AI Daily · 2026-09-16

Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two speech-to-speech models aimed at real-time conversation with parallel reasoning and visual context; the Extended Thinking variant leads several voice benchmarks and is available through the Gemini API, Workspace, the Gemini app, and Search. Google also used the day to spotlight its science and language agenda, from the AlphaGenome Atlas and WeatherNext 3 to Gemini-based live translation and transcription across hundreds of languages. OpenAI is testing Sponsored Agents in ChatGPT Ads and adding Ads Manager plus HubSpot and Shopify integrations, moving AI assistance deeper into ad creation and commerce. Mistral and Mozilla paired up to bring privacy-focused, multilingual models to Firefox Smart Window, while research posts from Google and IBM/Hugging Face addressed inference bottlenecks in query fan-out and the gap between one-shot agent success and repeatable performance.

North America · First-hand

OpenAI

⭐⭐ [Product Update] Reimagining advertising with AI

OpenAI introduced new AI-powered experiences for ChatGPT Ads. It is testing Sponsored Agents, which let users start a clearly labeled conversation with a business-sponsored agent after clicking an ad in ChatGPT; the conversation is distinct from ChatGPT's independent answers and from the user's original conversation, and the feature is being tested with select advertisers in the United States. For advertisers, an Ads Manager plugin in ChatGPT lets them create, update and analyze campaigns with natural-language prompts, while AI assistance in Ads Manager suggests copy and imagery based on the landing page and campaign objective, with advertisers able to review and edit before using them. An opt-in AI text customization feature adapts existing headlines and descriptions to the conversation context and automatically translates ad copy into a user's preferred language. OpenAI also integrated ChatGPT Ads with its first CRM partner, HubSpot, and first ecommerce partner, Shopify: HubSpot customers can connect a ChatGPT Ads account to create ads, track performance and follow up on leads, and US-based Shopify merchants can use the new ChatGPT Ads app in the Shopify App Store.

Google

⭐⭐⭐ [Model Release] Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking

Google DeepMind introduced two real-time dialogue models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, which it describes as its most advanced live dialogue models to date, with major upgrades in intelligence and parallel reasoning. The models handle complex reasoning, real-time visual context, and background task execution without interrupting a conversation. Gemini 3.8 Live is built for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding, while Gemini 3.8 Live Extended Thinking targets high-complexity tasks with multi-step reasoning. Gemini 3.8 Live Extended Thinking scores 82.6 to take the top spot on Artificial Analysis' Speech to Speech Quality Index, 68.6% on τ-Voice, 35.1% on Sierra's τ-Voice-banking, and 97.7% on Big Bench Audio; Gemini 3.8 Live placed second in the Speech Agent Arena. The models are available through the Gemini API, Google Workspace, the Gemini app, and Search.

⭐⭐ [Product Update] Building AI to accelerate science and improve lives

James Manyika, Google's SVP of Research, Labs, Technology & Society, describes how Google is applying AI to accelerate science and improve lives, saying Google technologies now support more than 300 languages spoken by 7 billion people, or 86% of the global population. He lists recent advances: mapping all 9 billion possible single-letter genetic changes in the human genome with AlphaGenome Atlas, released openly to researchers; WeatherNext 3, described as Google's most advanced global weather model, delivering 50% more accurate precipitation forecasts a day or more ahead and already used in its products; a Planetary Prediction Engine combining global health, food security and socioeconomic data, used in the Ebola outbreak in the Democratic Republic of the Congo and to identify vulnerable U.S. communities across 21 CDC health indicators; and scaled AI research aimed at cutting aviation's climate impact, applied in the U.K. with the government and in Asia. The post also notes that Nobel-Prize-winning AlphaFold has predicted all 200 million known protein structures and is used by 4 million researchers in 190 countries, that AlphaMissense predicts disease-causing mutations, and that Google released interactive insights with its AI & Economy ATLAS.

⭐⭐ [Product Update] AI for everyone in every language

Google says its technologies and products now power everyday interactions in more than 300 languages spoken by over 7 billion people, about 86% of the global population, and Google Translate has grown from a handful of languages in 2006 to more than 250 today. The post says Google is moving beyond text translation to native audio intelligence, training models such as Gemini to process audio directly and capture tone, pacing, emotion and context, including code-switching like Spanglish and Hinglish. Gemini 3.5 Live Translate powers real-time spoken translation across 70 languages and 2,000+ language pairs, while Gemini 3.5 Transcribe is described as Google's most precise speech-to-text model and drives the Rambler feature in Android Gboard. The 1,000 Languages Initiative aims to support the world's 1,000 most-spoken languages, and its Universal Speech Model was trained on 12 million hours of audio and uses cross-lingual transfer learning to improve speech understanding for low-resource languages.

