“The biggest individual talent move in AI since Andrej Karpathy joined Anthropic.” That is how Build Fast With AI described Noam Shazeer joining OpenAI on June 18th. That sentence contains a remarkable amount of information. The person who wrote the paper that made modern AI possible just left a $2.7 billion retention package at Google to join the company racing to an IPO. Sam Altman said it was a hire he had wanted since the very beginning of OpenAI. The market added 1.17% to Alphabet shares on the news. Someone calculated that correctly.
Ground Truth · Episode 12 · June 22 2026There is a 2017 paper by eight Google Brain researchers titled “Attention Is All You Need.” If you are unfamiliar with it, here is the summary: it introduced the transformer architecture that underlies every significant AI model deployed today — GPT-5.5, Claude, Gemini, Llama, every large language model, every image generation model, every AI system you interact with. Every one. The paper is the most consequential published work in the history of artificial intelligence. One of its eight co-authors is Noam Shazeer. On June 18th 2026, Shazeer left Google DeepMind — where Google had paid approximately $2.7 billion to retain him — to join OpenAI as Lead for Architecture Research. He lasted 22 months on the retention package. This is the context in which the rest of this week’s AI news arrived. Welcome to Episode 12 of Ground Truth.
The Man Who Invented the Transformer Just Left Google’s $2.7B Package. Here Is Why It Matters.
The Lead for Architecture Research role at OpenAI is not a research contributor role. It is the person responsible for the physical neural network structures underlying all OpenAI models — the architectural decisions that determine how efficiently a model learns, how much compute it requires, what its reasoning capabilities look like, and what its ceiling is. These decisions compound. An architectural innovation that improves compute efficiency by 20% across a model series of GPT’s scale is worth billions in training cost and represents a meaningful capability advantage over competitors who did not make the same innovation.
Shazeer has done this before. After leaving Google the first time in 2021, he co-founded Character.AI, whose model architecture attracted significant research attention before Google paid $2.7 billion to absorb him back into DeepMind. He lasted 22 months — and chose to leave for a competitor racing toward a public market debut and a role that, in Altman’s own description, he had wanted from the very beginning of OpenAI. The market’s reaction — Alphabet up 1.17% on the news of the departure — implies the market values Google’s $422 billion revenue base and compute commitments above any single researcher. That assessment may be correct today. The long-term implications of Shazeer’s architectural work at OpenAI will be visible in GPT model trajectories over the next three to five years, not in daily stock movements.
June 13: Andrej Karpathy resigns from Anthropic (OpenAI co-founder, former Tesla AI director). June 18: Noam Shazeer joins OpenAI from Google DeepMind (co-author “Attention Is All You Need,” the transformer paper). Google paid $2.7B to retain him. He lasted 22 months. Google response: Sundar Pichai promotes Demis Hassabis to oversee all Google AI research. The talent movements of June 2026 signal where people closest to the frontier believe the most consequential work is happening — and which organisations they believe are positioned to do it.
Google’s Most Capable Model Ever. The Benchmarks Are Rewriting the Leaderboard.
Today — June 22nd — Google launched Gemini 2.5 Pro with Deep Think reasoning mode. Deep Think is Google’s answer to OpenAI’s o-series and Anthropic’s extended thinking — a mode that allows the model to reason step by step before producing its final answer, dramatically improving performance on complex logic, mathematics, and code. Early benchmark results place it at or near the top of FrontierSWE and mathematics reasoning leaderboards, competing directly with Claude Opus 4.8 and ahead of GPT-5.5. Developers in the Vertex AI enterprise preview describe the model as qualitatively different on complex multi-step tasks. The “audible groan” from developers at Google I/O when the model did not launch as expected has been replaced with community reaction ranging from impressed to genuinely excited. Whether that enthusiasm survives contact with real-world workloads — rather than benchmark conditions — will become clear over the next two weeks.
For the First Time in Three and a Half Years, ChatGPT Does Not Have Majority Market Share.
Sensor Tower’s 2026 State of AI Report put ChatGPT’s global AI assistant market share at 46.4% by late May 2026 — the first time it has held less than half the market since its November 2022 launch. The share is being taken across multiple competitors rather than a single challenger: Claude growing in enterprise and technical segments, Gemini growing via Android pre-installation, Perplexity growing in search-adjacent use cases, Microsoft Copilot in enterprise workflows, Meta AI across WhatsApp and Instagram in international markets. The AI assistant market is fragmenting in the way that social media fragmented after 2012 — multiple platforms, different user bases, different use cases, none fully dominant.
