The thread for the whole block
For years software stayed behind glass — it happened inside the screen. This week it reached out of the rectangle and started doing real work in the world: signing off on cloud bills, shipping code, translating live. And when software starts acting on its own, your business gets a second customer that isn't human.
📰 The news — say this first
A few episodes back we said Anthropic's gated Mythos model would open to the public by the end of May. Where it actually landed: Mythos 5 is still locked — restricted to a vetted program, Project Glasswing, for cyber defenders and critical infrastructure. What opened to everyone on June 9 is Fable 5, the safer public version of the same model.
Why open here
Quick, honest, and straight into the news — the real story this week is what landed, not the date we called.
Quick callback before we dive in. A few episodes back we said Anthropic's Mythos would open to the public by the end of May.
We did.
Here's where it actually landed. Mythos is still gated — it's locked to a vetted program called Project Glasswing, for cyber defenders and critical infrastructure. What everyone got on June 9 is Fable 5, the safer public version of the same model. So let's talk about what actually shipped.
Producer note (backstage, not on air)
Verified vs anthropic.com June 12 — Mythos 5 is "not generally available... limited to approved customers in
Project Glasswing." Keep it light and quick; do NOT say "our AI prep got it wrong" on air. Just call the prediction and move into the news.
📰 The news — say this first
On June 9, Anthropic released two configs of the same model class. Claude Fable 5 — the public one — ships a 1-million-token context window, always-on adaptive thinking, and stronger safeguards in high-risk domains. Claude Mythos 5 — fewer safeguards, more cyber-capable — is NOT public; it's limited to vetted partners in Project Glasswing.
The signal that lands
Andrej Karpathy's post on Fable 5 pulled
2.48 million views, 25,000 likes, and over 6,300 bookmarks. In the builder world, a bookmark isn't applause — it's "I'm going to deploy this."
Let's define Fable 5's headline feature — "always-on adaptive thinking." Normally an AI just predicts the next word. Fable 5 runs reasoning in the background, maps out logic trees, and self-corrects before it ever prints a word. It thinks before it speaks.
And the engineers noticed. Karpathy's breakdown hit 2.48 million views, 6,300 bookmarks. That bookmark ratio is the tell — builders see a mechanical advantage they plan to ship.
But I want to push back on the million-token hype. A million tokens is a giant haystack. If you dump in disorganized junk, the model's attention gets diluted — you can get worse judgment, not better. The skill isn't "paste everything." It's curation.
Right — the raw horsepower keeps climbing, so the human job shifts to being the editor. And on Mythos: Anthropic kept the more cyber-capable one behind Project Glasswing on purpose. The open question is whether limiting it to trusted defenders actually contains the risk — or just delays it.
JUNE 9 · DUAL RELEASEFABLE 5 = PUBLICMYTHOS 5 = GLASSWING ONLYKARPATHY: 2.48M VIEWS · 6.3K BOOKMARKS
Why this matters · for your audience
The default AI just got more capable, and the frontier keeps ratcheting up. Don't wait for it to "settle." Run one real task through Fable 5 and through whatever you use now, side by side, and judge it on your work — not the benchmark headlines.
📰 The news — say this first
Announced June 4 and rolling out through the week, OpenAI's new ChatGPT memory — they call it "Dreaming" — runs in the background, synthesizing what it knows about you across your whole history instead of a manual list of saved facts. Plus and Pro in the US first.
The frame
The difference between a tool you re-explain yourself to every time and a real assistant is
context retention. One that updates its own understanding of you compounds — you stop explaining your baseline every new chat.
It's no longer a list of facts you dictate. It's reading the long tail of your chats, extracting patterns, and silently updating who it thinks you are.
Mechanically it's adjusting its own vector database in the background — re-weighting toward your preferences without you touching it.
Which is exactly what worries me. Say I spend three days asking about commercial plumbing for a friend's storefront. The system could decide I'm a plumber and start skewing every answer about my actual business toward pipe fittings. It can quietly drag my whole profile off course.
