Overview

July 2025 was the month the AI industry consolidated its position as a genuine economic force and began confronting the structural questions that scale creates. GPT-5 launched. Grok 4 reached the frontier. AI achieved gold-medal performance at the International Mathematical Olympiad — two months ahead of virtually any expert prediction. Walmart moved from agent experiments to a four-super-agent orchestration framework. And underneath all of this: a talent war of unprecedented scale, a geopolitical competition for AI infrastructure, and the first serious structural challenge to the startup equity model via “blitz hire” acquisitions.

Major Topics

IMO Gold Medal — A Genuine Inflection

OpenAI’s experimental reasoning model solved 5 of 6 IMO problems under the same constraints as human contestants (no tools, 4.5-hour sessions). Google DeepMind achieved the same result independently. AI safety researchers had assigned 8-16% probability to this outcome by 2025 even with tools permitted. Terence Tao predicted AI would not score highly on IMO as recently as June 2025. The result matters not because it signals AGI but because it confirms that new RL techniques applied to hard-to-verify tasks (long-form proofs) generalize to domains beyond competitive programming. Sam Altman predicted 2026 as the year AI begins contributing to actual scientific discovery.

Sources: 2025-07-22-ai-just-achieved-something-no-one-thought-it-would-until-years-from-now

Grok 4 at the Frontier

Grok 4 trained with 100x more compute than Grok 2 and 10x more reinforcement learning compute than any prior model. Artificial Analysis Intelligence Index: Grok 4 (73), O3 (70), Gemini 2.5 Pro (70). On ARC-AGI-2 — designed to test fluid intelligence and resistant to gaming — Grok 4 nearly doubled the previous high score. The result directly contradicts “scaling wall” narratives from late 2024. The competitive lead is expected to be short-lived as other labs deploy similar compute scales.

Sources: 2025-07-11-is-grok-4-the-best-llm-yet

ChatGPT Agent Launches

OpenAI launched ChatGPT Agent — combining Deep Research, Operator (browser interaction), and coding/terminal capabilities into a single general-purpose agent. On Humanity’s Last Exam: Agent scored 41.6% vs. O3’s 20.3%. On an internal OpenAI benchmark for complex knowledge work, Agent matched or exceeded human performance in roughly half of cases. Early use cases confirmed: customer feedback synthesis across 1,500+ emails, startup planning (38-minute autonomous run producing idea + financials + pitch deck), complex financial scenario planning. The paradigm shift: from prompting to delegating.

Sources: 2025-07-18-5-uses-for-the-new-chatgpt-agent, 2025-07-20-a-free-course-on-using-agents-created-by-chatgpt-agent

Walmart’s Agent Orchestration Framework

Walmart, under CTO Suresh Kumar, announced four super-agents: Sparky (customer-facing, live), Marty (supplier-facing, imminent), plus two more for employees and developers. The framework is built on Model Context Protocol (MCP) for interoperability. Evidence it is not vaporware: 900,000 associates interact with AI generating 3M questions/week; customer support resolution time cut by 40%; fashion production timelines cut by 18 weeks; shift planning time cut from 90 to 30 minutes. The episode’s framing: experimentation → orchestration is the natural enterprise AI arc, not a pivot.

Sources: 2025-07-29-walmart-blasts-past-agent-experimentation

The Norges Bank Playbook

Norway’s sovereign wealth fund ($1.8T, ~600 people) achieved 20% productivity gains and 213,000 hours saved in year one by: (1) making AI use mandatory as a condition of employment, and (2) building substantial support infrastructure — a 6-person AI enabler team, 40 AI ambassadors, repeated seminars and courses. Claude integrated with Snowflake for natural language querying. Earnings call analysis automated. The two-sentence playbook: make it mandatory; provide substantial support when you do.

