Markets pop on ceasefire headlines; Meta pivots to closed-source AI; Anthropic’s ‘too hot’ model; supply chain hacks surge; Kalshi wins in court; Eclipse raises $1.3B

Market Snapshot 🚀

• U.S. equities ripped higher: Dow +2.68%, S&P 500 +2.46%, NASDAQ +2.9%.
• Gains came on ceasefire headlines and hopes the “street might be opened”, though a later Wall Street Journal report suggested the “straight of moves might actually be closed again”—a recipe for near-term whipsaws.

“We got white suits on. You know what that means? The stock market is booming.”

Meta’s Closed-Source Pivot: Muse Spark Arrives

• Meta Platforms launched its first major model in over a year: Muse Spark, a closed model powering Meta’s AI assistant across apps. Shares jumped ~7.5% intraday and were described as “almost 8% today.”
• The move contrasts with the Llama open-weight era, aligning with a thesis articulated in 2024 by John Luttig: open-source serves until cost, safety, and ROI compel a shift. Luttig flagged the capex tipping point and shareholder pressure once training spend hits the “$10 billion or more” zone.
• Benchmarks were mixed. Observers noted a highlight-on-top chart where Muse Spark posted 86.4 on one benchmark and beat peers in some tests (e.g., healthbench hard), but underperformed on Arc AGI 2. The cultural signal: less fixation on leaderboard gaming after prior Llama 4 controversies and the shelved “Behemoth”.
• Personalization vs. privacy: early product anecdotes showed oddly specific suggestions (e.g., “Malibu-appropriate surf puns”), raising questions about whether cross-app signals inform replies. The AI insisted no special access; users pressed on perceived specificity.
• Token diet: Meta reportedly took down an internal token-usage leaderboard after over 60 trillion tokens were consumed in 30 days. A posted note said the dashboard was shuttered due to external sharing.

“The future of foundation models is closed source… when you reach $10 billion or more in capex spend for model training, shareholders will want clear ROI.”

Anthropic’s ‘Mythos’: Powerful, Gated, and Polarizing

• Model preview access was granted to roughly 50 companies that maintain critical infrastructure. Why? Mythos appears exceptionally good at finding zero-day exploits—high stakes if capabilities leak before patches.
• Partners listed on the project’s cybersecurity page include Apple, Google, Microsoft, Amazon, Nvidia, JP Morgan Chase, Broadcom, the Linux Foundation, Cisco, CrowdStrike, and Palo Alto Networks.
• Safety vs. scarcity: observers debated whether “too dangerous to release” is primarily a safety posture or a compute-and-competition moat. One notable quip captured the skepticism: “It’s amazing. It’s so powerful. It’s terrifying. And the best part is you can’t come.”
• Economics watch: George Hotz framed depreciation risk bluntly: “GPT‑4 cost $100 million to train two years ago…”—arguing that trained models are fast-depreciating assets, amplifying pressure to monetize quickly.
• Compute squeeze and market structure: Dean Ball warned the best models may increasingly go only to the highest bidder, turning compute into a seller’s market and rendering frontier capabilities “decreasingly legible to the general public.”

Apple, App Stores, and the ‘Based Act’

• Y Combinator’s public policy lead described a rising wave of “vibe-coded” apps running into the “worst DMV in the world”—the bottleneck of app review and platform rules.
• Policy push: A coalition of 275 startups and VCs backed the Based Act (bill “in 1074,” per remarks), a proposal banning anti-competitive self-preferencing by platforms over $1 trillion market cap or with 100 million U.S. users. Emphasis: sideloading and alternative app stores, pointing to Europe’s DMA and moves in Japan.
• Why it matters now: GitHub’s COO reportedly said if commit rates stay linear, they’re on track to be 14x last year. With AI accelerating software creation, distribution bottlenecks are becoming growth constraints.

Prediction Markets: Court Wins, Sports Dominance, and Federal Preemption

• Regulatory momentum: An appeals court judge in New Jersey gave Kali a green light to proceed. Expect eventual Supreme Court review; law “on Kali’s side,” with CFTC leadership described as supportive.
• Federal preemption appears to be the likely path. State-level opposition includes casinos (e.g., Nevada) and anti-gambling constituencies (e.g., Utah). The CFTC itself has intervened against some state bans.
• Category mix: In March, about 70% of Kali’s volume was sports, totaling roughly $13 billion for the month—down from February’s higher sports skew due to the Super Bowl.
• Structural differences vs. sportsbooks: each Kali contract requires CFTC approval within 24 hours, limiting granular in-play betting (e.g., next pitch). Notably, federal carve-outs ban onion futures and limit box-office trading—hence Rotten Tomatoes contracts, not revenue bets.
• Addiction optics: lawmakers may eye advertising restrictions even if platforms are regulated as commodities exchanges rather than gambling.

“Kali doesn’t care if you win or lose. They just care that you play. Sports books obviously want you to lose.”

