Category: News

OpenAI announced the acquisition of Statsig, a Seattle-based product experimentation platform, for $1.1 billion in an all-stock deal, marking one of its largest acquisitions to date. This strategic move, often described as an “acqui-hire,” brings Statsig’s founder and CEO, Vijaye Raji, into OpenAI as the new CTO of Applications, reporting to Fidji Simo, the former Instacart CEO who now leads OpenAI’s Applications division. The deal, pending regulatory approval, aims to enhance OpenAI’s ability to rapidly develop and deploy AI products like ChatGPT and Codex by integrating Statsig’s expertise in A/B testing, feature flagging, and real-time analytics.

Statsig, founded in 2021 by Raji, a former Meta engineering leader, has powered experimentation for companies like Atlassian, Notion, and OpenAI itself. Its platform enables data-driven product iteration, allowing teams to test features and make real-time decisions, which is critical in the fast-paced AI development landscape. By bringing Statsig in-house, OpenAI seeks to streamline its feedback loops, ensuring safer and more effective rollouts of AI features. Raji’s role will involve overseeing product engineering for ChatGPT and Codex, focusing on infrastructure and system integrity to deliver reliable, scalable applications.

The acquisition aligns with OpenAI’s shift from a research-focused lab to a product-driven powerhouse, competing with rivals like Google and Microsoft. Statsig’s tools address a key challenge in AI: transforming cutting-edge research into user-friendly, safe products. As AI-generated features proliferate, Statsig’s experimentation framework helps validate what works, reducing the risk of deploying untested capabilities. This move is seen as a response to the need for disciplined product development in an era where AI can generate countless feature variants but requires rigorous testing to ensure quality.

Alongside the acquisition, OpenAI announced leadership changes. Chief Product Officer Kevin Weil will transition to VP of AI for Science, focusing on AI-driven scientific discovery, while Srinivas Narayanan, former head of engineering, becomes CTO of B2B Applications, reporting to COO Brad Lightcap. These shifts reflect OpenAI’s strategic restructuring to balance research and product innovation. Statsig’s 150+ employees will join OpenAI but continue operating independently from Seattle, ensuring continuity for existing customers like Bloomberg and Figma.

The $1.1 billion valuation matches Statsig’s May 2025 funding round, suggesting investors are banking on OpenAI’s soaring stock value, with the company reportedly eyeing a $500 billion valuation. While some analysts view this as a brilliant move to secure first-mover advantages in AI product development, others note that Statsig’s integration must be carefully managed to avoid disrupting its customer base. This acquisition underscores the growing importance of experimentation in AI, positioning OpenAI to deliver more reliable, user-focused applications in a fiercely competitive market.

BMW Group and Qualcomm Technologies announced a groundbreaking collaboration to introduce the Snapdragon Ride Pilot, an advanced driver-assistance system (ADAS) set to debut in the all-new BMW iX3 electric SUV. This system, built on Qualcomm’s Snapdragon Ride system-on-chips (SoCs) and a co-developed software stack, enables hands-free highway driving, automatic lane changes, and parking assistance, marking a significant step in BMW’s push toward semi-autonomous driving. The technology, validated for use in over 60 countries with plans to expand to more than 100 by 2026, is now available to global automakers and Tier-1 suppliers, positioning Qualcomm as a key player in the growing ADAS market.

The Snapdragon Ride Pilot supports Level 2+ automation, allowing drivers to take their hands off the wheel on approved motorways while remaining responsible for supervision. It uses advanced camera systems, radar, and AI to provide a 360-degree view of the vehicle’s surroundings, enabling features like contextual lane changes triggered by subtle driver cues, such as mirror glances. The system also incorporates vehicle-to-everything (V2X) communication via Qualcomm’s V2X 200 chipset, allowing the car to interact with other vehicles and infrastructure to enhance safety. This “superbrain” computer, as described by BMW, evolves through a cloud-based data flywheel, using anonymous fleet data to refine AI models via over-the-air updates, improving performance over time.

