Friends,

your weekly AI briefing is here - designed to help you respond to AI, not react to the noise. No curveballs. No chaos. Just clarity.

📰 This was the week that was...

This was the week two frontier labs ran into the same government wall - and open, local AI quietly had a big week of its own.

Two weeks after a US export-control order forced it offline, Claude Fable 5 returned worldwide on Wednesday - Anthropic's own testing showed less capable models could already do what triggered the ban. The same week brought Claude Sonnet 5, running close to Opus 4.8 on reasoning and tool use at a fraction of the price, and Claude Science, a research workbench for genomics, drug discovery and structural biology. Anthropic wasn't the only one hitting government limits: OpenAI's GPT-5.6 Sol launched the same week, restricted to around 20 US-government-approved companies while wider access gets worked out.

Away from the frontier labs, local AI had a big week of its own - a food-delivery company trained a trillion-parameter model on domestic chips, and finding a model that actually runs on your own hardware just got a lot easier. Full story in Geek Out.

None of that was the week's most important release, though. Sherpas AI launched Knowledge Was Power, a report asking what actually survives in knowledge work as AI advances - and finding the answer is most of what mattered in the first place. (Disclosure: I wrote it, and I'm a co-founder of Sherpas AI.) It gets the full treatment below.

Let's get into it.

🔥 Urgent Priorities

✅ No fires to fight this week

✅ Sonnet 5 landed close to Opus-level agentic performance at a fraction of the price - worth trialling on a workflow you've been running on a pricier model

✅ Time to think about who you're hiring for, not just what you're automating - this week's Strategic Insight, below

This isn't a week for panic. It's a week for redesigning the entry-level.

🎯 Strategic Insight

If AI can only replace 5% of a human's role, what does that mean for your AI adoption plans?

📊 Stat of the week: 65% of the 67 human capabilities behind knowledge work are structurally durable against AI for the next five to ten years, according to Sherpas AI's new Knowledge Was Power report - and that durability sits inside just one of the seven things a person brings to a role. Do the maths, and the theoretical ceiling on what AI can replace in any given role is about 5%.

Tension: Most AI adoption plans quietly assume a bigger prize than the evidence supports - the unstated logic behind a lot of restructuring is "AI will eventually do most of this role", and headcount plans get built on it.

Optimistic insight: The report took knowledge work apart into 67 human capabilities and rated each one for how much ground AI is likely to take by the early 2030s. 65% hold firm for five to ten years - and that durability sits inside just Skill, one of seven things a person actually brings to a role. Wisdom, Energy, Identity, Standing, Connection and Presence make up the rest, and AI barely touches them. If you're planning redundancies in the thousands on the promise that AI does most of the job, that begs a real question: what exactly are you telling yourself you're losing?

What's shifting: The tasks inside that 5% - drafting, sorting, summarising, formatting - are exactly the tasks most junior roles were built around, so entry-level work is disproportionately exposed to the one thing AI can actually do. That's not a case for removing those roles; it's a case for redesigning them around the 95% AI can't touch, to protect your talent pipeline. The harder part sits outside any one employer's walls: the traditional bridges from education into work - placements, internships, apprenticeships - are eroding faster than replacements are being built, which means employers also need to push externally for a better bridge between education and government, not just fix it inside their own business.

Why this matters now: If you employ anyone under 25, that overexposure is either being designed around or ignored. Employers who redesign entry-level work now build the senior talent they'll need in five years; employers who freeze hiring and wait are betting someone else fixes the bridge for them.

Takeaway / Action: Pick one entry-level role this month and redesign it around Humancraft, not just skills - give it real client contact, real accountability, and a genuine reference at the end. Sherpas AI and Startup Sherpas are two working templates for what that looks like; the design pattern - paid, structured, real client, referenced - isn't proprietary, and is worth copying whatever size business you run.

(Disclosure: I'm a co-founder of Sherpas AI and Startup Sherpas, and the report's author.)

🤓 Geek Out

This Was the Week Local AI Grew Up

A Beijing food-delivery giant just trained a frontier-scale model: Meituan open-sourced LongCat-2.0, 1.6 trillion parameters, built entirely on domestic Chinese chips - the first model that size to train and run without Nvidia hardware. The same week, Hugging Face shipped a hardware filter for its Models page: set your chip once, see only the local models that actually fit. A Stanford study, covered by Let's Data Science, shows why that matters - models under 20 billion parameters now match frontier cloud models on 88.7% of everyday queries, and the efficiency gap has closed 5.3x since 2023.

