AI Daily · 2026-09-24

Google DeepMind released Gemini 3.8 Flash TTS and Flash-Lite TTS, enabling custom voice creation via natural language or 30-second samples, plus director-style control over pacing and emotion; Simon Willison's playground demonstrated multi-character dialogue generation at about 2.74 cents for 78 seconds. LangChain made a broad set of agent-development updates: LangSmith Engine v2 adds red teaming and automated verification, Trajectories provides readable session views, and Custom Apps, Fine-Tuning, and Managed Deep Agents 0.8 cover custom interfaces, trajectory-based fine-tuning, and production hosting. Google Beam expanded to six countries with HP hardware and an Industrious experience network. Elsewhere, LiquidAI shipped a DSpark speculative-decoding draft model for LFM2.5-VL-3B with up to about 3x on-device decoding speedup, NVIDIA highlighted Warp/MJWarp for GPU-scale robot simulation and joined an open-science coalition for pandemic preparedness, and TypeSafe AI introduced Jev, a non-generative classification model.

North America · First-hand

Google

⭐⭐⭐ [Model Release] Gemini 3.8 text-to-speech says hello

Google DeepMind introduced two new text-to-speech models: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. Users can create custom character voices from scratch with natural language prompts or replicate an existing voice from a 30-second audio sample, backed by built-in consent verification. The models allow line-by-line direction of audio — controlling pacing, emotion and realistic conversational sounds — for use in audiobooks, podcasts and real-time voice agents. Gemini 3.8 Flash TTS targets deep creative direction and character design across more than 100 languages and dialects, while Flash-Lite TTS targets high-volume, cost-efficient scale. Both are available across Google AI Studio, Gemini API, Gemini Enterprise, Gemini Notebook and Google Vids, with built-in safety tools such as watermarking and a library of 2,000+ production-ready voices.

⭐⭐ [Product Update] Google Beam expands with new regions, partners, and customers

Google Beam is expanding to six countries — the U.S., Canada, the U.K., France, Germany and Japan — supported by 18 channel partners, with hardware delivered by HP as HP Dimension with Google Beam, and integration with both Google Meet and Zoom. An eight-week internal Google study found that teams using Beam felt 50% more connected, found it 33% easier to ensure their feedback was clearly understood, and saw a 21% drop in the need for follow-up meetings. Google also announced its first extended network, partnering with Industrious so that Beam units can be booked at select Industrious locations starting in October. The post also notes that customers such as Netflix and Bain are already using Beam to improve remote interviews.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐ [Model Release] Accelerating vision-language models with LFM2.5-VL-DSpark

LiquidAI released an experimental DSpark draft model for its vision-language model LFM2.5-VL-3B, adding a speculative decoding path that trades a minimal memory increase for faster inference without changing output quality. The drafter follows the same architecture as the text LFM2.5-DSpark drafters: an attention-only drafter with 4 layers and a block size of 9, totaling roughly 280M parameters (193.0M decoder stack, 21.0M hidden-state projection, 65.5M Markov head), an 8.9% increase over the 3B target model. Reported decode speedups reach up to 3.13x on device (MLX on M5 Max) and up to 2.66x on an H100, with end-to-end gains up to 2.62x and 2.27x, evaluated on six vision tasks following the MMSpec benchmark. The drafter ships with day-one DSpark integrations for llama.cpp, MLX-VLM and SGLang. The post notes that speculative decoding accelerates only decode, not vision encoding or prefill, so end-to-end gains are capped by Amdahl's law when those stages dominate wall time.

⭐⭐ [Model Release] Gemini 3.8 TTS Playground

Google released two new Gemini text-to-speech models today, gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts, which come with a library of over 2,000 voices and the ability to create a custom voice from just a 30-second audio sample. Simon Willison vibe coded a bring-your-own-key playground with GPT-6 Astra that composes single-voice narration or multi-speaker conversations, previews the generated audio, shows request and response details, and saves compose settings to bookmarkable URLs. The API makes it easy to define a full conversation between multiple characters, each with different voices and voice style instructions. In his example, generating 1m18s of audio with Gemini 3.8 Flash TTS (not the cheaper Flash-Lite) took about 20 seconds and cost 2.74 cents.

Product Update

⭐⭐ [Product Update] New in LangSmith Engine: red teaming and automated testing

LangChain released LangSmith Engine v2, adding red teaming and automated testing. Red Teaming analyzes an agent's production traces and repos to understand its purpose and behavior, then tests for weaknesses and surfaces issues that have not yet appeared in production, such as hallucinations and violations of system prompts; it is available today in Private Beta to existing LangSmith Deployment users. Engine v2 also detects harder-to-spot issues, including inefficient agent trajectories such as incorrect or repetitive tool calls, and trends in error rate, latency, and cost. The new version automatically validates its proposed prompt and code fixes before human review and deployment. LangChain says Engine has analyzed more than 70M traces and diagnosed tens of thousands of issues since its May launch.

