Another AI Newsletter: Week 29
Agentic assistants take the spotlight as OpenAI and AWS unveil powerful new tools, Google advances graph and medical models, while billion-dollar investments and global policy shift reshape the future
Breakthrough Research or Papers
Introducing ChatGPT agent: bridging research and action
July 2025 | OpenAIOpenAI’s latest release turns ChatGPT into an autonomous agent that can carry out complex, multi-step tasks. As OpenAI explains, “ChatGPT can now do work for you using its own computer, handling complex tasks from start to finish.” For example, users can ask it to “look at my calendar and brief me on upcoming client meetings based on recent news,” “plan and buy ingredients to make Japanese breakfast for four,” or “analyze three competitors and create a slide deck.” ChatGPT will then “intelligently navigate websites, filter results, prompt you to log in securely when needed, run code, conduct analysis, and even deliver editable slideshows and spreadsheets that summarize its findings” (openai.com).
This unifies previous ChatGPT plugins (like web browsing and code execution) into one agentic model that reasons and acts in sequence.
Why it matters: This is a major step toward highly capable AI assistants. By giving ChatGPT “its own computer” (a simulated environment to browse, compute, and interact), users can delegate real-world workflows instead of just asking questions. The model still keeps users in control (it requests permission for consequential actions) but now can autonomously execute tasks “from start to finish” (openai.com). This demonstrates significant progress in agentic AI – we’re now seeing LLMs that bridge pure chat with practical automation.
Graph Foundation Models for Relational Data
July 2025 | Google ResearchGoogle unveiled a new class of graph foundation models (GFMs) designed to learn from relational database tables as graphs. Unlike traditional tabular ML, GFMs can generalize across different database schemas. As the announcement notes, “Similar to frontier language and vision models like Gemini, a GFM is a single model that learns transferable graph representations that can generalize to any new, previously unseen graph, including its schema, structure, and features” (research.google).
In practice, the team trained a Transformer-based GFM on billions of database rows and tested it on tasks like spam detection across dozens of linked tables. Compared to tuned single-table baselines, the GFM achieved dramatic improvements – yielding 3×–40× gains in average precision by leveraging the graph structure of data (research.google).
Why it matters: Relational databases power most enterprise systems, but standard ML can’t easily use relationships between tables. Graph foundation models change that by treating tables as nodes and foreign-key links as edges. The result is one model that can “push the frontiers of graph learning” across varied schemas (research.google). The huge precision boosts (3–40×) demonstrate that incorporating graph structure is crucial for many tasks. In short, GFMs extend the idea of large pre-trained models (like Transformers) to multi-table data, unlocking far better performance on real-world databases.
MedGemma & MedSigLIP: Health AI Foundation Models
July 2025 | Google ResearchGoogle announced two new open AI models tailored for healthcare. MedGemma 27B Multimodal handles complex, longitudinal patient data by combining text and images (e.g. interpreting electronic health records). MedSigLIP is a lightweight image+text encoder for medical classification and search.
The blog states: “MedGemma 27B Multimodal… adds support for complex multimodal and longitudinal electronic health record interpretation. The second new model is MedSigLIP, a lightweight image and text encoder for classification, search, and related tasks” (research.google).
These models complement Google’s earlier 4B and 27B text-only MedGemma models. Despite their size, the models run efficiently (a single GPU) and even fit on mobile. Early benchmarks are strong: the 27B variants rank among the best open models on medical QA tasks (MedQA), with the text model scoring 87.7% accuracy – within 3 points of the top closed-source systems.
Why it matters: High-quality AI for medicine has been held back by data and privacy issues. MedGemma and MedSigLIP lower that barrier. They provide “robust starting points” for healthcare AI that developers can run and modify freely (research.google). With these models, tasks like automated report generation, diagnosis assistance, or image analysis become more accessible. As specialized LLMs for health, they promise to accelerate medical research and clinical tools while preserving efficiency and privacy.
Real-World Use Cases and Demos
Pegasystems & AWS launch generative-AI modernization platform
July 2025 | ITProPegasystems announced a five-year partnership with Amazon Web Services to “accelerate IT modernization through generative AI” (www.itpro.com). The deal brings AWS’s Bedrock LLMs and a new “agentic AI” platform (AWS Transform) into Pegasystems’ Pega Blueprint modernization tool. The aim is to tackle technical debt and legacy systems – “key barriers that hinder AI adoption and modernization efforts” (www.itpro.com) – by enabling dynamic, AI-driven workflows and smoother cloud migrations.
Why it matters: By integrating generative AI directly into enterprise modernization tools, this collaboration shows how companies can use AI to overhaul outdated infrastructure. It signals a move from proof-of-concept toward real production use of AI to drive efficiency and agility in large organizations.
