AI intelligence for the real world. Curated weekly by humans.
Read time: ~5 min
What’s new and interesting in AI/ML this week.
Multiple operators reported real value from AI — but only when deployments were systematic, not experimental. MIT research shows a 95% failure rate for AI projects done without a structured approach. Accenture's Hector Rocha put it bluntly: "Not everything has to be AI." The companies leading aren't the ones with the most pilots — they're the ones that started with a business problem.
Fortune ran a piece today on what senior leaders are calling the "SaaSpocalypse" — AI agents replacing per-seat software, collapsing margins, and eroding moats. Cloudflare dropped 12%, Snowflake 9%, and ServiceNow 7% in a single session last week.
The shift is subtle but significant — AI search is increasingly keeping users inside the answer rather than sending them to sources. Changes how authority, discoverability, and fact-checking work for everyone.
The Machines Are Getting Smarter. The Scoreboard Just Dropped.
Stanford released its annual AI Index this week — 300+ pages of where the technology actually stands. The numbers that matter: Agents are real now. Cybersecurity agent accuracy went from 15% to 93% in two years. Real-world agent task success jumped from 20% to 77%. These aren't demos. These are benchmarks on production-grade tasks. Adoption is outpacing everything. Generative AI hit 53% population adoption in three years — faster than the personal computer or the internet. Global corporate AI investment hit $581.7 billion in 2025, up 130% year-over-year.
The costs are scaling too. Grok 4's training run produced emissions equivalent to 17,000 cars running for a year. AI data center power capacity hit 29.6 GW — enough to power New York State at peak demand. And some things AI still can't do: tell time, fold laundry (12% success rate), generate coherent video, or do reliable financial analysis. The full report is worth bookmarking.
The Stanford AI Index 2026 is the most comprehensive annual assessment of where AI actually stands — from agent capabilities to environmental costs to adoption rates. The numbers are staggering in both directions.
Quick hits worth your attention this week.
Nearly double the 2022 vehicle. First major raise under new co-stewards Alfred Lin and Pat Grady. The money is going to late-stage AI infrastructure.
An in-house alternative to OpenClaw-style autonomous capabilities. Every major platform is building its own agent framework now.
Both cited AI. Block's stock surged on the news. Stanford's data shows software developer employment for ages 22–25 has dropped nearly 20% since 2024. The entry-level squeeze is now a quarterly earnings story.
Anthropic just dropped Claude Opus 4.7, featuring top-tier reasoning and much sharper instruction-following. This version is a beast at handling long projects, plus it actually verifies its own work before getting back to you. Just a heads-up: this is its own thing and totally separate from that unreleased Mythos Preview.
Stanford's Foundation Model Transparency Index dropped from 58 to 40. The companies building the most powerful systems are sharing less about how they work and what the risks are.
The 95% Problem
CERAWeek surfaced a number this month that deserves more attention than it got: MIT research showing a 95% failure rate for AI deployments done without a systematic approach.
Not 95% of bad ideas. Not 95% of underfunded pilots. Ninety-five percent of all unsystematic AI efforts. The operators reporting real value from AI at CERAWeek had one thing in common — they didn't start with the technology. They started with the problem.
That's the gap. Not between companies that use AI and companies that don't — but between companies that deploy it with intent and companies that deploy it with hope. The difference isn't the model. It's the method.
The best tools disappear into the workflow. You don't think about them — you just make better decisions because the right information was in front of you at the right time. No one on the rig floor cares what's running under the hood. They care whether the answer was right and whether it was there when they needed it.
We've been quiet for a reason. The infrastructure we're assembling doesn't have a waitlist — it has an invite list. More soon.
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.