|
|
|
|
This week an AI didn't summarise, sort or chat. It found something true
that no human had ever found, and proved it well enough that a
Fields Medalist backed it.
|
|
|
Together with Nocturne
Know someone who nods along in AI conversations and understands none of
it? Send them the glossary.
Ten words, ninety seconds each.
|
|
|
The big one
An AI just did real, new mathematics
OpenAI's next model family, Astra, produced machine-checkable proofs for
ten open problems spanning group theory, geometry, coding theory and
cryptography — published on 1 August alongside the actual proof files, so
anyone can verify them by running the checker, not by trusting a press
release.
The one that matters: back in May,
the same model disproved the Erdős unit distance conjecture — an
80-year-old geometry question nobody had cracked since 1946. Fields
Medalist Tim Gowers said he'd have recommended it for a top journal
without hesitation, then helped write a version humans could follow.
Compute cost: about $2,000.
What it means for you: nothing
changes on your laptop today. But "AI just remixes what it's read" was
always the safe rebuttal to the hype, and it's getting harder to say
with a straight face — even narrowed down to one model, one field, and
output that still needed human mathematicians to make it readable.
The full story, told well (5 min) →
OpenAI's own announcement →
|
|
Quick hits
|
|
Claude just got 50% more expensive
Anthropic's Claude Sonnet 5 launch pricing ends 31 August; from 1
September the rate jumps from $2/$10 per million tokens to $3/$15.
So what: that's a
developer-side number, not something you'll see on a consumer app bill
directly — but if a tool you pay for runs on Sonnet, expect the price
to follow eventually.
Anthropic's own pricing page →
|
|
...and DeepSeek just got dramatically cheaper
DeepSeek's V4-Flash left preview on 31 July at $0.14/$0.28 per million
tokens — independent testing put a full benchmark run at 3 cents versus
$3.15 for a comparable Anthropic model.
So what: "cheap"
no longer means "worse." Worth shopping around before assuming the
pricier option is the better one.
Quartz on the benchmark →
|
|
Microsoft is bundling Copilot into one super app
Announced 29 July: every Copilot experience, consumer and work, merging
into a single app.
So what: most
people won't adopt a new AI tool on purpose — they'll get one dropped
into software they already open every day. This is that, at Microsoft
scale.
Gizmodo's writeup →
|
|
Try this
Ask for the three strongest counterarguments
After any answer, ask:
"What are the three
strongest arguments against what you just told me?" Works on
anything — a business plan, a diagnosis, a decision you've already
made. Fastest way to catch a confident wrong answer.
|
|
|
Word of the week
System prompt
Standing instructions the tool gives the AI before you type anything —
its job, its tone, its limits. You usually can't see it. Why care: it's
almost always the reason a chatbot refuses something odd or has a
strangely specific personality.
|
|
Tool of the week
What I actually use — Notion
Every project this newsletter comes out of lives in Notion — notes,
plans, drafts, one place, with an AI layer for search built in. Free
plan is genuinely usable solo; costs the moment a second person
joins. No affiliate link — their partner programme is closed to new
applicants right now.
Read the full write-up →
|
|
|
Also
News outlets are asking a judge to sanction OpenAI over evidence in the copyright fight (4 min)
Why OpenAI buried a huge model announcement inside a maths post (3 min)
Every model released this month, tracked live (website)
[Ad]
The stack I actually use to make this — most of it free
|
|
See you Tuesday,
|
|
Links marked Ad are affiliate
links — if you sign up through one I earn a commission at no extra cost
to you. I never claim to have used something I haven't.
Nocturne · Glasgow, United Kingdom · reply to reach a real person
Unsubscribe ·
View in browser ·
YouTube
|
|