June 19, 2026 — The Trend Tribune

By tech_tanaka · June 19, 2026

OpenAI Acquires Transformer Co-Inventor, Accelerating Major Talent and Funding Push Ahead of IPO — (Photo: illustrative)

"The Father of Transformers" Joins OpenAI — A Blockbuster Hire Ahead of IPO

The AI industry has been rocked by a major announcement. Noam Shazeer, one of the co-inventors of the Transformer architecture — the design philosophy underpinning large language models, introduced in the landmark 2017 paper "Attention Is All You Need" — publicly announced on June 18 that he is joining OpenAI.

Shazeer spent many years as a researcher at Google and is also well known for his more recent tenure at Google DeepMind. His move is more than a simple talent acquisition. It is a symbolic moment: one of the chief architects of the very architecture that gave birth to ChatGPT and GPT-4 is now stepping inside the company that has done the most to commercialize it.

In the same week, OpenAI also brought on Dean Ball, a former AI policy official in the Trump administration. This back-to-back hiring of heavyweights in both technology and policy leaves little doubt that the moves are part of a broader effort to solidify the organization ahead of a public offering (IPO).

The "Inference Gold Rush" Shows No Signs of Slowing — Baseten Closes In on $1.5B

While OpenAI stacks up talent, enormous capital is also flowing into the AI infrastructure space. Baseten, a startup providing cloud infrastructure specialized in inference — the process by which a trained model actually generates responses — is reportedly on the verge of closing a funding round of approximately $1.5 billion, with a projected valuation of around $13 billion.

What is particularly striking is that only a few months have passed since its last major round. As the cost of actually using AI models has exploded, demand for fast, cost-efficient inference infrastructure continues to grow with no ceiling in sight. Baseten's meteoric rise is a microcosm of exactly that reality.

Snap Makes the Call to "Spin Off" Its AI Video Team

A contrasting move has also emerged. Snap, the company behind Snapchat, announced that it is spinning off its in-house AI video development team into a new independent company called Dotmo. The reason: cost. Developing and operating generative AI video requires enormous computational resources, and that model proved to be a poor fit with Snap's social media business economics.

Dotmo will be launched by current Snap employees who will transfer directly to the new entity. By spinning out technology that had become "too heavy to keep in-house," the goal is to position it as a focused startup better suited to raising dedicated funding. This is a noteworthy example of what "focus and specialization" can look like in the age of AI.

YC Spring 2026 Demo Day: The Startups Poised to Ride the Next Wave

At Y Combinator (YC)'s Spring 2026 Demo Day — the premier launchpad for Silicon Valley startups — eleven companies were highlighted as ones to watch by VCs. Among them are firms already valued at over $175 million, and competition across AI-driven application domains is intensifying further.

In Summary: The Race Over How AI Gets Used Is Heating Up

Looking across this week's news, a common theme emerges: how to put AI to use. OpenAI is locking in talent with foundational knowledge of its core architecture as it eyes an IPO. Baseten embodies the reality that enormous capital is now gravitating toward inference infrastructure — the engine that makes AI run. Snap's Dotmo spinoff represents one answer to the structural question of where, organizationally, to house the heavy lifting of AI development.

Even as the race to build AI models themselves continues, the battlefield is rapidly shifting toward the implementation layer — where, by whom, and how AI is actually deployed. What Noam Shazeer will create from inside OpenAI, and what Baseten will build with hundreds of billions of yen in fresh capital, are set to become the defining questions of the second half of 2026.

References: techcrunch.com / twitter.com / techcrunch.com / techcrunch.com / techcrunch.com

OpenAIAIIPOスタートアップ推論インフラ

Can Robots Learn from Failure and AI Discover 'Zero'? — Reading the Frontline Papers of June 2026

By thesis_yamamoto

From the latest papers posted on arXiv, two major trends have emerged: autonomous learning and physical manipulation in robotics, and the mathematical discovery capabilities of AI. The groundwork for real-world applications is steadily being laid.

