Welcome, humans.

We’re LIVE NOW discussing our predictions for 2026! Come hang!

We’re breaking down our boldest AI predictions for 2026—who's winning, who's losing, and what wildcards nobody's talking about yet.

Drop your own 2026 predictions in the chat and join the debate!

Just click the image below to join:

Here’s what happened in AI today:

• Gmail launched AI Inbox features and made AI writing tools free.

• Chinese AI companies MiniMax and Zhipu AI went public in Hong Kong.

• Dell shifted away from AI-first marketing.

• Deepseek V4 coming on lunar New Year.

Today’s Prompt Tip of the Day is THE Story of the Day

Context engineering was THE AI buzzword of 2025: how to structure information so agents understand exactly what you need. But here's the thing: we should absolutely be doing this with humans too.

I (Grant) wrote a whole thing about this on the website and how it changed my life, but here’s the TL;DR and how to apply it to yours.

Recent example: Someone I worked with asked me and another coworker a question. I didn't know the answer, but I went straight to the one person who definitely would have it. Got it in 5 minutes. If I'd just pinged whoever was nearby? Could've taken days.

That's human context engineering.

Now let’s turn this into a Framework. Two TED talks changed how I ask for help (from both AI and humans):

• Dr. Heidi Grant: We think our needs should be obvious. They're not. 90% of workplace help only happens when someone explicitly asks.

• Barbara Sher: Always state your wish (what you want) AND your obstacle (what's stopping you). “If you don't say both, nothing happens.”

We can apply this to AI as well as humans. 

For AI:

• ❌ “Help me analyze this data”

• ✅ “Help me analyze customer feedback—I'm terrible at stats and need patterns for tomorrow's presentation, but can't spot correlations between demographics and satisfaction”

For Humans:

• ❌ “Anyone know how to update the dashboard?”

• ✅ “I need to update our customer dashboard by EOD—I've never touched the API and the docs aren't clear on authentication. @ProjectOwner, can you point me to the right setup guide or hop on a quick call?”

OR

• ❌ “Does anyone know about contracts?”

• ✅ “I need a lawyer who handles tech startup contracts in California—I'm signing my first SaaS agreement and don't understand the IP clauses, but my budget is under $2K. Does anyone have a referral?”

How This Changed My Life: In July 2024, Noah posted in this very newsletter asking for writers (clear wish) who could write like him (obstacle). I barely checked my email back then, but happened to see it and jumped in with exactly what I could offer: experience running a different 500K+ audience, AI knowledge, and the right writing style.

We'd never met before. But because Noah did good context engineering—being specific about what he needed and why—I knew immediately if I was the right person. Ten months later, here we are.

Barbara Sher is, to quote Claude, absolutely right about this:

That's true for AI (which has access to the entire internet's worth of written knowledge) AND for humans (who each hold unique experiences and connections).

So that’s the take: Context engineering isn't just an AI technique; it's a life skill. Start asking for what you want with complete context. State your wish, name your obstacle, and watch how much faster you get real help.

Whether you're prompting Claude or reaching out to a colleague, good context engineering is the difference between getting what you want… or not.

FROM OUR PARTNERS

LTX-2: Multimodel AI Generative Video Model

LTX-2 by Lightricks, a production-ready AI video model designed to move at the speed of your ideas.

Built for creators and teams who need both speed and control, LTX-2 delivers high-quality motion, consistent outputs, and flexible style tuning, so you can iterate quickly without sacrificing precision.

From content pipelines to visual prototyping and format exploration, LTX-2 helps teams move from concept to output with less friction and far fewer retries.

See what production-ready AI video really looks like.

Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 600K readers here!

1. *Wispr Flow turns your speech into clean, final-draft writing across email, Slack, and docs. It matches your tone, handles punctuation and lists, and adapts to how you work on Mac, Windows, and iPhone. When writing stops being a bottleneck, work flows. Start flowing for free today.

2. Gmail launched an AI Inbox showing to-dos (bills due, appointments to confirm) and catch-up topics grouped by category, plus AI Overviews answering natural language questions like "Who was the plumber that quoted me last year?" without opening emails—and made "Help Me Write," thread summaries, and replies free for all 3 billion users (video).

3. Situation Monitor (created by Reggie James) gives you a live global map tracking conflicts, military bases, nuclear facilities, strategic waterways, and the Pentagon Pizza Index—which spikes when officials order late-night delivery during crises—free.

4. MiniMax builds multimodal models generating text, audio, video, and music—their products include Hailuo (text-to-video), Talkie (character conversations), and the M2 language model—and went public in Hong Kong with shares jumping 109% (raised $619M).

5. Zhipu AI (Z.ai internationally) builds the GLM language model series—their latest GLM-4.7 tops open-source coding benchmarks with multi-step reasoning—and went public in Hong Kong with shares rising 13.2% (raised $558M at $6.55B valuation).

