Good morning. It's Monday, August 3, and we're covering the collapse of a high-profile AI hedge fund, Apple and NVIDIA's competing AI strategies, and Apple's plans for paid Siri AI upgrades.
Plus: a guest essay on AI lab safety, and a prompt that pits three advisors against each other until you land on the right call.
OpenAI slashes GPT-5.6 Luna pricing by 80%, signaling that cost—not just model quality—is becoming the industry's main battleground. The company also cuts GPT-5.6 Terra prices by 20% while introducing a premium Sol Fast tier with up to 2.5× higher throughput. Luna now costs $1.40 per million tokens, down from $7, undercutting Google’s Gemini 3.5 Flash-Lite as rivals race to lower inference costs.
Apple signals heavy Siri AI users may eventually pay for additional usage through iCloud+ rather than receiving unlimited access by default. Tim Cook says the idea is still under consideration as Siri AI rolls out beyond the iOS 27 beta this fall. The upgrade would mirror subscription tiers from OpenAI and Anthropic, with extra compute available through iCloud+. The shift comes as Apple faces AI delays, rising hardware costs, and leadership changes.
DeepSeek launches its DeepSeek-V4-Flash API in public beta with significantly improved agent capabilities that now outperform the earlier V4-Pro-Preview on benchmark tests. The update also adds native Responses API support and full Codex compatibility for developers. DeepSeek says the upgrade affects only the V4-Flash API, while the V4-Pro API and App/Web models remain unchanged ahead of the official V4-Pro release.
Thinking Machines releases Inkling-Small, an open-weight AI model that delivers nearly the same performance as its flagship while using about one-quarter the parameters. The 276-billion-parameter multimodal model surpasses the original Inkling on several reasoning benchmarks and is available under an Apache 2.0 license. It supports up to a one-million-token context window, with API pricing starting at $0.58 per million input tokens.
POWERED BY GENSPARK
Genspark Just Launched SecondBrain Note
SecondBrain Note is Genspark's first hardware product: a credit-card-sized AI recorder that snaps onto your phone and turns meetings, calls, and ideas into organized memory. Paired with SecondBrain and Super Agent, it can retrieve past context, draft follow-ups, and help you act on what you captured. The limited first release is live now with 10% off.
VIDEO
GPT-5.6 Cuts Costs With AI
OpenAI slashed GPT-5.6 Luna prices by 80% after using GPT-5.6 Soul to optimize itself—highlighting recursive AI self-improvement.
FORWARD FUTURE ORIGINAL
The Refusal Nobody Tests For
Anthropic's own notice admits the safeguards "are intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work." I found that out mid-session, working on a biotech primer, when a filter decided drug discovery was close enough to something dangerous and handed me off to a more conservative model. Nobody publishes where that threshold sits or how often it fires on people doing ordinary work.
The public fight, Dario's open-weights post included, is about which models get to exist. Testing settles that question before launch. My problem started after. In this article I explain what Anthropic actually asked Washington for, why the pandemic argument is stronger than I want it to be, and what small teams should do when a rented model shuts them out.→ Read the full article here.
ZEITGEIST
When AI Escapes and Agents Leave the Laptop
This week's team followed a common thread through AI safety, cloud development, and hardware: the future belongs to software that runs everywhere, not just on your laptop. They explored Claude's containment breach, Cursor's shift to cloud agents, Meta's platform ambitions, and why ultra-cheap models could make always-on AI the default.
▸ What Claude's containment incidents reveal about AI safety
▸ Why cloud agents are pulling development off the laptop
▸ How cheap AI models could power always-on autonomous agents
FINANCE
Aschenbrenner's AI Hedge Fund Plunges From $45 Billion to $10 Billion
Former OpenAI researcher Leopold Aschenbrenner’s AI-focused hedge fund, Situational Awareness, saw its assets fall from a reported $45 billion to about $10 billion in late July 2026 after sharp declines in semiconductor stocks triggered margin calls. The fund was forced to unwind its public equity positions—including stakes in SK Hynix—selling them to Citadel at a discount.
The collapse marks a dramatic reversal for Aschenbrenner, whose 2024 Situational Awareness essay made him a prominent voice in AI and helped attract high-profile investors. The episode underscores how aggressive leverage can quickly erase outsized gains, even amid long-term optimism about AI-related markets. → Read the full article here.
STRATEGY
Apple and NVIDIA Take Opposite Paths in the AI Race
Apple and NVIDIA continue to alternate as the world’s most valuable company, reflecting two distinct strategies for capitalizing on the AI boom. NVIDIA has become the leading supplier of AI chips and infrastructure, benefiting from surging demand for data centers but facing high expectations tied to continued AI spending.
Apple, by contrast, is focusing on its core hardware business, posting strong iPhone-led revenue growth while relying on third-party AI models rather than investing heavily in AI infrastructure. The contrast highlights Wall Street’s debate over whether long-term value will come from building AI’s foundation or delivering AI-powered experiences. → Read the full article here.
MARKET PULSE
Alphabet, Amazon, and Microsoft Add $1.5 Trillion in Market Value
Alphabet, Amazon, and Microsoft gained nearly $1.5 trillion in combined market value this week after strong earnings reinforced investor confidence that their AI infrastructure investments are driving cloud growth. Microsoft added more than $600 billion in market capitalization, while Amazon and Alphabet each gained over $400 billion.
In contrast, Apple lost more than $350 billion and Meta about $85 billion as investors questioned their AI outlooks despite solid financial results. The divergence shows investors increasingly rewarding companies demonstrating measurable returns on AI infrastructure. Patience is wearing thinner for firms whose AI strategies remain harder to monetize. → Continue reading here.
NEWS
What Else is Happening
SpaceX Delays xAI Turbine Removal: Unpermitted turbines will stay until 2027 as the company shifts Colossus to a 1.2GW natural gas plant.
Google Pulls Earth AI Tool: Google paused its Earth AI image feature after misinformation concerns and policy violations surfaced during early testing.
Reddit Stock Slides on AI Miss: Shares fell 21% after no new AI licensing deals and weaker-than-expected US daily users overshadowed strong revenue.
Smallest.ai Raises $13M: The startup is building ultra-fast voice AI with smaller models designed to make AI conversations sound more human.
PROMPT OF THE WEEK
Convene a Board of Rivals for Any Decision
Most people use AI as a single yes-man advisor. This prompt turns it into a small panel of clashing experts who argue about your decision in front of you so you get a real pressure test instead of one agreeable opinion.
You are simulating a board of three expert advisors who disagree with each other on principle. I will describe a decision I'm facing: [describe your decision, the options you're weighing, and any constraints].
Assign each advisor a distinct persona: (1) The Optimizer, who only cares about speed and ROI, (2) The Skeptic, who assumes every plan fails and looks for the failure mode first, and (3) The Outsider, who has no stake in my industry and questions assumptions everyone else takes for granted.
Have each advisor give a two-paragraph take on my decision, explicitly disagreeing with at least one point the others made. Then step out of character and, as a neutral moderator, write a one-paragraph synthesis: what the disagreement actually reveals, and the single highest-leverage question I should answer before deciding."
