AI Daily · 2026-09-25

The standout release is Google DeepMind's Gemini 3.8 Live with Live Avatar, which adds near-real-time visual avatars, lip sync, and 97-language support to its live voice model and ships to enterprise users with SynthID watermarking. LangChain's LangSmith updates form a broader production-agent workflow: Fine-Tuning turns agent trajectories into SFT datasets, Trajectories provides readable session views, Engine v2 adds red teaming and automated fix validation, and Deep Agents plus Custom Apps extend deployment and bespoke interfaces. In world and video generation, Runway's WorldPrompt makes real-time simulated worlds promptable but not scriptable, while Google Research introduces a multi-agent「AI video co-director」framework for more coherent long-form video narratives and character consistency. DSPy 3.4.0 enters its LM transition with Jev probabilistic decisions and deprecation warnings, and a Latent Space essay frames biotech's two paths as Foundries that industrialize measurement versus Navigators that embed AI into everyday decisions.

North America · First-hand

Google

⭐⭐⭐ [Model Release] Introducing Gemini 3.8 Live with Live Avatar

Google DeepMind introduced Gemini 3.8 Live with Live Avatar, which adds near real-time visual presence to its native live dialogue models, following last week's Gemini 3.8 Live launch. The feature pairs low-latency streaming video generation with speech, offering precise lip-syncing, natural expressions and fluid turn-taking. Asynchronous tool calling lets it trigger tools and fetch data in the background while dialogue continues uninterrupted. It supports native multilingual speech-to-speech synchronization across 97 languages, adapting lip-sync and expressions accordingly. It is available today in Gemini Enterprise, with a preset avatar library plus custom avatars generated from a reference image (custom creation currently via enterprise allowlisting only), and all outputs are watermarked with SynthID.

⭐⭐ [Research] Automating coherent long-form video generation

Google researchers introduce an AI video co-director, a unified multi-agent framework that autonomously generates temporally consistent long-form video narratives, targeting identity drift and cascading failures in existing linear pipelines. Built as an orchestration layer on top of Gemini and Veo, it natively inherits safety mechanisms such as SynthID watermarking. The work frames long-form generation as a global optimization and world-state tracking problem, breaks it into four foundational pillars, and presents a suite of frameworks: Co-Director (to appear at COLM 2026), CANVAS (to appear at EMNLP 2026), A²RD, and VQQA. The approach uses hierarchical parameterization in which a multi-armed bandit globally searches for promising creative directions that are then injected into sub-agents' system prompts for top-down steering. The authors report substantial gains in multi-shot narrative consistency and character persistence, generating minutes-long videos while mitigating visual drift and pipeline error propagation.

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

Model Release

⭐⭐⭐ [Model Release] Runway’s WorldPrompt and the Engineering of Real-Time Worlds

Runway introduced GWM Worlds 2 earlier this month, a research preview that turns high-fidelity video and audio generation into real-time interactive simulation; Runway calls it an «autoregressive diffusion» model. Its new WorldPrompt feature is an input format for specifying a generated world and the actions within it: you can fix some aspects of the simulated environment, including the first frame, then create a series of timestamped events, and actions can even be prompted in real time. In a Latent Space interview, Runway CTO Kamil Sindi and Principal Research Scientist Robin Kahlow said WorldPrompt is a prompting mechanism rather than a programming language, offering no scripting capabilities or state control. Kahlow said the research preview is not perfect, with reliability depending on how difficult the action is — movement, for instance, works quite reliably — while Sindi said more training plus scaling data and models yields better following. Runway is reportedly valued at $5.3 billion based on its $315 million raise in February, its first release GWM Worlds launched last December, and the space also includes Google DeepMind's Genie 3, Odyssey-2 Pro and World Labs' RTFM — all with limitations, such as Genie 3 currently supporting a few minutes of continuous interaction rather than extended hours.

Product Update

⭐⭐⭐ [Product Update] Introducing LangSmith Fine-Tuning

LangChain has launched LangSmith Fine-Tuning along with smithtune, a CLI tool that turns agent trajectories stored in LangSmith into custom fine-tuned models in one end-to-end workflow. smithtune handles dataset creation and preparation from LangSmith trajectories, trains the model via Fireworks or Baseten, and runs evaluation with LangSmith, uploading results back for analysis; it targets post-training and currently supports supervised fine-tuning (SFT). LangChain says the LangSmith trajectory format records exactly what the model saw at each turn, including tool definitions and cases like deferred tool loading, so each action can be paired with its true context, while held-out trajectories provide context and reference actions for evaluation. The CLI can be run directly or driven by a coding agent via an installed skill. The post also quotes leads from Fireworks and Baseten on linking production traces to managed training and deployment.