This three-year collaboration, involving 1,400 specialists across Germany, the U.S., Sweden, and BMW’s Autonomous Driving Test Center in the Czech Republic, integrates Qualcomm’s computing expertise with BMW’s automotive engineering. The system meets the highest Automotive Safety Integrity Levels (ASIL) and includes cybersecurity measures to protect against threats. While not fully autonomous (Level 5), it competes with systems like Tesla’s Full Self-Driving, GM’s Super Cruise, and Ford’s BlueCruise, offering a balance of convenience and safety. However, safety concerns persist, as studies suggest drivers may over-rely on such systems, potentially increasing crash risks.

The debut in the BMW iX3, part of BMW’s Neue Klasse platform, aligns with the automaker’s vision of combining electrification, connectivity, and automation. Qualcomm’s automotive revenue, which reached $2.9 billion in fiscal 2024, is projected to hit $8 billion by 2029, reflecting its strategic shift from smartphones to automotive tech. The partnership not only strengthens BMW’s position in the premium EV market but also signals a broader industry trend of tech-automotive collaborations to meet consumer demand for intelligent, safe vehicles. As BMW rolls out this technology, it remains committed to driver engagement, ensuring its cars remain “fun to drive” even as automation advances.

Cognition AI, the San Francisco-based startup behind the AI-powered coding assistant Devin, has raised $400 million in a funding round that values the company at $10.2 billion, more than doubling its $4 billion valuation from March 2025. Announced on September 8, 2025, the round was led by Founders Fund, with participation from existing investors like Lux Capital, 8VC, Elad Gil, Definition Capital, and Swish Ventures. This significant capital infusion highlights the intense investor enthusiasm for AI-driven software development tools, despite market volatility and regulatory scrutiny.

Cognition’s rapid ascent is driven by the explosive growth of Devin, an AI agent designed to autonomously write, debug, and deploy code. The startup’s annual recurring revenue (ARR) surged from $1 million in September 2024 to $73 million by June 2025, a 73x increase in less than a year. This growth, coupled with a disciplined net burn rate under $20 million since its 2023 founding, underscores Cognition’s capital efficiency. Major clients, including Dell Technologies, Cisco Systems, Goldman Sachs, and MongoDB, have adopted Devin for tasks like bug scanning and project automation, signaling strong enterprise demand.

A pivotal factor in Cognition’s growth was its July 2025 acquisition of Windsurf, a rival AI coding startup, shortly after Google acquired Windsurf’s leadership team and licensed its technology for $2.4 billion. The acquisition doubled Cognition’s ARR by integrating Windsurf’s developer tools and enterprise customer base, with less than 5% overlap in clientele. “Combining Devin’s rapid adoption with Windsurf’s IDE product has been a massive unlock,” said CEO Scott Wu, emphasizing the strategic move’s impact on scaling enterprise offerings.

The funding will fuel Cognition’s expansion of its go-to-market and engineering teams, alongside the development of new AI models tailored for programming tasks. Devin’s enterprise edition, which offers customized versions with enhanced cybersecurity controls, is poised to capture a larger share of the market. However, the company faces challenges, including a high-pressure work culture—evidenced by 80-hour work weeks and recent layoffs of 30 staffers alongside buyout offers to 200 employees. Critics also note that Devin, while promising, sometimes struggles with complex real-world tasks, requiring human oversight.

The broader AI coding sector is fiercely competitive, with players like Microsoft’s GitHub Copilot, Google’s Gemini Code Assist, and Amazon’s CodeWhisperer vying for dominance. Yet, Cognition’s valuation—29 times its prior fundraising multiple—reflects investor confidence in its potential to disrupt traditional software engineering. As AI agents evolve, Cognition’s focus on automation could reshape workflows, though regulatory pressures on AI ethics and competition from open-source alternatives pose risks. With this $400 million raise, Cognition is well-positioned to accelerate its roadmap, setting a benchmark for AI startups navigating a dynamic landscape.

On September 8, 2025, Nebius Group, an Amsterdam-based AI infrastructure company, announced a landmark $17.4 billion deal to supply Microsoft with GPU infrastructure capacity over a five-year term, with the potential to reach $19.4 billion if Microsoft opts for additional services. This agreement, one of the largest in the AI infrastructure sector, sent Nebius shares soaring by over 47% in after-hours trading, with some reports noting a peak surge of 60%. The deal underscores the surging demand for high-performance computing resources as tech giants race to bolster their AI capabilities.

