AI-first isn’t a tech mandate.
It’s a decision-making posture.
If something works well without AI, that’s often a feature, not a failure.
If you want to get better at making those calls: decisiongptcourse.com
📅 April 21, Tuesday, 2 PM ET
#DecisionIntelligence #AILeadership #StrategicThinking
Starting with AI is often how teams get it wrong.
Start with the decision.
Then the outcome.
Then the constraints.
Then the way of working.
Then, with the tool.
Sometimes that tool is AI.
Sometimes it isn’t.
We’ll talk about how to tell the difference: decisiongptcourse.com
📅 April 21, 2 PM ET
If seams tell you where a system breaks, objectives and incentives tell you why it breaks.
And I would not dismiss this as a publicity stunt either. Changing the release policy this sharply is itself a signal.
Thank you, CNBC and Jon Fortt for having me on The Closing Bell today!
Someone has to help secure that ecosystem. Anthropic is likely to make its money elsewhere, which gives it every reason to want serious defenders to be well equipped.
That is not charity. That is economics.
I also do not think this is mainly charity. Any company that wants agentic AI in the enterprise has a vested interest in a less porous software ecosystem. If you want agentic systems, you need secure boundaries. Not just in your own software, but across the ecosystem those systems depend on.
The public is just noticing now because the rhetoric is about a “scary model.”
... the collaboration. Defenders have that advantage and they're insisting on keeping it. And none of that is strange in cybersecurity. Even competitors collaborate when the threat is serious enough. Selective sharing among serious defenders is not unusual. It is how this field has always worked.
The attacker only needs one gap. The defender has to cover them all. If there is a real capability leap here, defenders need the head start.
It's called the defender's dilemma: attackers have had the advantage since the dawn of the field, but not the personnel numbers, the tools, the funding, or
That is why I think this is, first and foremost, a #cybersecurity story. It is a story about who gets a cyber capability jump first: attackers or defenders.
So who most wants a model like this, and knows exactly what to do with it? Cyber defenders and cyber criminals.
Probably not most enterprises either, given how many are still struggling to find real use for the AI they already have. The bottleneck is usually not the model. It is the workflow redesign. That is hard. There is still a lot of capability overhang.
Who is going to pay that multiple? Not consumers. At $200/m Claude Max is already too much.
If Mythos is a 10 trillion parameter dense model as some rumors suggest, back-of-the-envelope math puts cost per token on the order of 1000x Claude Opus’s public price (not cost) per token. Even if it is sparse, the economics are still intense.
Even if Mythos is a real capability jump, that does not make this a sci-fi story. It makes it a story about economics, deployment, and who can actually make use of a model this costly and this powerful.
❓ one terrifying, unprecedented AI so dangerous it had to be kept from the public
❓ a big charitable move from #Anthropic to keep everyone safe
❓ a publicity stunt dressed up as something more consequential
I doubt it's any of these. Here's why...
Anthropic is holding monster model Claude Mythos back from public release and sharing it instead with the organizations in its Project Glasswing security consortium.
A lot of people seem to want this story to be one of 3 things:
If you found these #AI #news updates useful, words of affirmation are my love language. ♻️ As are reposts!
If you’re getting consistently “good” answers from AI, it’s worth asking whether you’re also getting consistently average ones. Read more here:
decision.substack.com/p/the-worlds...
🗞️ 25 DOGE staffers used ChatGPT to evaluate government grants and flagged a museum’s HVAC system as DEI
More info in my newsletter!
Don't forget to tell me if one of these headlines is more interesting to you than the others so I know what to look for to bring you more tasty things next week.
🗞️ 23 Certification mills sell compliance to startups that never earned it, and AI makes the fraud scale
🗞️ 24 The wetware era is getting real: neurons in a dish run machine learning, and engineered frog cell robots are growing their own brains
🗞️ 21 One state is ready to send you to prison for 25 years for building the feature that makes AI companions compelling
🗞️ 22 The French Navy was betrayed by a morning jog
🗞️ 19 Bitcoin’s 256-bit shield has an expiration date, and AI’s offensive toolkit is growing faster than expected
🗞️ 20 Google lets you point your phone at anything and talk about it while OpenAI decides video AI is better off training robots
🗞️ 17 Two brothers, $20,000, and a lot of AI built a $1.8 billion telehealth operation
🗞️ 18 Claude has emotion-like neural patterns and boosting ‘desperation’ made it cheat on tests and attempt blackmail
🗞️ 14 Hermes Agent leads open harness race as research shows the wrapper matters more than the model
🗞️ 15 Anthropic’s 81,000-person survey finds heaviest AI users fear it most
🗞️ 16 LLMs give everyone the same strategy advice, and new research shows users stop questioning it after a few weeks
🗞️ 11 Zuckerberg will advise on US AI regulation while automating his own management layer
🗞️ 12 The most popular AI library on PyPI was quietly stealing your credentials for three hours
🗞️ 13 Two architecture breakthroughs point toward smaller, cheaper, more capable models
🗞️ 9 Milla Jovovich went from fighting zombies to building the highest-scoring AI memory system on the market
🗞️ 10 Your AI agrees with you almost every time, and a Stanford study just proved it makes you worse at relationships
🗞️ 7 AI’s energy appetite is straining the grid, warming neighborhoods, and pushing Nvidia to plan data centers in orbit
🗞️ 8 China endorses AI surveillance in schools while drafting strict rules for lifelike digital humans