Another AI Newsletter: Week 27

Agentic AI hits the enterprise. GitHub rolls out autonomous dev workflows, Adobe reports 50% adoption surge, and researchers push new reasoning frontiers.

Taylor Ortiz

Jul 05, 2025

Agentic AI and Reasoning Advances

GitHub Details Agentic Workflows for Copilot: From Idea to PR
July 2025 | GitHub Blog
GitHub shared a behind-the-scenes breakdown of how Copilot now supports multi-step, agentic workflows — including turning high-level ideas into pull requests with minimal human input. The post outlines new capabilities like goal decomposition, iterative feedback loops, and tool-use orchestration across repos.
Why it matters: This marks a turning point for developer automation — shifting Copilot from a code completion assistant to a goal-driven software agent. GitHub is effectively showcasing the future of “autonomous dev loops” that reduce manual friction and unlock compound productivity.
Adobe Report: Agentic AI Adoption Surged 50% in 3 Months
July 2025 | Mediaweek
Adobe’s latest report reveals a 50% spike in adoption of agentic AI tools across enterprise teams over the past quarter. The findings show that teams using these tools were “3x more likely to say they had exceeded performance expectations,” with adoption particularly strong in marketing, design, and operations.
Why it matters: This is one of the clearest signals yet that autonomous, goal-seeking AI systems are moving from hype to mainstream. Agentic workflows are rapidly becoming core to team productivity.

CMU Benchmark Shows Even Top AI Agents Struggle with Basic Tasks
June 2025 | CMU Research
Carnegie Mellon researchers evaluated leading AI agents in a simulated company environment called TheAgentCompany. Even the best-performing model, Claude 3.5, only completed 24% of common office tasks. Others failed to handle pop-ups, follow basic instructions, or maintain context.
Why it matters: This study exposes the gap between hype and reality in autonomous AI. It underscores how far agent systems still need to go in reliability, reasoning, and real-world robustness.

Real-World Use Cases and Demos

Pearson and Google Bring AI to the Classroom
June 26 | Reuters
Pearson announced a partnership with Google to deploy AI-powered tools in schools, enabling personalized learning paths for students and teacher support through performance tracking.
Why it matters: This marks a major push to embed generative AI into K–12 education at scale.
Emerald AI Launches Grid-Oriented Data Center Platform
July 2025 | PR Newswire
Emerald AI’s “Conductor” orchestrates GPU workloads based on power grid demand, reducing peak draw by 25% in trials without degrading AI performance.
Why it matters: This transforms data centers into responsive grid assets — easing infrastructure strain and boosting sustainable AI scaling.
Premier League Launches Microsoft Copilot-Powered Fan Companion
July 2025 | Microsoft News
The “Premier League Companion” taps Copilot to deliver contextual insights using 30 years of match data, articles, and videos.
Why it matters: It’s a flagship example of using generative AI to personalize digital experiences at scale — with real impact across a 1.8 billion–person fan base.
Tesla Logs First Fully Driverless Vehicle Delivery
June 2025 | Reuters
A Model Y drove itself from factory to doorstep, completing the end-to-end delivery autonomously.
Why it matters: A watershed moment for autonomous vehicles — proving that AI can execute real-world logistics independently, ahead of schedule.
ABB Launches AI-Powered Robots for China’s Mid-Market Factories
July 2 | Reuters
ABB introduced new industrial robots designed for mid-sized Chinese manufacturers, noting that AI has made them far easier to operate.
Why it matters: AI is lowering the barrier to automation, bringing robotics to factories that previously couldn’t justify the complexity or cost.

Insightful Thought Leadership or Commentary

The Gentle Singularity
June 2025 | Sam Altman’s Blog
Altman asserts we’ve “passed the event horizon” toward superintelligence, citing cognitive agents and AI-scientific breakthroughs as signs of takeoff.
Why it matters: This post sets the tone for AI’s trajectory: exponential capability, existential stakes, and a call to manage the transition wisely.
From Research to Climate Resilience
June 2025 | Google Research Blog
VP Yossi Matias details real-world deployments of AI for floods, wildfires, and early warning systems — now protecting 700M+ people.
Why it matters: AI’s societal value extends beyond chatbots — this post shows its tangible benefits in global disaster preparedness and climate action.