Nebius, which emerged from a 2023 split of Russian tech giant Yandex, will provide Microsoft with dedicated GPU capacity from its new data center in Vineland, New Jersey, starting later in 2025. The company specializes in delivering Nvidia GPUs and AI cloud services, offering computing, storage, and tools for AI developers to build and run models. The agreement is expected to significantly accelerate Nebius’s AI cloud business, with CEO Arkady Volozh stating, “The economics of the deal are attractive, but significantly, it will help us accelerate growth in 2026 and beyond.” Nebius plans to finance the data center’s construction and chip purchases using cash flow from the deal and potential debt issuance.

This partnership highlights the intense competition for AI infrastructure, with Microsoft, a key player in AI through its Azure platform and OpenAI partnership, seeking to secure robust computing resources. Unlike its competitor CoreWeave, which saw a modest 5% stock increase, Nebius’s deal positions it as a major player in the AI supply chain. The contract’s scale—exceeding Nebius’s projected 2025 revenue of $625 million by nearly three times annually—has fueled optimism among investors, with some on X predicting a potential 10x growth for Nebius due to ongoing GPU supply constraints.

However, the deal also raises questions about market dynamics. Microsoft, previously CoreWeave’s largest customer, had not signed a comparable long-term contract with them, and CoreWeave denied earlier reports of contract cancellations. Nebius’s agreement could signal a strategic shift for Microsoft, possibly diversifying its AI infrastructure partners. Meanwhile, Nebius’s U.S. expansion, with offices in San Francisco, Dallas, and New York, and its doubled market cap of $15 billion in 2025, reflect its aggressive push into the American market.

Despite the enthusiasm, some analysts remain cautious. Posts on X and reports note that while Nebius benefits from current GPU shortages, its long-term viability depends on innovation beyond merely supplying Nvidia chips. The company’s additional ventures in autonomous driving (Avride) and edtech (TripleTen) may diversify its portfolio, but financial health concerns, including declining revenues and weak cash flow, persist. As the AI infrastructure race intensifies, Nebius’s ability to capitalize on this deal and secure further contracts will be critical to sustaining its meteoric rise.

Researchers at the University of Florida (UF) have unveiled a groundbreaking silicon photonic chip that leverages light instead of electricity to perform complex artificial intelligence (AI) tasks, achieving up to 100 times greater power efficiency than traditional electronic processors. Published on September 8, 2025, in Advanced Photonics, this innovation marks a significant step toward sustainable AI computing, addressing the escalating energy demands of modern machine learning models.

The chip, developed by a team led by Volker J. Sorger, the Rhines Endowed Professor in Semiconductor Photonics at UF, focuses on convolution operations—a core component of AI tasks like image recognition, video processing, and language analysis. These operations are notoriously power-intensive, straining global power grids as AI applications proliferate. By integrating optical components, such as lasers and microscopic Fresnel lenses, onto a silicon chip, the team has created a system that performs convolutions with near-zero energy consumption and significantly faster processing speeds. Tests demonstrated the chip’s ability to classify handwritten digits with 98% accuracy, matching the performance of conventional electronic chips while drastically reducing power usage.

A key advantage of this photonic chip is its use of wavelength multiplexing, allowing multiple data streams to be processed simultaneously using different colors of light. “We can have multiple wavelengths, or colors, of light passing through the lens at the same time,” said Hangbo Yang, a research associate professor and co-author of the study. This capability enhances data throughput and efficiency, making the chip ideal for high-demand applications like autonomous vehicles, healthcare diagnostics, and telecommunications. The chip’s design, built using standard semiconductor manufacturing techniques, ensures scalability and potential integration with existing AI systems, such as those by NVIDIA, which already incorporate optical elements.

The research, conducted in collaboration with the Florida Semiconductor Institute, UCLA, and George Washington University, addresses a critical challenge in AI: the unsustainable energy consumption of traditional electronic chips. As AI models grow more complex, they push conventional hardware to its limits, with data centers projected to consume vast amounts of electricity by 2026. The UF team’s photonic chip offers a solution by performing computations at the speed of light, reducing both power demands and heat generation. “Performing a key machine learning computation at near zero energy is a leap forward for future AI systems,” Sorger noted, emphasizing its potential to scale AI capabilities sustainably.

