Big week for AI announcements — GPT-5.5, DeepSeek V4, OpenAI workspace agents. But the headline number isn't the parameter count this time. It's the token count. Read on for what that means for the rig floor.
Read time: ~5 min
What’s new and interesting in AI/ML this week.
The two companies announced an expanded partnership at CadenceLIVE Silicon Valley last week, combining Cadence's multiphysics simulation engines with NVIDIA's Isaac robotics libraries and Cosmos world models. The target: the long-standing performance drop robots take when moving from virtual training to the real world. End-to-end agent-orchestrated workflow from world-model training through physics sim to deployment feedback. Cadence shares rose 4% on the news. Sim-to-real is the bottleneck for every physical AI deployment — including the ones eventually showing up on rigs.
PatSnap published a technology landscape report showing how ML, deep learning, and signal processing are converging to automate what used to take teams of geophysicists weeks. The tools are getting faster. The question is whether the interpretation quality keeps up.
GPT-5.5 Just Dropped. Here's What Actually Changed.
OpenAI released GPT-5.5 on Wednesday — and for once, the benchmarks back up the marketing. Coding got serious. Terminal-Bench 2.0: 82.7% (up from 75.1%). SWE-Bench Pro: 58.6%. Engineers with early access say it has "conceptual clarity" — it understands why something is broken, not just what is broken. Computer use is real now. OSWorld score: 78.7%. The model can plan multi-step tasks, use tools, check its own work, and keep going through ambiguity.
Same speed, fewer tokens. GPT-5.5 matches GPT-5.4 latency while using significantly fewer tokens to complete the same tasks. More intelligence, same cost. That's the part that matters for production deployments. Cybersecurity got its own access tier. First time OpenAI has created a tiered access policy for a specific domain. The model found decades-old zero-days at scale during testing. NVIDIA gave 10,000+ employees early access. One engineer said it was "blowing my mind."
The efficiency story is the real story. Fewer tokens for better answers means production AI deployments just got cheaper. For operations teams running real-time inference on sensor data, that math changes everything.
Quick hits worth your attention this week.
The Chinese lab's latest uses Mixture-of-Experts and cuts memory usage by 90% at 1M token context. The gap between US and Chinese frontier models continues to narrow.
The list: humanoid data, LLMs+, supercharged scams, world models, the new war room, weaponized deepfakes, agent orchestration, China's open source bet, artificial scientists, and resistance. Agent orchestration and artificial scientists are the two your team should read.
Gartner's latest report highlights a significant shift: 62% of oil and gas firms are now using AI operationally. By 2028, autonomous AI agents are expected to boost efficiency by over 10%, with frontline staff like drilling engineers formally training them as a core job duty. This marks a clear move towards real-time operational autonomy in the sector.
Emory University researchers used an AI pipeline to extract previously unknown interaction laws from experimental plasma data — 99% accuracy, non-reciprocal forces that challenge existing theories. When AI finds things physicists missed, the tool is no longer just a tool.
Persistent agents that live inside your team's workspace, maintain context across conversations, and can be assigned ongoing responsibilities. The shift from "chat with AI" to "work with AI" is happening faster than most orgs are ready for.
The Quiet Part
Everyone's talking about GPT-5.5 this week. The benchmarks, the coding demos, the NVIDIA engineers losing their minds. Fair enough — it's impressive.
But here's what caught our attention: the model uses fewer tokens to reach better answers. Not more compute. Not bigger context. Fewer tokens. That's an efficiency story, not just an intelligence story.
It matters because the same principle applies on the rig floor. The best decisions don't come from more data. They come from the right data, structured correctly, delivered at the right moment. You can drown a drilling engineer in sensor readings or you can surface the three things that actually matter for the next 50 feet of hole.
The AI industry is learning what operations people have always known: brute force doesn't scale. What scales is context. Knowing which signal matters, when it matters, and why.
That's the quiet part of this week's news. Not that the models got smarter — but that they got more efficient. Intelligence without efficiency is a science project. Intelligence with efficiency is a tool.
We're building tools.
We're rolling out the DrillSense connector for Claude. It brings your drilling data, real-time ops, geosteering and more right into your Claude sessions. Reach out for early access!
Data ≠ Decisions. Context changes everything. DrillSense is the intelligence layer for drilling operations, built for the people who make the calls.
Know someone who should be reading this? Send them the archive, or subscribe from drillsense.com.