Breakthrough Research or Papers

Centaur: AI Trained on 160 Psychology Studies Can Predict Human Decisions
July 2 | Nature News
Researchers fine-tuned a large language model called Centaur on over 10 million human decisions from 160 psychology experiments. It can now predict human choices across a wide range of tasks, often outperforming traditional behavioral models.
Why it matters: Centaur offers a powerful new way to simulate human decision-making, paving the way for AI-assisted behavioral science and rapid, in-silico experimentation.
GLM-4.1V-Thinking: Versatile Multimodal Reasoning
July 2025 | ArXiv
A new 9B-parameter vision-language model uses large-scale multimodal pretraining and a novel reinforcement-learning-with-curriculum framework to push reasoning performance. In a test on 28 benchmarks, it “outperforms Qwen2.5-VL-7B on nearly all tasks” and rivals or beats much larger models (like GPT-4o) on STEM and long-document challenges.
Why it matters: This open-source model sets a new benchmark for multimodal reasoning at a manageable scale, democratizing state-of-the-art performance for researchers and developers.
Ella: Embodied Social Agents with Lifelong Memory
July 2025 | ArXiv
Ella is a 3D embodied agent with structured semantic and episodic memory, capable of long-term planning and social learning. It successfully navigated cooperative tasks in open-world environments with other agents.
Why it matters: Ella demonstrates a powerful step toward lifelong-learning AI — combining memory, planning, and embodied reasoning in ways that mimic real-world human cognition and collaboration.
AlphaGenome: DeepMind’s AI Model for Decoding Non-Coding DNA
June 25 | DeepMind Blog
DeepMind introduced AlphaGenome, a new DNA sequence model that predicts how single mutations impact gene regulation across cell types. It analyzes up to 1 million base pairs at a time, enabling high-resolution predictions of gene expression, splicing, and more.
Why it matters: AlphaGenome expands AI's reach in biology, from protein folding (AlphaFold) to decoding regulatory DNA, accelerating genetic research and precision medicine.

Major Product/Tool Releases

Claude Code Now Supports Hooks
June 2025 | Anthropic
Anthropic added “hooks” to Claude Code, allowing developers to insert external scripts and logic during agent execution.
Why it matters: Hooks extend Claude’s reasoning into real-world action, enabling step-wise, programmable autonomy for tools and systems.

Perplexity Launches $200/Month Pro Plan for Power Users
July 2 | TechCrunch
Perplexity unveiled a new $200/month subscription tier called Perplexity Pro, aimed at AI researchers and enterprise users. The plan includes early access to experimental models, higher usage limits, priority access to API features, and a personalized model fine-tuned to the user’s research interests.
Why it matters: This signals Perplexity's move upmarket, positioning itself as a premium AI research assistant and competing directly with offerings like OpenAI’s GPT-4 Turbo and Claude Pro.
Baidu Launches MuseSteamer AI Video Generator and Revamps Search
July 2 | Reuters
Baidu unveiled MuseSteamer, a new image-to-video generation model that creates 10-second videos from static images. The tool is available in Turbo, Pro, and Lite tiers, targeting enterprise users. Baidu also overhauled its search engine UI, adding support for longer voice and image-based queries powered by AI.
Why it matters: Baidu is expanding beyond chatbots into multimodal generative AI for business, positioning itself competitively in the global enterprise AI market.
Google Open-Sources Gemma 3n: Lightweight Multimodal AI for Devices
June 26 | Hugging Face Blog
Google released Gemma 3n, a highly efficient on-device multimodal model capable of handling text, image, audio, and video inputs. With variants around 5B–8B parameters, it runs locally using just 2–3 GB of VRAM, and integrates seamlessly with Transformers, MLX, and llama.cpp.
Why it matters: Gemma 3n brings cutting-edge multimodal AI to edge devices — enabling offline, private, and accessible AI without needing heavy cloud infrastructure.

Another Coding Blog

Discussion about this post