While the chip represents a major advancement, challenges remain, including integrating photonic systems with existing electronic infrastructure and scaling the technology for broader applications. However, its compatibility with commercial foundry processes suggests a viable path to market. As Sorger predicts, “In the near future, chip-based optics will become a key part of every AI chip we use daily.” This breakthrough not only paves the way for more efficient AI but also signals a paradigm shift toward optical computing, promising a greener, faster future for technology.

On September 5, 2025, authors Grady Hendrix and Jennifer Roberson filed a proposed class-action lawsuit against Apple in the U.S. District Court for the Northern District of California, accusing the company of illegally using their copyrighted books to train its “OpenELM” large language models. The lawsuit alleges that Apple relied on a dataset of pirated books, specifically from “shadow libraries” like Books3, without consent, credit, or compensation, violating intellectual property rights. The plaintiffs claim their works were included in this dataset, and Apple’s actions undermine authors’ rights by creating AI outputs that compete with original works. The lawsuit seeks damages, restitution, and potentially the destruction of models trained on pirated content, with one source estimating damages at $2.5 billion.

This case is part of a broader wave of legal actions against tech companies, including Microsoft, Meta, and OpenAI, for similar misuse of copyrighted materials in AI training. Notably, Anthropic recently settled a related lawsuit for $1.5 billion, described as the largest copyright recovery to date, setting a precedent that could influence Apple’s case. Apple has not publicly responded, but it may argue fair use or technical distinctions about its OpenELM models. The outcome could shape AI development and copyright law, especially as Apple pushes its AI initiatives, including an overhaul of Siri.

Language models, despite their remarkable advancements, often produce hallucinations—plausible but false statements delivered with unwarranted confidence. A recent OpenAI research paper, published on September 5, 2025, delves into why these errors persist and how current evaluation practices inadvertently exacerbate the issue. This article explores the key findings, shedding light on the mechanisms behind hallucinations and proposing solutions to mitigate them.

Hallucinations occur when models generate incorrect answers to seemingly straightforward questions. For instance, when queried about a person’s birthday or the title of a PhD dissertation, a model might confidently provide multiple incorrect responses. This stems from the way models are trained and evaluated. Unlike spelling or syntax, which follow consistent patterns, factual details like birthdays are often arbitrary and lack predictable structures in training data. This randomness makes it nearly impossible for models to avoid errors entirely, especially for low-frequency facts.

The root of the problem lies in pretraining, where models learn by predicting the next word in vast text corpora. Without explicit “true/false” labels, models cannot easily distinguish valid from invalid statements. They rely on patterns in fluent language, which works well for consistent elements like grammar but falters for specific, unpredictable facts. As a result, hallucinations emerge when models attempt to fill in gaps with plausible guesses rather than admitting uncertainty.

Current evaluation methods further aggravate this issue. Most benchmarks prioritize accuracy—rewarding correct answers while ignoring whether a model guesses or abstains when uncertain. This setup mirrors a multiple-choice test where guessing might yield points, but admitting “I don’t know” scores zero. For example, the SimpleQA evaluation shows that models like OpenAI’s o4-mini achieve slightly higher accuracy (24%) than gpt-5-thinking-mini (22%) but have a significantly higher error rate (75% vs. 26%) due to excessive guessing. This incentivizes models to prioritize lucky guesses over cautious abstention, undermining humility—a core value at OpenAI.

To address this, the paper proposes rethinking evaluation metrics. Instead of focusing solely on accuracy, scoreboards should penalize confident errors more heavily than expressions of uncertainty. Partial credit for abstaining or acknowledging uncertainty could discourage blind guessing, aligning model behavior with real-world reliability. This approach draws inspiration from standardized tests that use negative marking for wrong answers, a practice that could be adapted to AI evaluations.

