How Claude Code's simpler prompts beat traditional engineering

Amazon projects $220B capex in 2026 with majority for AI; claims sub-3-year payback on server hardware

Amazon has disclosed the most concrete ROI underwriting case any hyperscaler has provided for AI infrastructure investment: sub-3-year payback on server hardware with 5-6 year useful life and 5-year customer contracts. The company expects $220B total capex in 2026 with the majority allocated to AI. This framework is significant for understanding whether massive capex bets will actually generate returns, though it assumes sustained demand visibility and no acceleration in hardware obsolescence.

Starlink now dominates SpaceX launch manifest at 79%, squeezing out commercial satellite operators until 2028-2029

Starlink's share of Falcon 9 missions has grown to 79% in 2026, up from 54% in 2020, with at least seven spacecraft companies told launches are fully booked until 2028 or 2029. SpaceX is prioritizing its own constellation as it transitions to Starship, which analysts estimate could generate tens of millions more in revenue per mission than commercial customer payloads. Starlink accounted for roughly 60% of SpaceX's $15.5B revenue in 2025 ($11.4B from Starlink itself), creating a powerful economic incentive to reserve launch capacity for internal satellites rather than external customers. SpaceX's IPO prospectus explicitly states it "may prioritize our own launch payloads over additional U.S. government contracts or third-party customers," signaling a structural bottleneck for the broader commercial space industry and raising barriers for smaller players without their own launch capability.

Google consolidates AI strategy as Hassabis steps down; four senior researchers depart to found Discovery Loop

Demis Hassabis is stepping down as DeepMind CEO to become chairman and Alphabet chief scientist, with Koray Kavukcuoglu becoming SVP leading DeepMind. The move signals Google's effort to streamline internal AI competition and improve compute allocation across fragmented units—a structural problem acknowledged internally for years. Simultaneously, four of Google's most senior AI researchers—Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals—are departing to found Discovery Loop, a Public Benefit Corporation focused on automating machine learning, science, and engineering discovery loops. The founders are backed by Radical Ventures, Khosla Ventures, Lightspeed, Kleiner Perkins, and Alphabet itself. Discovery Loop's founding statement emphasizes automating experimental loops at scale and cites different infrastructure requirements than Google's consumer-focused systems, suggesting potential preference for non-TPU compute. Together, these moves signal both Google's recognition of execution challenges within its current AI structure and a strategic shift toward AI-for-science as a primary frontier, with top talent voting with their feet.

AI agents handling 8-hour tasks now represent 25% of Codex requests, up from 2% in December 2025

Roughly a quarter of Codex users are now making monthly requests for work that would take humans a full work day to complete, up from 2% in December 2025. This concrete growth signal indicates meaningful progress in autonomous task duration and complexity within a major AI platform, suggesting agents are moving beyond narrow, short-horizon tasks into sustained multi-hour workflows.

Boris Cherny: Building Claude Code

Boris Cherny (Claude Code creator), Diana Hu (interviewer) in Y Combinator Startup School 2026 published an interview titled "Boris Cherny: Building Claude Code". Main takeaways:

• Opus 5 runs for extended periods (days/weeks/months) in Auto Mode without scaffolding, representing a step-change in agent autonomy and persistence.
• Claude Code deleted 80% of its system prompt for Opus 5 because the model's improved capabilities made previous corrective instructions redundant; simpler prompts sometimes yield better model performance.
• Successful harness design requires empirical ablation on each model release: delete system prompts/tools entirely, observe failures, and add back only what the model demonstrably needs—a radical departure from traditional software engineering practices.
• Prompt injection resistance has reached a new frontier in Opus 5 via three-layer defense: alignment training, mechanistic interpretability-based neuron detection, and auto-mode classifiers; the model no longer appears vulnerable to indirect instruction injection.
• "Product overhang" describes unrealized capabilities in current models that products haven't yet exposed; building involves "unhobbling" models by removing unnecessary constraints rather than adding features.

Google in Talks to Acquire Mechanize Talent and Technology for $1.5B+ in Coding AI Push

Business Insider reported on Google's discussions with San Francisco startup Mechanize for a potential $1.5B+ deal involving talent acquisition and non-exclusive technology licensing. Main takeaways:

• Google is acquiring Mechanize talent focused on model evaluation and development to improve coding capabilities, where it lags OpenAI and Anthropic by ~6 months
• Deal structured as hybrid transaction (talent + licensing) to mitigate antitrust scrutiny, following Google's pattern with Windsurf and Character AI acquisitions
• Mechanize's technology addresses Google's documented weakness in AI coding performance, a critical competitive battleground consuming most current compute spending