Microsoft saves ~$500M using AI in call centers
July 2025 | Windows CentralMicrosoft reported saving roughly $500 million in 2024 by integrating AI into its customer support operations (www.windowscentral.com). AI-powered chatbots and tools handled routine inquiries at scale, allowing Microsoft to cut thousands of support jobs while maintaining customer satisfaction. Chief Commercial Officer Judson Althoff noted that the new AI systems have kept customers happy and “improved productivity” even as headcount was reduced (www.windowscentral.com).
Why it matters: This striking cost saving is one of the largest real-world ROI examples announced publicly. It illustrates that enterprise AI – here in the form of generative support assistants – can substantially boost efficiency. Other companies will take notice that AI can have massive tangible impacts on operating costs and service capacity.
Google licenses AI coding tech from startup Windsurf (≈$2.4B deal)
July 2025 | ReutersAlphabet’s Google struck a roughly $2.4 billion agreement to license code-generation technology from AI startup Windsurf (www.reuters.com). As part of the deal, Google “acqui-hired” Windsurf’s CEO and R&D team into Google DeepMind to boost its in-house coding AI (Gemini) initiatives. The license gives Google non-exclusive use of Windsurf’s generative coding agent, while most Windsurf engineers remain at the startup (www.reuters.com).
Why it matters: This deal highlights how AI is transforming software development. By acquiring cutting-edge AI coding tools and talent, Google is embedding generative assistants into its engineering workflow. It signals that AI-driven code completion and agentic coding are moving from research into real-world production at scale, profoundly changing enterprise software development practices.
Agentic AI Announcements
OpenAI unveils ChatGPT agent for complex tasks
July 2025 | ReutersOpenAI on July 17 launched a new ChatGPT “agent” that can autonomously execute multi-step web tasks (www.reuters.com). The agent runs in a virtual browser toolkit with access to the web and user apps (like Gmail and GitHub), letting it handle prompts end-to-end. For example, it can plan and purchase an outfit for a specific event – considering “variables like weather and dress code” – by researching and acting on the user’s behalf (www.reuters.com).
Why it matters: This marks a shift from passive chatbots to active AI agents that complete complex workflows. By reducing constant human oversight, it shows how generative AI is moving toward genuinely autonomous task handling.
AWS introduces Bedrock AgentCore for enterprise agents
July 2025 | AWS News BlogAmazon Web Services (AWS) announced Bedrock AgentCore, an enterprise-grade platform to simplify building and running AI agents at scale (aws.amazon.com). As AWS explains, AgentCore “enables developers to deploy and operate AI agents with the scale, reliability, and security critical to real-world applications” (aws.amazon.com).
The suite offers modular tools – such as sandboxed Runtimes for isolated sessions (up to 8-hour tasks), Memory stores for context, a Gateway for API integration, a cloud Browser tool, and a secure Code Interpreter – plus observability dashboards and identity controls (aws.amazon.com).
Why it matters: By providing turnkey infrastructure (session management, security, long-term memory, etc.), AgentCore helps businesses bring AI agents into production more safely and efficiently. It bridges the gap between research demos and real-world deployments of autonomous AI services.
AWS AI leader declares agents a ‘tectonic change’
July 2025 | TechRadarAt AWS Summit New York 2025, AWS VP Swami Sivasubramanian heralded AI agents as a “tectonic change” in software development – a leap comparable to the birth of the internet (www.techradar.com).
He presented AWS’s agentic roadmap: Bedrock AgentCore (above), a new AWS Marketplace for pre-built AI agents and tools, and Amazon S3 Vectors (cloud storage with native vector support for embeddings). Together, these initiatives give developers an “end-to-end” ecosystem for agents – from modular components to data storage.
Why it matters: This signals that major cloud providers are fully embracing agentic AI. By building marketplaces, vector databases, and production tools for agents, AWS aims to make its platform the default choice for launching autonomous AI applications, accelerating the shift toward AI-driven workflows.
Thought Leadership and Commentary
Your guide to AI: July 2025
Jul 2025 | Air Street Press
Nathan Benaich’s latest AI newsletter notes a major industry shake-up: Meta “has launched a significant restructuring of its AI initiatives,” creating a new Meta Superintelligence Labs led by top AI executives. He also highlights AI’s massive economic footprint – OpenAI’s revenue run rate has hit $10 billion and Anthropic’s $4 billion.
