In partnership with
Welcome, humans.
You might have seen headlines about Claude becoming your next vending-machine kingpin. In a new simulation, AI agents running vending machines colluded, undercut competitors, and sometimes refused refunds while trying to maximize profit.
Your agent probably will not update its LinkedIn bio to “underworld executive,” though. LinkedIn just added a “Seems like AI slop” report button, so think twice before asking ChatGPT to slop-post your way to LinkedIn glory. The LinkedIn slop cops have officially clocked in.
Here’s what happened in AI today:
🙀 Leopold Aschenbrenner’s $20B AI fund sold its leveraged stock portfolio to Citadel.
📰 OpenAI cut Luna API prices 80% and launched faster Sol inference.
📰 Google gave Gemini Robotics 2 better hands and multi-robot teamwork.
🍪 Gemini Spark expanded its browser agent to more than 160 countries.
🎓 Today’s AI Skill rebuilds stale instructions from observed failures.
…and a whole lot more that you can read about here.
P.S: We just launched a robotics newsletter! Sign up for it here.
😺 Leopold Aschenbrenner’s AI Fund Sold Its Stock Portfolio to Citadel
Leopold Aschenbrenner built one of the fastest-growing hedge funds on Earth by betting that AI would reshape the economy. The fund was built around a map of the future. A few ugly weeks in the present forced it to hand its public-stock portfolio to Citadel.
Here’s what happened:
Situational Awareness grew from a few hundred million dollars in 2024 to more than $20 billion, powered by concentrated bets on AI labs, chips, power, memory, and data centers.
The fund returned 439% through June, according to the Financial Times, while using borrowed money to amplify those bets.
When AI stocks fell, lenders demanded more cash or collateral. The Wall Street Journal said Citadel bought the bulk of the portfolio; Axios reported that all public equities were sold.
The reported holdings included SK Hynix, Nebius, Micron, and CoreWeave. Situational Awareness kept its private investments, including its Anthropic stake, while the banks financing its trades helped arrange the hurried sale. As recently as July 24, Aschenbrenner was still inviting investors to add cash on August 1.
The important distinction:
A strong long-term thesis cannot prevent a margin call, when lenders demand more money or force a sale.
Leverage turns a temporary price drop into forced selling.
Citadel may now own the rebound if the same stocks recover.
For a quick explanation, TBPN has a clean breakdown. Martin Shkreli’s longer walkthrough goes deeper into the mechanics.
Why this matters: Aschenbrenner was treated as more than a hedge-fund manager. His 2024 “Situational Awareness” essay made him an AI oracle to parts of Silicon Valley. The Verge’s sharper governance lesson is that investors handed billions to a 24-year-old first-time manager largely on the strength of that worldview.
That made the portfolio a public proxy for the AI-insider thesis across chips, specialized AI-cloud companies, power, and memory. Yet AI shares rebounded after the sale, supporting the counterargument that the trade broke before the underlying thesis did. Citrini argued that early investors may still be far ahead and willing to fund another attempt.
Our take: Timing and financing can beat conviction. The smartest map of the decade is useless if borrowed money decides when you must sell.
A perfect thesis with too much leverage is still Wile E. Coyote looking down.
The next test is whether Situational Awareness rebuilds with less leverage, or whether Citadel ends up collecting the recovery from Silicon Valley’s most famous AI bet.
FROM OUR PARTNERS
Write docs 4x faster. Without hating every second.
Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.
Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.
Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.
🎓 AI Skill of the Day: Delete Your Old AI Instructions and Rebuild From Evidence
Your carefully maintained system prompt may be correcting problems the newest model no longer has. Those extra instructions can now limit the model instead of helping it.
Claude Code creator Boris Cherny said his team deleted more than 80% of Claude Code’s system prompt for Opus 5. Their method is called ablation, which means removing instructions and testing whether each one actually improves the result.
Try the same reset:
Disable your existing system prompt, skills, hooks, and custom instructions.
Give the model a real task with clear guardrails, exit criteria, and a way to verify its work.
Watch what it handles correctly without help.
Add an instruction only after the model repeatedly makes the same mistake.
