1 2026, and the aforementioned job cuts occurred shortly thereafter. I have said earlier that corporate greed is a disease.
