Google AI clones voices in seconds

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🎙️ Google AI clones voices in seconds

🔒 Google builds AI memory it can't read

🤖 Open robotics model beats Nvidia

🖼️ Alibaba 7B AI beats Google model

🧠 DeepSeek reveals how it trains AI agents

Plus: 💡 5 strategies & tactics, 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🎙️ Google AI clones voices in seconds LINK

Google launched Gemini 3.8 Flash TTS and Flash-Lite TTS, rolling out today in the Gemini API and AI Studio, letting users design custom voices from plain-language descriptions or clone one from a 30-second sample across 100+ languages.

Flash TTS topped Hume AI's Voice Design Benchmark at 71.4 and led the Overall Quality Index ahead of Flash-Lite, while both won blind Voice Arena preference tests in Japanese, Hindi, and Mexican Spanish, with a 2,000+ ready-made voice library.

Both models support line-by-line stage directions, two-speaker scenes, and hours-long audio, and every clip carries a SynthID watermark, though voice cloning requires a verbal consent recording that the system verifies against the reference speaker before generating.

🔒 Google builds AI memory it can't read LINK

Google DeepMind detailed a persistent-memory architecture for Private AI Compute yesterday that retains personal context across devices while keeping decryption keys on users' hardware, leaving stored data unreadable to Google itself.

When a model needs context, an end-to-end encrypted connection sends data into a secure enclave for temporary decryption, then re-encrypts before storage, using per-user databases and device-derived keys with server-authenticity verification.

A Trail of Bits review from Sept. 21 logged 10 findings and confirmed no employee-access path, though one high-severity issue stayed open, deleted memory could be restored under privileged conditions, and no rollout date or supported devices were announced.

🤖 Open robotics model beats Nvidia LINK

Black Forest Labs released FLUX 3 Action, a 7B open-weight World Action Model that turns camera frames, robot state and language instructions into physical actions, topping every model on NVIDIA's RoboLab-120 leaderboard.

The model hits a 42.92% success rate on RoboLab-120, beating NVIDIA's 16B Cosmos3-Nano-Policy by 6.1 points while running 1.43x faster on 44% as many parameters, and it also flew drones and cleared Doom in early experiments.

BFL is shipping weights, code, a fine-tuning recipe and an SO-101 plus LeRobot implementation, teams can fine-tune on their own robot demos, though there's no Action API, pricing or commercial weights license announced yet, making self-hosting the only route for now.

🖼️ Alibaba 7B AI beats Google model LINK

Alibaba Cloud has released Qwen Image 2.1, a 7B open-weight image model that its developers claim outperforms closed-weight rivals like Google's Nano Banana 2.0, while running on consumer GPUs as old as the RTX 3090.

The model adds native transparency and multi-image editing with up to 10 reference images, and scored 60.2 on Alibaba's internal Qwen Image Benchmark versus 59.82 for Nano Banana 2.0, though GPT Image 2.5 Sunburst led at 67.

Early adopters report 1MP generation in roughly 5 seconds on an RTX 4090 and 2K images in about 50 seconds on an RTX 3060, though the revised license bars commercial resale of the model without a separate Alibaba license.

🧠 DeepSeek reveals how it trains AI agents LINK

DeepSeek published a paper called DeepSeek Elastic Compute detailing the platform it uses to train AI agents at scale, running about 3 million isolated sandboxes a day, or roughly 380,000 at once, Bloomberg reported.

The system sustains over 5,000 sandbox creations per second across four isolation tiers ranging from function calls to full virtual machines, and reallocates processor capacity to agents only while they actively run.

The authors, ~130 people including founder Liang Wenfeng, concede "no single mechanism can prevent all agent misbehavior," leaning on observability instead, though the arXiv paper posted September 19 has not been peer reviewed.

💡 Strategies & Tactics

> How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows: Explains how to scale robot simulation from one to thousands of parallel copies on NVIDIA graphics chips using MJWarp, so reinforcement learning collects experience far faster.

> Kilo Swarm: a 39% cost cut on the task it's built for, inconclusive on cost everywhere else: Kilo Swarm's shared message board lets subagents coordinate mid-task, cutting cost 39% on decomposable jobs where duplicated work drives spending, but proving inconclusive elsewhere.

> How SWE-Serve Exposes the Gap Between Local Tests and Live Serving: SWE-Serve tests AI coding patches through a live model server, revealing that about a third pass local checks but fail real serving.

> If You Think You Can Do Real-World Text-to-SQL: AI tools that turn plain-English questions into database queries score above 90% on public tests but crash to 10-30% on real company data, because messy schemas and jargon defeat them.

> Validate GPU Cluster Readiness Before AI Workloads Land: Run real distributed test workloads on GPU clusters before production, so NVIDIA's tool pinpoints the specific failing nodes that health checks miss.

Other news you might like

• Claude discovers a novel enzyme system with CRISPR-like repeatsLINK

• Australian prime minister says OpenAI agent accessed government health website, raises 'extreme concern'LINK

• Thinking Machines signs $65m-a-year inference deal with CrusoeLINK

• Introducing MentalHealthBenchLINK

• ChatGPT mobile app gets voice-based agentic featuresLINK

• Africa's AI moment has arrived — this 1.5B model built from scratch for 12 African languages beats Google, Meta, and Alibaba while being 8x smallerLINK

🧰 Trending tools

Superset Mobile: control coding agents from your phone, running each in an isolated workspace so you can start tasks, watch progress, and review diffs before merging.LINK

CtrlOps: manages servers with an AI terminal that generates commands for your approval, one-click GitHub deploys, visual file manager, and real-time monitoring, all running locally.LINK

Minicart: run an online store by chatting with AI agents that handle storefront setup, marketing, inventory, shipping, and customer replies in plain language.LINK

CC: google's experimental hub for testing early-stage AI products like search enhancements, Workspace assistants, and generative tools such as NotebookLM before wider releaseLINK

slop-grader: a rule-based CLI that scores text against custom rulesets, flagging filler and buzzwords line-by-line so AI agents can clean up copy.LINK

discord-mcp: provides AI agents with 208 typed tools for Discord operations, featuring safety controls, resumable server builds, and activity tracking.LINK

📚 Trending papers & reports

The capability manifold offers a way to predict a model's real skills like reasoning and planning from the resources spent building it, not just its raw accuracy, unifying today's scattered scaling rules into one map.LINK

Reasoning training keeps the scoring system that grades an AI's answers in sync with the AI itself, cutting the reward gaming that makes such training unstable, and beats existing methods across reasoning tasks.LINK

Machine unlearning often hides data rather than deleting it, and this work shows why forgotten information stays recoverable inside a model even after removal, a warning for anyone relying on unlearning for privacy or compliance.LINK

Robot camera blind spots get fixed by turning a few real demo videos into many synthetic camera angles, making robots ~5x better at tasks when viewed from angles they never saw during training.LINK

Federated learning security catches hidden sabotage in shared AI training even when attackers try to stay quiet, cutting successful attacks to ~1.5% while keeping ~95% accuracy, versus rival defenses that failed almost completely.LINK

See you tomorrow for a new dose of ☕️ AIpresso!