Sources: 2025-07-25-how-one-company-saved-213000-hours-with-ai

America’s AI Action Plan and Infrastructure Imperative

The White House AI Action Plan (three pillars: accelerate innovation, build infrastructure, lead internationally) and Anthropic’s companion report converged on the same diagnosis: American AI dominance is a physical infrastructure challenge. Anthropic projects needing 2-5GW data centers to train a single advanced model in 2027-2028; the US AI sector will need at least 50GW total. China added over 400GW of power capacity in the past year; the US added several dozen. Open-source models were endorsed as geopolitical tools — American-values-based open models as the global default.

Sources: 2025-07-24-americas-ai-action-plan, 2025-07-29-is-global-ai-cooperation-even-possible

Jensen Huang’s 9 Predictions

NVIDIA CEO’s forward-looking thesis: AI will create more millionaires in 5 years than the internet did in 20. Every company will have two factories — a physical one and an AI factory (digital twin). “Everything that moves will be autonomous. Every industrial company will be an AI company, or it won’t be an industrial company.” The infrastructure build-out is in its early stages: “We are reinventing computing for the first time in 60 years.” The American tech stack (CUDA, NVIDIA) must remain the global developer default — that is the real competitive moat.

Sources: 2025-07-26-9-ai-predictions-from-jensen-huang

Worker Preferences — The Four Quadrants

Stanford HAI study of 15,000 workers across 100+ occupations mapped worker desire for automation against technical feasibility: Green light (high feasibility + high worker demand), R&D Opportunity (workers want it but AI can’t yet), Low Priority (neither workers nor AI ready), Red Light (AI can but workers resist due to high error cost). 69% of workers welcome automation that frees them for higher-value work. Only 2% want full automation with no human input. Primary concern is not job replacement (23%) but distrust of AI accuracy (45%). 41% of YC startups are reportedly building in the red light zone.

Sources: 2025-07-17-how-much-ai-do-workers-actually-want

The Talent War

Meta personally recruited eight OpenAI researchers including an entire Zurich office, with offers reported at $200M-$1B over four years. OpenAI’s chief research officer issued an internal memo comparing the situation to a home invasion. Zuckerberg maintains a personally curated list of the highest-priority AI researchers. The host’s thesis: this is not a compensation story — it is a question of who builds transformative AI and at what cost to organizational culture and incentives. Elite AI researchers are being repriced as premium capital goods.

Sources: 2025-07-01-how-the-war-for-talent-will-shape-ai, 2025-07-26-9-ai-predictions-from-jensen-huang

Blitz Hire Acquisitions and the Startup Equity Problem

Google acquihired Windsurf’s CEO and select engineers while leaving hundreds of other employees with little to nothing (many had not crossed the one-year vesting cliff). The pattern — license IP + hire select talent + leave nominal shell independent — was identified as a new M&A structure designed to circumvent antitrust review. Precedents include Inflection/Microsoft, Character AI/Google, Scale AI/Meta. The consequence: early employees who accepted below-market salaries in exchange for equity are being systematically excluded from liquidity events. Critics argue this converts startup culture from missionary to mercenary.

Sources: 2025-07-15-are-ai-acquihires-screwing-up-startups

Vibe Coding Matures

Lovable reached $100M ARR in 8 months (fastest ever) with 45 employees and 2.3M users. 75% of tech-oriented survey respondents use vibe coding tools; 88% report satisfaction. Use cases have expanded from prototypes to production consumer apps, niche scientific tools, and agentic business automation where coding agents query accounting systems, generate quotes, and update backends autonomously. Claude Code grew 300% in user base in two months.

Sources: 2025-07-18-all-the-cool-things-people-are-vibe-coding

AI Browsers — The New Browser Wars

Perplexity Comet (generally available), The Browser Company’s Dia (beta), and OpenAI’s forthcoming browser all launched or were announced in July. AI browsers differ from sandboxed agents by having native access to all open tabs, logged-in sessions, and the ability to seamlessly hand off control between user and agent mid-task. The strategic framing: controlling the AI browser = controlling how humanity interfaces with AI-mediated computing. Analogized to the 1990s browser wars.

