DeepSeek V3.2 dropped yesterday. Open-source. MIT-licensed. Gold medal at the 2025 Math Olympiad. Cost per 128K tokens: $0.70. GPT-5 charges $15. Meanwhile, MIT built a digital twin of 151 million American workers—and found AI can already replace 11.7% of them.

That's $1.2 trillion in wages. The tech layoffs dominating headlines? Just 2.2% of the real exposure. This edition: why collapsing inference costs change everything, a constraint technique that forces AI to be specific, and a meeting-prep workflow you can steal today.

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. 

Two models released: V3.2 (everyday reasoning) and V3.2-Speciale (high-compute). Speciale won Gold at the Math Olympiad, Informatics Olympiad, and ICPC World Finals—no internet access during testing. Both MIT-licensed.

The breakthrough: A new "Sparse Attention" architecture slashes inference costs 70%. They achieved this despite U.S. chip export controls.

[Read the full story]

MIT and Oak Ridge simulated 151 million workers, 32,000 skills, 3,000 counties. Finding: AI can already perform tasks worth $1.2 trillion in wages—11.7% of the labor market.

The iceberg: Tech layoffs are the visible tip—just 2.2%. Below the surface: routine HR, logistics, finance, and admin, spread across all 50 states. Rural areas aren't insulated. They're just not paying attention.

Read more:
[CNBC] • [Fortune]

Amazon launched AgentCore: memory, identity, and observability for AI agents. The tell is in the specs—"longest session runtime in the industry for asynchronous workloads." They're not building for chatbots. They're building for agents that run for days.

Also shipped: Nova Multimodal Embeddings (unified search across text, images, video, audio) and Transform upgrades that modernized Air Canada's Windows stack in days, not months.

Read more:
[AWS News]

Connect three dots: DeepSeek proves GPT-5-class reasoning costs $0.70 per 128K tokens. AWS launches infrastructure for agents that run continuously. MIT shows 11.7% of work is already automatable.

Here's what nobody's saying: When inference costs drop 70% in a single release, the economics of always-on agents flip overnight. At $0.70 per 128K tokens, running an agent 24/7 costs less than a Slack seat. AWS isn't building AgentCore for chatbots. They're building it for employees that never clock out.

The strategic frame: Companies still debating "should we use AI?" are asking last year's question. The 2025 question is: what happens when your competitor deploys 50 agents at the cost of one intern? MIT's Iceberg Index isn't a warning. It's a menu.

Your move: Identify one process that runs on human attention but doesn't require human judgment. That's your first agent candidate—and the math just got 70% better.

Open-ended prompts produce verbose mush. Constraints force precision.

The Prompt:

You have exactly 3 sentences to [explain/summarize/recommend] [topic]. Each sentence must: (1) contain a specific number or concrete example, (2) share zero words with the previous sentence, and (3) end with something the reader can act on today.

Why it works:

"No repeated words" forces vocabulary diversity. "Specific number" kills vagueness. "Actionable ending" converts passive information into usable output. Together, they route around the model's default verbosity.

Example output (improving cold emails):

"Personalized first lines boost response rates 36%—reference something specific from their latest LinkedIn post. Subject headers under 7 words outperform lengthy alternatives by 2.4x, so cut every unnecessary modifier. Tuesday through Thursday at 10am local time is optimal; schedule your next batch accordingly."

Browser automation for people drowning in tabs.

What: Chrome extension that builds workflows using AI "Playbooks." Unlike Zapier (API-based), Bardeen works at the browser level—it sees what you see.

Cost: Free (100 credits/month) / $10/month Pro / Enterprise on request

Steal This Workflow — Meeting Prep in 30 Seconds:

•       Trigger: Calendar event fires 30 min before meeting

•       Action 1: Scrape attendee LinkedIn (title, company, 2 recent posts)

•       Action 2: Pull last 3 email threads with each attendee from Gmail

•       Action 3: Generate one-page briefing → save to Google Drive

•       Action 4: Ping Slack with the briefing link

Setup: ~45 min. Saves 10-15 min per meeting—of research you weren't doing anyway. That's the real win: it makes you prepared for meetings you'd have walked into cold.

Honest downsides: Learning curve is steeper than the marketing suggests. Scraping breaks when sites update layouts. Occasional data accuracy hiccups.

Verdict: 4/5 — If you copy-paste between tabs more than twice daily, this pays for itself in a week.

→ bardeen.ai

Few-shot prompting can take accuracy from 0% to 90%. But most people include examples that are too similar—the model learns a narrow pattern and fails on edge cases.

The fix: Three example types—one clear, one edge, one boundary.

Task: Classify feedback as Positive, Negative, or Neutral.

Clear: "Love this product!" → Positive

Edge: "It works, I guess" → Neutral (no complaint ≠ endorsement)

Boundary: "Would be great if it didn't crash" → Negative (criticism disguised as wish)

The boundary example does the heavy lifting. Without it, "Would be great if it didn't crash" often lands as Neutral or Positive. The third example teaches the model where ambiguity resolves—and that's where classification actually happens.

🏆 Model Wars

DeepSeek V3.2-Speciale took gold at IMO, IOI, and ICPC—matching GPT-5 performance under MIT license at 70% lower cost. This isn't a incremental improvement. It's proof that compute moats erode faster than anyone expected.

OpenAI acquired a stake in Thrive Holdings. Translation: they're buying distribution because the model edge alone isn't enough anymore.

⚡ Infrastructure

AWS shipped AgentCore and Nova Multimodal Embeddings at re:Invent. The strategic tell: "longest session runtime in the industry." They're not optimizing for chatbots. They're building payroll for digital workers.

Banks are negotiating $38B in loans for OpenAI data centers. When lenders bet this big, they're not pricing in chatbot subscriptions. They're pricing in infrastructure monopoly.nters (FT).

🛡️ Security

IBM found 13% of organizations breached via AI—and 97% had no access controls. The threat isn't hypothetical anymore.

Anthropic's threat report confirmed Claude was weaponized for ransomware development. Unit 42 discovered "agent session smuggling"—injecting instructions into ongoing agent conversations. If you're deploying agents without security review, you're running with the doors unlocked.

Here's the uncomfortable math: DeepSeek V3.2 at $0.70 per 128K tokens means running a reasoning agent for 8 hours costs roughly $2.50. That's not a typo. The bottleneck just shifted from "can we afford to experiment?" to "can we afford not to?"

Your one move this week: Take your most tedious recurring task—the one you've been "getting to" for months—and run it through DeepSeek's API once. Not to automate it. Just to see what $0.70 buys you now. That number will rewire how you think about everything else.

— R. Lauritsen

P.S. Forward this to someone who thinks AI is the next Big Thing.