Today's Key Points: Robot × Autonomous Learning and AI's "Mathematical Discovery" Advancing in Parallel

On the morning of June 19, 2026, scanning the latest papers posted to arXiv, two major currents emerge. The first is an acceleration of research into robots autonomously acquiring capabilities through failure, video, and language as cues. The second is a question that is almost philosophical in nature—one that re-examines the intellectual limits of AI itself: can it even discover "zero"?

Robots Get Smarter from "Failure": RDSRC and Guava

Among robot research this morning, the most eye-catching work is the RDSRC (Recovery-Driven Synthesis of Relational Concepts) framework proposed by Carnegie Mellon University and collaborators (arXiv:2606.18328).

In conventional reinforcement learning, every failure required learning a separate "policy dedicated to that specific failure," which was highly inefficient. RDSRC enables robots to simultaneously acquire abstract concepts and skills from failure, preventing similar failures from occurring in the future. Its defining feature is a three-stage loop of "Recover → Discover → Plan," and the idea of turning failure into fuel for learning feels refreshingly novel.

In the same vein, Guava (arXiv:2606.18363) is also worth noting. By leveraging large-scale vision-language models (VLMs) as a "general-purpose harness" and combining perception, planning, and control modules as external tools, it demonstrates greater generality than end-to-end VLA (Vision-Language-Action) models. Its design philosophy of not depending on any specific robot hardware could open doors to industrial applications.

Learning Navigation from Video Alone: VEGA

VEGA (arXiv:2606.18426) takes an approach of learning navigation capabilities from "unlabeled first-person-view videos" that abound on the internet. It makes use of cluttered indoor footage captured in everyday human life and real-world environments full of obstacles, leveraging geometric trajectories as supervision signals.

The high cost of collecting quality training data for robots has long been a challenge, but VEGA attempts to break through that bottleneck by tapping the inexhaustible resource of "existing human-captured video." It could fairly be called a revolutionary approach in terms of scale.

A World Model with 3D Consistency: PAIWorld

PAIWorld (arXiv:2606.18375), a world foundation model specialized for robot manipulation, directly tackles a challenge that has often been overlooked: 3D consistency across multiple camera viewpoints (egocentric, wrist, and receding-hand perspectives). It resolves the problem of 3D consistency breaking down when multi-view video is simply concatenated, improving the accuracy of simulators used for robot policy learning.

"Can AI Discover Zero?"—Probing the Limits of Mathematical Discovery

On the AI research side, a more fundamental question has been raised. "Nothing from Something" (arXiv:2606.17289) is a paper that examines whether neural network-based AI can independently discover mathematical concepts beyond its training data.

The specific test case is set as "the independent discovery of the concept of 0 (zero)." Just as it took a long time in the history of human mathematics for the concept of zero to be discovered and established, zero is not a simple number for AI either—it is an abstract concept representing "the absence of something." This research presents a new axis for evaluating AI capabilities by questioning how far current LLMs can go in making genuine mathematical discoveries outside the training distribution.

A Transformer Efficiency Revolution: Gaussian Mixture Attention

On the architecture side, Gaussian Mixture Attention (GMA) (arXiv:2606.18283) has appeared. By replacing the all-pairs token computation (O(n²)) of standard dot-product attention with routing to K learned Gaussian mixture distributions, it achieves linear time complexity. As a practical approach to eliminating the bottleneck in long-context processing, its adoption in future work will be worth watching.

In Summary: AI Research Relentlessly Expanding Its Interface with the Real World

The theme common to this morning's papers is "expanding the interface with the real world." Robots are absorbing data from failure, video, and 3D space, while language models are under evolutionary pressure on both their mathematical discovery capabilities and attention efficiency. Even if individual papers appear as isolated dots, their accumulation points toward a larger wave moving in the direction of "general-purpose intelligent robots." Tomorrow's arXiv will be just as hard to look away from.

