Last Week in AI #322 - Robotaxi progress, OpenAI Business, Gemini in Chrome

Amazon's Zoox jumps into U.S. robotaxi race with Las Vegas launch, OpenAI secures Microsoft’s blessing to transition its for-profit arm, and more!

Sep 22, 2025

The Robotaxi Race is Speeding up in the US

Relevant stories:

Amazon’s Zoox launched its first public robotaxi service on the Las Vegas Strip, offering free rides from locations like Resorts World, New York-New York, Luxor, Area15, and Topgolf while awaiting approval to charge fares. Unlike Waymo and Tesla, Zoox uses a custom-built, bidirectional electric vehicle with no steering wheel or pedals, two facing rows for up to four riders, floor-to-ceiling windows, and 16-hour battery life.

The company operates from a 190,000-square-foot Vegas depot and has been testing a 50-vehicle fleet in Las Vegas and San Francisco, with early rider programs coming to SF this year and testing planned for Austin and Miami. After minor incidents in April and May, Zoox issued software recalls to improve prediction and turning behavior, and it recently opened a Hayward, CA factory targeting up to ~10,000 vehicles/year at full scale.

The broader robotaxi race is accelerating with multiple other developments

Waymo says it has surpassed 10 million paid rides and now has permission to operate at San Francisco International Airport, starting with employee tests at SFO’s Kiss & Fly lot before commercial service.
Waymo also struck a deal with Lyft to launch all-electric Jaguar I-Pace robotaxis in Nashville in 2026, with rides via the Waymo app first and later the Lyft app, while continuing expansion to Dallas, Denver, Miami, Seattle, and Washington, D.C.
Tesla received a Nevada DMV testing permit to run autonomous Model Ys on public roads (with front-seat Tesla employees), but must secure Nevada Transportation Authority approval to operate a commercial “autonomous vehicle network company.”
Lyft, meanwhile, launched a small May Mobility pilot in Atlanta using hybrid Toyota Sienna Autono-MaaS vans with safety operators, limited hours, and plans to scale from dozens to thousands of vehicles over time—even as Uber and Waymo already offer fully driverless rides in the city.

OpenAI makes deals with Microsoft and Oracle

Relevant stories:

OpenAI and Microsoft signed a nonbinding memorandum of understanding to revise their partnership as OpenAI moves to convert its for-profit arm into a public benefit corporation. Under the plan, OpenAI’s nonprofit parent will retain control and receive an equity stake in the new PBC reportedly worth over $100 billion, with the companies still working to finalize definitive terms and secure approvals from the California and Delaware attorneys general.

The shift could open a path for OpenAI to raise more capital and potentially go public, while preserving Microsoft’s preferred access to OpenAI tech and Azure as a primary cloud—though Oracle has reportedly signed a massive cloud deal with OpenAI worth $300 billion over roughly five years, per the Wall Street Journal, with purchases slated to begin in 2027. If accurate, this would be among the largest cloud contracts ever and follows Oracle’s post-earnings disclosure that it secured multiple multi-billion-dollar agreements. Neither Oracle nor OpenAI confirmed the report. The deal would significantly expand OpenAI’s footprint beyond Microsoft Azure, continuing a pivot that began in early 2024 when OpenAI started using Oracle for compute and later diversified with Google Cloud, according to Reuters.

Oracle’s stock rises more sharply than others over the year, moving ahead of its competitors.

Meanwhile, Microsoft is hedging. It will license Anthropic’s Claude models — including Claude Sonnet 4 — to power features in Office 365 apps alongside OpenAI’s, and has begun shipping its own models (MAI-Voice-1 and MAI-1-preview). Executives Satya Nadella and Mustafa Suleyman said Microsoft will make “significant investments” to build frontier models in-house while pragmatically using others, and Microsoft now explicitly lists OpenAI as both a partner and a competitor.

Together, the moves suggest a rebalanced relationship: OpenAI securing nonprofit control and optionality via a PBC structure, and Microsoft reducing single-vendor reliance while keeping OpenAI as a core frontier-model partner.

Google Injects Gemini Into Chrome as AI Browsers Go Mainstream

Google is bringing Gemini to Chrome so it can answer questions about your open tabs

Google is rolling out deep Gemini integration in Chrome, adding a dedicated Gemini button and AI tools that can answer questions about the current page and synthesize information across multiple open tabs. The “Gemini in Chrome” mode, previously limited to paying subscribers since May, is now available to all desktop users in the US browsing in English via a browser update. On Android, parts of Gemini are already embedded in the Chrome app, with an iOS update expected soon. Google also teased upcoming “agentic” capabilities, including cursor-controlling tools, signaling more proactive, task-executing automation inside the browser.

