Energy & Technology

Ai2 представила Olmo 3: Відкриті моделі, що конкурують з Meta, DeepSeek та іншими

The Allen Institute for AI (Ai2) представила нову генерацію своїх флагманських великих мовних моделей, розроблені для більш прямої конкуренції з галузевими та академічними лідерами. Seattle-based некомерційна організація представила Olmo 3, колекцію відкритих мовних моделей, яка, за заявами, перевершує повністю відкриті моделі, такі як Stanford’s Marin та комерційні відкриті моделі, як Meta’s Llama 3. Раніше версії Olmo були в основному призначені для наукових цілей, для розуміння того, як будуються AI моделі. З Olmo 3, Ai2 розширює свій фокус, позиціонуючи моделі як потужні, ефективні та прозорі системи, придатні для використання у реальному світі, включаючи комерційні застосування.

«Olmo 3 доводить, що відкритість та продуктивність можуть розвиватися разом», – сказав Ali Farhadi, CEO Ai2, у прес-релізі, опублікованому в четвер, що оголошував про нові моделі.

Це частина ширшої еволюції в AI світі. Протягом останнього року все більш потужні відкриті моделі з компаній та університетів — включаючи Meta, DeepSeek, Qwen, та Stanford — почали конкурувати з продуктивністю власних систем великих технологічних компаній.

Багато з останніх відкритих моделей розроблені для демонстрації їхнього міркування крок за кроком — часто звані «моделі, що «думлять»», що стало ключовим критерієм у цій галузі.

Ai2 випускає Olmo 3 у кількох версіях: Olmo 3 Base (основний базовий фундамент); Olmo 3 Instruct (настроєний для дотримання інструкцій користувача); Olmo 3 Think (розроблений для демонстрації більш явного міркування); та Olmo 3 RL Zero (експериментальна модель, навчена за допомогою навчання з підкріпленням).

Відкриті моделі набирають обертів завдяки стартапам та бізнесу, які хочуть більше контролю над витратами та даними, а також чіткішу видимість того, як працює технологія. Ai2 йде далі, випускаючи повний «потік моделі» позаду Olmo 3 — серію знімків, що показує, як модель прогресувала на кожному етапі навчання. Крім того, оновлений OlmoTrace інструмент дозволить дослідникам пов’язувати міркування моделі назад до конкретних даних та рішень щодо навчання, які вплинули на них.

Що стосується енергії та ефективності витрат, Ai2 заявляє, що новий базовий Olmo 3 в 2,5 рази більш ефективний для навчання, ніж Meta’s Llama 3.1 (на основі GPU-годин на токен, порівнюючи Olmo 3 Base з Meta’s 8B post-trained modelem), та навчався на значно менше токенів, в деяких випадках в шість разів менше, ніж у аналогічних моделях.

Крім того, Olmo 3 може читати або аналізувати набагато довші документи одночасно, з підтримкою вхідних даних до 65 000 токенів, що приблизно дорівнює довжині розділу короткої книги.

Заснована в 2014 році пізнім засновник Microsoft Paul Allen, Ai2 протягом тривалого часу функціонувала як дослідницька некомерційна організація, розробляючи відкриті інструменти та моделі, тоді як більші комерційні лабораторії домінували в центрі уваги. Інститут зробив серію кроків цього року, щоб підняти свій профіль, зберігаючи при цьому свою місію розвитку AI для вирішення найбільших проблем світу. В серпні Ai2 було обрано Національним Науковим Фондом та Nvidia для ініціативи вартістю 152 мільйони доларів для створення повністю відкритих багатомодальних AI моделей для наукових досліджень, позиціонуючи інститут для того, щоб бути ключовим учасником національного AI-скелету. Він також є ключовим технічним партнером для Cancer AI Alliance, допомагаючи Fred Hutch та іншим провідним центрам раку в США навчати AI-моделі на клінічних даних, не розкриваючи пацієнтських записів.

Olmo 3 доступний зараз на Hugging Face та Ai2’s model playground.

