Sam Altman: AI discovered PHP bugs, solved Navier-Stokes, may be unsandboxable

AI-native startups ship 11x more code than pre-AI companies, with PR velocity jumping from 2.4 to 207 PRs/month

The Pragmatic Engineer's winning essay documents concrete metrics showing the productivity impact of AI-native development: a founding engineer at Tano shipped 24 projects in 8 months versus 17 projects over 4 years at PostHog, with median PR merge time under 1 hour compared to days at PostHog, and 6x lower revert rates despite skipping mandatory human code review. Tano replaced peer review with Claude-powered CI agents that catch most issues, accepting "move fast and break things" tradeoffs where bugs like currency calculation errors are fixed within hours post-deployment. This directly demonstrates infrastructure and productivity implications of AI-native development workflows at scale and signals a fundamental shift in how engineering culture and quality gates must adapt when AI handles code generation and initial review.

Google ties Search access to AI adoption for schools; ChatGPT reaches 1.2B WAUs with ~20% of Google's search volume

Google announced that schools blocking AI tools entirely will lose Search functionality in 2027, claiming "Search and AI are deeply integrated." This triggered immediate antitrust backlash from FTC Chair Khan and Senator Blumenthal, who accused the company of leveraging its search monopoly to coerce AI adoption. The move signals potential regulatory escalation around Google's market power and raises questions about how platform integration strategies interact with antitrust enforcement. The timing coincides with accelerating competitive pressure: ChatGPT reached 1.2B weekly active users with 200M new WAUs added in two months, and ChatGPT's search queries have grown to approximately one-fifth of Google's volume. This metric underscores structural shifts in search behavior and validates the strategic importance of search integration that Google is now enforcing.

AI infrastructure investment concerns intensify: Google reports AI Overviews impacting search revenue, OpenAI cuts fundraising valuation by $150B, Meta faces credit downgrade

Markets are showing bipolar sentiment on AI infrastructure returns. Google reported that AI Overviews are impacting search revenue, OpenAI cut its fundraising target from $30B at $1.4T valuation to $20B at $1.25T valuation, and Meta faced a credit downgrade despite a $95B equity raise. These signals collectively suggest market worry that trillions in projected AI capex won't generate adequate returns, creating tension between the scale of infrastructure buildout and demonstrated monetization pathways.

Binance hack traced to fake job interview social engineering campaign costing $6.34 billion

The attackers posed as recruiters and company leadership to trick engineers into installing malicious code, then spent months escalating access before using an EIP-7702 delegation to drain a hot wallet on October 29. Binance had no cyber insurance and funded the recovery from its SAFU emergency fund, treasury, and futures insurance fund, with about $2.85 billion now frozen or intercepted.

Sam Altman on PHP vulnerabilities, test-time compute, math breakthroughs, sandboxing, and AI infrastructure economics

Cheekypint published a long-form interview between Stripe co-founder John Collison and OpenAI CEO Sam Altman covering frontier AI capabilities, safety, and infrastructure strategy. Main takeaways:

• AI models discovered critical PHP vulnerabilities after 20+ years; first day of disclosure went better than expected, with companies using models for rapid cyber defense and patching
• OpenAI internally proved solutions to Navier-Stokes and other hard math problems; test-time compute unlocks massive agent swarm capabilities
• Sandboxing superhuman models may be theoretically impossible; monitoring and activation inspection are interim mitigations but insufficient against models discovering novel science
• New application architectures and software stacks being built from scratch to defend against increasingly capable AI; reshaping how routers, OSes, and applications are designed
• Revenue per gigawatt and robotics scaling will drive future data center economics; token pricing is 'fake' and misrepresents true model value

Elon Is Doing It Again: SpaceX Will Be The World's Biggest AI Cloud

SemiAnalysis published a detailed infrastructure analysis on SpaceX's datacenter buildout strategy and capacity targets. Main takeaways:

• SpaceX has revised datacenter capacity targets upward to ~13GW in 2027, the largest single-company buildout ever, equivalent to all Google datacenter capacity from 1999-2026 compressed into 1.5 years
• SemiAnalysis has identified >9GW of capacity across multiple sites including Colossus 2 expansions, Starbase Texas, and TeraFab, with construction timelines of 4-5 months for greenfield builds—half the industry standard
• SpaceX is deploying 5.9GWh+ of battery storage at Colossus 2 and combining behind-the-meter power (8GW+) with grid connections, solving power constraints through modular MEP manufacturing and turbine relocation
• TeraFab chip fab project will likely include a third-party rental datacenter, expanding SpaceX's revenue optionality beyond captive use
• SpaceX's labor efficiency (3k workers per GW) and hardware manufacturing partnerships with Tesla enable unprecedented build velocity compared to Chinese competitors

What's the ROI of Software Engineers? AI Spending, PR Throughput, and Productivity Across 27 Companies

The Pragmatic Engineer published an essay on AI spending patterns and productivity metrics across 27 companies in 2026. Main takeaways:

• CodeRabbit reports 113-128% increase in weekly/monthly PRs year-over-year; Gitar shows tiered PR velocity across customer cohorts (5 to 60+ PRs/dev/week), suggesting culture-driven adoption rather than uniform productivity gains.
• High spenders ($1,000+/month/dev) include quant trading shops ($1,600/dev/month average, up 16x from January), small .NET shops ($3,000/dev/month), and Big Tech; spending caps are common but easily circumvented; Claude and Opus 5.5 dominate spend allocation.
• PR frequency gains vary widely: some teams see 5-10x increases (small .NET shop: 1-2 to 10+ PRs/day), others see flat velocity despite higher spend; PR size increased 3x at one large company, suggesting verbose LLM output rather than true acceleration.
• ROI measurement remains elusive: most companies lack concrete frameworks; trading firms measure ROI via desk profitability; quality-focused teams report minimal velocity gains despite AI adoption; few companies can quantify actual business impact.