Cathie Wood: AI Inference Costs Could Fall 99.99% Annually, Unlocking Era of 'Benign Deflation'

Beating AI Insight News Brief: DDoS defense pioneer and Prolexic founder Barrett Lyon has launched doxx.net, securing $38 million in Series A funding led by a16z. The platform can be initially understood as Tailscale built specifically for AI agents—it connects computers, servers, and AI agents into a single private network, then controls which entities can connect and access what resources. For example, Tailscale lets you securely link a MacBook and Mac mini, then restrict cross-device and cross-port access; doxx.net goes further by enabling isolated management of agents running on a Mac mini, such as allowing only that agent to connect to a specific server and service, while blocking all other machines entirely. Its core difference from Tailscale is that agents are not merely integrated into the network—they can autonomously modify network configurations: onboarding devices, adjusting firewalls, setting up internal domains, and even completing full network setup. The platform’s developers even let users delegate configuration access directly to agents, letting them handle these tasks independently. This design is tailored for long-running AI agents, which will increasingly reside on computers and servers to deploy services, call APIs, and gain broader permissions over time. doxx.net’s goal is to manage these permissions in isolation, ensuring each agent only interacts with the machines and services it is authorized to use, with permissions that can be revoked immediately when no longer needed. Additionally, doxx.net operates its own IP addresses, DNS, certificate systems, and network infrastructure, while providing encrypted communication between devices. The product is now available for open testing.

The drop in $PUMP 8 hours ago liquidated 2 traders, with a total of 707.6M $PUMP ($3.61M) liquidated.

According to Yu Jin Monitoring, the largest short seller that opened a short position on Changxin Memory the day before its IPO closed all positions 10 hours ago. The address shorted at $6.5, held the position for two months before closing at $8.5, resulting in a total loss of $10.98 million, including $5.18 million in funding fee losses and $5.8 million in price fluctuation losses.

Beating AI Insight News Flash: After Jev’s launch, AWS, Cloudflare, and Databricks have all recently rolled out their own decision models. AWS introduced the open-source Strands Decider 2B, built on Qwen3.5-2B. It removes the text generation module, replaces it with a decision module dedicated to answer selection, and is fine-tuned via LoRA. Official tests show its median single decision latency on an RTX 3090 is approximately 115ms, with full training taking around 11 hours. Cloudflare released Clef and Clef-flash simultaneously, modified from Qwen3.8-27B and Qwen3.5-9B respectively, and both are fully open-source. In Cloudflare’s tests on Jev’s official evaluation set, the Clef series outperformed Jev in 3 out of 4 task categories; Clef-flash has a median latency of 38.8ms, compared to Jev’s 524.1ms. Databricks has integrated this capability directly into its platform. The newly launched ai_decide can be called via SQL or REST API, used for data classification, scoring, and agent routing, and its interface is fully compatible with TypeSafe AI API. The company notes a single decision returns in under one second. From developers’ own builds to follow-ups by AWS, Cloudflare, and Databricks, decision models are rapidly evolving from a new concept to a standard component in agent infrastructure.

Beating AI News Flash: Supabase Acquires Database Startup Turso Turso has re-implemented SQLite in Rust to build cloud databases that can be created at scale and on-demand. Supabase plans to leverage Turso to serve AI Agents, enabling each agent to quickly access an independent database. The acquisition targets the explosive growth in database volumes driven by AI Agents. Supabase states it currently creates over 1 million databases weekly, with approximately 70% of new databases originating from Agents or AI tools. Turso’s architecture allows a single server to manage millions of databases—loading them when needed and pausing them during idle periods—making it well-suited for the large number of small, on-demand Agent tasks. The two companies will not immediately merge their products. Supabase will continue development around Postgres, while Turso will remain focused on SQLite; existing databases, APIs, and workflows will stay unchanged, and Turso Database will continue to be open-source. Moving forward, the pair will gradually integrate their offerings: small Agent tasks will initially use Turso, and as demand scales, they will transition to Supabase’s Postgres ecosystem. Turso founder Glauber Costa will join Supabase to lead Agent-related services. The acquisition terms were not disclosed.

Beating AI Express News: Microsoft has rolled out three audio models in one go: MAI-Transcribe-2-Streaming, MAI-Voice-2.1, and MAI-Voice-2.1-Flash. The first is designed for real-time speech-to-text, while the latter two handle text-to-speech, primarily targeting low-latency voice agents. MAI-Transcribe-2-Streaming can output text continuously before a speaker finishes talking, supporting 60 languages and automatic language detection. Per Artificial Analysis’ streaming speech-to-text leaderboard, it ranks first in both final transcription accuracy and first partial transcription accuracy, with a word error rate (WER) of 2.5%. The final text is returned approximately 0.13 seconds after the speaker concludes, compared to Grok Voice Transcribe 2.0’s 2.7% WER and 0.49-second latency. MAI-Voice-2.1 supports 23 languages, enabling the same voice to switch between different languages. The Flash version is optimized for low latency, with model inference taking around 45ms and priced at $15 per million characters; the standard version has ~550ms latency and costs $22 per million characters. Both models support voice matching via a short reference audio clip. This suite addresses the waiting issue that most impacts voice agent user experience. The system can initiate transcription and inference while the user is still speaking, then use the Flash model to quickly generate voice responses, eliminating the need to wait for a full sentence to end before starting operations.