Samwise Tech/AI/Robotics Newsletter
Saturday, June 20, 2026
OpenAI Hires Transformer Co-Author Noam Shazeer From Google DeepMind Ahead of IPO
Noam Shazeer, a co-author of the 2017 “Attention Is All You Need” paper that introduced the transformer architecture underlying every major language model, announced June 18 that he is leaving Google DeepMind to join OpenAI. Shazeer rejoined Google two years ago via a $2.7 billion licensing arrangement with Character.AI and became VP of engineering and co-lead of the Gemini program. OpenAI CEO Sam Altman called him “one of the people I have most wanted to work with since the very beginning of OpenAI.” The hire arrives as OpenAI advances toward a public offering. Former Trump White House AI policy official Dean Ball also joins as head of policy.
Sources: TechCrunch
European Commission Selects EUROPA Consortium to Build Open-Source Frontier AI in All 24 EU Languages
The European Commission selected the EUROPA consortium, led by Italian company Domyn, to build a frontier artificial intelligence model covering all 24 official EU languages, the Commission announced June 19. The project, awarded through the Frontier AI Grande Challenge, calls for a model exceeding 400 billion parameters — on par with the most capable commercial systems — and grants the consortium one year of access to EuroHPC supercomputing infrastructure. The initiative aims to give European businesses, researchers, and public institutions access to powerful AI built on European infrastructure, reducing dependence on large private AI firms based outside the bloc. Language accessibility is the program’s central design objective.
Sources: European Commission
VLC Media Player Creator Jean-Baptiste Kempf Raises $5M for Kyber, a Real-Time Robot Control Infrastructure Startup
Jean-Baptiste Kempf, lead developer of VLC Media Player — which runs on roughly four billion devices worldwide — has turned his attention to robots. His Paris-based startup Kyber has raised a $5 million seed round led by Lightspeed Venture Partners, TechCrunch reported June 19. Kyber builds real-time infrastructure for controlling remote devices, applying the low-latency networking principles Kempf developed at VLC to the challenge of reliable robot teleoperation at scale. Lightspeed has also backed Anthropic and Mistral AI. The funding reflects growing investor interest in robotics middleware companies that solve infrastructure problems rather than building robots directly.
Sources: TechCrunch
Sina Weibo’s VibeThinker-3B Claims Frontier-Level Reasoning From Just 3 Billion Parameters
Nine researchers at Sina Weibo posted a 14-page technical report to arXiv on June 15 claiming VibeThinker-3B, built on the Qwen2.5-Coder-3B base using their Spectrum-to-Signal post-training pipeline, matches or exceeds reasoning performance of flagship systems from Google DeepMind, OpenAI, Anthropic, and DeepSeek — models hundreds of times larger. VentureBeat reported the claim immediately triggered debate over whether benchmark results reflect genuine capability or narrow test-set optimization. The model requires just 6.7 GB of VRAM, making it deployable on consumer hardware. If results hold under independent scrutiny, they suggest parameter count is a poor proxy for reasoning capability.
Sources: VentureBeat
XDOF Turns Unglamorous Robot Training Data Collection Into a Business, Supplying Major AI Labs
The race to build capable robots has created an unglamorous but essential market: humans who physically collect manipulation data that robotic AI models need to generalize. XDOF, a startup supplying robot training data to AI labs, operates a network of workers performing repetitive object-handling demonstrations — folding laundry, sorting items, opening packages — captured by cameras feeding training pipelines. TechCrunch reported June 17 that several AI robotics programs, including OpenAI’s relaunched robotics initiative, are paying XDOF for training data they cannot collect cost-effectively themselves. Without diverse physical demonstrations across varied environments, robotic foundation models fail to generalize from lab conditions to real-world deployment.
Sources: TechCrunch
Anthropic’s Claude Design Hits 1 Million Users in First Week, but Extreme Token Consumption Draws Backlash
Anthropic’s Claude Design, released in April 2026 as a research preview, attracted more than one million users in its first week — but its token consumption has triggered backlash among paying subscribers, VentureBeat reported June 18. A PCWorld reviewer burned through approximately 80 percent of their weekly Claude Pro allowance in roughly 25 minutes producing just three variations of a single webpage prototype, describing it as “another token-hungry Claude product” that Pro users would “barely be able to use.” The episode highlights recurring tension in AI product design between showcasing generative capability and delivering sustainable value to consumers paying $20 per month.
Sources: VentureBeat
Pixi Launches iOS App That Sends AI-Powered Augmented Reality Characters Through iMessage
Pixi launched a messaging-native iOS app June 18 on the App Store that allows users to send AI-powered augmented reality characters through iMessage. The characters come to life through the recipient’s iPhone camera, reacting to physical surroundings, interacting with nearby people, and responding in real time — with no additional app download required from recipients. TechCrunch reported the launch marks a new approach to AI-native social communication, embedding interactive agents directly into existing iMessage infrastructure rather than requiring a dedicated platform. The app combines generative AI character animation with real-time AR, positioning interactive AI characters as a new format for casual mobile messaging.
Sources: TechCrunch
Tech Pulse
Top Frontier Models (SWE-Bench Pro): Claude Mythos 5 (80.3%) | Claude Fable 5 (80.0%) | GPT-5.4 (59.1%)
Top Open Source Models (SWE-Bench Pro): DeepSeek-V4-Pro-Max (80.6%) | GLM-5.2 / Zhipu AI (62.1%) | MiniMax M3 (59.0%)
Top Small Models (15–50B): Qwen3-32B | Llama 4 Scout 17B MoE | Mistral Small 3.2 (22B)
Top Edge Models (0–15B): VibeThinker-3B (6.7GB VRAM) | Gemma 3 9B | Qwen3-8B
AI Leaders: NVIDIA $5.2T | Alphabet $4.6T | Microsoft $3.1T
Robotics Leaders: Intuitive Surgical $175B | FANUC $30B (est.) | Symbotic $10B (est.)
Curated by JD · samwise.agency