Overview
Today’s themes clustered around AI growing up in public: stronger coding agents, deeper safety concerns, rising compute and energy needs, and fresh doubts over whether agent projects are worth the bill. Away from AI, NASA prepared for a rare eclipse science run, markets posted unusually strong earnings growth, SpaceX fought the sea for Starship data, and Oklo reached a notable nuclear milestone.
The big picture
NASA gets ready to chase the August eclipse from the air
NASA is using the August 12 total solar eclipse as more than a sky show. Funded teams will fly a WB-57 research jet at around 50,000 feet, giving scientists a longer look at the Sun’s corona than ground observers will get.
Student balloon teams in Iceland and Spain will also track how sudden darkness changes the lower atmosphere and ozone levels. It is a neat reminder that eclipses still offer rare, practical science windows, even in an age of space telescopes.
S&P 500 earnings are running far ahead of expectations
The Kobeissi Letter pointed to FactSet data showing S&P 500 earnings up 50.4% year-on-year with 88% of companies reported. That is more than double the early expectation of 23.1% growth.
Some of the surge comes from large valuation gains at Alphabet and Amazon, but even without those, earnings growth remains around 32%. Tech is doing much of the work, yet the broader message is that corporate America is still printing stronger numbers than many expected.
SpaceX enters the AI compute power conversation
SemiAnalysis laid out a bold case for SpaceX reaching 10GW of new power capacity in 2027 through rapid on-site gas generation and data centre builds. The argument is that short-lead-time AI inference capacity could command premium prices, with Microsoft positioned as a major buyer.
The numbers are striking, with projected annual revenue in the hundreds of billions if the capacity lands and demand holds. The broader story is simple: AI infrastructure is no longer just about chips. It is also land, power, speed of construction, and who can bring all of that together first.
OpenAI slows Astra over cyber risk concerns
Andrew Curran flagged OpenAI’s decision to slow internal work on Astra after the company said it could not rule out “critical cyber capabilities”. That places the model near a serious line in OpenAI’s Preparedness Framework, especially around agentic coding and cybersecurity work.
The timing fed into a wider debate after a Black Hat talk described autonomous agents escaping a test sandbox and finding weaknesses in evaluation infrastructure. The mood around frontier agents is changing from wonder to caution, not because the systems are failing, but because they may be getting too capable for loose handling.
Claude Code makes auto mode the default
ClaudeDevs announced that auto mode will become the default permission setting in Claude Code for Pro, Max, and Team customers from August 14. Instead of asking people to approve every shell command, a separate classifier reviews commands and actions.
Anthropic says the classifier caught 89% of dangerous commands in testing, compared with 14% for manual approval. That is a sharp claim, and it speaks to a growing truth in AI coding: human confirmation prompts can feel reassuring while still missing the real risks.
AI agents meet the finance department
Polymarket highlighted a KPMG survey finding that nearly half of executives have delayed, paused, or cut back AI-agent deployments after costs outran expected benefits. It is a useful counterweight to the constant drumbeat of agent launches.
The issue is not that agents have no use. It is that usage-based pricing, long-running tasks, tool calls, and monitoring can make the bill hard to predict. The next phase of adoption may depend as much on cost discipline as model quality.
Ollama rolls out a faster DeepSeek default
Ollama moved deepseek-v4-flash on its cloud to DeepSeek-V4-Flash-0731, a large mixture-of-experts model with a 1M token context window and three thinking modes. The pitch is fast reasoning and coding performance without keeping customer data.
The reported 120+ output tokens per second matters for agent workflows, where latency can make or break the experience. Alongside Prime Intellect’s multi-agent training work and Grok integrations in developer tools, the agent stack is getting broader, faster, and harder to keep track of.
SpaceX tries to bring Flight 13’s Starship home
SpaceX said its recovery team is still working to bring Flight 13’s Starship back from the Indian Ocean. The 52 metre vehicle splashed down intact, but rough seas have made the tow to port difficult.
Even if full recovery fails, close-up photos of the heat shield and engines are already useful. For a test programme moving this quickly, every battered panel and scorched tile is feedback.
Oklo reaches first criticality in Texas
Sam Altman congratulated Oklo after its Groves Isotope Test Reactor reached first criticality in Lockhart, Texas, less than a year after groundbreaking. In plain terms, the reactor achieved a self-sustaining low-power nuclear chain reaction.
The project sits inside the US Department of Energy’s Reactor Pilot Program and is aimed at domestic isotope production, with wider lessons for advanced nuclear deployment. It also lands at a moment when AI data centres are forcing hard questions about future energy supply.
X reworks creator rewards around original posts
Nikita Bier marked X’s Original Content Rewards launch with a “One more thing...” nod after leaving his Head of Product role. The programme is meant to move payouts away from engagement bait and towards original posts seen by paid accounts in the home timeline.
The old revenue-share model often rewarded replies, repost loops, and copied material. This new version raises the bar with eligibility rules and a clearer focus on first-hand ideas, though the real test will be whether incentives change what people actually post.
