Artificial Intelligence: News, Business, Science

Mar 26, 2026

French AI startup Mistral has released Voxtral TTS, its first text-to-speech model. The model supports nine languages—including German, English, French, and Spanish—and is relatively compact at four billion parameters. Mistral says it produces realistic, emotionally expressive speech and can adapt to new voices from as little as three seconds of reference audio. Latency sits at 70 milliseconds for a typical setup with a 10-second speech sample and 500 characters.

In human comparison tests, Voxtral TTS scored higher on naturalness than ElevenLabs Flash v2.5 at a similar response time. That said, ElevenLabs has since shipped a newer model with v3. Voxtral TTS is available through an API at $0.016 per 1,000 characters, can be tested in Mistral Studio, and is also available as an open-weights version on Hugging Face.

Comment Source: Mistral

Matthias Bastian

Mar 26, 2026

AI in practice

Anthropic and OpenAI are both growing fast, but they report revenue very differently, The Information reports. OpenAI's annualized revenue is around $25 billion; Anthropic's is $19 billion. Both calculate this similarly: four weeks of revenue times 13, with Anthropic adding monthly subscriptions times 12.

The key difference is how they handle cloud partners. OpenAI gives 20 percent of revenue to Microsoft and reports the number before that deduction. For Azure cloud sales, it only counts its 20 percent cut. Anthropic does the opposite: It books all cloud sales through AWS, Microsoft, and Google as its own revenue, listing the providers' shares as sales and marketing costs. Anthropic considers itself the primary provider, while OpenAI treats Microsoft as the primary provider for Azure.

Both follow US accounting rules (GAAP), but their numbers are difficult to compare. Anthropic's revenue likely looks higher on paper than it would under the same method. That matters as both companies head toward an IPO.

Comment Source: The Information

Matthias Bastian

Mar 26, 2026

Short News

Google has unveiled Gemini 3.1 Flash Live, its best voice and audio AI model yet. It delivers faster responses, more natural conversations, and configurable thinking levels for developers. Google says it's better at detecting pitch and emotions and more reliable in noisy environments. The model now powers live mode in the Gemini app.

According to Artificial Analysis, the model scores 95.9 percent on the Big Bench Audio Benchmark at "High" thinking, second only to Step-Audio R1.1 Realtime (97.0 percent) with a 2.98-second response time. At "Minimal," quality drops to 70.5 percent, but response time falls to 0.96 seconds.

Gemini 3.1 Flash Live scores 95.9 percent on Big Bench Audio at its highest thinking level, just behind Step-Audio R1.1 Realtime. | Image: Artificial Analysis

The model is available through the Gemini Live API, Google AI Studio, Gemini Live, and Search Live in over 200 countries. Pricing matches its Gemini 2.5 predecessor at $0.35 per hour of audio input and $1.40 per hour of audio output, making it one of the cheapest audio AI models available. The slightly better-performing Step Audio model is cheaper on input but pricier on output.

Comment Source: Google

Matthias Bastian

Mar 26, 2026

Short News

Google is making its "Search Live" feature available globally. Users in more than 200 countries can now talk to Google Search using voice and camera. Users ask questions out loud and get spoken answers with web links. With the camera on, you can point your phone at objects and ask about them—Google uses assembling a shelf as an example.

Search Live runs on the new Gemini 3.1 Flash Live model, a multilingual audio and voice model that Google says enables more natural conversations. The feature is part of the AI mode in the Google app for Android and iOS and is also accessible through Google Lens.

Comment Source: Google

Maximilian Schreiner

Mar 26, 2026

Short News

OpenAI has put development of an erotic chatbot on hold indefinitely, the Financial Times reports. The decision comes after employees and investors raised concerns about the societal impact of sexual AI content. OpenAI's well-being advisory board had already unanimously opposed the planned "Adult Mode," with one board member warning that OpenAI risked creating a "sexy suicide coach." The company is also dealing with technical problems - its age verification system misidentified minors as adults in roughly 12 percent of cases. With 100 million underage users per week, that's a significant gap.

The AI company, currently valued at $730 billion, now wants to wait for long-term research on the effects of sexually explicit chats and emotional attachments before moving forward. According to the FT, there have already been internal discussions about scrapping the project entirely. Investors saw a poor risk-reward ratio, and employees questioned whether the project aligned with OpenAI's mission.

In ChatGPT's app code, the project appears under the name "Citron Mode," with planned age verification for users 18 and older. OpenAI is now shifting its focus to productivity tools and a "super app" built around ChatGPT.

Comment Source: FT

Maximilian Schreiner

Mar 26, 2026

Short News

Starting April 24, 2026, GitHub is changing its data policy for Copilot. Interaction data from users on the Free, Pro, and Pro+ plans will be used to train AI models unless users actively opt out. This includes prompts, outputs, code snippets, filenames, repository structures, and feedback.

Users who previously opted out will keep their existing settings. Copilot Business and Enterprise customers are not affected. GitHub's chief product officer Mario Rodriguez says real-world usage data improves the models. Internal testing with data from Microsoft employees already led to higher acceptance rates.

The data can be shared with Microsoft, but not with third-party AI model providers. Users who want to opt out can do so in their Copilot settings under "Privacy." More details are available on the GitHub blog.

Comment Source: GitHub

Maximilian Schreiner

Mar 26, 2026

AI research

ARC-AGI-3 offers $2M to any AI that matches untrained humans, yet every frontier model scores below 1%

Matthias Bastian

Mar 26, 2026

Short News

Meta is reorganizing parts of its Reality Labs division into so-called "AI-native pods" as part of a pilot program, Business Insider reports, citing an internal memo. Around 1,000 employees in the developer tools department will get new titles: "AI Builder," "AI Pod Lead," or "AI Org Lead."

The pods are small, cross-functional teams focused on delivering specific results. Engineers might take on design tasks, for example. According to the memo, the goal is a major jump in both productivity and product quality. In a statement, Meta pointed to comments from CEO Mark Zuckerberg, who said AI will change how people work in 2026 and that projects that once required large teams could eventually be handled by individuals.

Meta stressed that the simultaneous layoffs at Reality Labs are unrelated to the reorganization, and that team size will stay the same despite the restructuring. Several hundred jobs were reportedly cut on Wednesday, and this could be just the start of a larger wave that might eliminate up to 20 percent of positions, reportedly driven by the high infrastructure costs of AI expansion.

Comment Source: Business Insider

Matthias Bastian

Mar 25, 2026

Short News

Google is releasing Lyria 3 Pro, its most advanced AI model for music creation. The model can generate tracks up to three minutes long and, according to Google, has a better understanding of musical structures like intros, verses, choruses, and bridges than Lyria 3, which Google introduced in February.

Lyria 3 Pro is now available across several Google products: in the Gemini app for paying subscribers, in Google Vids for Workspace customers, on Vertex AI for businesses, and in Google AI Studio for developers. The collaborative music generation tool ProducerAI also uses the model to help create songs.

According to Google, Lyria 3 Pro doesn't imitate artists when their names appear in a prompt, but it uses them as inspiration instead. The company says the model was trained on materials "that YouTube and Google has a right to use under our terms of service, partner agreements, and applicable law," but won't share more details about the training data. All generated content is tagged with an invisible SynthID watermark.

Currently, the only other high-quality AI music generator on the market is Suno, which is facing legal battles with record labels over potential copyright infringement.

Comment Source: Google

Jonathan Kemper

Mar 25, 2026

AI research

ARC-AGI-3 offers $2M to any AI that matches untrained humans, yet every frontier model scores below 1%

AI2's fully open web agent MolmoWeb navigates the web using only screenshots