Security Wave: Axios npm Breach and Supply-Chain Reality

• Axios npm compromised. A North Korean actor socially engineered the maintainer via a fake Microsoft Teams call and prompted a malware install. The poisoned versions dropped a remote access Trojan.
• Scale: Axios sees about 100 million weekly downloads. The malicious window lasted ~3 hours.
• Detection: Socket detected the malicious update within 6 minutes and reported a surge of interest: ~2,000 organizations signed up in the following 24 hours.
• Campaign context: The incident fits a broader wave tied to a team referred to as PCP/PCB, whose canister worm reportedly stole 300 gigabytes of compressed credentials. Prior compromises touched Aqua Security’s Trivy scanner, Light LLM, and Checkmarx.
• Why this keeps happening: software supply chains run on blind trust, mass-installing code from strangers with broad permissions—no review, no granular controls. Defenders must block every pathway; attackers need one.

“In an AI world where agents install dependencies, you need guardrails before the community even knows a package is poisoned.”

Enterprise AI for Defense and Offense

• Depth First raised $80 million (Series B) from Veritech and launched DFS Mini 1, focused on vulnerability detection and smart contracts. The team uses GPT‑4‑OSS as a base, plants flags, and trains with an RL loop; they reported outperforming Opus 4.6 on one benchmark “at one the cost.”
• Mythos’s bug-finding prowess was cited as validation that security is a prime AI use case. Depth First said it also runs models on open-source code and responsibly discloses issues, noting findings in widely used software such as Chrome and Linux.
• Agent Charlie (founded by a former Meta trust & safety lead) is building on-device, small language model defenses to intercept real-time scams and social engineering—using LLM-as-judge evals that pit attacker vs. victim role-plays to measure uplift.

AI GTM: Mutiny’s Agent for Revenue Teams

• Mutiny introduced an AI agent that helps companies like Rippling and Snowflake create customer-facing assets across the entire deal cycle: personalized vertical campaigns, prospecting pages, curated case studies, ROI reports, pricing proposals, and post-sale expansion materials.
• Why it matters: cold email performance is deteriorating, while reps reportedly spend ~30% of time selling and ~70% creating bespoke materials and follow-ups. Personalization and quality content are becoming the differentiators.

Deep Tech, Defense, and the Next IPOs

• Eclipse closed a $1.3 billion raise and now manages about $10 billion AUM. The first fund was $125 million.
• Cerebras: the wafer-scale story—linking cores across the entire wafer—was a 2015 bet on physics beyond Moore’s law as nodes approach 2nm and possibly 1nm.
• VulcanForms: high-precision metal parts via 160 fiber lasers in a single head. The company booked multi-billions in deals last year and a $1 billion deal the year before; four U.S. factories are being built.
• Robotics: adoption is crossing the chasm from custom PLCs toward general-purpose, physical-AI systems. Caution flagged on 2021-style valuation behavior creeping into hard-tech.
• Defense tech: with ceasefire headlines and no escalation to morally indefensible strikes, the sector’s momentum looks intact. Either way, munitions restocking is inevitable.

Key Quotes

“There’s a crazy bull case for Nvidia… arguing it should be worth what $22 trillion.”

“Trained AI models are the fastest depreciating asset in history. GPT‑4 cost $100 million to train two years ago…”

“In the age of AI… security is going to be extremely important.”

Crypto Note: The Satoshi Chase Returns

• John Carreyrou’s New York Times feature reignited the Satoshi hunt, focusing on Adam Back—who responded, “I am not Satoshi.” A 2024 HBO documentary had previously suggested Peter Todd, who also denied it.
• Beyond the intrigue, a re-identification would alter core Bitcoin lore; status quo skepticism persists.

Watchlist & Takeaways

• Macro/Geopolitics: Ceasefire optimism vs. “straight of moves” closure angst. Expect headline-driven volatility.
• AI Platforms: Closed-source pivots accelerate. Expect gated rollouts, compute allocation games, and premium pricing for strategic use cases (e.g., zero-day discovery).
• Security: Supply-chain risk is now board-level. Harden CI/CD, add pre-install guardrails for agents, and run internal drills for social engineering.
• Distribution: App-store self-preferencing and policy battles intensify. Europe’s DMA is the test bed; U.S. proposals like the Based Act push the Overton window.
• Venture & IPOs: Eyes on SpaceX/Anthropic/OpenAI. Real-asset IPOS—semis, space, AI infra, and data centers—look primed as public markets seek defensible, capital-intensive moats.
• Consumer Investing: Beware structural premiums: one public vehicle targeting private tech was cited as trading at 6x NAV—a potential pain point if/when that normalizes.

Programming note: Model releases are accelerating. One claim detailed seven models in training at xAI—V2 variants around 1T, two variants at 1.5T, plus 6T and 10T. Meanwhile, Meta said Muse Spark is an early data point, with larger models in development. The race to rumored 10T scale is very much on.