⭐⭐ [Product Update] LangSmith Custom Apps: Build custom interfaces around your agent data

LangChain launched Custom Apps in LangSmith, letting teams build, publish, and run custom UIs on top of LangSmith data directly inside LangSmith, avoiding the hosting, auth, and permissions work of self-built frontends. Users can generate an app by describing the desired data and visualization to LangSmith Chat, or start from provided templates and build against the LangSmith API with a coding agent. Stated use cases include tailored annotation interfaces that show different context to engineers versus subject matter experts, and custom experiment comparison views across prompt versions, model swaps, customer segments, and failure categories. The Chat-based app generation is available on Plus and Enterprise plans.

⭐⭐ [Product Update] Introducing LangSmith Fine-Tuning

LangChain launched LangSmith Fine-Tuning and smithtune, a CLI that turns agent trajectories stored in LangSmith into custom fine-tuned models. smithtune handles the whole workflow in one command line: building and preparing a dataset from LangSmith trajectories, training on Fireworks or Baseten, and evaluating the fine-tuned model in LangSmith with results uploaded back for analysis. It currently supports supervised fine-tuning (SFT), which trains a model to imitate examples of good behavior. The post explains that LangSmith's trajectory format records exactly what the model saw at each turn, including tool definitions and context changes such as deferred tool loading, so each action can be paired with its true context; trajectories held out from training provide the context and reference actions for evaluation. The tool is available on GitHub along with a smithtune skill that lets coding agents drive the process end to end.

⭐⭐ [Product Update] Trajectories now in LangSmith: A readable view of every agent session

LangChain launched Trajectories in LangSmith, a chronological, conversational view of an agent session. A trajectory aggregates messages from humans, AI, and tools across the main agent and any subagents, showing them in the order they first appeared and once each. Users can now view threads as trajectories, score them with online evaluators, and route them to annotation queues or datasets. Trajectories work out of the box for traces from LangChain, LangGraph, and Deep Agents, agent SDKs such as OpenAI and Claude, and coding agents including Codex, Claude Code, and Cursor. LangChain says this helps teams find where agent behavior deviated before diving into the full trace, and lets subject-matter experts review behavior in a readable format.

⭐⭐ [Product Update] Managed Deep Agents delivers a better user experience for agents in production

The LangChain blog post presents Managed Deep Agents as the simplest way to build, deploy, and run agents in production. The 0.8 release adds support for user-owned credentials, user-level memory, HTTP channels, file transfer in Slack, and a pre-built tool for web search.

⭐⭐ [Product Update] What Is Jev? A Guide to TypeSafe AI’s System One Model

LangChain's blog introduces Jev, a new model released by TypeSafe AI. Jev is not a traditional LLM and does not generate text; TypeSafe AI calls it a System One model that evaluates a state and returns typed answers and probabilities that software can use directly, trained with reinforcement learning for calibrated decisions (RLCD). The company reports up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks. Jev supports three question types — Choice, Score, and Noul (yes/no) — and evaluates multiple questions about the same state in parallel, so adding questions barely changes response time. LangChain exposes Jev through TypeSafeClassifier in the langchain-typesafe package, where .invoke() returns classification results instead of a chat response, and the state can be text, structured data, or LangChain messages.

Other

⭐⭐ [Other] How Open Science Can Help Researchers Prepare for the Next Pandemic

An NVIDIA blog post notes that when COVID-19 emerged, scientists had decades of prior coronavirus research to draw on, so they understood the virus's key proteins well enough to design vaccines in record time. The next pandemic may not offer that same head start. To improve the odds, NVIDIA has joined a coalition of global research organizations, including Google and others. The post is tied to an open protein dataset effort (the body text is truncated at this point).

⭐⭐ [Other] How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows

This second article in the State of Simulation for Physical AI series shows how to use NVIDIA Warp and MuJoCo Warp (MJWarp) to accelerate robotics simulation and learning workflows. Classic MuJoCo provides fast CPU-based robot simulation and can parallelize sampling across CPU cores, while MJWarp, built on NVIDIA Warp, brings compatible MuJoCo models to GPU scale. The article moves an SO-101 follower arm from a standard MuJoCo workflow to as many as 2,048 parallel MJWarp environments and walks through the technology and validation steps that make the transition possible. It describes NVIDIA Warp as a Python framework for authoring statically typed, GPU-accelerated kernels compiled for CPU or CUDA, with three value pillars: performance, ease of use, and capability (differentiable kernels and DLPack-style interop). A small robotics kernel example that advances point positions under gravity is included, and the series' later Newton and Isaac Lab installments are noted; no policy is trained.