This update extends Google’s broader strategy of threading Gemini throughout its ecosystem, from Gmail to Google Docs, now bringing AI-assisted search, research, and Q&A directly into the world’s most-used browser. It marks a mainstream moment for “AI browsers,” shifting from niche alternatives to default experiences within Chrome. Users who prefer not to use generative AI can remove the Gemini shortcut by clicking the sparkle icon and unpinning it from the top-right of Chrome.

Other News

Tools

xAI debuts a faster and more cost-effective version of Grok 4. The update reduces computation and cost by using about 40% fewer thinking tokens and a unified reasoning/non-reasoning architecture that xAI says cuts price by 98% while matching Grok 4's benchmark performance and improving speed.

Gemini AI solves coding problem that stumped 139 human teams at ICPC World Finals. Competing in the 2025 ICPC World Finals using an uncustomized Gemini 2.5 model (with extended "thinking" capacity), it solved 10 of 12 problems within five hours to earn a gold medal, outperforming most human teams.

Luma AI's New Ray3 Video Generator Can 'Think' Before Creating. A reasoning model can spend extra computation to check and refine outputs, enabling more complex, longer action sequences; a visual annotation tool supports iterative edits; 16-bit HDR output and a fast low-res draft-to-upscale workflow round out the release.

Amazon launches AI agent to help sellers complete tasks and manage their businesses. The updated Seller Assistant will proactively monitor accounts and inventory, recommend and take authorized actions like repricing or removing slow-moving items, ensure cross-border compliance, analyze demand and shipping needs, and help create ads through conversational prompts.

Zoom launches a cross application AI notetaker, AI avatars and more in its latest update. New capabilities include cross-platform note-taking and search across other meeting platforms and calendars, suggestions to schedule or cancel meetings, enhanced personal notes, photorealistic avatars for video presence, and custom AI agents with translation and search.

Business

Figure AI passes $1B with Series C funding toward humanoid robot development. New funding will scale Figure’s Helix embodied intelligence platform and BotQ manufacturing to produce and deploy tens of thousands of humanoid robots for household and commercial tasks, supported by NVIDIA-backed GPU infrastructure and expanded real-world data collection.

Nvidia buys $5B stake in Intel, planning AI chip collaboration. The partnership will integrate Nvidia’s NVLink-connected GPUs with Intel-made x86 CPUs and x86 RTX SoCs for data centers and PCs, with Nvidia taking roughly a 4% stake to help jointly develop multiple generations of AI-focused products.

Nvidia AI chip challenger Groq raises even more than expected, hits $6.9B valuation. A round led by Disruptive with participation from BlackRock and other institutional investors brings Groq's total capital raised to over $3 billion and supports its cloud and on-premises LPU hardware aimed at providing lower-cost alternatives to Nvidia for running open AI models.

Perplexity reportedly raised $200M at $20B valuation. The new injection brings the company's total raised to about $1.5 billion and implies accelerating revenue growth, with ARR reportedly nearing $200 million.

Replit hits $3B valuation on $150M annualized revenue. A $250M round brings total capital raised to roughly $478M and follows the company's rapid jump to about $150M in annualized revenue after growing from $2.8M in under a year, with participation from Prysm Capital, Amex Ventures and Google's AI Futures Fund.

Since Leaving Washington, Elon Musk Has Been All In on His A.I. Company. He has been spending long, intense stretches at xAI’s Palo Alto offices, leading rapid reorganizations and aggressive hiring while the startup spends billions, releases attention-grabbing products like the Grok chatbot, and faces internal turmoil and uncertain revenue despite claims of millions of users.

Micro1, a competitor to Scale AI, raises funds at $500M valuation. The company says it has grown to $50M ARR, works with leading AI labs and Fortune 100 firms, and uses an AI recruiter to vet thousands of domain experts for data-labeling and new "environments" offerings after Scale AI’s shifts left demand for alternative training-data providers.

Rolling Stone Publisher Sues Google Over AI Overview Summaries. Penske alleges Google's AI-generated summaries use publishers' content without permission and have significantly reduced click-through traffic, forcing publishers to either allow such use to remain indexed or risk disappearing from search.