Why it matters: Meta’s pivot toward “superintelligence” R&D and the staggering revenue figures underscore how central AI is becoming to tech strategy and the economy.Where AI Provides Value
Jul 2025 | Schneier on Security
Security expert Bruce Schneier breaks down AI’s edge: it isn’t better at everything, but it has clear advantages when tasks emphasize “speed, scale, scope and sophistication.” He notes AI can do some jobs blazingly fast or replicate tasks millions of times simultaneously – for example, upscaling images or targeting ads – even if humans understand these tasks well.
Why it matters: This view shifts the narrative from blanket AI threat to strategic alignment: we should leverage AI where it outpaces humans (massive scale or speed) rather than expect it to replace human versatility across the board.More advanced AI capabilities are coming to Search
Jul 2025 | Google Blog
Google announced that Search is gaining powerful AI features: subscribers will get access to Gemini 2.5 Pro and a new “Deep Search” mode. Deep Search is touted as “our most advanced research tool” – it can issue hundreds of queries, reason over the results, and generate a “comprehensive, fully-cited report in minutes.” Google also added agentic features like using AI to call local businesses for pricing or availability.
Why it matters: These updates turn Search into an AI-driven assistant, automating complex research and errands. This signals a shift toward embedding intelligent agents into everyday tools, reshaping how we find and use information.
Major Product/Tool Releases
Introducing ChatGPT Agent: bridging research and action
July 2025 | OpenAI BlogOpenAI announced on July 17, 2025 that “ChatGPT now thinks and acts, proactively choosing from a toolbox of agentic skills to complete tasks for you using its own computer.” This update adds an “agent” mode to ChatGPT so it can autonomously browse the web, fill forms, research data, generate reports, and even control apps to execute user requests. In effect, ChatGPT becomes an active personal assistant rather than just a reactive chatbot.
Why it matters: By giving ChatGPT independent reasoning abilities, this release blurs the line between conversation and action. It significantly expands the AI’s real-world utility – making it a true task-completing assistant and signaling a major step toward autonomous AI companions.
AWS launches Bedrock AgentCore and AI agent marketplace
July 2025 | Amazon News (AWS)At the AWS Summit New York on July 16, 2025, Amazon unveiled a suite of agentic AI tools. Chief among them is Amazon Bedrock AgentCore, which “enables organizations to deploy and operate secure AI agents at enterprise scale with seven core services” (covering runtime, memory, planning, skills, etc.). AWS also expanded its AI agent ecosystem with new offerings in the AWS Marketplace that let businesses quickly find, buy, and deploy pre-built agents and tools from leading providers.
Why it matters: AWS is positioning itself as a one-stop platform for building and scaling AI agents. By providing a managed infrastructure (AgentCore) and an agent marketplace, Amazon makes it easier for enterprises to adopt AI-driven agents and automation. This move could accelerate development of custom AI assistants and enterprise bots, potentially reshaping how businesses integrate generative AI.
Cornerstone Galaxy introduces AI assistants in July 2025 release
July 2025 | Cornerstone OnDemandCornerstone’s talent-management platform received a major AI-powered update in July 2025. The company says this release “introduces a host of new AI capabilities in Cornerstone Galaxy, including an intelligent in-platform assistant that empowers administrators with AI capabilities for content curation, compliance monitoring, report generation, and more.” In practice, this means HR and training teams gain AI-driven assistants that can automatically curate learning content, flag compliance issues, generate reports, and tailor employee development paths.
Why it matters: This feature rollout shows how generative AI is expanding into enterprise HR and learning tools. By embedding AI assistants directly into the workflow, Cornerstone is equipping organizations to work smarter – reducing manual tasks for admins and delivering personalized learning at scale. It exemplifies the trend of AI augmenting business software for improved productivity.
AI Safety and Ethics Developments
Global Faith Leaders Endorse ‘Rome Call for AI Ethics’
July 2025 | Financial TimesAt a Hiroshima summit, leaders of 11 major religions signed a “Rome Call for AI Ethics” advocating humane technology development. The Vatican framed AI as a transformative social force and emphasized the importance of ethically managing AI’s influence to safeguard human dignity.
Why it matters: This interfaith initiative signals a broad moral consensus on responsible AI, underscoring that diverse communities worldwide demand human-centered, ethical guidelines as AI advances.
California Courts Propose Statewide AI Use Policy
July 2025 | ReutersCalifornia’s Judicial Council voted to require all 65 state courts to adopt model policies on generative AI (or ban it) by Sept 2025. The draft rules focus on confidentiality, privacy, bias mitigation, safety, and transparency – for example, barring input of confidential data into public AI tools and mandating verification and disclosure of AI-generated content.
