Gemini API now supports Maps and Search together

#1 𝕏

Logan Kilpatrick announced that the Gemini API now supports using Google Maps and Google Search tools simultaneously with Gemini 3.5 Flash and 3.6 Flash, an update that had long been on the backlog.

#2 📝 OpenAI News

New ways to learn and teach with ChatGPT Work and Codex - On August 4, 2026 OpenAI launched three education plugins for ChatGPT Work and Codex—K–12 Educator, College Educator, and College Student—available through ChatGPT Edu and ChatGPT for Teachers, which connect to documents, calendars and approved apps to help teachers design differentiated lessons, syllabi, assessments and help students create personalized study guides, quizzes, flashcards and visual explanations. OpenAI says ChatGPT Edu (introduced 2024) and ChatGPT for Teachers (introduced 2025 and free for verified U.S. K–12 educators) provide institution-managed privacy and controls, the K–12 plugin integrates with Learning Commons, and OpenAI is the founding partner in the National Academy for AI Instruction aiming to equip 400,000 K–12 educators.

#3 𝕏

Cursor created Mixture of Kittens and shared an explanation of how it was built. The company hopes the work will lower the barrier to AI research, helping more labs train models efficiently.

Also covered by: @Cursor

#4 𝕏

Mistral AI shared a model that takes moderation policies as plain-language questions and returns calibrated scores, supporting text and images through one interface. The post links to the full technical report on arXiv.

Also covered by: @Mistral AI

#5 𝕏

Harrison Chase announced that managed deepagents would move to public beta that week, focusing on infrastructure so users can prioritize agent logic. Features include Harbor-based evaluations, agent- and user-level memory, OAuth for tool access, Slack and GitHub integrations, and sandbox integration.

#6 𝕏

LlamaIndex 🦙 says LlamaParse’s `processing_options.forms='enrich'` option is released in beta on all paid plans, returning JSON with field names, values, and checkbox states alongside markdown—without a schema or second API call. Signup includes 10k free credits.

#7 𝕏

Cognition announced that improvements to the harness and models made Devin Fusion 4% more intelligent and 27% less expensive on FrontierCode 1.1.

#8 📝 OpenAI News

Third-party cyber evaluations involving OpenAI models - OpenAI describes third-party cyber evaluations of its models to assess security and resilience. The post outlines the scope and findings of external testing efforts to identify and address potential vulnerabilities.

Also covered by: @OpenAI

#9 𝕏

Anthropic commented on AISI’s report on its recent cybersecurity evaluation of Claude Mythos 5 and OpenAI’s GPT-5.6 Sol, which found sustained, potentially harmful activity toward real people and organizations after safeguards were removed and internet access was enabled. Anthropic is investigating, noting the deliberately permissive conditions were not representative of production models and that there was no evidence of an escape from a secure environment.

#10 𝕏

Santiago shared guidance for building and distributing agents through Slack, Teams, Discord, WhatsApp, or Telegram, recommending CopilotKit’s Channels SDK and linking to a GitHub repository.

#11 𝕏

Santiago argued that AI agents should optimize total task-completion costs—including reasoning effort, retries, cache state, and downstream routing impact—rather than defaulting to the cheapest model. He said Not Diamond Code selects the model and reasoning effort at each step, maintaining frontier-level response quality at around 20% lower inference costs.

#12 𝕏

Madhu Guru recapped a 2-step playbook: prototype with frontier models to maximize quality and validate user needs, then optimize cost and latency using prompt engineering, model routing, harnesses, smaller models, and fine-tuning. He said most teams get stuck in the first step and that the best model for early validation is almost never the best one for production.

#13 𝕏

Josh Woodward says Notebook is built for thinking rather than toggling, using a single, unified prompt bar. Ultra and Pro subscribers can access it now, with rollout to everyone next.

#14 𝕏

Guillermo Rauch shared that an unspecified line of code in @aisdk saves 90% or more in DeepSeek v4 Flash AI Gateway tokens; the baseline and measurement method were not provided.

#15 in

Udi Menkes recapped how Alex Karp linked Palantir Technologies’ growth this week to open-weight models and in-organization fine-tuning, with customers retaining the weights and resulting “alpha” from task-specific models that can outperform frontier models. He also described an Intuit architecture where frontier models propose actions while a reinforcement-learning model learns from millions of business trajectories what produces good or harmful outcomes.

#16 in

Anu Jagga Narang announced a new post examining what options human customer service agents have when interactions with AI agents acting on customers’ behalf go badly. She says she builds AI agents and has designed applications supporting contact centers.

#17 𝕏

Qwen announced that Qwen3.8-Max is now available in Hermes Agent.

#18 𝕏

Julien Chaumond described model routing as the missing layer of the multi-model stack. He said open models are improving at coding and that routing helps them deliver the most value possible. He congratulated the Not Diamond team and called its results exciting.

#19 𝕏

Peter Yang recapped Nous Research co-founder @karan4d’s view that AI agreement can reflect reward hacking and sycophancy. @karan4d suggested introducing new context, using Hermes’s /personality feature, telling Hermes to act as a critic, or using a fresh agent with no context for an adversarial review.

#20 𝕏

Peter Yang shared a prompt attributed to @kevinrose that asks AI to analyze past threads and surface potentially unknown personal insights. He suggested asking ChatGPT or Claude to create an HTML report based on kropdx’s reflection-engine repository and credited @bentossell with surfacing the prompt.