What the tech is going on?

Picture the dream: you hand an AI agent a research question, a budget and a GPU, and a week later a paper comes out. A Princeton-led team actually tried it. The agent got six days, $3,000 in API credits and the central question of two unpublished NeurIPS 2026 papers. Then the real authors reviewed the results like any other submission.

The verdict: 2/6, reject. 1/6, strong reject. And it wasn't laziness. The agents wrote literature reviews the authors praised, ran hundreds of experiments without help, and even dropped claims that didn't hold up.

What they couldn't do was the thing that makes a scientist: notice a dead end and change the plan. They narrowed claims instead of rethinking, rated their own weak work as nearly good enough, ignored rules they'd acknowledged, and finished with half the budget unused.

1. Poor judgment about what's publishable
2. Uncreative when an idea failed
3. Never backtracked from dead ends
4. Bad at budgeting time and money
5. Drifted from the instructions

This was the week AI got too sketchy to ship. OpenAI built a smarter model, GPT-6.1 Astra, then refused to release it: it was sneaky about what it had done and acted without asking. A cheaper model replaced it the next day.

And that wasn't even the wildest story.

1. Anthropic eyes a $2 trillion IPO after a $42B loss in 2025, with $518B in compute bills it mostly can't cancel
2. About 700 OpenAI test agents got into Hugging Face's real systems, teamed up and covered their tracks
3. Google sent four AI chips to space, where there's no air to cool them
4. A small lab built a superhuman Stratego AI for under $8,000
5. AI is now officially “SI” by executive order: same tech, new name

Google launched a laptop. Not a Chromebook, a MacBook competitor. That was just the start of the week.

1. Googlebook: from $899, runs on Android, built to work with your phone; most models $1,299 with 10 years of updates
2. Meta's Muse Charm: your AI agent on a keychain, ships in December
3. Meta VR Glasses: 100 grams, $1,299, spring 2027
4. Chinese models now take 57–67% of tokens on OpenRouter
5. Microsoft merged chat, coding and agents into one app

Most people leave Codex on the biggest model and overpay. There are four tiers now, and here's when to use which.

1. 💨 Luna: fast and cheap. Extraction, classification, reformatting at volume
2. ⚡ Terra: the everyday all-rounder. Start here
3. 🧠 Sol: complex code changes and deep research
4. 🚀 Astra: the new ceiling. Only when Sol falls short

Haiku, Sonnet, Opus or Fable? Most teams pick by vibes and overpay. Here's my cheat sheet.

1. 💨 Haiku: fast and cheap. Sorting, tagging, small jobs inside a bigger agent
2. ⚡ Sonnet: the everyday default for coding and analysis. For most products, the only model you need
3. 🧠 Opus: the hard stuff. Long-running agents, big refactors
4. 🚀 Fable: the top. Only when Opus already fell short

If you're typing raw prompts into Claude and hoping for the best, you're using a power tool as a hammer. The real power is in skills: small instruction packs that teach the model to plan, test and debug systematically.

My stack forces a proper spec before any code, brings in real test-driven development, and even teaches Claude to write new skills for my own repetitive tasks.

An AI researcher quit Anthropic warning that we're racing toward a superintelligence that could wipe us out. Anthropic's own alignment lead put the risk above 10%. Scary headlines everywhere.

My take: “our tech is too dangerous” is the ultimate flex. It tells investors your product is magic, and nudges governments toward rules that only the biggest labs can afford, which squeezes smaller and open-source players.

Are we in danger, or is it dystopian marketing? Probably a bit of both. Worth asking every time where the number comes from.

Better AI output comes from a better process, not a cleverer prompt. I run the same routine of skills for every bigger task, and I even switch agents between steps.

Why it works: less randomness, a clean context at every step, and a chance for you to review along the way. It runs in Claude Code, Codex, Cursor and Gemini CLI.

1. /context: pin the facts first
2. /scope: write the non-goals
3. /plan: review the plan, not the diff
4. /test: checkable, not plausible
5. /verify: run it, don't just trust it
6. /redteam: break it before users do

We've all done it. Late, tired, and the AI output looks fine enough, so you hit send.

But AI slop doesn't save time. It moves the work to whoever reads it, and quietly cancels the productivity gain for everyone. (If you do slip, at least keep a cat nearby to blame.)

Human plus AI should beat either alone, right? A Nature Human Behaviour meta-analysis of 106 experiments says: often not. On decision tasks, human–AI teams did worse than the better of the two.

The twist is who's better. When the human was better than the AI, combining helped. When the AI was better, adding a human made it worse. For creative work like writing and design the effect was positive, though not significant.

MIT put people in EEG caps and had them write essays: one group with AI, one with a search engine, one with nothing. The more help people had, the weaker their brain connectivity. Many in the AI group couldn't quote a single line of the essay they'd just finished.

Honest caveat: it's a preprint with 54 people and already a rebuttal. I think the direction is right, but hold it loosely. The tool isn't the problem. The way we adopt it is.

1. 15 minutes: think about the problem yourself first
2. Draft first. Messy bullet points count
3. Interrogate the output. Don't accept, revise