So auditability becomes your new job. When the AI edits its own memory, you become the auditor — you check its assumptions so it hasn't invented a false version of you from a one-week anomaly.
Callback — direction HIT
Chris, Ep41:
"Self-improvement feedback loops are going to become part of AI as we know it." Now it's spreading platform to platform. That direction hit.
Try this today · for your audience
Ask ChatGPT: "What do you remember about me and my work?" Read it like the notes of a brand-new assistant you just hired. Keep what's right, tell it to delete what's stale or made-up. That hygiene habit is about to be mandatory.
📰 The news — say this first
On June 9, AWS put its FinOps Agent into public preview — for AWS customers, to tame their own AWS bill. FinOps is the tedious art of deciphering cloud billing. This isn't a chatbot giving advice — it's a digital worker that watches your costs, investigates a spike, traces the cause, and files a report into Slack or Jira.
Why it matters even if you'll never touch it
Most of you don't manage an AWS bill — and that's fine. Watch this as the
clearest example yet of where AI is going: from chatbot to worker. Not "ask it a question," but "hand it a recurring job and let it run."
Here's why I care about this even though most of us aren't running an AWS bill. It's the cleanest picture of AI as a worker instead of a chatbot. It watches the bill, catches a spike, traces the cause, and files the report — the boring, expensive job a person used to dread on a Friday afternoon.
And that points to a real restructuring of work. Once an agent takes the repetitive operating task, your job doesn't disappear — it elevates. You go from digging through logs to supervising: you review the agent's evidence and hit approve or deny.
One honest note — it files a report ticket; it's not out there autonomously demanding refunds. But the principle holds: the second you let an agent touch anything money-related, you want hard guardrails and a human approving.
AWS FINOPS AGENT · JUNE 9PUBLIC PREVIEW · AWS CUSTOMERS ONLYWATCHES → INVESTIGATES → FILES REPORT
Try this week · for your audience
Even if you never touch AWS, the pattern is yours. Pick the most repetitive "check these numbers and flag the weird one" task in your week and write it out step by step, like a job description for an intern. You just wrote the spec for your first AI agent.
📰 The news — say this first
At WWDC on June 8, Apple unveiled a ground-up Siri for iOS 27 — deep personal context across messages, mail, and photos, plus on-screen awareness: it reads the actual layout of your apps and can take actions across them. The catch: you cannot pick or swap the underlying model.
The image that lands
It's like buying a high-performance car with the hood
welded shut. Promised to be perfectly tuned — but you can't swap a part. Smoothness over control.
On-screen awareness is the leap. The OS isn't screenshotting — it's parsing the interface, so it knows where the reply button is and what field is active, and acts across apps without you tapping. And Apple itself says Siri's built in collaboration with Google's Gemini models, alongside Apple's own — but it's a sealed box. You can't pick or swap the model.
And I have to put my hands up — this is a predictive miss for me. Earlier this year I strongly argued Apple would go bring-your-own-model at the OS layer. They did the exact opposite. They decided the friction of choice was too high for the average person, and chose a locked, unified ecosystem.
So does "bring your own model" even survive as a consumer idea?
For mass-market devices, the invisible model wins. Most people don't care about the architecture. They just want their phone to find the invoice in an email thread and pay it Friday.
Honest scorecard — resolves a prediction as a MISS
Chris called Apple going
bring-your-own-model at the OS layer (Ep41). Apple went locked, unified, no model choice. Own it on air.
(Verified: the Gemini tie is CONFIRMED — Apple's own newsroom says Siri is custom-built "in collaboration with Google and its Gemini models.")
📰 The news — say this first
On June 9, Google released Gemini 3.5 Live Translate — low-latency speech-to-speech translation in 70+ languages that keeps your tone, pace, and pitch, switches languages automatically, and doesn't wait for you to finish your sentence. It expands Google Meet from 5 languages to 70+.
