AI News of the Day: Mathematics, Voices, Deepfakes

Today’s focus is the kind of AI news that doesn’t just generate headlines, but has real consequences: for research, creative workflows, and security. What’s especially interesting is the mix of a possible mathematical coup from OpenAI, new speech models from Google, and another warning sign around deepfakes.

At the same time, the day makes it pretty clear where the market is heading: AI is becoming more useful, but also more uncomfortable. More agents, more automation, more responsibility — and unfortunately also more cases where platforms suddenly realize that “we’ll deal with it later” is not a strategy.

🧠 OpenAI model reportedly solved an open Erdős problem

According to The Decoder, GPT-5.4 Pro is said to have solved an open problem from the Erdős list, specifically #1196, in about 80 minutes. If that holds up, it would be more than a PR stunt: it would mean an LLM is not just a computational aid, but a real contribution to mathematical research. Terence Tao is said to have classified the approach as sensible.

Why does this matter? Because mathematics has long been considered one of the last bastions where LLMs can explain well but not really discover anything new. A case like this would show that a model can do more than recombine known patterns — it can also work in open proof spaces. Of course, one report is not yet a scientific breakthrough. But if it is confirmed, it significantly shifts the discussion around AI agents, reasoning, and math. And for anyone who likes to argue that “AI can only generate text”: that argument doesn’t get much more stable from this.

🔐 OpenSSL 4.0 makes TLS more private and ready for post-quantum crypto

With OpenSSL 4.0, an important security foundation is getting a major update. The cryptography library clears out legacy baggage, introduces Encrypted Client Hello (ECH), and prepares for post-quantum cryptography. ECH is especially interesting because it hides metadata in the TLS handshake more effectively — meaning it doesn’t just protect content, but also reveals less about where you’re actually connecting.

For the AI world, this is indirectly very relevant. Wherever models are connected via APIs, agents, or inference services, secure transport encryption is not a luxury, but a requirement. As AI systems become production-ready, the attack surface grows too: prompt exfiltration, API abuse, man-in-the-middle scenarios, the whole package. OpenSSL 4.0 is therefore not a sexy product launch, but exactly the kind of infrastructure update without which everything else is built on sand. In other words: less glamour, more substance — rare, but valuable.

📱 Grok almost got kicked out of the App Store

Two reports are about the same issue: Apple’s pressure on X’s AI app Grok over sexualized deepfakes. heise online reports that Grok was close to being removed from the App Store; The Verge describes the same conflict as a quiet display of Apple’s power. The core issue: a platform that cannot get a grip on deepfake and safety problems risks losing its distribution.

That’s an important signal for ai-policy, platform rules, and deepfakes: the debate is shifting away from theoretical ethics questions toward very practical levers. Apple doesn’t have to solve every AI problem in the world — but the App Store remains a powerful gatekeeper. For product teams, that means safety is not just compliance, but a distribution risk. And yes, sometimes the best model doesn’t decide the outcome — instead, it’s whether your product accidentally interprets the policy on non-consensual content as a feature.

🎙️ Google brings more flexible voices with Gemini 3.1 Flash TTS

Google is expanding its audio offering with Gemini 3.1 Flash TTS, a text-to-speech model that promises more natural, more dynamic voices in over 70 languages. Especially interesting: speech output can be controlled more precisely via audio tags. That’s a pretty clear step away from “one text, one default voice” and toward genuinely controllable speech generation.

For products, podcasts, learning apps, or voice agents, this is more than a nice-to-have. As TTS becomes more emotional, more precise, and more controllable across languages, the value rises significantly — but so does the responsibility. The more realistic AI voices sound, the more important labeling, abuse prevention, and clean workflows become. For ambitious teams, that means voice is no longer just a demo feature, but a real product channel. And probably also the next area where users suddenly realize that “Please sound neutral and serious” is, unfortunately, a wish rather than a law of nature for AI voices.

🛠️ Tool tip of the day: Adobe Firefly AI Assistant

With the Firefly AI Assistant, Adobe is introducing an AI agent that is supposed to control creative workflows across apps like Photoshop and Premiere via chat. That’s especially interesting for teams juggling images, video, and variant production: instead of operating each tool separately, you can bundle steps more tightly and delegate tasks to an assistant.

The appeal lies less in the “magic prompt” and more in the workflow: if the assistant orchestrates repetitive steps, you save time and reduce friction between applications. For creators, agencies, and marketing teams, that’s a realistic productivity lever. If you evaluate such tools, it’s worth taking a look at #. And if you want to go deeper into automated creative processes, # and # are also useful comparison points.

🔬 Transformer research: signal flow better understood at initialization

The new arXiv paper Subcritical Signal Propagation at Initialization in Normalization-Free Transformers analyzes how signals and gradients behave in transformers at initialization. At the center is the “Averaged Partial Jacobian Norm” (APJN) — a measure of how strongly gradients are amplified across layers. The work extends the analysis to bidirectional attention and permutation-symmetric token configurations, and derives recurrence relations for activation statistics.

Why is this relevant? Because good models don’t just depend on more data or more parameters, but also on being trainable in a stable way in the first place. Especially in normalization-free transformers, the question of signal propagation, stability, and gradient dynamics is crucial. Studies like this are not the kind of news that sets social media on fire — but they are exactly the foundation on which better models are later built. In short: without clean mathematics, “let’s just scale it” quickly turns into “why is the loss exploding again?”

Don’t want to miss any news? Subscribe to the newsletter