⭐⭐ [Research] Bypassing inference bottlenecks: Accelerating complex AI search with Retrieve-for-Train

Modern search and recommendation applications are increasingly expected to return a coherent set of results rather than a single best match; for a query like 「camping gear」, users want a complementary slate such as a tent, sleeping bag, portable stove and headlamp. Systems use query fan-out to break a broad prompt into related sub-queries, but teaching an LLM to do database-aware query decomposition dynamically drains a large thinking budget, because zero-shot LLMs are general autoregressive text predictors that are not optimized for the specific geometric manifold of a target corpus and need extended test-time computation to optimize set-level properties (diversity, coverage, complementarity, coherence) while staying grounded to a fixed database. In their ICML 2026 paper 「Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion」, Google Research introduces the Retrieve-for-Train framework: offline reinforcement learning discovers reward-aligned fan-outs that are compiled into supervision, and these optimized exploration behaviors are distilled into a lightweight diffusion retriever, enabling efficient single-pass query fan-out at inference without test-time thinking tokens. The post also describes two challenges of relying on generic models: paraphrastic collapse, where zero-shot LLMs generate redundant near-synonymous queries (e.g. 「bohemian festival fashion」 and 「bohemian festival clothes」 for the prompt 「Bohemian festival style」, missing directions such as fringe jackets, crochet dresses or suede boots), and the autoregressive latency bottleneck, where models typically need hundreds of intermediate chain-of-thought tokens to plan their expansion before emitting actual search terms.

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

Model Release

⭐⭐ [Model Release] Gemini Live audio

Google released Gemini 3.8 Live and 3.8 Live Extended Thinking, two new speech-to-speech models roughly similar in shape to OpenAI's GPT-Live family. Simon Willison had GPT-6 Astra Extra High read the documentation and build him a web UI for trying the new models: you pick a model and voice preset, enter an optional system prompt, and start a voice conversation in the browser, including the ability to interrupt the model while it is speaking. The implementation uses no libraries; it connects to the wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent WebSocket endpoint and uses a Web Audio API AudioContext for both capture and playback. He also links a Gemini Live tutorial for getting started with that WebSockets API.

Product Update

⭐⭐ [Product Update] Mistral and Mozilla are bringing open, private and multilingual AI to your web browser

Mistral and Mozilla announced a partnership under which Firefox Smart Window (beta), Mozilla's AI browsing assistant, is now powered by Mistral models. Smart Window helps users make sense of complex searches, remember something they clicked away from, and source information from their browser tabs. Mistral will power Smart Window for users in France and North America, with the United Kingdom and Germany expected to follow later this year. The companies say privacy protections are built in: conversations are not saved on Mozilla's servers by default, and partners such as Mistral agree to zero data retention. Mistral also says its models are fine-tuned on regional languages, dialects and cultural context, and frames the deal as extending its traditionally enterprise-focused work to consumers.

Research

⭐⭐ [Research] Your Agent Aced the Task. Will It Do It Again?

In a Hugging Face blog post, IBM Research argues that average success rates hide agent unreliability: a ReAct agent backed by GPT-4.1 on AppWorld posts a Mean@5 of 77.4%, but succeeds on all five repeated runs for only 53.0% of tasks (Pass^5), a 24.4-point consistency gap that reaches 30 points on hard tasks. Most benchmarks report only Mean@k, while Pass^k answers whether the same request will succeed again. The authors introduce the Consistency Analyzer, which takes a single recorded trajectory and no ground truth, resampling each decision point with one call requesting k completions (k=5 by default) to find flip-prone steps instead of re-running the task end to end. Adding consistency guidelines on top of their earlier ALTK-Evolve halves the gap from 24.4pp to 12.0pp (same-task Pass^5 +16.0pp, similar-task +13.0pp) with no loss in average accuracy; full methodology and evaluations are in an arXiv technical report.

⭐⭐ [Research] Can Skills Learned in Games Transfer to Real-World Work?

Good Start Labs trains AI models on games, having spun out of media and tools company Every last October with $3.6 million in funding from General Catalyst, Inovia, Every and angel investors. The idea came from a 2025 Twitch stream in which frontier models played Diplomacy; co-founder and CEO Alex Duffy noticed that OpenAI's o3 won every game by planning a future betrayal, while Claude Opus 4 refused to lie and lost. The company recently trained a 30B model inside the game 1830: The Game of Railroads and Robber Barons and then tested it on financial research tasks. Both the single-turn question-answering design and a multi-turn terminal agent that uses tools to explore, plan a strategy and adapt in real time improved their respective in-game objectives, but only the terminal-agent design improved performance on the Finance-Agent benchmark. Duffy's larger idea is that the way a game is presented to an AI determines which skills it learns and whether those skills carry into work outside the game.

Other

⭐⭐ [Other] Emerald AI, Google and NVIDIA Launch Alliance to Advance Flexible AI Data Centers

Emerald AI, Google and NVIDIA announced the launch of the AI Energy Management Alliance (AEMA). The source describes it as a first-of-its-kind coalition advancing data centers that can dynamically manage their electricity use. The source also states that AI factories are the infrastructure of the intelligence era, and that scaling them responsibly will depend as much on innovation across the grid as inside the data center.