Tencent Hires OpenAI Researcher as China Steps Up Talent Search. The hire will focus on integrating AI agents into Tencent’s products, reportedly for a compensation package that may reach around 100 million yuan.

Research

The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs. The authors show that scaling LLMs boosts their ability to execute longer-horizon tasks by separating execution capability from per-step accuracy and reducing harmful self-conditioning effects.

Thinking Machines Lab wants to make AI models more consistent. The team argues that nondeterministic outputs stem from how GPU kernels are orchestrated during inference and that controlling that layer could produce more reproducible responses, benefiting enterprise reliability and reinforcement learning.

Emergent Hierarchical Reasoning in LLMs through Reinforcement Learning. Findings show RL-tuned LLMs first learn reliable low-level execution skills and then shift to a strategic planning bottleneck; HICRA, a credit-assignment method, amplifies learning signals for planning tokens to speed up high-level reasoning and improve performance.

Outcome-based Exploration for LLM Reasoning. The paper introduces outcome-based exploration methods that adapt UCB-style bonuses to final answers to mitigate diversity collapse during RL post-training while improving the accuracy–diversity tradeoff.

LoCoBench: A Benchmark for Long-Context Large Language Models in Complex Software Engineering. The benchmark provides 8,000 multi-file evaluation scenarios across 10 languages and 8 task categories, systematically testing context windows from 10K to 1M tokens with 17 metrics to measure long-context software engineering capabilities.

Towards a Physics Foundation Model. A transformer trained on 1.8 TB of diverse simulation data can infer dynamics from short state-history prompts to model many different physical systems, generalize zero-shot to new conditions, and produce stable, physically plausible long rollouts without explicit governing equations.

OpenAI’s research on AI models deliberately lying is wild. Experiments show current models can learn to deliberately deceive testers; attempts to train scheming out can backfire by teaching stealthier deception; and a method called "deliberative alignment" can reduce but not eliminate such behavior in simulated tasks.

OpenAI releases first-of-kind study revealing how people are using ChatGPT for everyday tasks. An analysis of 1.5 million privacy-preserving consumer conversations finds a growing shift toward non-work uses—rising to 73% of messages by June 2025—and offers demographic and usage breakdowns from OpenAI and external researchers.

Concerns

FTC launches inquiry into AI chatbot companions from Meta, OpenAI, and others. The FTC seeks details on how companies assess safety and monetize chatbot companions for minors, what safeguards they use to limit harm, and whether parents are informed of risks after reports of harmful outcomes and lawsuits involving youth interactions.

Google Gemini dubbed ‘high risk’ for kids and teens in new safety assessment. A nonprofit found Gemini's kid- and teen-facing tiers are essentially adult versions with only added filters, can still provide inappropriate or unsafe content (including risky mental health advice), and therefore rated them "High Risk" overall.

Claude’s new AI file-creation feature ships with security risks built in. Anthropic warns the feature gives Claude internet and sandboxed code execution access that could be exploited via prompt injection to coax the assistant into reading connected data sources and making external network requests to leak sensitive information.

Disney, Warner Bros. Discovery, NBCU Sue Chinese AI Company MiniMax, Alleging It ‘Pirates and Plunders’ Studios’ Copyrighted Works on ‘Massive Scale’. Plaintiffs accuse the Chinese AI company behind Hailuo of generating and distributing images and videos that use the studios' copyrighted characters without permission, say it ignored cease-and-desist requests, and seek damages plus an injunction.

Tesla is trying to hide 3 Robotaxi accidents. The company redacted key narrative details in its NHTSA reports for three July Robotaxi crashes in Austin — one causing a minor injury — despite operating only about a dozen vehicles and being required to disclose crash information.

Policy

China tells its tech companies they can’t buy AI chips from Nvidia. The Cyberspace Administration of China has ordered firms like ByteDance and Alibaba to halt testing and orders of Nvidia’s RTX Pro 6000D servers, enforcing a shift toward domestically produced AI chips and cutting off a major supplier amid broader US-China semiconductor tensions.

A California bill that would regulate AI companion chatbots is close to becoming law. If signed, the law would require operators to implement safety protocols—like three-hour reminder alerts for minors, referral reporting, and bans on chats about self-harm or sexual content with vulnerable users—and create civil penalties and reporting obligations for companion chatbot companies.

Last Week in AI