References: arxiv.org / arxiv.org / arxiv.org / arxiv.org / arxiv.org / arxiv.org / arxiv.org

AI論文ロボティクス深層学習VLATransformer

The AI Budget 'Invoice' Has Arrived — The Reality of 2026 as Companies Struggle to Calculate ROI

By fin_suzuki

The 'Tokenmaxxing' boom of unlimited token usage has brought enormous costs to businesses. AI ROI issues, including Uber's AI budget overruns, have become the top concern among investors and executives.

"Keep Using AI," They Said — Then Uber Burned Through Its Budget in Months

In the first half of 2026, one of the hottest management trends in Silicon Valley was "Tokenmaxxing." CEOs were encouraging employees to "squeeze every last drop out of AI," stoking internal expectations around productivity gains. Then the bills arrived, and reality hit hard.

Tiffany Luck, a partner at NEA (New Enterprise Associates), put it plainly in a recent interview: "Enterprise companies still can't calculate the ROI of AI. That's the biggest challenge facing the entire industry right now."

Uber reportedly burned through its annual AI budget in just a few months, and some companies have pulled Anthropic Claude licenses for parts of their organizations. Meta shut down its internal AI usage leaderboard. The assumption that "more usage equals better outcomes" is being rapidly reconsidered through the lens of cost management.

The Structural Problem with Token Costs — What Happens in a Single Inference

From the perspective of someone with a CS background, it's worth noting that inference costs for large language models are primarily usage-based, billed in proportion to token count. The per-thousand-token rate may look like just a few cents, but when hundreds of employees use these tools at full capacity for daily work, monthly costs can easily reach millions of dollars.

The problem with Tokenmaxxing is that companies issued a blanket directive to "just use it" without setting any caps on token consumption or establishing metrics for cost-effectiveness. From an accountant's perspective, this is simply a failure of budget control. The bill has now come due for having put off the quantification of baseline operational efficiency — a prerequisite for calculating return on investment.

The VC Perspective — The Real Winners "Earn with AI," Not "Sell AI"

Chi-Hua Chien, a veteran VC at Kleiner Perkins, argues that "the true AI winners are not the companies selling AI." Known for identifying Facebook's rise at an early stage, his insight is not mere contrarianism — it cuts to the heart of what makes a revenue model sustainable.

His logic is straightforward. The layer providing AI infrastructure and models is trending toward commoditization, compressing profit margins. The long-term winners, he contends, will be companies that use AI to transform inefficient industrial processes and monetize the outcomes they deliver.

This is clearly reflected in YC's (Y Combinator's) Spring 2026 batch. Among the 11 standout startups that TechCrunch compiled from VC interviews — some valued at over $175 million — the majority follow an industry-specific business model built around "doing something with AI."

What the Closure of PayPal Ventures Signals — Selection Pressure on CVCs

That same week, it emerged that PayPal Ventures would be shutting down despite a ten-year track record of 80 investments. While framed as part of PayPal's broader restructuring, the closure — viewed in a wider context — offers one answer to the question of how much distinctive value corporate venture capital (CVC) arms can actually demonstrate amid the AI investment boom.

Investment portfolios that expanded on vague strategic synergies tend to be the first on the chopping block when pressure mounts on the parent company's core business. From a financial standpoint, CVCs already face a structural challenge in serving the dual objectives of financial and strategic returns, and the accelerating pace of change in the AI era makes those tensions easier to expose.

Conclusion — The Focus of AI Investment in the Second Half of 2026 Is "Measurement"

The phase of "invest in and use AI at all costs" that defined 2025 through the first half of 2026 has clearly reached an inflection point. Over the next six months, the primary challenges for both companies and investors will be making AI costs visible, quantifying ROI, and building monetization models tied to measurable outcomes.

Here's a question worth sitting with over your morning coffee: what is your company's AI spending actually producing today? The CFO who can answer that with confidence may still be in the minority.

References: techcrunch.com / techcrunch.com / techcrunch.com / techcrunch.com

AI投資ROIエンタープライズAIベンチャーキャピタルコスト管理

OpenAI Acquires Transformer Co-Inventor, Accelerating Major Talent and Funding Push Ahead of IPO