⭐⭐⭐ [Product Update] 3.4.0

DSPy has released version 3.4.0. The release adds Jev integration through the TypeSafe client, along with three experimental decision types — Noul, Choice, and Score — that bring probability evidence into ordinary DSPy signatures, with ReAnchor calibrating those decisions against the program's metric. It also introduces native LM engines, a local CPython interpreter for trusted code, async ReActV2, and versioned documentation. DSPy 3.4 is the LM transition release and 3.5 is the migration deadline: the experimental LM types introduced in 3.3 are now replaced, while legacy custom-LM integrations and OpenAI-style messages= calls still work with deprecation warnings, and ordinary DSPy programs and lm(「hello」) calls remain supported. RLM's call-time interpreter API changes so that a factory is passed via the keyword-only interpreter_factory= option rather than a live interpreter instance or positional factory. Jev is used through the same lm= interface as other backends, with thresholds, Score cuts, and Choice weights applied locally rather than sent to the backend; decision streaming, RLM decision outputs, and generative optimizers or fine-tuning with TypeSafe are unsupported.

⭐⭐ [Product Update] New in LangSmith: Engine v2, Managed Deep Agents, Fine-Tuning, and more

At its Interrupt NYC event, LangChain announced a set of LangSmith updates organized around runtime and observability/evaluations. Engine v2 adds red teaming, which uses production traces and repos to generate and test hypotheses about issues that may not yet have appeared in production; broader issue detection covering performance trends in error rate, latency and cost as well as inefficient agent work such as repetitive tool calls or overly long trajectories; and automatic validation of proposed fixes for agents running on LangSmith Deployment, with one-click PR creation for deployment. A future self-hosted LangSmith release will support BYOK for Engine, and the post says Engine has analyzed more than 60M traces since its May launch. Managed Deep Agents v0.8 introduces user-level memory, identity-scoped auth, channels and built-in web search. Trajectories offers a conversational view of traces that LangSmith Fine-Tuning uses to tune models for specialized tasks, and Custom Apps lets teams build and publish custom LangSmith interfaces for workflows like annotation queues and experiments.

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

LangChain released LangSmith Engine v2, adding Red Teaming and automated fix testing. Red Teaming analyzes an agent's production traces and repos to proactively surface issues such as hallucinations and system prompt violations before they reach production, and is available in Private Beta to existing LangSmith Deployment users. The v2 update also detects harder-to-spot issues like inefficient agent trajectories (incorrect or repetitive tool calls) and trends in error rate, latency, and cost. Engine now automatically tests and validates its proposed prompt and code fixes before human review. According to the post, Engine has analyzed more than 70M traces and diagnosed tens of thousands of issues since launching in May.

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

A LangChain blog post describes 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 and user-level memory. It also introduces HTTP channels, file transfer in Slack, and a pre-built tool for web search. The post is a short feature-update announcement with no benchmark data or customer examples.

⭐⭐ [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 each message once in the order it first appeared; it is a projection over the traces in a thread that removes the nested run structure. It works out of the box for traces sent from LangChain, LangGraph, and Deep Agents, from agent SDKs like OpenAI and Claude, and from coding agents like Codex, Claude Code, and Cursor. Users can view threads as trajectories, score them with online evaluators, and route them to annotation queues or datasets, as well as save useful sessions to datasets and export examples for supervised fine-tuning.

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

LangChain launched Custom Apps in LangSmith, which lets users build, publish, and run custom UIs on top of LangSmith data directly inside LangSmith. Users can start from a chat prompt — on Plus and Enterprise plans, LangSmith Chat turns a description of the desired data and visualization into a working app — or build in code from provided templates against the LangSmith API with a coding agent. Once published, an app becomes a shared interface in the workspace, sparing teams the hosting, auth, permissions, and sharing work. The post highlights uses such as annotation interfaces tailored to reviewers (full traces with tool calls and metadata for engineers; user request, model response, context, and rubric for subject matter experts) and experiment-review interfaces that slice results by prompt version, model swap, customer segment, failure category, model provider, or internal evaluation rubric. LangSmith already offers default UIs including prebuilt dashboards for latency, cost, and error rates plus a comparison view for regressions across experiments.

Opinion

⭐⭐ [Opinion] Foundries vs Navigators: Lowering the Cost of Science

This Latent Space guest post by Adrian Sanborn argues that in science thinking has become cheap while doing has not: knowledge work around experiments is dramatically faster, but the throughput of the physical experiments themselves, which take days or weeks to verify, has barely improved. He describes two ways biotech has adapted: foundries industrialize measurement (through next-generation sequencing and multiplexing, high-throughput microscopy, or physical automation) to generate data an order of magnitude faster, with the experimental data itself as the differentiating asset, while navigators put AI models into the ordinary machinery of the company to drive better decisions and faster processes without needing a proprietary model or a massive dataset. Foundries require capital, years and a bet on a particular technology and are easy to see, whereas navigation is invisible by comparison yet available to every company and runs fastest at early-stage startups with no legacy to shed. The impact shows up as experiments iterating faster when analysis takes an hour instead of a week, six-figure software becoming a one-day build, and disease programs chosen from 500 candidates where a team could ordinarily evaluate five. He notes that in experimental science changing requirements are the objective, something software engineering has no real equivalent for.