14 summarised stories about Safety Research, each linking back to the original source. Browse all topics →
OpenAI Blog 2 years ago
OpenAI described how its Voice Engine text-to-speech model generates realistic synthetic voices from brief audio samples and text inputs. The system requires only a 15-second voice sample to create a matching synthetic voice. The company emphasized safety considerations including detection systems for identifying synthetic speech and working with policymakers on responsible deployment guidelines.
The daily briefing
TLDRocket reads 60+ sources, removes duplicate coverage, and summarises the day in two minutes. Free, no spam, unsubscribe anytime.
TLDRocket
23 July 2026
Poolside AI's release of Laguna S 2.1 crystallizes a shift away from scale-at-all-costs. The 118-billion-parameter model, with only 8 billion active parameters per token, achieves 70.2% on Terminal-Bench 2.1 and outperforms Deepseek v4 Pro while costing less than v4 Flash—a direct challenge to the assumption that bigger models always win. Poolside's systematic approach, running 10,000 to 20,000 monthly experiments across fewer than 70 researchers in eight-week cycles, suggests that engineering rigor and data efficiency matter as much as raw compute. This matters because it hints at a possible end to the AI industry's winner-take-all consolidation: smaller teams with disciplined workflows can still build competitive open-weights models.
Read the full briefing →