AI at the edge of politics, research, and energy
Meta, Google, and new AI research shape the day: from Pentagon deals to efficient serving, unlearning, and edge AI on microcontrollers.
Inhaltsverzeichnis
Today once again shows how broad the AI field has become: here a secret Pentagon deal, there new approaches to more efficient recommendation systems, and in between the question of how to properly forget models at all. Add to that topics that reach directly into everyday life — from media production to energy supply. In short: AI is long past being just model chat in the browser, and is now infrastructure, politics, and product design all in one.
🛰️ Meta orders solar power from space
According to TechCrunch, Meta has signed a contract with Overview Energy to receive solar power at night, beamed down to Earth from space. It sounds like science fiction with a spreadsheet attached, but it is a serious building block for the next phase of data center energy supply. Large AI infrastructure in particular consumes electricity on an industrial scale, which is why companies are looking for reliable, low-CO₂ sources.
Why does this matter? Because energy is no longer just a cost factor, but a strategic bottleneck for AI data centers. If power comes from alternative sources at night or during peak load times, operations become more predictable. For the AI industry, that means: if you want to scale models, you also need to solve the energy question — otherwise the beautiful future vision ends pretty quickly in the transformer room.
Source: TechCrunch
🤝 Google and the Pentagon: AI “for all lawful purposes”
According to The Verge, Google has signed a secret deal with the US Department of Defense that reportedly allows the Pentagon to use Google AI for “any lawful government purpose.” This is politically sensitive because the debate about military use of AI has been simmering at Google for years — and, as we know, the line between “legitimate” and “problematic” is rather elastic.
For the industry, this sends a signal: enterprise AI and government use are continuing to converge. At the same time, the report shows how large the tension has become between product business, ethics, and employee protest. Once AI systems end up in security-critical or military contexts, it is no longer enough to talk only about accuracy. Then it becomes about governance, abuse prevention, and responsibility — in other words, the parts of the business that PowerPoints often print a little smaller.
Source: The Verge
🧠 MTServe: Faster inference for generative recommenders
With MTServe, researchers on arXiv present a serving system for generative recommendation models. The key point: generative recommender systems deliver strong models, but inference is expensive because long user histories must be re-encoded again and again. MTServe uses hierarchical caches to better share reusable states across requests.
Why is this important? Recommendation is one of the most expensive and at the same time most productive AI areas of all. Even small efficiency gains can translate directly into real infrastructure savings at massive user scale. For ambitious beginners, this is also a good reminder: the most interesting research results are often not just new models, but better system designs. In production, it is not only “How good is the model?” but also “Can I afford to run it millions of times?”
Source: arXiv
🔁 Reward models may in fact be more than final scores
A second research highlight also comes from arXiv: Reward Models Are Secretly Value Functions argues that reward models in RLHF should not only evaluate the final token. Instead, their output at each position can be understood as a kind of expected terminal value. This is an important shift in perspective, because current reward models often use only part of the signal and therefore generate token-level noise.
The practical significance: if reward models capture temporal structure better, training could become more stable and more interpretable. This matters for alignment, RLHF, and generally for the question of how to evaluate models more reliably. Anyone working on evaluation today gets a useful reminder here: many “errors” in models are actually errors in our measuring instruments. Or, less romantically: garbage in, benchmark out.
Source: arXiv
🧹 Machine unlearning gets geometric
With Shape of Memory, another arXiv paper takes a geometric perspective on machine unlearning — specifically for second-order optimizers. The authors argue that current unlearning definitions are blurry for such optimizers and do not model model memory cleanly enough. But that is exactly what matters when data is meant to be truly deleted instead of merely waved away symbolically in memory.
The topic is especially relevant for regulation, privacy, and compliance. As AI usage grows, so does the pressure to remove training data cleanly on request. In practice, that is much harder than it sounds on paper — especially for large models and complex optimizers. For companies, unlearning is therefore not a luxury feature, but increasingly a mandatory component of trustworthy AI systems.
Source: arXiv
🎙️ Netflix, AI, and the matter of voices
At Netflix, according to heise, there is a dispute over AI clauses and the series “Lupin”: in the fourth season, the main cast is reportedly to receive new German voices. At first glance this sounds like a niche entertainment issue, but it touches a larger conflict: how far can AI intervene in creative production chains, and who retains control over voice, performance, and identity?
For the media industry, this is highly relevant because voice cloning, dubbing, and generative audio tools can dramatically reduce costs — but also intensify conflicts with voice actors, unions, and audiences. The case shows that AI in content is not just a question of efficiency, but one of fairness and acceptance. And sometimes it is only when your favorite character is affected that you realize automation is not neutral.
Source: heise
🛠️ Tool tip of the day
If you are experimenting with LLMs, evaluation, or agents, it is worth taking a look at a good observability and benchmark setup. Especially with research prototypes, clean tracking often determines whether you see real progress or just reproduce lucky hits. Pay attention to tools that make prompts, outputs, costs, and latency visible together. Especially practical: #
Don’t want to miss any news? Subscribe to the newsletter