Why it matters: If enacted, California would become the largest U.S. court system to formally regulate AI. This sets a precedent for ensuring that AI use in legal settings is fair, transparent, and subject to oversight, addressing key safety and ethical concerns.
OpenAI Board Calls for Nonprofit Oversight and ‘Common Sector’ Governance
July 2025 | AP NewsAn independent advisory panel convened by OpenAI recommended the company remain under nonprofit oversight and adopt a “common sector” governance model that supports broader democratic participation. The board’s report urges democratizing AI by empowering communities already affected by AI, improving transparency, and investing in public-interest initiatives (e.g., arts, health, AI literacy).
Why it matters: The recommendations highlight growing calls for inclusive, societally accountable AI governance. They push OpenAI toward greater transparency and community involvement, reinforcing that advanced AI development should serve the public good.
Industry Investment
Mira Murati’s Thinking Machines Lab Raises $2B Seed Round at $12B Valuation
July 2025 | TechCrunchMira Murati’s new AI startup, Thinking Machines Lab, has raised $2 billion in a seed round at a stunning $12 billion valuation. The round was led by Andreessen Horowitz and included Nvidia, Accel, ServiceNow, Cisco, and AMD as backers (source). Observers note that the deal “marks one of the largest seed rounds — or first funding rounds — in Silicon Valley history,” despite the company not yet disclosing any product details.
Why it matters:
This record-setting seed round signals unprecedented investor enthusiasm for elite AI founders—pushing funding boundaries even before any commercial release.
Google Hires Windsurf Execs in $2.4B AI Coding Talent + Licensing Deal
July 2025 | ReutersGoogle has hired top executives and researchers from Windsurf, a leading AI coding platform, through a $2.4 billion licensing and acquihire deal. CEO Varun Mohan, co-founder Douglas Chen, and core R&D personnel will join Google DeepMind to accelerate Gemini’s AI coding capabilities (source). The remainder of Windsurf will continue independently, offering liquidity to existing investors (source) and transitioning to new leadership (source).
Why it matters:
This deal exemplifies how Big Tech is accelerating internal AI efforts through strategic talent acquisitions—often using licensing to avoid the red tape of full acquisitions.
Cognition AI Acquires Windsurf to Expand AI Developer Platform
July 2025 | ReutersIn a follow-up move, Cognition AI has acquired Windsurf outright. The startup—valued at $1.25 billion with $82 million in annual revenue (source)—offers a popular integrated development environment for AI. Cognition will absorb all of Windsurf’s IP, brand, and engineering teams to enhance its agent-based tooling (source).
Why it matters:
This acquisition highlights the intense competition to control foundational AI tooling. With both Google and Cognition securing major pieces of Windsurf, the startup has become a cornerstone in the evolving AI coding landscape.
Regulatory Policy
California Courts Propose AI Governance Rule for All State Judges
July 2025 | ReutersCalifornia’s Judicial Council is preparing to mandate that all 65 state courts either adopt formal policies for generative AI use—or prohibit it entirely. The proposed rules, led by Chief Justice Patricia Guerrero’s task force, would enforce confidentiality, privacy safeguards, bias mitigation, and transparency in AI-assisted court operations. The proposal includes a prohibition on entering confidential court data into public AI tools, mirroring growing concerns seen in other states.
Why it matters:
If enacted, California would become the largest U.S. judicial system to formally regulate AI. It sets a legal standard for AI use in the courts and may influence national adoption of AI governance in sensitive institutions.
UN Urges Global Deepfake Regulation and Verification Standards
July 2025 | ReutersA new UN report presented at the AI for Good Summit in Geneva warns of the growing impact of AI-generated deepfakes on elections and financial fraud. It calls for robust authentication and verification mechanisms across digital platforms, including watermarking and content-scanning tools. The International Telecommunication Union emphasized that public trust in media is rapidly declining—making authenticity safeguards a global priority.
Why it matters:
The report signals urgent international momentum around deepfake detection policy. It may catalyze standards for AI-generated media transparency across platforms and governments, particularly ahead of key global elections.
EU Study Pushes for ‘Opt-In’ Copyright Model for AI Training
July 2025 | PC GamerA EU-commissioned study led by Professor N. Lucchi calls for a shift away from the bloc’s “opt-out” AI copyright regime, arguing it “effectively treats silence as consent.” The report urges a pivot to an “opt-in” framework requiring creators’ explicit permission, along with clearer guidelines on data mining, fair compensation, and stronger legal protections.
Why it matters:
This study reflects growing pressure within the EU to modernize copyright law for the AI era. If adopted, its recommendations could shape new legislation requiring stricter IP protections for content creators across Europe—and influence global AI training norms.
Powered by OpenAI Deep Research API