Agentic AI
🧪 Agents Start Rewriting Scientific Infrastructure
What happened
OpenAI published a field report on eight agent-assisted scientific computing projects, mostly in life sciences, showing coding agents helping modernize research software through maintenance, optimization, language migration, and GPU-native redesigns. The report says researchers are shifting from hands-on implementation toward verification and orchestration, while agents take on more of the engineering work.
Why it matters
This is one of the clearest signs yet that agentic coding is moving beyond app demos and into the domain software that actually bottlenecks research. Just as important, OpenAI’s own takeaway is that validation, stewardship, and scientific judgment remain the hard part, so the winning setup is still human-led and agent-accelerated.
What’s next
Expect more labs to use coding agents for software upkeep, migrations, and performance work before trusting them with open-ended scientific reasoning. The practical race now is not just better agents, but tighter evaluation loops and clearer long-term ownership of the code they produce.
Generative & Enterprise AI
📝 Google Turns Docs Into a More Active AI Coworker
What happened
Google rolled out two Workspace updates for Docs: Gemini can now generate and edit visuals directly inside documents, and it can summarize comment threads, draft replies, add comments, and suggest edits grounded in reviewer feedback. Both rollouts began July 28 for eligible business, enterprise, education, and consumer tiers.
Why it matters
This is a platform move, not a novelty feature drop. Google is pushing Gemini from “help me write” into multimodal document production and collaboration management, bringing it closer to owning the full workflow inside enterprise productivity software.
What’s next
Watch whether these features become habitual enough to change how teams review documents, not just how they draft them. If adoption sticks, the next enterprise AI battleground will be who controls the review loop, the feedback loop, and the final edit inside core workplace tools.
☁️ Compute Spending Keeps Getting More Extreme
What happened
Recursive Superintelligence signed a $410 million multiyear compute deal with AWS, a commitment that TechCrunch reported represents the bulk of the company’s funding to date. CEO Richard Socher said the company expects to ship its first tangible products within months as it pursues open-ended self-improving systems.
Why it matters
This is a sharp reminder that frontier AI competition is increasingly constrained by infrastructure access, not just model ideas. It also shows how some labs are reorganizing around “agent count” and automated development loops, putting more capital directly into compute instead of traditional headcount.
What’s next
Expect more bespoke cloud deals as labs try to lock in capacity and shape infrastructure around their workloads. The near-term test for Recursive is simple: whether expensive compute turns into real products fast enough to justify this kind of spending.
Physical AI
🚕 London’s Robotaxi Fight Gets a New Entrant
What happened
Baidu began testing autonomous vehicles in London through its partnership with Lyft and Freenow, with human safety operators onboard and dozens of vehicles operating in Brent. The companies are targeting public ride-hailing in 2027, subject to regulatory approval.
Why it matters
Physical AI markets are consolidating around city-by-city deployment races, and London is becoming one of the most contested proving grounds in Europe. Baidu is not just exporting software here; it is trying to secure a place inside a live Western mobility network before the market hardens around a smaller set of robotaxi winners.
What’s next
The immediate watchpoints are regulatory approvals, operational safety data, and whether hybrid fleets of human drivers plus robotaxis can scale without political backlash. If London works, it strengthens the case that robotaxi expansion will be decided by local partnerships and permitting, not by model quality alone.
💡 Bottom Line
AI is moving deeper into the systems that matter—scientific software, workplace collaboration, cloud infrastructure, and physical transportation. Agents are taking on more execution, but advantage is shifting toward the organizations that can validate the work, control the workflow, secure the compute, and navigate real-world deployment. Capability is accelerating and orchestration is becoming the moat.
⚙️ Try It Yourself
Put an Agent Through a Review Loop
Choose a small piece of code, analysis, or technical documentation that needs improvement. Use a coding agent to:
Identify outdated, inefficient, or fragile sections.
Propose and implement a cleaner version.
Explain every meaningful change and flag anything it could not verify.
Then move the explanation into Google Docs and use Gemini to summarize reviewer comments, draft responses, and suggest revisions based on the feedback.
Do not judge the experiment only by whether the agent completed the task. Track how much time you spend validating its work, resolving uncertainty, and deciding what should remain under human ownership.
The goal is to experience the emerging workflow firsthand: agents execute, humans verify, and the review loop becomes the real operating system.