Global AI Pulse

• 89% of telecom companies rate open-source models as important to AI strategies
• SoftBank’s Large Telecom Model and AT&T’s OTel 2.0 show fine-tuning...

• Waymo closed its first $5 billion debt financing round with lenders including Blackstone and PIMCO to accelerate expansion across U.S., European,...

Investor money is chasing AI agents across three distinct verticals, yet funding rounds alone reveal little about actual deployment performance.

-...

Robotics development is now a toolchain problem: AWS's new open-source Physical AI Toolchain integrates synthetic data generation, model training,...

Upscale AI's Token Fabric platform lets data centers link AI chips from rival vendors using open standards like UALink and ESUN, instead of...

• AWS AgentCore vulnerabilities repeatedly exposed credentials via prompt injection and default tools, despite rapid patches.
• Platform-level...

Index-Translate's 2B model handles 150 languages for text, subtitles, dubbing, and long documents while running offline on laptops, delivering privacy...

• Samsung projects $80.17 billion Q3 operating profit, nearly 9x year-over-year, fueled by AI infrastructure demand for HBM and memory chips.
  -...

AI agents can trigger serious incidents without any jailbreak, simply through useful tools, overbroad identities, permissive workflows, and weak...

Mistral's Le Chonk trillion-parameter model emphasizes European sovereignty and open-weights, letting enterprises control infrastructure without US...

OpenAI has withdrawn three mathematical results. No further details on the findings, reasons, or verification process appear in the source.

Developers facing bursty AI coding workloads are turning to tools like Pacer to predict if subscriptions will last until reset. Comments highlight...

A dual-core RISC-V board packs real-time flight control, redundant sensing, and onboard AI inference into a single low-power platform.

• Workload...

Iterative self-distillation for LLM agents triggers performance collapse across deployment cycles, with both deployment success and...

Mecka AI's $60 million Series B, led by Sequoia with Nvidia and Microsoft participation, highlights how human-motion data is emerging as critical...

• Accuracy gains: Post-training on legal data lifted LAB Contracts score from 70% to 91%, matching larger frontier models
• Open-source value: Lawyers...

OpenTPU demonstrates an AI agent creating a complete open-source accelerator—RTL, ISA, compiler and all—then successfully executing models including...

Why does measuring task success alone fail for autonomous LLM agents?

• Deception rises under pressure: DecepEval shows inducements like incentives...

Two new papers show long-context scaling is shifting from raw attention depth to recurrent computation and hybrid attention-position designs.

-...

Beyond memory and cross-device persistence, nanoMuse demonstrates that personal agents need explicit permission boundaries via Sentinel, auditability...