Why it's a leap
The old way was a clunky three-step pipeline — transcribe, translate, then synthesize robotic audio. This does
audio-to-audio directly: it maps the emotion, timbre, and inflection of your voice onto the translation.
You could interview a structural engineer in Tokyo — you both speak your own language and the friction just evaporates. It keeps your voice, not a robot's.
But watch confidence versus accuracy. Nuance, idioms, sarcasm can be beautifully synthesized and still semantically wrong. Game-changer for customer support or casual conversation — risky for a legally binding contract.
Right — you don't want a tone slip to accidentally commit your company to a thousand units you don't need.
GEMINI 3.5 LIVE TRANSLATE · JUNE 970+ LANGUAGESMEET: 5 → 70+
Callback test
Chris (Ep33): Google's entries usually land
"a rung below." Watch whether this one breaks the pattern.
Try this week · for your audience
Next time you'd avoid a conversation because of a language gap, run it through live translate. Feel how much friction just disappears.
📰 The news — say this first
On June 9, the city of Seattle enacted an emergency one-year moratorium on large new data centers (those above a 20-megawatt power threshold), citing pressure on the local power grid and land use. A major US city just told the AI buildout to wait.
The bigger shift
AI stopped being only software this week. It's now
gigawatts of electricity, millions of gallons of cooling water, and city politics. The "just spin up more compute" era officially collided with thermodynamics.
The software feels invisible, but the hardware to run it is violently physical. Millions doing real-time audio translation, hundreds of millions of phones parsing app layouts, agents scraping cloud databases — it all runs on heat.
But ask whether a local ban actually constrains AI or just forces geographic arbitrage. The global demand for compute doesn't vanish. Companies export the data centers to regions with cheaper power and looser rules. You move the footprint — you don't shrink the demand.
SEATTLE MORATORIUM · JUNE 91-YEAR · LARGE DATA CENTERS (>20MW)POWER + LAND PRESSURE
Try this week · for your audience
If your business leans on a cloud AI tool, ask where its compute lives and what happens to your costs if that supply gets politically squeezed.
📰 The news — say this first
In remarks aboard Air Force One on June 5 (Reuters reported it that day; the New York Times followed June 10 with more), President Trump floated the idea of the public — or the government on its behalf — holding equity stakes in major AI companies, so Americans share in the wealth. This was floated, not announced. No deal, no policy, no legislation.
Say this out loud first
We're looking at this
strictly mechanically, not politically. We're not endorsing it and not taking a side — we're reporting the source because it raises a structural economic question.
The argument floated is that these models are trained on the public's collective data and powered by the public's electrical grid — so society holding equity is logically consistent from a pure resource standpoint.
But the mechanics are a regulatory nightmare. If the government takes a stake in private AI labs, how does it objectively regulate them? The regulator becomes the shareholder — a profound conflict of interest.
Whether or not it ever becomes policy isn't the point. The fight over who captures AI's upside just moved from dorm-room philosophy to the highest level of government.
Verify-before-you-cite — non-negotiable
Say
"floated" or "raised the idea." Do NOT say "Trump announced." Third-party reporting (NYT + Reuters); keep the mechanical-not-political guardrail.
The real takeaway · for your audience
Nothing to do — just don't let anyone in your feed turn "he floated it" into "it's happening." Notice how fast a remark becomes a fake headline.
So: frontier models with million-token contexts that demand curation, and a memory that updates itself silently in the background.
An agent at AWS quietly doing the forensic cloud accounting a person used to dread on a Friday afternoon.
A zero-latency universal translator that bypasses text. And a city slamming the brakes because the thermodynamic limits are real.
The software breached the container. It's operating in the physical world and executing in the real economy.
The mull-on-it close → hands to the deep dive
Here's the thought to leave you with. If autonomous agents are out there right now evaluating data and acting for companies — then they're also out there discovering, evaluating, and buying. The next entity that lands on your pricing page and decides whether to engage your business might not be human. Your business has a second customer now. Is your digital presence built to sell to an algorithm? Let's get into it.