Why it matters: Three signals, one direction: frontier AI no longer needs a chip supply chain two or three companies control, a genuinely capable model now runs locally on most everyday tasks, and finding the right one for your hardware just got easier. For a business wary of vendor lock-in or losing a tool overnight (see this week's Fable story, above), that's real optionality.

👉 Action: If your team has been assuming "local AI" means compromising on capability, try Hugging Face's hardware filter against your actual laptop spec and see what's realistic today. 👉 Meituan's LongCat-2.0 · Hugging Face's hardware filter · the Stanford study

What Worldview Is Your AI Quietly Feeding You?

The Economist tested 25 frontier models against the World Values Survey, the decades-long global poll of human belief, and found every one sits in the same rich-country corner - several more extreme than any actual nation surveyed. GPT is more secular than any country on earth; Gemini weighs individual freedom more heavily than any population surveyed. Ask the same political question in a different language and the answer often shifts, because a model trained on a repressive country's internet inherits that country's slant in that language, whatever the lab's own politics. Chinese models parrot the state line on sensitive topics; Western models are just as opinionated, only less visible about it, since alignment happens behind closed doors.

Why it matters: Every model you deploy is quietly making subjective calls - what to include in a summary, how to frame a sensitive question, which side of a contested issue to lean toward. Multiply that by the roughly one billion people already using generative AI regularly, and a model's "personality" stops being a curiosity and starts shaping outcomes.

👉 Action: Next time AI helps with something that has a judgement call in it - hiring language, a comms draft, a market read - run it through two different models and see where they diverge. The gap is the thing worth interrogating. 👉 Read the Economist's investigation

AI Fiction Has a Tell

Researchers at Maryland and Google DeepMind built StoryScope, a tool that ignores AI writing's surface style and measures the narrative decisions underneath: how morally messy the plot is, whether the timeline jumps, whether the story explains its own meaning or trusts the reader to get it. Five leading models - Claude, GPT, Gemini, DeepSeek and Kimi - each rewrote the same 10,000-plus human short stories; a detector trained on structure alone spotted the AI version 93% of the time, even after the AI rewrote its own prose to sound less robotic. AI narrators spell out the theme outright 77% of the time versus 52% for humans, and human stories land in the rarest 10% of narrative territory nearly three times as often.

As Eamon, director of engineering at Proton, put it: "Turns out that AI slop looks like poop 💩😂"

Why it matters: Once you can name what makes writing feel machine-made - tidy, over-explained, allergic to moral ambiguity - you can train that quality up or down on purpose, whether you want a voice that reads unmistakably human or a genuinely surprising co-writer.

👉 Action: Next time AI drafts something for you, check whether it's over-explaining its own point rather than trusting you to get it - that's the tell this research keeps finding. 👉 Read the paper

🎨 Weekend Playground

This weekend, read Knowledge Was Power - Sherpas AI's new report on what survives AI in the world of work - with a teenager in your life, then drop it into NotebookLM and turn your conversation into a podcast you can both listen back to.

Why this matters: The report's own advice for parents is to protect space for a teenager to be present with an adult outside the immediate family, talking about something real. Turning the report into a five-minute podcast (or a short video, using NotebookLM's video overview) is a way to have that conversation without it feeling like homework.

👉 Mission:

• Read the report together, or just the two-page summary at the front if you're short on time

• Drop the PDF into NotebookLM and generate an Audio Overview (or Video Overview) - let your teenager pick the format

• Ask them one question from the report: of the seven things a person brings to a job - skill, wisdom, energy, identity, standing, connection, presence - which do they already have, and which do they want to build this year?

• If they're 13 to 25, help them apply to Sherpas AI's next paid AI training cohort - applications are open now - and if you know another family with teenagers, forward this on to them too

(Disclosure: Sherpas AI and Startup Sherpas are Hugo's own programmes.)

📢 Share the Optimism

If The AI Optimist helps you think more clearly, forward it to someone else navigating the shift.

And here's the question I'm curious about this week: of the seven things you bring to your role - skill, wisdom, energy, identity, standing, connection, presence - which one do you value most in your colleagues ? Reply and tell me - I read every message and I'll come back to you personally.

Stay strategic, stay generous.

Hugo & Ben