Sources: 2025-07-12-are-ai-browsers-the-next-big-ai-trend

Key Trends

• AI achieved gold-medal IMO performance — benchmark saturation progressing from grade-school math to Olympiad-level proof in roughly one year
• General-purpose agents (ChatGPT Agent, Manus) are now capable of matching human performance on roughly half of complex knowledge-work tasks
• Enterprise agent orchestration is becoming standard at large organizations — Walmart, Citigroup, Norges Bank represent the production-scale template
• Vibe coding has crossed from novelty to permanent structural shift in who can build software
• Worker attitudes: 69% welcome AI augmentation but only 2% want full automation; trust in output quality is the primary barrier
• AI talent repricing is unprecedented — elite researchers being valued as premium capital goods at $100M-$1B packages
• “Blitz hire” acquisitions are systematically undermining the startup equity compact
• Physical AI infrastructure (data centers, power generation, chips) is the binding constraint on AI capability growth — geopolitical competition is intensifying
• AI browsers are emerging as the next platform layer battle, analogous to 1990s browser wars

Emerging Ideas

• Ambient agents and the “on the loop” paradigm: Moving from humans approving every agent step to humans observing and able to intervene — agents trigger from events, run in background, reach out when needed
• Context engineering as production discipline: “Prompt engineering is for hobby projects; context engineering is for production” — building the full informational environment around a model, not just writing prompts
• Agent experience (AX) replacing user experience (UX): Software must increasingly be designed for AI agents as the primary user, not human users navigating screens
• AI factories as mandatory second factory: Jensen Huang’s thesis that every company will operate a physical factory and an AI factory (digital twin), with the AI factory handling simulation, prototyping, and optimization
• Financialization of compute: Proposals for spot and forward markets for GPU compute, endorsed in the White House AI Action Plan, to improve price discovery and access

Sources

• 2025-07-31-can-ai-trade-stocks_instructions.md
• 2025-07-31-ai-starting-to-self-improve-says-zuckerberg_instructions.md
• 2025-07-29-walmart-blasts-past-agent-experimentation_instructions.md
• 2025-07-29-is-global-ai-cooperation-even-possible_instructions.md
• 2025-07-27-ambient-agents-and-6-other-big-ideas-coming-out-of-ai_instructions.md
• 2025-07-26-9-ai-predictions-from-jensen-huang_instructions.md
• 2025-07-25-how-one-company-saved-213000-hours-with-ai_instructions.md
• 2025-07-24-americas-ai-action-plan_instructions.md
• 2025-07-22-ai-just-achieved-something-no-one-thought-it-would-until-years-from-n_instructions.md
• 2025-07-20-a-free-course-on-using-agents-created-by-chatgpt-agent_instructions.md
• 2025-07-18-all-the-cool-things-people-are-vibe-coding_instructions.md
• 2025-07-18-5-uses-for-the-new-chatgpt-agent_instructions.md
• 2025-07-17-how-much-ai-do-workers-actually-want_instructions.md
• 2025-07-16-does-ai-secretly-slow-developers-down_instructions.md
• 2025-07-15-are-ai-acquihires-screwing-up-startups_instructions.md
• 2025-07-13-15-ways-i-use-ai-and-the-models-i-use-for-each_instructions.md
• 2025-07-12-are-ai-browsers-the-next-big-ai-trend_instructions.md
• 2025-07-11-is-grok-4-the-best-llm-yet_instructions.md
• 2025-07-10-everything-we-know-about-gpt-5-so-far_instructions.md
• 2025-07-09-the-7-types-of-ai-agents_instructions.md
• 2025-07-08-the-latest-ai-job-loss-predictions_instructions.md
• 2025-07-06-the-state-of-ai-mid-2025_instructions.md
• 2025-07-04-velvet-sundown-or-how-scared-should-we-be-of-ai-music_instructions.md
• 2025-07-02-how-ai-eats-consulting_instructions.md
• 2025-07-01-how-the-war-for-talent-will-shape-ai_instructions.md