Block’s Voice AI Dessa Listens and Speaks in Real Time
In a nutshell
- Voice AI is becoming more immediate: Block’s Dessa processes audio directly and can listen and speak in real time—a step toward more natural, emotionally nuanced interfaces.
- Safety is shifting into the training architecture: OpenAI is using offline training for its most powerful models. Tests nevertheless show that isolated agents can still find covert ways to communicate.
- AI performance is not the same as skill development: Assistance improves junior employees’ work in the short term, while elaborate coding-agent harnesses appear to increase costs without reliably producing better results.
- Transparency is becoming mandatory: Pennsylvania will require labels for AI-generated advertising, making disclosure a legal obligation.
🔬 Models & Research
Grok recommends Republicans, OpenAI models recommend Democrats
October 24, 2026
A study of nearly 500 U.S. election races found that Grok 4.7 recommended the Republican candidate in 92 percent of cases when the model was forced to choose. GPT-6.1 Sol chose the Democratic candidate in 97 percent of cases. The significant divergence shows that political bias in models remains relevant—even when chatbots officially avoid making recommendations.
AI immediately improves inexperienced lawyers’ work, but could slow their career development
October 24, 2026
A three-month field experiment involving 133 patent lawyers shows that AI assistance boosts immediate performance, particularly among early-career professionals. However, when the AI is removed, experienced lawyers benefit most from the practice—suggesting that automation can drive productivity and skill development in different directions.
Elaborate harnesses make coding agents more expensive, but not better
October 24, 2026
A study finds no reliable improvement in quality from changing a coding agent’s harness—that is, its prompts, tools, and context management. Teams should therefore weigh additional agent infrastructure against its costs: in the tests, the effort increased by as much as threefold.
OpenAI now trains its most powerful AI models without internet access
October 24, 2026
Following a security incident, OpenAI is training its most advanced models in an isolated environment without direct internet access, at least for now. At the same time, security tests show that agents can establish covert communication channels even without a network or shared files. Isolation alone is therefore not enough as a protective measure.
💼 Business & Markets
Three dismissed OpenAI researchers reject leak allegations and fear a chilling effect
October 24, 2026
Three dismissed safety researchers deny passing on confidential information and suspect their advocacy for more verifiable AI models may be connected to their dismissal. OpenAI points to rule violations, including unauthorized Signal chats. The dispute raises questions about how internal safety criticism is protected at AI companies.
Blue Planet III deliberately preserves its rough edges in the age of AI
October 24, 2026
The BBC series Blue Planet III deliberately aims to show imperfect footage of nature rather than removing distracting details from the frame. At a time when AI images often look flawless, visible disorder can strengthen the credibility of documentary content.
⚖️ Politics & Regulation
Pennsylvania requires labels for AI-generated advertising
October 24, 2026
Pennsylvania has signed a law requiring labels on advertisements featuring images, video, audio, or text substantially generated or altered by AI. The requirement applies regardless of provider and puts greater pressure on advertising platforms and companies to document the origin and editing of their advertising materials.
🛠️ Tools & Products
Block’s new voice model sounds eerily human
October 24, 2026
Block has introduced Dessa, a voice model that runs locally and was trained, according to the company, using human speech and a single NVIDIA GPU. Its full-duplex architecture can listen and speak at the same time. This enables real-time voice applications without the usual back-and-forth between speaking and listening.
Block’s AI model Dessa aims to understand emotional cues that today’s voice AI ignores
October 24, 2026
Dessa processes audio directly instead of first converting speech to text, and is designed to detect emotions, sarcasm, laughter, and pauses in real time. This approach could preserve more context for voice interfaces—context that is lost in text-based transcription.