AI Daily · 2026-08-05

Today's highlights center on autonomous driving and AI agents: NVIDIA released the commercially usable Alpamayo 2 Super model for robotaxis, while Google launched a suite of production-grade Gemini models and its most capable embodied reasoning model, Gemini Robotics ER 2. On the tooling side, LLM 0.32 introduced major upgrades like reasoning traces and server‑side tools, and LangChain's ecosystem showcased practical agent patterns—from Stripe's Kai built in one week to CX agent lessons from Lyft and Vodafone, plus guidance on voice agent evaluation. Meanwhile, OpenAI's ChatGPT Work surpassed ten million users in three weeks, offering a preview of an agent product aimed at a billion‑user scale.

North America · First-hand

OpenAI

⭐⭐ [Product Update] Third-party cyber evaluations involving OpenAI models

OpenAI addresses recent cybersecurity evaluation incidents involving third parties, explains the issues encountered, and announces new safeguards to strengthen AI model testing and evaluation. The measures are designed to improve the reliability and transparency of the assessment process.

Google

⭐⭐⭐ [Model Release] The latest AI news we announced in July 2026

In July Google released three new Gemini models—Gemini 3.6 Flash, 3.5 Flash‑Lite, and 3.5 Flash Cyber—optimized for production AI agents with higher efficiency and lower latency. It also launched Gemini Robotics ER 2, its most capable embodied reasoning model yet, enabling robots to understand physical surroundings and collaborate naturally with people. Alongside Samsung Galaxy Unpacked 2026, it rolled out the first Gemini Intelligence capabilities on new foldable phones and improved native data migration in Android 17, while also introducing creative tools and wildfire detection satellites.

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

Model Release

⭐⭐⭐⭐ [Model Release] NVIDIA Alpamayo 2 Super, the Frontier Open Model for Robotaxis and Autonomous Vehicles, Now Available for Commercial Use

NVIDIA has released Alpamayo 2 Super, a frontier open model for robotaxis and autonomous vehicles, now available for commercial use. The model targets rare, complex long-tail events by going beyond object detection and motion prediction to enable situational understanding, causal reasoning, and appropriate action selection. This open commercial release is expected to accelerate autonomous driving development.

⭐⭐ [Model Release] Deploy local agents everywhere with LFM2.5-2.6B

LiquidAI released LFM2.5-2.6B, a 2.6B-parameter agentic model designed for on-device deployment, supporting tool calling and multi-step workflows. It was trained via pre-training, mid-training extending context to 128K, SFT, multi-domain teacher distillation, and Agentic RL within real agent harnesses. Benchmarks show it outperforms or matches 4× larger models on instruction following and tool use, topping IFBench and Multi-IF. It achieves 220 tok/s on an Apple M5 Max and requires under 2.5 GB of memory, with day-one support for llama.cpp, MLX, vLLM, and other inference backends.

Product Update

⭐⭐⭐ [Product Update] New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter logging

Simon Willison released LLM 0.32, the most significant update since the project’s launch. It shows reasoning traces on stderr by default, integrates the OpenAI Responses API and sets GPT-5.6 Luna as the default model, while adding native support for server-side tools like CodeInterpreter and WebSearch. The Anthropic plugin gains WebSearch, WebFetch, CodeExecution, and AnthropicMCP for in-request MCP calls. The Python API now accepts a messages parameter and offers stream_events() to handle reasoning, text, and tool calls; a new 'llm openai endpoint' command runs one-off prompts against any compatible endpoint.

⭐⭐ [Product Update] How to Evaluate Voice Agents with LangSmith

The article introduces a three-dimensional framework for evaluating voice agents: execution (adherence to instructions and tool call sequences), outcome (whether the business goal was achieved), and experience (conversation smoothness). It explains how to use LangSmith’s tracing, deterministic code evaluators for explicit rules, and LLM judges for semantic requirements like policy compliance or clarification handling. The post provides practical examples for each dimension, emphasizing the need for evaluating beyond just the transcript.

⭐⭐ [Product Update] How Stripe Built Kai on Deep Agents in 1 Week

Stripe launched Kai, its company-wide AI knowledge platform, in just one week using LangChain and Deep Agents. Kai provides a context-aware assistant for all employees, deeply integrating with internal data warehouses, Slack, and Google Suite, enabling multi-turn chat with document and dashboard generation. Deep Agents served as the agent harness, handling tool-calling loops, middleware, streaming, and state management, allowing Stripe’s team to concentrate on domain-specific workflows. The platform already hosts over 1,000 skills from more than 100 teams, making an agentic coding experience accessible to non-engineers.

⭐⭐ [Product Update] llm-anthropic 0.26

Simon Willison released llm-anthropic 0.26, which adds support for the Anthropic Claude 5 model series (Fable 5, Sonnet 5, Opus 5). Built on LLM 0.32, it introduces server-side tools (WebSearch, WebFetch, CodeExecution, AnthropicMCP) accessible via the -T interface or Python tools parameter, and removes the old web_search options. Reasoning now streams as typed events, and extended thinking configuration is simplified; Claude 5 models think by default.

⭐⭐ [Product Update] PipeNetwork/minimax-h3-mlx

MiniMax released MiniMax-H3, an omni-modal generative system that accepts text, image, audio, and video to produce up to 15-second video clips with audio. The community pipe network ported it to MLX for Apple Silicon. Simon Willison tested it on an M5 Max MacBook Pro, downloading ~115 GB of models and generating a video in about 45 minutes; the visual was impressive but the audio was gibberish without prompt guidance. The port demonstrates local video generation feasibility on Mac, and the official prompting guide provides detailed control options.

⭐⭐ [Product Update] llm 0.32

Simon Willison released version 0.32 of his command-line LLM tool. The update adds support for reasoning traces, integration with the OpenAI Responses API, server-side tools, and smarter logging. Detailed release notes are available on his blog.

Opinion

⭐⭐⭐ [Opinion] Unpacking ChatGPT Work: the Agent for a Billion Users

On July 9, 2026, OpenAI launched ChatGPT Work, an agent product for knowledge work that exceeded 10 million users in three weeks and is set to merge with ChatGPT by year’s end. The post externally reconstructs Work’s inner workings—memory, proactivity, scheduling, browser use, and plugins—showing it runs on the Codex harness in cloud VMs or local desktop mode, integrates with Slack, email, calendars, and hundreds of plugins, and generates interactive artifacts. It also examines the product’s relationship with Codex and ChatGPT, framing Work as a preview of the agent experience for ChatGPT’s billion weekly users and discussing design tensions and future direction.

⭐⭐ [Opinion] As AI Increases Demands on Memory, Storage Steps Up

Surging AI demands are driving the need for massive datasets and context windows that surpass system memory limits. Simply adding more storage capacity is not enough; what's needed are useful insights from AI factories and efficient, secure storage architectures that enable them. NVIDIA discusses these topics at the recent Future of Storage event.

⭐⭐ [Opinion] Customer Experience (CX) Agents in Production: Lessons from Lyft, Vodafone, and LATAM Airlines

This article distills production lessons from Lyft, Vodafone, and LATAM Airlines on customer experience (CX) agents. Lyft built a self-serve platform for non-technical teams to configure and launch agents; Vodafone’s Super TOBi and Super Agent serve both customers and internal reps; LATAM Airlines uses Concierge and Compass to improve semantic routing and convert unstructured chats into structured signals. It also highlights patterns like frontline copilots, scalable self-serve platforms, and evaluation as a shared language, while showing how LangSmith, Deep Agents, and LangGraph support continuous improvement throughout the agent development lifecycle.