France’s Côte-d’Or, Akila and ThinkDeep built sovereign digital-twin agents across more than 500 public buildings, cutting document searches from weeks to minutes and total operating costs by 10%; they target 30% at scale.
Separately, LFM2.5-24B-A2B extends its context window from 32K to 131K tokens and decodes 184 tokens per second on an M5 Max, versus 114 on a Ryzen AI Max+ 395. It leads or matches competitors on five of eight tool-use and agentic benchmarks and scores more than three times the next-best model on AutomationBench.
The connection is architectural, not a disclosed joint deployment: Côte-d’Or’s project uses a sovereign digital twin, while LFM2.5 supports local agents. The benchmark results are vendor-reported.
Why it matters: Municipal CIOs evaluating local AI can separate evidence from analogy: Côte-d’Or reports a 10% cost reduction across 500 buildings, while LFM2.5’s vendor benchmarks describe on-device performance—not a city deployment.
YouTube Music is adding labels showing when generative AI helped create a song, including its composition or vocals. The labels will appear on player pages and in the main YouTube app; clicking one reveals AI’s role.
Partners can classify tracks as “Fully Gen AI,” “Partly Gen AI,” or “No Gen AI.” Auto-Tune edits qualify as “No Gen AI.” YouTube combines partner disclosures with detection tools, but AI involvement alone will not block monetization.
That disclosure-first approach contrasts with Deezer, which reported receiving nearly 90,000 fully AI-generated tracks daily and excludes detected tracks from recommendations and editorial playlists.
Why it matters: Music partners now need platform-specific release strategies: YouTube lets AI-involved tracks remain eligible for monetization, while Deezer can remove detected fully AI-generated tracks from discovery, reducing their reach on the service.
OpenAI’s flagship still needs alignment work and testing infrastructure before it can move forward. Employees say one or two models will likely ship by year’s end, but it’s unclear whether the major upgrade is among them.
The model uses looped transformers, or “Recurrent Depth,” passing tokens through the same layers multiple times to improve math and coding while lowering costs. The tradeoff: some “thinking” moves into numerical representations reviewers can’t inspect, undermining chain-of-thought monitoring.
After safety evaluations, an executive order from President Trump gives the U.S. government access up to 30 days before outside partners and longer before a public launch. After the Hugging Face incident, OpenAI moved roughly 10% of its compute toward alignment and safety.
Why it matters: The practical effect is a staggered rollout: the Trump administration reviews OpenAI’s model first, outside partners wait up to 30 days, and public users wait longer—giving safety evaluations direct control over when customers can use the flagship.
Anthropic says small AI models could eventually power self-replicating worms on IoT devices, including smart fridges. Gemma4-26B-A4B established inference across 15 Android TV boxes with 4GB of memory each; Qwen3.5-9B configured other LLMs on eight of nine hardware types.
The models adapted inference software and dependencies for unfamiliar devices, unlike malware written for specific platforms. No lightweight model completed the full chain of finding targets, gaining access, installing a payload, configuring inference and propagating. GLM-5.2 orchestrated the test while smaller Llama models ran on targets with unauthenticated Android Debug Bridge exposed.
Gemma ran at 0.39 tokens per second, taking 71–150 hours to set up inference. Anthropic says small weight-based replicators may be possible soon after year-end improvements.
Why it matters: IoT defenders face a future containment problem: a replicator that adapts across unfamiliar hardware could spread without human intervention, and Anthropic says a decentralized worm would have no single shutdown point.
Ricursive Intelligence is building AI that designs chips, learns from each iteration, and improves the next one. Founders Anna Goldie and Azalia Mirhoseini say the system could make chips use data-center bandwidth, wafers, and fabs more efficiently as compute demand rises.
At TechCrunch Disrupt on Wednesday, Goldie said compute demand could grow at least 10,000x in five years—a projection she called potentially conservative. Ricursive is optimizing entire chips rather than splitting them into dozens of smaller blocks, enabling unconventional wave-like layouts.
Goldie said AI could cut chip-design timelines from six months, a year, or two years to about a week. That would make custom chips financially practical for applications including space, AR, VR, and biological computing.
Why it matters: For Ricursive, a one-week design cycle could make custom chips financially practical for space, AR/VR, and biological-computing applications that six-month-to-two-year design timelines force companies to choose selectively.
Nvidia CEO Jensen Huang announced $2 billion in new European investments, with details due the following week, and pledged $250 million in compute over five years to UK AI-safety labs. Bloomberg reported the pledge during Huang's European tour.
In Paris, Huang disclosed Nvidia stakes in French startups Doctolib, Genesis AI, Pigment and Entalpic. He promoted open-weight models as a way for Europe to close its AI gap and pushed back against calls to slow AI development.
The compute will support BOLD at Oxford, SOFAIR at UCL and Imperial College London on AI-agent safety, open-ended learning, open-source architectures and robotics safety.
Why it matters: BOLD, SOFAIR and Imperial College London are set to receive five years of Nvidia-backed compute for safety and robotics research, while Nvidia's stakes give it a direct financial position in Doctolib, Genesis AI, Pigment and Entalpic.
Google made Gemini 4 Argon globally available at an introductory $2 per million input tokens and $10 per million output tokens. Developers and enterprises can access it through the API in AI Studio and the Gemini platform. Google AI Ultra subscribers get Argon in Antigravity; Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search.
Google says Argon ranks first on its Vals Index, which measures performance across finance, coding, legal and tax work. In Antigravity, Google engineers increased agentic code submissions by more than 35% and cut rollbacks by nearly a third, according to company figures.
Argon solved all 12 simulated end-to-end cyberattack scenarios, up from five for the previous generation. Google is expanding its Fairwind Program to hundreds of additional partners.
Why it matters: For developers and enterprises, Google’s 95% cached-input discount could lower the cost of repeated API workloads, while Argon’s code and safety gains remain Google-reported results rather than independent tests.