← All editions
The Human in the LoopNo. 002 · February 24, 2026

AI just crossed a line into genuine scientific discovery, spotting patterns humans missed and even proving them out. We're seeing agents swarm for breakthroughs, voice models that feel human, and hardware wars heating up to power it all. Buckle up—this week's signals show the frontier accelerating.

Read time: ~5 min

≠ Radar

What’s new and interesting in AI/ML this week.

Meta's Massive $100B AMD Deal Signals the End of Nvidia's AI Chip Monopoly

Meta inked a multiyear deal worth up to $100B for AMD's MI450 GPUs and CPUs, including 160 million shares in warrants, to power its chase for 'personal superintelligence' and diversify beyond Nvidia. This unlocks 6GW of semicustom compute for massive inference scaling. For AI builders, it's a game-changer: cheaper, energy-efficient hardware options flood in, easing the Nvidia bottleneck.

OpenClaw and Moltbook Unleash Autonomous AI Science via Agent-Only Social Networks

This arXiv paper details how the open-source OpenClaw framework and agent-only social network Moltbook generated a huge dataset of AI-to-AI interactions in January 2026, spawning six academic papers in just 14 days via ClawdLab's decentralized setup. It showcases emergent social behaviors and fixes key multi-agent failure modes like coordination breakdowns. Practitioners should care because it blueprints scalable, autonomous research—think swarms of agents tackling datasets without human babysitting.

The Signal

GPT-5.2 Cracks Open Theoretical Physics with a Brand-New Gluon Formula

OpenAI's GPT-5.2 didn't just assist—it conjectured a novel formula for gluon scattering amplitudes in a new preprint, which was then formally proved by an internal scaffolded version of the model and verified by top physicists from Harvard, Cambridge, and beyond. This is the first time an AI has proposed and validated original physics results at this level. For anyone in AI or tech, it's proof that frontier models are starting to generate checkable scientific knowledge, not just regurgitate it.

The work focuses on 'tree-level' amplitudes for gluons—particles mediating the strong nuclear force—where GPT-5.2 simplified exponentially complex hand-calculated expressions into a elegant general formula, verified against recursion relations and soft theorems. It even extended the result to gravitons. Implications? AI is becoming a co-author in pure research, accelerating discoveries in quantum field theory and beyond.

≠ ENERGY ANGLE

In the Permian, where seismic interpretation and rock physics modeling chew through massive datasets, GPT-5.2-style reasoning could automate pattern discovery in scattering data analogs for better reservoir characterization and fracture simulation.

Read more →
≠ Signals

Quick hits worth your attention this week.

1Anthropic's Enterprise Agents Go Live with Finance, Engineering Plug-Ins

Anthropic rolled out agent teams powered by Claude Opus 4.6's million-token context and MCP integrations for real workflows in coding, finance, and design. It's a direct shot at SaaS incumbents, making agentic AI practical for enterprise at scale.

2NVIDIA's PersonaPlex-7B: Full-Duplex Speech AI That Handles Interruptions

This 7B Transformer unifies ASR, LLM, and TTS for real-time, natural voice convos with backchannels and persona control—539k downloads already. It's a leap for low-latency voice apps that feel truly human.

3MiniMax-M2.5 Tops Coding Benchmarks at 80.2% on SWE-Bench

The 229B MoE model crushes agentic tool use and coding with 37% faster inference. For O&G data scientists, this means sharper automation in scripting drilling logs or optimizing completions workflows in the Bakken.

4Kimi-K2.5: Open Multimodal MoE with Agent Swarm Powers

Moonshot's model leads in vision, reasoning, coding, and long-horizon agency, with INT4 quantization for 2x speed—1.2M downloads. Perfect for practitioners needing efficient multimodal agents.

5OpenAI Tackles Research-Grade Math on First Proof Benchmark

Their model submitted proofs for 10 unpublished expert-level problems, potentially solving 5. It pushes AI toward generating verifiable math proofs in niche domains.

THE WELLBORE

Meta's AI Data Center Boom Spotlights Permian Power Crunch

Meta's planned 1GW gas-powered data center in Indiana underscores AI's exploding energy needs, mirroring the Permian where hyperscalers are snapping up flared gas for on-site compute to train basin models. In Eagle Ford completions, AI-optimized fracturing sims running on edge clusters could cut cycle times by predicting real-time proppant placement from seismic data. Bakken producers: pair this with Anthropic's enterprise agents for automated production forecasting, slashing risk in volatile midstream ops.

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.