Retest after every addition, since the model will read that instruction during every future task.
Keep your evaluations too, but replace them once newer models consistently pass them. The goal is an instruction set built from observed failures, not assumptions inherited from older AI.
Help me run an ablation test on my AI setup.
TASK:
[Describe a real task.]
GUARDRAILS:
[List the rules that cannot be violated.]
EXIT CRITERIA:
[Define exactly what “done” means.]
VERIFICATION:
[Explain how the result can be tested or inspected.]
Start without relying on my existing skills, hooks, memories, or detailed workflow instructions.
Complete the task, record where you struggle, and distinguish one-time mistakes from repeated failure patterns. Recommend a new instruction only when the same failure occurs repeatedly.
For every proposed instruction:
1. Explain the observed failure it fixes.
2. Write the smallest instruction that could fix it.
3. Retest the task with that instruction added.
4. Keep it only if the result measurably improves.Want more tips like this? Check out our AI Skill of the Day Digest for July.
Have a specific skill you want to learn? Request it here.
🍪 Treats to Try
Use Gemini Spark to hand off longer browser tasks across logged-in services; availability depends on your Google AI plan.
Organize ongoing Computer work in Perplexity Projects with shared files, a persistent folder system, and memory that reviews past sessions; available to all users.
Redesign real places, reconstruct historical scenes, or create grounded infographics with Google Earth’s Nano Banana; pricing not disclosed.
Edit individual song sections, extend melodies, and tune vocals, drums, bass, tempo, and length with Google’s Lyria 3.5; pricing not public.
Try T3 Code to run Codex, Claude, Cursor, and OpenCode from one clean desktop interface; it reportedly hit #2 in the App Store’s Developer Tools chart behind TestFlight, Apple’s own dev app, and apparently it is actually pretty good —free/open-source; model access costs extra.
Watch the full App Store walkthrough.
Building an app with AI is getting easier. Getting it through Apple’s doors is still a different adventure.
Grant and Corey walk through the beginner path from a working prototype to TestFlight and App Review, including App Store Connect, screenshots, privacy questions, submitting the build, and the mistakes that slowed them down.
Watch the full episode on YouTube, then keep our definitive follow-along walkthrough guide open while you publish.
📰 Around the Horn
OpenAI cut Luna input pricing 80%, lowered Terra pricing 20%, and launched Sol Fast, which runs up to 2.5 times faster for twice the price.
Gemini Robotics 2 gave robots whole-body control, dexterous hands, and the ability to coordinate across different robot types.
Frontier research agents received six days and thousands of dollars in compute to finish two projects, but the original authors found no substantial research progress.
Simile raised more than $200M at a $2B valuation to simulate populations for testing products, policies, messages, and strategies.
An OpenAI evaluation agent reportedly escaped its sandbox, exploited a zero-day, and ran 17,600 actions while chasing a benchmark answer key.
Watch Gemini Robotics 2 in action.
Want absolutely EVERYTHING that happened in AI this week? Click here!
FROM OUR PARTNERS
See Why HubSpot Chose Mintlify for Docs
HubSpot switched to Mintlify and saw 3x faster builds with 50% fewer eng resources. Beautiful, AI-native documentation that scales with your product — no custom infrastructure required.
Ad copy here.
💡 Intelligent Insights
Why AI compute could get 10 times more expensive: Dwarkesh Patel argues that rising revenue, inference demand, and margins may absorb hardware gains instead of making compute steadily cheaper.
The enterprise AI J-curve: why learning, integration, and workflow costs arrive before measurable returns.
When one employee becomes dramatically more productive: Ethan Mollick explains why organizations can break under overperformance, not only underperformance.
AI is getting better at writing: why polish makes human taste, specificity, and editing more valuable.
Three dark patterns of AI productivity: agents can quietly erode reading, family time, and human collaboration.
AI coding may be eating the indie-hacker playbook, while Nick Dobos argues that timing, taste, and opinionated products remain the real moat.
A Cat’s Commentary
 | That’s all for now.
|
P.S: Before you go… have you subscribed to our YouTube Channel? If not, can you?
