welcome
SMARTGIRLS Tech · Session 6

What just happened
in AI.

The last few weeks in the AI world, in one hour. The parts that actually matter for your business.

SMARTGIRLS Tech
Why we're doing this

You don't need every update

The goal is not to keep up. It's to know which wave is worth catching.

2 / n
The month in five lines

The headlines

We'll take the ones that matter for you.

3 / n
The word behind all the noise

What everyone's arguing about: "AGI"

AGI, short for Artificial General Intelligence, means AI that's as good as a person at almost any task, not just one narrow thing. Today's AI is a brilliant specialist. AGI would be a capable generalist.

The whole field is stuck on one question: are we heading there, and how fast? Two camps, next.

4 / n
The two camps

Will AI keep going all the way?

The optimists

In late July, Sam Altman said we're "now, in the singularity." And the progress curve really does look almost exponential.

The skeptics

Serious scientists (Yann LeCun, Gary Marcus) say today's methods won't reach human level. The gains are shrinking.

Nobody actually knows. The skill isn't picking a side. It's judging it yourself, from the benchmarks, not the hype.

Altman: Relentless podcast (Jul 2026) · LeCun · Marcus4 / n
The story of the year

The timelines just flipped

Last autumn a top AI engineer called agents "slop, a decade away." Two months later: "I've never felt this behind. The profession is being refactored."

Andrej Karpathy, who built AI at Tesla and OpenAI. When the skeptics turn, that's the news.

Karpathy on X · 80,000 Hours (2026)4 / n
This week · the price war

The models keep getting cheaper

DeepSeek · V4-Flash

China shipped its fast, cheap model this week and cut prices again. The whole market follows.

Anthropic · Claude Opus 5

Top quality for about a third of the old flagship price, with a big memory for long tasks. This is what runs under Claude Code.

OpenAI · GPT-5.6

Reportedly cut the price of its cheapest tier hard, after using AI to make itself more efficient.

The pattern, every few weeks: better and cheaper. You benefit without lifting a finger.

Anthropic · Caixin / DeepSeek (2026)4 / n
What it costs you

So what does it actually cost?

The top model, summarising 100 customer reviews: about 7 cents.

You won't count cents, though. For you it's a flat subscription: about €20 a month, or €100 for the top model all day.

The models keep getting cheaper. Your bill barely moves.

Anthropic pricing (2026)8 / n
What might land next

An unreleased OpenAI model, Astra, reportedly solved ten problems mathematicians were stuck on for years. For about $2,000.

OpenAI calls Astra its next major model, and Sam Altman demoed it in Washington. A prediction market gives it about a 56% chance of launching within a month.

You can't use it yet. But "the next big one" might be weeks away, not years. Take it with a grain of salt.

Forbes · The Next Web (Aug 2026)5 / n
The big shift this year

From answering to doing

Reconcile these invoices into a sheet and draft the client email.
Done. Here's the spreadsheet and a draft email to review.
✓ It works for hours, pulling from your own apps.
✓ You review the result, like you'd check an intern.

Tools like ChatGPT Work and Notion's agents do this now.

Bloomberg · OpenAI (2026)6 / n
The money behind it

The AI industry is spending about $600 billion on data centers this year. Every year, they need to earn that back.

And it's not all hype: OpenAI and Anthropic together are on track for over $200 billion a year. Demand is so high that a four-year-old AI chip now costs more to rent than when it was new.

The whole industry is betting the house. You get the upside without the bill.

80,000 Hours (2026)7 / n
You already pay for these

Your everyday tools grew AI

Canva Code 2.0

Describe a booking page or a quiz, get a real working site. Now on free accounts too.

Google Workspace

Gemini builds a full, on-brand deck, or edits images and diagrams, right inside Slides and Docs.

Microsoft 365 Copilot

Now shows images from your own files in its answers, and lets you pick GPT-5.6 or Claude inside Office.

Shopify Sidekick

Now voice-controlled on every screen. Ask it about stock, let it draft the email.

You don't need new tools. The ones you already have just got smarter.

Canva · Google · Microsoft · Shopify (2026)8 / n
This week, and it's about you

The EU AI Act just went live

You stay in charge, and now you also say clearly when it's AI.

European Commission (Aug 2026)9 / n
Step back for a second

So what does all this mean for you?

New models every week. Cheaper every month. Altman in Washington talking "singularity." It's a lot of noise.

The real question underneath: when a new model lands, what can it actually do for your business, and what can't it? To answer that, you need to know how we even measure AI.

13 / n
How we measure AI

A benchmark: one test, every model

what an AI drew: "a pelican on a bicycle"

One playful test: ask every model to draw a pelican on a bicycle, in code. Same task each time, so you can watch them get better.

The serious ones (MMLU, SWE-bench) are mostly maths and coding, though. A model can ace those and still not do your job.

Topping a maths test isn't the same as running a business. So how good is it at that?

simonwillison.net/2025/Jun/6/six-months-in-llms14 / n
The honest part

It still can't run the business

what a person does ~90% Maths & coding 3% A whole real job

the best model, on a month of real product & strategy work ("AA Briefcase")

Aces the tests it can practise on. Lost on a real job. (A $500 vending machine? It makes $10k; a sharp human, $60k.)

80,000 Hours (2026) · Vending-Bench 2 · AA Briefcase10 / n
A little theory · how AI learns

It learns by trying, a million times

Nobody hand-codes the answers. Like a dog with a treat, the AI chases a reward and repeats whatever earns it.

1 · Try It writes an answer. 2 · Score Right or wrong. 3 · Reward Good tries stick. repeat about a million times

That's reinforcement learning. The catch: it can only learn this way when the answer can be checked, right or wrong.

DeepSeek-R1 · Tülu 3 (Allen Institute)13 / n
The key word: verifiable

Why it's brilliant at maths

A maths answer is verifiable: you can check it instantly, right or wrong. So the AI solves millions of maths problems, gets scored each time, and gets brilliant at it.

Verifiable
Maths, code. A clean right or wrong the AI can train on, over and over.
Not verifiable
Strategy, a client call, a month of work. No instant right or wrong, nothing clean to score.

What's checkable, it masters. What's fuzzy and human, it doesn't. That gap is the whole story.

DeepSeek-R1 · Tülu 3 (Allen Institute)12 / n
So here's what matters for you

Your job: make the work checkable

AI gets reliably good only at the clear, checkable parts. So the real skill is yours: break your work into those parts, hand them off, and keep the judgment.

That's why building good skills and learning to articulate your work clearly matter so much. You're not letting AI do your job. You're handing it the right pieces.

You stay the boss. You just get very good at explaining the task.

19 / n
Homework · this week

Hand off one whole task.

Not a question. A finished thing.

Post how it went on the forum. We'll dig in together.12 / n
How we do this together

Between now and Friday

See you Friday

SMARTGIRLS Tech13 / n
← → navigate · F fullscreen
EN NL