For OpenAI’s IPO narrative, the trajectory matters as much as the absolute number. ChatGPT still has more users than any other AI assistant by a significant margin. It also has a declining share in a growing market, at exactly the moment it is asking public investors to price a near-trillion-dollar valuation. The S-1 will need a compelling answer to the question of which of those two facts investors should weight more heavily.
A New Attack Class Hijacking AI Coding Agents. 85% Exploitation Rate. 2,388 Organisations.
Agentjacking — disclosed this week — targets the Sentry error-tracking platform used by hundreds of thousands of development teams. Attackers craft fake error reports containing markdown injection instructions that appear, to AI coding agents like Claude Code, Cursor, and OpenAI Codex, as legitimate debugging guidance. When the agent reads the injected report, it interprets the malicious instructions as part of its workflow and executes them. The exploitation rate is 85%. 2,388 organisations have been affected. There is no universal patch.
The reason this is particularly alarming is the trust model that makes it work. Developers have specifically trained themselves — and their entire automated workflows — to trust their coding agents’ responses to error reports. When Claude Code says to run a command to fix a bug, you run the command. Agentjacking exploits exactly that conditioned trust. Immediate mitigation: treat all error-tracking platform output as untrusted input before allowing agents to act on it. Add a human review layer between error reports and autonomous agent execution. This is not a solution. It is a stopgap until the security community produces platform-level patches.
Disable automatic agent execution on Sentry error reports. Route all error-tracking output through human review before agent action. Treat it as untrusted input — the same model you apply to user-submitted data. Do not run agent-generated commands sourced from external error reports without reviewing the command origin. Watch for fix suggestions unrelated to the error being debugged — that is the injection signature. Audit every external data source your agents read. Sentry is the disclosed vector. It is unlikely to be the only one.
The Most Consequential Analysis of Fable 5 Published This Week
The Economist’s June 20th cover story frames the Fable 5 recall not as a response to a technical jailbreak but as a deliberate geopolitical assertion: the United States government intends to treat frontier AI models as it treats advanced weapons and semiconductors — subject to export controls requiring government clearance for international distribution. If that framing is correct, AI labs deploying globally would need export licences for their most capable models in the same way semiconductor manufacturers need export licences before shipping advanced chips to restricted countries.
The article also notes the self-defeating dynamic this creates. The policy designed to protect American AI superiority drives international developers toward Chinese alternatives like GLM-5.2 and Qwen — models without equivalent export restrictions. The geopolitical logic of frontier AI export controls is real. The second-order effects may not serve the interest they are designed to protect. Fable 5 and Mythos 5 remain offline today — ten days since the recall. The free trial window closes tonight. Every day the models remain offline, the geopolitical doctrine framing gains evidence over the narrow technical vulnerability framing. A genuine software jailbreak would have been patched in days. Ten days of silence suggests something more structural is being negotiated.
The Economist put Fable 5 on its cover and called it a weapons export. If that framing holds — if frontier AI models are now treated under export control law like advanced semiconductors — the entire AI industry’s global deployment strategy just changed permanently. Ten days in, with the model still offline, the doctrine framing is gaining evidence. A genuine software jailbreak would have been patched by now.Neal Lloyd · Ground Truth, Episode 12
ChatGPT invented the consumer AI assistant category, grew to a billion users faster than any product in history, and no longer commands a majority of the market it created. Gemini 2.5 Pro just launched today and is rewriting the benchmark leaderboard. The man who co-wrote the transformer paper just joined OpenAI. Agentjacking is hitting 85% of targets it attempts. And Fable 5 is still offline. This was not a slow news week.Neal Lloyd · Ground Truth, Episode 12
Ground Truth, Episode 12 · June 22 2026
Neal Lloyd covers the real-world impact of AI — money, power, geopolitics, and the stories behind the headlines. Ground Truth is his daily AI news and analysis series on emdexter.blogspot.com.
- Ep 01The Gold Rush
- Ep 02ChatGPT Knows Everything
- Ep 03Siri Is Now Google
- Ep 04America’s AI Law Is a Mess
- Ep 05Is AI Taking Your Job?
- Ep 06Microsoft vs Everyone
- Ep 07SpaceX Is Trading
- Ep 08The Government Pulled Fable 5
- Ep 09Trump and Bernie Want to Own AI
- Ep 10SpaceX Buys Cursor for $60B
- Ep 11The Fable 5 Truth
- Ep 12The Week That Changed EverythingYou are here