The paper also dispels common misconceptions. Hallucinations are not inevitable; models can abstain when uncertain. Nor are they exclusive to smaller models—larger models, despite knowing more, may struggle to gauge their confidence accurately. Most critically, achieving 100% accuracy is unrealistic, as some questions are inherently unanswerable due to missing information or ambiguity. Simply adding hallucination-specific evaluations is insufficient; primary metrics across all benchmarks must reward calibrated responses.

OpenAI’s latest models, including GPT-5, show reduced hallucination rates, particularly when reasoning, but the challenge persists. By refining evaluation practices and prioritizing uncertainty-aware metrics, the AI community can foster models that balance accuracy with humility, ultimately making them more reliable for real-world applications.

Reports confirm OpenAI is advancing its first custom AI chip, focused on inference (running trained models for predictions and decisions), in collaboration with Broadcom for design and intellectual property (IP) and TSMC for manufacturing on a 3nm process node. Mass production is targeted for 2026, aligning with the details in your query. The project is led by former Google engineer Richard Ho, who heads a team of about 40 specialists, many with experience from Google’s Tensor Processing Units (TPUs). This initiative aims to reduce OpenAI’s heavy reliance on Nvidia GPUs, which dominate the AI hardware market but face shortages and high costs.

Key Developments from Recent Reports (September 2025)

• Partnership Confirmation and $10B Deal: On September 5, 2025, the Financial Times and Reuters reported that OpenAI is finalizing the chip design in the coming months, with Broadcom providing engineering support and TSMC handling fabrication. Broadcom’s CEO Hock Tan disclosed a $10 billion order from a new AI client (widely identified as OpenAI) during an earnings call, boosting Broadcom’s AI revenue projections for fiscal 2026. This deal focuses on custom “XPUs” (AI processors) for internal use, not commercial sale, emphasizing inference workloads with potential for scaled training. OpenAI has scaled back earlier ambitions to build its own foundries due to costs exceeding hundreds of millions per iteration, opting instead for this partnership model.
• Team and Technical Specs: Led by Richard Ho (ex-Google TPU head), the team includes engineers like Thomas Norrie. The chip features a systolic array architecture (similar to Google’s TPUs for efficient matrix computations), high-bandwidth memory (HBM, possibly HBM3E or HBM4), and integrated networking. It’s optimized for OpenAI’s models like GPT-4 and beyond, with initial small-scale deployment for inference to test viability. Analysts note risks, including potential delays or underperformance on the first tape-out (design finalization for production), as seen in other custom chip efforts by Microsoft and Meta.
• Market Impact: Broadcom shares surged over 10% on September 5, reaching a $1.7 trillion market cap, while Nvidia and AMD dipped ~2-3% amid concerns over custom silicon eroding Nvidia’s 80%+ market share. HSBC analysts predict the custom AI chip market could surpass Nvidia’s GPU business by 2026. OpenAI’s move ties into broader AI infrastructure pushes, including the $500B Stargate project (with Oracle) and collaborations like Microsoft’s Maia chips.

Broader Context and Challenges

OpenAI’s compute costs are massive—projected $5B loss in 2024 on $3.7B revenue—driving this diversification. The company is also integrating AMD’s MI300X chips via Azure for training, complementing Nvidia. Geopolitical risks (e.g., TSMC’s Taiwan base) and high development costs (~$500M+ per chip version, plus software) loom, but success could enhance bargaining power and efficiency. No official OpenAI statement yet, but industry sources indicate tape-out soon, with prototypes possible by late 2025.

This positions OpenAI alongside Google, Amazon, and Meta in the custom silicon race, potentially reshaping AI hardware dynamics. Updates could emerge from upcoming tech conferences or earnings.

Qwen3-Max-Preview is the preview release of Alibaba’s Qwen3-Max, the flagship model in the Qwen3 series developed by Alibaba Cloud’s Qwen team. It’s a massive Mixture-of-Experts (MoE) large language model with over 1 trillion parameters, designed for advanced reasoning, instruction following, and multimodal tasks. Key features include:

• Improvements over prior versions: Major gains in math, coding, logic, science accuracy; better multilingual support (100+ languages, including strong Chinese/English handling); reduced hallucinations; higher-quality open-ended responses for Q&A, writing, and conversation.
• Optimizations: Excels in retrieval-augmented generation (RAG), tool calling, and long-context understanding (up to 256K tokens, extendable to 1M). It lacks a dedicated “thinking” mode but focuses on efficient, reliable outputs.
• Architecture: Built on Qwen3’s MoE framework, pretrained on trillions of tokens with Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF). It’s positioned as a high-capacity model for complex, multi-step tasks, competing with top closed-source LLMs like GPT-4 or Claude 3.5.

