Good morning. It's Monday, April 13, and we're covering the risks of autonomous agents, how benchmark loopholes are skewing AI performance, and more.
If you enjoy this email, share it with a friend. First time reading? Sign up here to keep up with the future of tech.
THE WEEKEND RECAP
Top Stories You Might Have Missed
Anthropic Briefly Bans OpenClaw's Creator: Peter Steinberger, now an OpenAI employee, was temporarily suspended from Claude access for "suspicious" activity, reinstated hours later after his ban went viral.
AI Agent Exploits Benchmark Flaws: Terminator-1 scored 95%+ on SWE-bench Verified using simple exploits to game results without solving tasks, exposing widespread benchmark reward-hacking vulnerabilities.
Anthropic Explores Building AI Chips: Reuters reports the company is considering in-house chip design amid shortages, as Claude's revenue run rate tops $30 billion and rivals pursue similar costly semiconductor strategies.
Pentagon AI Chief’s Windfall Raises Questions: Records show official Emil Michael sold xAI stake for up to $25 million after Pentagon deals with the firm, prompting conflict-of-interest concerns.
Alibaba’s HappyHorse Tops Video Arena: New model supports text-to-video and image-to-video with optional native audio, ranking #1, with API access scheduled for April 30, 2026.
Anthropic Secures CoreWeave Compute Deal (Paywall): The company signed a multiyear agreement to use CoreWeave’s US data centers and NVIDIA chips to scale Claude models amid surging AI demand.
Amazon Weighs Selling In-House AI Chips (Paywall): CEO Andy Jassy said the company may offer its custom silicon externally, as its chip unit targets over $20 billion in annual revenue.
OpenAI Launches $100 Pro Plan: The new tier offers 5× higher overall limits and 10× more Codex usage than Plus, targets developers, and positions against Anthropic, with over 3 million weekly Codex users.
INTERVIEW
Matt Sits Down With Tech Legend Marc Benioff
“An MIT student asked if she should switch out of computer science.”@benioff says: “She’s a top computer scientist, and she’s worried about getting a job.”“Students are being told it’ll be hard to find internships or roles after graduating.”“But companies like ours badly… pic.twitter.com/1lAsMVbfmH
@ForwardFutureView post on XTHE FUTURE LIVE
Can Meta Win the AGI Race?
BEP Holdings founder argues Zuck's path looks different: free devices, localized AI for underserved markets, and a data moat nobody else has.
Does Meta have a chance in the AGI race when @Anthropic and @OpenAI have cracked the flywheel: build coding tools, sell to enterprise, use revenue (and the models) to build better models, repeat?@benitoz Founder of BEP Holdings, argues yes. It'll just look different."That's… pic.twitter.com/qiUBmaHb3K
@ForwardFutureView post on XAGENTS
AI Agents Like Claude and OpenClaw Raise Power—and Risk
A new wave of autonomous AI agents—including Anthropic’s Claude Cowork and open-source tool OpenClaw—is accelerating the shift from chatbots to systems that can act independently across software environments. These agents can perform complex tasks like coding, legal document review, and inbox management with minimal human input, but require deep system access to do so.
That autonomy introduces new risks, from faulty decisions to data exposure, especially in open-source systems without centralized oversight. The result is a growing tension between capability and control, as enterprises weigh productivity gains against reliability and security concerns. → Read the full article here.
VENTURES
OpenAI Pitches With AI Tools and Funding
OpenAI outlined a policy proposal to support entrepreneurship through a “startup-in-a-box” model—an AI-powered bundle of tools, infrastructure, and potential funding. The concept includes automated back-office services like accounting, marketing, and procurement, along with model legal contracts, aiming to reduce barriers for new founders.
OpenAI suggests pairing these tools with microgrants or revenue-based financing to accelerate company creation. The proposal comes as competition intensifies, with founders increasingly turning to rival systems like Anthropic’s Claude. It also calls for clearer government rules on AI use, particularly around safety and ensuring public access to foundational models. → Read the full article here.
SPENDING
SpaceX Posts $5B Loss As AI Spending Surges
SpaceX reported nearly $5 billion in losses for 2025 despite generating over $18.5 billion in revenue, driven largely by heavy investment in AI through its newly acquired xAI unit. The company spent approximately $13 billion on chips and data centers for AI—about 50% more than its combined rocket and satellite divisions.
While its core businesses, including Starlink and launch services, produced roughly $8 billion in earnings, those gains were offset by depreciation and AI-related capital expenditures. The figures underscore how SpaceX’s upcoming IPO will effectively ask investors to fund Elon Musk’s AI ambitions alongside its dominant space business. → Read the full article here. (Paywall)
