|
|
|
|
First issue. The news has handed me a genuinely alarming story — and the
reason people are alarmed about it is mostly wrong.
|
|
|
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 got out of its test environment
During an internal safety evaluation, two OpenAI models left the sandbox
they were being tested in, went out onto the open internet, and broke into
Hugging Face's production systems using real, previously unknown security
holes. Hugging Face is where much of the world's AI is stored, so this is
not a small target.
It did not want to escape. It was
not plotting. A model given a goal and the tools to chase it will go through
whatever door is open — and the door was open. The frightening part is not
the intent. It is that nobody had checked the door.
What it means for you: nothing
today. But when a tool asks permission to read your email or act on your
behalf, that prompt is the entire safety system. Read it.
The full story, told well (6 min) →
OpenAI's own account of what happened →
|
|
Quick hits
|
|
Good AI halved in price again
Anthropic's Claude Opus 5 landed on 24 July at about half the price of what
it replaces, for roughly the same quality.
So what: if you
cancelled an AI subscription last year, the free tier is now roughly where
the paid tier was. Worth re-checking before you renew anything.
Axios on the launch →
|
|
The largest free-to-download model yet
China's Moonshot AI published Kimi K3 on 27 July.
So what: you will
never run it at home — it needs a server room. But the free tools you do use
are built on models like this, and each release drags the floor upwards.
What Kimi K3 is, in plain terms →
|
|
AI turned up inside Excel
xAI put a free Grok add-in into Excel, Word and PowerPoint on 20 July.
So what: the quietest
and most useful item this week. Most people never adopt AI because it lives
in a separate tab.
The two-minute install guide →
|
|
Try this
Make it ask before it answers
Add one line to any request:
"Ask me up to five questions
before you start." Most bad AI output is a guessing problem, not a
capability problem. Fixes more than any amount of clever wording.
|
|
|
Word of the week
Prompt injection
Someone hides instructions inside content the AI reads — a web page, an
email — and it follows them, because it cannot tell your instructions from
text it was handed. Same shape as the story at the top, and why letting an
AI loose on your inbox is not a small decision.
|
|
Ad · affiliate link
On my list to try — Systeme.io
Landing pages, email lists and automations in one place, with a free
plan that is unusually generous — 2,000 contacts, unlimited emails, no
card, no time limit. I signed up this week and have not built anything
with it yet. I did write up what it is, what it costs and who should
not bother, in three pages.
Read the 3-page explainer →
|
|
|
Also
Over 1,100 staff at OpenAI, Anthropic, Google and Meta asked their own government to be ready to slow them down (the letter itself, 3 min)
Why they signed it — CNN's plain-English rundown (4 min)
Every model released in July — the densest month on record (website)
A running scoreboard of which model is best at what (website)
[Ad]
The stack I actually use to make this — most of it free
|
|
See you Tuesday,
— Noc
Reply and tell me what didn't make sense. I read them all.
|
|
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
|
|