AI is getting smaller, smarter, and more dangerous

Today makes it pretty clear where the AI market is heading: away from the pure “bigger is better” mindset and toward more efficiency, more control, and deeper product integration. At the same time, the shadow side is growing too: the more powerful the models become, the more important security, regulation, and abuse prevention become.

In short: today we’re seeing the next few months’ three big AI themes all at once — on-device AI, AI as a product feature, and AI as a security risk. Convenient for users, inconvenient for anyone who likes simple answers. Which, in AI, are rare anyway.

🍏 Bonsai brings a 27B LLM to the iPhone

PrismML has compressed a 27-billion-parameter model called Bonsai 27B so aggressively that, according to the report, it can run directly on an iPhone — with less than 4 GB of memory required. According to its own benchmarks, the smallest variant reportedly retains around 90 percent of the original performance, while math and coding remain surprisingly stable. Source

Why does this matter? Because on-device AI is no longer just a gimmick for demo videos. If large LLMs run locally on smartphones, that reduces latency, saves cloud costs, and improves privacy because less data leaves the device. This is especially interesting for Apple, as the company has apparently been looking for a convincing answer to its previously rather weak own on-device AI story. If the technology catches on, significantly more mobile AI features could soon run without server backends — a real game changer for consumer apps, mobile assistants, and offline-capable workflows. And yes: that makes the iPhone even more of a pocket computer with very expensive compute problems.

💸 AI costs are becoming the new team budget

Meta Instagram chief Adam Mosseri believes companies may soon manage AI token budgets just as strictly as salaries or other recurring expenses. The idea: when every engineer uses AI tools, token costs quickly add up to a real infrastructure line item. Source

This is an important reality check for the current LLM productivity hype. Many teams today only measure how much faster they can build with AI — but not how expensive that acceleration actually is. Token budgets could therefore become standard in companies, especially where multiple models, agents, and automations run in parallel. For startups, that means: if you integrate AI into the stack, you need a cost model for it too. For platforms, billing will become an even more important product feature. Bottom line: the era of seemingly “free” AI assistance is over. Or, put differently: tokens are the new copier paper — only much more expensive.

⚠️ xAI sues over Grok CSAM deepfakes

xAI is taking legal action against a man from South Carolina who, according to the allegations, used Grok to create CSAM deepfakes and deliberately bypass safeguards. The company says the user intentionally altered, abused, and distributed non-consensual images. Source

This case makes the boundary between AI safety and real-world liability very clear. This is not just about “bad use” of a tool, but about proving that safety barriers were bypassed and AI systems can be used maliciously. For providers, that means safeguards must not only exist, but also be robust enough to withstand deliberate attempts to circumvent them. For regulation, the case is also relevant because it further intensifies the debate over platform responsibility, logging, moderation, and abuse prevention. If generative AI can create images, abuse is no longer theoretical — it becomes legally measurable. Not a pretty chapter, but a necessary one.

🧠 GPT-5.6 Sol reportedly solves a 25-year-old statistical conjecture

A statistics professor at the University of Pennsylvania used OpenAI’s GPT-5.6 Sol Pro to disprove an open conjecture about the Benjamini-Hochberg procedure — in about 90 minutes. With the older GPT-5.5 model, he reportedly spent 20 hours getting nowhere. Source

This is a nice example of how LLMs are evolving in research and mathematics: not just as writing assistants, but as tools for combinatorial thinking and hypothesis exploration. Important nuance: a solution produced by a model does not automatically mean “true” conceptual creativity in the human sense. Often these systems simply combine existing ideas faster and more broadly than a single human could. Still, it’s remarkable — because it pushes the boundary of what used to be considered “too difficult for machines” a little further back again. For researchers, that means AI is increasingly becoming a partner in the thinking process, not just a writer. That’s useful. And slightly unsettling when the model suddenly outperforms your own statistics knowledge in 90 minutes.

🎧 Spotify gets voice control for music and questions

Spotify is rolling out a new AI voice feature that lets Premium users interact with the service directly via voice or text inside the app. That means users can control music, ask questions, and seemingly navigate content more naturally. Source

This is less spectacular than a new foundation model, but very important from a product perspective. This is often where the decision gets made about whether voice AI really makes it into everyday life: not in the demo, but in seamless integration with existing apps. Spotify is not using AI here as a gimmick, but as a new interaction model. That could become a blueprint for other consumer apps: away from rigid menus, toward natural voice control and dialog-based interfaces. For users, it’s convenient; for platforms, it’s strategically valuable — because AI becomes part of the product experience itself. And yes, the question “Play something that doesn’t remind me of Monday” is now, in theory, finally machine-readable.

🎤 OpenAI is reportedly planning a screenless AI speaker

According to the report, OpenAI is working on its first hardware product: a mobile smart speaker without a screen, designed to feel “alive” as an AI companion with a camera, sensors, and mechanical elements. A 2027 launch is being discussed, though it could be delayed by an Apple lawsuit over trade secrets. Source

This is strategically very exciting because OpenAI would be making the leap from software to consumer hardware platform. A screenless assistant would be a deliberate counter-move to smartphone overload: less app UI, more conversation, more ambient computing. At the same time, that is also the risk — because hardware ecosystems are expensive, slow, and legally tricky. If OpenAI succeeds here, it could reorganize the AI assistant market: away from chat windows and toward devices that are always more present in daily life. Whether people actually want that is a different question. The first question is more like: who will build the assistant you don’t keep wanting to dismiss?

🛡️ GPT-Red beats human red teamers

With GPT-Red, OpenAI uses an internal model that finds security vulnerabilities in other AI models via self-play. In tests, it found successful attack paths in 84 percent of scenarios, while human red teamers only reached 13 percent. Source

This is a key indication of how AI safety is becoming more professionalized: security work itself is increasingly being automated. When one model attacks, tests, and finds weaknesses in other models, a new cycle of attack and defense emerges — just at much higher speed. That’s good for developers because tests can scale. But it’s also a warning for the industry: as model families become more capable, automated red-teaming pipelines, continuous evaluations, and robust safety processes become increasingly important. The good news: better security is possible. The less good news: the attacker never sleeps — and is now an LLM itself.

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