6. Lambda rents on-demand access to Nvidia GPUs for training models (from $0.50/hour to clusters of thousands)—Nvidia became their biggest customer through a $1.5B leaseback deal—and is raising at least $350M ahead of a potential H2 2026 IPO (raising $350M).

7. Spangle AI generates custom storefronts in real-time based on where shoppers arrive from—click an Instagram ad and see social-browsing layouts, search for dresses and get event-related products—using ProductGPT to adapt without user history, delivering up to 50% conversion lifts for REVOLVE and Steve Madden (raised $15M).

8. Livedocs connects databases/CSVs and uses an AI agent to answer data questions—ask "Which features drive retention?" and get automated SQL queries, charts, and insights without code.

Around the Horn

• DeepSeek reportedly plans to release V4, its next flagship AI model, around Lunar New Year, with internal tests showing it outperforms Claude and GPT in coding tasks.

• X restricted Grok's image generation to paying subscribers only after international backlash over non-consensual nude images.

• Microsoft's Copilot web market share stayed at just 1.1% while ChatGPT dominates at 64.5% and Gemini surged to 21.5%.

• Elon Musk's lawsuit against OpenAI will go to trial in March after a judge found evidence supporting his claims.

• Epoch AI found Chinese AI models have lagged US models by seven months on average since 2023.

• Dell shifted away from AI-first marketing at CES 2026, with its head of product Kevin Terwilliger bluntly stating consumers "aren't buying based on AI" and "AI probably confuses them more than it helps," marking a major change from a year ago when Dell was “all about the AI PC.”

• Surge AI called LMArena “a cancer on AI,” claiming it rewards formatting over accuracy after disagreeing with 52% of votes analyzed.

• Security researchers found IBM's 'Bob' AI coding agent vulnerable to prompt injection attacks that bypass command validation and download and execute malware without human approval.

FROM OUR PARTNERS

The Practical Guide to Agent Observability

When LLMs call tools in loops, things can break—silently. Agent observability helps you trace every decision, catch drift early, and deliver agents your business can trust.

Download The Practical Guide to Agent Observability to learn how leading teams monitor and debug agents in production.

Get the Guide

Intelligent Insights

1. Jacob Rintamaki (featured above) wrote the The Final Offshoring (pdf), an awesome (and beautiful!) website essay (we need a good word for this… webessay?) which argues robotics will finally work due to video models and scaling creating data flywheels (with robots costing closer to iPhones than cars thanks to “robots making robots” and cheaper actuators), leading to:

   1. Commoditized physical labor causing broad deflation that requires new economic frameworks (Fed inflation target updates, robot futures markets, American sovereign wealth funds),

   2. Geopolitical instability risks (remittance-dependent countries collapsing without immigration pathways)

   3. Existential questions about human meaning in abundance, warning that without a compelling vision that includes ordinary Americans, tech faces the same fate as Boston's collapse, while individuals risk “wasting away” without purpose, as illustrated through two short stories about a man building a sailboat and an engineer realizing he's no longer needed, ultimately suggesting meaning comes from human connection rather than pure optimization.

   4. P.S: If you like this, you should read 1. Android Dreams and 2. Grant’s reaction to Android Dreams (which admittedly needs a rewrite and its own webessay…)

2. Regarding Boston’s collapse, Will Manidis argues Boston collapsed as a tech hub (producing just $100B in enterprise value versus San Francisco's $14T despite having MIT, Harvard, and Y Combinator), due to three factors:

   1. Progressive taxation (including delayed QSBS adoption until 2022, a millionaire's tax, and 6.25% SaaS sales tax).

   2. Organized venture capital misconduct that elite networks were too intertwined to police.

   3. An “inputs-focused mentality” prioritizing universities over execution.

   He then warns San Francisco is repeating these same mistakes while the broader tech industry fails to articulate why innovation matters to voters who increasingly view AI as wasting resources and enabling scams.

3. Cursor introduced dynamic context discovery based on a core insight (according to Malte Ubl): since AI models are trained on coding tasks that navigate large filesystems, agents work better when everything is represented as files—tool responses, chat history, MCP tools, terminal sessions, and critically, pre-compaction context all go into the filesystem so they remain accessible after summarization.

   1. Malte said, “we all migrated our agent inputs to have file system representation” because “agents are great at filesystems,” though he cautions “this may not be how we build agents forever, but it's the right starting point now.”

4. MIT researchers proposed the “Platonic representation hypothesis”, or the idea that vision models and language models are converging on similar internal representations of reality as they grow more capable, with experiments showing stronger models develop more aligned representations than weaker ones even when trained on completely different data types (images vs. text), though the hypothesis remains contested and has practical implications for training multi-modal AI systems.

5. Dan Shipper of Every published a comprehensive technical guide (coauthored with Claude) on building agent-native software architecture, covering five core principles including parity between UI and agent capabilities, granular atomic tools, and emergent capability, alongside practical patterns for using files as a universal interface, agent execution with code samples, and mobile-specific agent design patterns.

A Cat’s Commentary

P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.