This preview allows early testing before full release, emphasizing production usability over experimental features.

News: Now Live on OpenRouter

As of September 5, 2025, Qwen3-Max-Preview became available on OpenRouter, a unified API platform for 400+ AI models. Alibaba’s official Qwen account confirmed the launch, highlighting its strengths in reasoning and tool use. OpenRouter integration enables easy access via OpenAI-compatible APIs, with token-based pricing (e.g., tiered by input/output length; specifics vary by provider but start low for previews). Users can route requests through OpenRouter for vendor-agnostic setups, avoiding lock-in.

• Access Details: Available at openrouter.ai/models (search “Qwen3-Max”) or directly via API endpoint. Free tiers may have limits; paid starts at ~$1.60/M input tokens. It’s also accessible via Qwen Chat (interactive UI) and Alibaba Cloud (enterprise IAM).
• Community Buzz: Early X posts praise its potential for coding/programming (e.g., “saves my programmer life?”), with calls for benchmarks. No major issues reported yet, but expect high compute costs due to scale.

This rollout positions Qwen3-Max-Preview as a key player in the open-weight AI race, with full Qwen3 updates (e.g., thinking modes) expected soon.

As of September 6, 2025, there’s active speculation in the AI community that Google’s upcoming Gemini 3.0 model (or an early version of it) has been quietly released on OpenRouter under disguised names. OpenRouter, a platform aggregating access to hundreds of AI models via a unified API, announced two new “stealth models” yesterday: Sonoma Alpha and Sonoma Sky Alpha (also referred to as Sonoma Dusk Alpha in some posts). These are free to use, support a massive 2 million token context window, and are described as “maximally intelligent” with prompts logged by the creator for training—features that align closely with expected Gemini 3 specs.

Key Details from the Announcement and Speculation

• Announcement: OpenRouter posted about these models on September 5, 2025, calling them “stealth” (implying anonymity to avoid direct attribution). They emphasize high intelligence, 2M context (double the 1M seen in Gemini 2.5 Pro), and free access, but note that the provider handles logging for improvement.
• Why Gemini 3.0?
  • Leaks from July-August 2025 referenced “gemini-beta-3.0-pro” and “gemini-beta-3.0-flash” in Google’s internal code (e.g., Gemini CLI repo), hinting at variants with enhanced reasoning (“Deep Think”) and multi-million token contexts—matching the 2M here.
  • Community tests and posts suggest strong performance in reasoning, speed, and multimodal tasks, outperforming current Gemini 2.5 models but falling short of full Gemini-like polish in some outputs (e.g., one user called it “disappointing” compared to known Gemini quality).
  • The “Sonoma” naming (evoking California’s wine country, near Google’s HQ) fuels the theory, as does the free tier—Google has previously offered experimental Gemini models for free on OpenRouter to gather data (e.g., Gemini 2.5 Pro Experimental in March 2025).
• Alternative Theories: Not everyone agrees—some speculate it’s an xAI Grok variant (due to “maximally intelligent” phrasing echoing xAI’s ethos) or a new Chinese model. However, the 2M context and free logging point more toward Google testing pre-release.

Broader Gemini 3.0 Context

Google hasn’t officially announced Gemini 3.0, but rumors from mid-2025 predict a late 2025 release (preview in December, full in early 2026), building on Gemini 2.5’s “thinking” mode with:

• Trillion-parameter scale for superior reasoning in code, math, and multimodality (text, images, video, 3D).
• Integrated self-correction to reduce hallucinations.
• On-device variants like Gemini Nano 3 for Pixels.

These stealth models could be betas, allowing Google to benchmark against rivals like GPT-5 or Grok 4 without fanfare. OpenRouter’s history (e.g., hosting “Quasar Alpha” in April 2025, speculated as GPT-5/Gemini 3) supports this pattern of anonymous drops.