A Chinese lab trained a frontier AI model for $6 million. American labs burned through billions.

The $6 million model matched their benchmarks—then went open-source. Geoffrey Hinton spent Christmas warning that progress "exceeded his expectations" and 2026 will eliminate "many, many jobs." OpenAI is now paying someone $555,000 just to worry about what they're building. The hype era got a vibe check.

Here's what actually mattered in 2025—and the five bets to make before everyone else catches on.

In Today’s 5-Minute AI Digest. You will get:

1. The MOST important AI News & research
2. AI Prompt of the week
3. AI Tool of the week
4. AI Tip of the week

…all in a FREE Weekly newsletter. 

Sam Altman posted a job for "Head of Preparedness"—someone to predict AI's harms before they ship. The role has burned through three leaders in 18 months. Altman called it "a stressful job" where you "jump into the deep end immediately." Mental health impacts, cybersecurity exploits, and deceptive models are now official concerns. They're building faster than they understand.

Geoffrey Hinton told CNN that AI now excels at "reasoning and deceiving people." His 2026 prediction: capabilities double every seven months, replacing "many, many jobs." He puts odds of AI "taking over" at 10-20%. The Nobel laureate who built the foundations is losing sleep over what's coming.

MIT's August study: 95% of enterprises saw zero ROI from AI investments. VCs surveyed by TechCrunch predict 2026 will finally be "the year"—the same prediction they made in 2024 and 2025. The 5% who cracked it aren't talking. They're compounding.

AI-generated junk now exceeds half of all English web content. Merriam-Webster made "AI slop" Word of the Year. Meta killed Instagram AI search after backlash. Humane AI Pin flopped. 2025 proved that shipping AI features isn't the same as shipping useful ones.

DeepSeek built R1 for $6 million—1/16th of GPT-4's training cost—matched the benchmarks, and open-sourced it. Nvidia stock cratered. Every AI budget assumption died overnight.

The lesson: algorithms beat brute-force compute. Anthropic paid $1.5 billion for pirating books—but the same judge ruled training on legally acquired works is fair use. Legal playbook: settled. Cursor hit $29.3B valuation building for agents, not autocomplete.

The pattern across all three: leverage over scale. Better algorithms. Legal clarity. Practical tools. That's your 2026 edge.

[Read the full story]

What it does: Forces AI to assume your project already failed—then diagnose why. Surfaces risks your optimism buries:

It's six months from now. [Your project/decision] failed completely. As a brutally honest advisor, write a post-mortem: What went wrong? What warning signs did we ignore? What should we have done differently?

Why it works: "What could go wrong?" triggers defensive optimism. "What DID go wrong?" from the future bypasses it. Psychologist Gary Klein found pre-mortems catch 30% more risks than traditional planning.

Real result: Ran this on a product launch plan. AI flagged onboarding friction and week-3 churn risk we'd glossed over. Both would've hurt.

What it is: AI-native code editor that understands your entire codebase—not just the file you're in.

Why now: $1B ARR. $29.3B valuation. Two years old. Half the Fortune 500 uses it. This isn't early-adopter territory—it's the new baseline.

One-liner: "Copilot suggests lines. Cursor understands projects."

Rating: ⭐⭐⭐⭐⭐

Standout features: Multi-file Composer • 8 parallel agents per prompt • Background tasks • Claude/GPT/Gemini/DeepSeek support

Link: cursor.com — Free tier available, Pro $20/mo

The tip: Replace "Can you...?" with direct commands. Not "Can you summarize this?" → "Summarize this." Not "Would you explain...?" → "Explain."

Why it works: Questions trigger hedging. The model considers whether it CAN, adds caveats, softens claims. Declarations skip deliberation and execute. Outputs get more direct and confident.

Limitation: Won't fix unclear requests. If the task is ambiguous, no phrasing saves you.

Pro move: Combine with output priming. Instead of "Can you give three options?" write "Three options:" and let the model complete. You set the frame.

1. Multi-agent becomes standard. Google admitted single-model safety failed. One AI watching another isn't clever—it's required.

2. Content licensing explodes. Anthropic's $1.5B settlement proved piracy costs. Legal training is fair use. Smart money flows to licensing deals.

3. Non-coders ship code. Cursor, Claude Code, Copilot agents—the skill floor drops. Product managers prototype. Founders skip the dev hire.

4. Efficiency beats scale. DeepSeek proved $6M and better algorithms beats $6B and brute force. More labs chase architectural wins.

5. Safety becomes a sales pitch. After ChatGPT mental health lawsuits, enterprises want guardrails. Robust safety wins deals.

You just learned:

• 2025's winners found leverage—efficiency, legal clarity, practical tools—not scale

• 95% of enterprises got no ROI from AI; the 5% are now lapping competitors

• 2026 bets: multi-agent, licensing, coding for everyone, efficiency over scale, safety that sells

Now pick one and move.

Most people will skim this, nod, and change nothing. The ones who spend 30 minutes this week—testing Cursor, drafting a licensing strategy, auditing their AI stack—will enter 2026 with a head start their competitors won't understand until Q2.

— R. Lauritsen

P.S. Know someone still asking "what should I do about AI?" Forward this. It's the only 2025 recap they need.

P.P.S. Next week: The first major AI drop of 2026 lands January 6. We'll have the breakdown—and what it means for your stack—before the news cycle catches up.