All articles
Token-based AI billing: how it works and the Tokenpocalypse
From June 1, 2026, GitHub Copilot charges by token consumption rather than a flat fee, and Uber burned through its annual AI budget in four months. We explain what token-based billing is, how it turns prompts into cost, and why the industry calls this shift the Tokenpocalypse.
ChatGPT Is No Longer Just a Chatbot — OpenAI Rebuilds Before IPO
OpenAI is rebuilding ChatGPT into an AI agent platform combining Codex, image generation, and partner apps. The strategic shift is directly tied to a planned IPO and the growing priority of enterprise revenue.
The EU Technological Sovereignty Package: what it is and how it works
On 3 June 2026 the European Commission unveiled its Technological Sovereignty Package: four initiatives binding chips, cloud, AI and open source into a single value chain. We explain what it is, what it contains and how it is meant to work.
Humanoid robot in the Himalayas — Unitree G1 summits Chimborazo
A modified Unitree G1 named Pemba summited the 6,268-meter Chimborazo in Ecuador on June 5, 2026. It is the first leg of a Triple Crown expedition, with Mount Everest as the ultimate goal. Behind the project is conservation engineer Pablo, who envisions mobile robots monitoring wildlife in places where humans cannot reach.
Microsoft breaks free from OpenAI — MAI models, Maia chip and the superintelligence goal
Mustafa Suleyman, Microsoft AI CEO, revealed that roughly six months ago the company formally obtained the right to pursue superintelligence independently after renegotiating its deal with OpenAI. At Build 2026 Microsoft unveiled seven MAI models including MAI-Thinking-1 (35B active parameters), the Maia 200 chip and Frontier Tuning for enterprise.
GUARD Act: US Bill to Ban Chinese Robots Could Cripple AI Research
The bipartisan GUARD Act, introduced June 3, 2026, would block imports of Chinese humanoid and quadruped robots. The problem: the US has no ready hardware alternative for research labs.
LLMs learn false claims — even when training data labels them "false"
New research on "Negation Neglect" shows that language models absorb false claims from training data even when those same data explicitly warn that the claims are untrue. The effect proved nearly as strong as fine-tuning without any warnings at all.
Do Language Models Need Sleep? Offline Recurrence and Memory Consolidation
Researchers from CMU and the University of Maryland propose "LLM sleep" — a phase where the model repeatedly processes context and writes it into fast weights before clearing its attention cache. Longer sleep improves reasoning over evicted context without increasing response latency.
1X goes all-in on world models, hires Luma AI head researcher
1X Technologies launches the 1X World Model Lab and hires Samarth Sinha from Luma AI. CEO Bernt Bornich declares that fine-tuning LLMs for robotics is a dead end. The company plans to ship thousands of NEO robots in 2026.
OpenAI launches in-house robotics division with custom actuator design team
OpenAI announced on May 31, 2026, the launch of an internal robotics division led by DALL-E creator Aditya Ramesh. The company is hiring for 11 specialized hardware roles and plans to design custom actuators, collect physical interaction data, and build robots from the ground up.
OpenAI AI model disproves Erdős conjecture open for 80 years
An internal OpenAI AI model has disproved the Erdős unit distance conjecture — a discrete geometry problem that stumped mathematicians for 80 years. It is the first instance of an AI system autonomously resolving a major open mathematical conjecture.
Sutton & Rafiee: AI Needs a Body and Experience, Not Just Data
Turing Award laureate Richard S. Sutton and Banafsheh Rafiee published a paper on May 22, 2026 proposing to ground AI in enactive cognition — four pillars: experience, action-perception inseparability, autonomy, and embodiment. Reinforcement learning comes closest to this model, but falls short of full equivalence.