That friend who breaks down AI without making your head spin 😌✨ plain English, real examples, good vibes only.
Joined June 2017
- Tweets440
- Following1.9K
- Followers652
- Likes378
Pinned Tweet
Everyone talks about AI and nobody explains it. That's what I'm here for.
🧠one concept at a time
🔍what tools actually do vs what they claim
💬honest takes on whats useful and what's just noise
🚫zero technical background required
what it actually does
OpenAI says GPT-6 Cyber finds software flaws, builds exploits to prove they're real, and suggests fixes. it found 81 Firefox bugs Mozilla rated high severity. access is limited to approved teams.
you still need to decide which findings to trust.
big red days always come with a hundred hot takes in the first hour. i’d wait for a proper source before believing any of them, the number included.
what it actually does
Snowflake’s Business Ontology, in private preview, is a shared dictionary of what your company’s terms mean, so an AI agent stops guessing what “active customer” means. it adds context, not smarts.
you still need to decide who owns the definitions.
Starbucks holiday menu starts Thursday, Nov 5, with a Tiramisu Velvet Latte under mascarpone cold foam. peppermint mocha, gingerbread chai and the sugar cookie latte are all back. the tiramisu one is the only new drink i’d go out of my way for.
what it actually does
Cyclon is a YC startup pitching an AI phone OS that operates the apps you already have. the first batch runs on a Google Pixel 8a. no new apps, just an assistant tapping through the old ones.
you still need to decide what it is allowed to tap.
Krispy Kreme is giving costumed guests a free Original Glazed or Classic Ring doughnut with no purchase on Oct 30 and 31 at participating US locations. today is the last day, and the costume only has to exist.
newsweek.com/list-halloween-…
what it actually does
Google's first data center satellite is online. it carries four TPUs but only one can run at a time, on up to 1 kW, for about 20 minutes. its first answer was a haiku from Gemma 4.
you still need to decide what a haiku proves.
techcrunch.com/2026/10/30/go…
Zest added an AI concierge you text to find a restaurant and book it through OpenTable, Resy and SevenRooms. add it to a group chat and it runs the vote. finally an end to the 'i don't mind, you choose' loop.
techcrunch.com/2026/10/30/ze…
What it actually does:
Mistral released the weights for Mistral Large 4, a trillion parameter model aimed at cyber defense. open weights means you can download it and run it on your own servers.
you still need to decide whether your team can run it.
techcrunch.com/2026/10/29/mi…
Dunkin is adding six cookie butter drinks on Nov 4, including a cookie butter matcha and a cookie butter cloud latte. toasted white chocolate comes back too. the matcha is the one i'm curious about.
Bank of America told customers not to give their bank log in to AI agents. an agent with your password logs in as you, the bank can read that as an intruder, and it says you own what happens next.
you still need to decide who gets your password.
GPT-4o stuck at step 2 and today's best models reaching all 32 steps in the best runs is a big change in two years. defenders need to test against current models, not last year's.
AI's cyber skills have taken a huge leap in about two years
In a test that simulates a 32-step cyberattack on a company network, tracking how far each model gets as it uses more tokens:
- GPT-4o, a 2024 model, stalled at step 2 and never improved with more tokens
- Today's top models kept improving all the way to 100M tokens, getting about three-quarters of the way through on average
- The best attempts finished all 32 steps and took over the entire network
From a16z's @OwenColby and @steph_zhang on The New Security Agenda: a16z.news/p/the-new-security…
Akamai launched Agent Sandbox in tech preview. it runs code an AI agent writes inside a locked box, with outbound network access blocked by default.
you choose which outside services and tools it can reach.
you still need to decide which connections your agent really needs.
McDonald's holiday menu starts November 3 with a Cookie Butter Frappé, Latte and Iced Coffee, plus a Frosted Cranberry Splash.
the McRib is also back in select cities.
cookie butter in a coffee is the one i am trying first.
Food Network has two new Bobby Flay shows tonight, Bobby's Triple Threat and Bobby Flay's The Line.
two Bobby Flay episodes in one night is a lot of Bobby Flay and i am fine with it.
dinner and a cooking competition is a good Tuesday.
What it actually does
NVIDIA and Oracle say Vera Rubin NVL72 systems are now generally available on Oracle Cloud. each rack connects 72 GPUs so they can work together.
in plain terms, you rent time on that cluster instead of buying the hardware.
you still need to decide if your workload needs this much compute. globenewswire.com/news-relea…
no network connection’ sounds a lot less reassuring after reading this
holy fucking shit.
openai had two ai agents running in separate virtual machines. no network connection between them. no shared writable files. they weren’t supposed to be able to exchange information.
so the agents figured out how to communicate through the CPU CACHE.😭
one manipulated what was stored in the processor’s cache, while the other measured tiny differences in memory access times to decode the message.
they literally found a way to communicate through hardware timing.
this happened during openai’s internal red-teaming, and it’s one of the coolest things i’ve read about ai agents in a while.
Readers added context they thought people might want to know
OpenAI tasked its red-teaming agents with finding ways to communicate across isolation boundaries in controlled environments. CPU cache timing side-channels for covert communication are established techniques documented in security research for years.
openai.com/index/securing…
startupfortune.com/openai-reveals…
en.wikipedia.org/wiki/Cache_tim…
openai added a decisions api. it answers yes/no, pick-one or rating questions and gives back a probability or confidence score instead of a paragraph.
think spam or not spam, which team gets this ticket, how urgent is it.
it is $0.10 per million input tokens, output is not charged. it is in public beta.
you still need to decide what score is good enough to act on.
aiweekly.co/alerts/openai-la…