Amazon’s Surprise Indie Hit: Kiro Launches Broadly in Bid to Reshape AI-Powered Software Development

Can the software development hero conquer the “AI Slop Monster” to uncover the gleaming, fully functional robot buried beneath the coding chaos?

That was the storyline unfolding inside a darkened studio at Seattle Center last week, as Amazon’s Kiro software development system was brought to life for a promotional video. Instead of product diagrams or keynote slides, a crew from Seattle’s Packrat creative studio used action figures on a miniature set to create a stop-motion sequence. In this tiny dramatic scene, Kiro’s ghost mascot played the role that the product aims to fill in real life — a stabilizing force that brings structure and clarity to AI-assisted software development.

No, this is not your typical Amazon Web Services product launch.Kiro (pronounced KEE-ro) is Amazon’s effort to rethink how developers use AI. It’s an integrated development environment that attempts to tame the wild world of vibe coding, the increasingly popular technique that creates working apps and websites from natural language prompts.

But rather than simply generating code from prompts, Kiro breaks down requests into formal specifications, design documents, and task lists. This spec-driven development approach aims to solve a fundamental problem with vibe coding: AI can quickly generate prototypes, but without structure or documentation, that code becomes unmaintainable.

It’s part of Amazon’s push into AI-powered software development, expanding beyond its AWS Code Whisperer tool to compete more aggressively against rivals such as Microsoft’s GitHub Copilot, Google Gemini Code Assist, and open-source AI coding assistants.

The market for AI-powered development tools is booming. Gartner expects AI code assistants to become ubiquitous, forecasting that 90% of enterprise software engineers will use them by 2028, up from less than 14% in early 2024. A July 2025 report from Market.us projects the AI code assistant market will grow from $5.5 billion in 2024 to $47.3 billion by 2034.

Amazon launched Kiro in preview in July, to a strong response. Positive early reviews were tempered by frustration from users unable to gain access. Capacity constraints have since been resolved, and Amazon says more than 250,000 developers used Kiro in the first three months.

The internet is “full of prototypes that were built with AI,” said Deepak Singh, Amazon’s vice president of developer agents and experiences, in an interview last week. The problem, he explained, is that if a developer returns to that code two months later, or hands it to a teammate, “they have absolutely no idea what prompts led to that. It’s gone.”

Kiro solves that problem by offering two distinct modes of working. In addition to “vibe mode,” where they can quickly prototype an idea, Kiro has a more structured “spec mode,” with formal specifications, design documents, and task lists that capture what the software is meant to do.

Now, the company is taking Kiro out of preview into general availability, rolling out new features and opening the tool more broadly to development teams and companies.

As a product of Amazon’s cloud division, Kiro is unusual in that it’s relevant well beyond the world of AWS. It works across languages, frameworks, and deployment environments. Developers can build in JavaScript, Python, Go, or other languages and run applications anywhere — on AWS, other cloud platforms, on-premises, or locally.

That flexibility and broader reach are key reasons Amazon gave Kiro a standalone brand rather than presenting it under the AWS or Amazon umbrella. It was a “very different and intentional approach,” said Julia White, AWS chief marketing officer, in an interview at the video shoot. The idea was to defy the assumptions that come with the AWS name, including the idea that Amazon’s tools are built primarily for its own cloud.

White, a former Microsoft and SAP executive who joined AWS as chief marketing officer a year ago, has been working on the division’s fundamental brand strategy and calls Kiro a “wonderful test bed for how far we can push it.” She said those lessons are starting to surface elsewhere across AWS as the organization looks to “get back to that core of our soul.”

With developers, White said, “you have to be incredibly authentic, you need to be interesting. You need to have a point of view, and you can never be boring.” That philosophy led to the fun, quirky, and irreverent approach behind Kiro’s ghost mascot and independent branding. The marketing strategy for Kiro caused some internal hesitation, White recalled. People inside the company wondered whether they could really push things that far.

Her answer was emphatic: “Yep, yep, we can. Let’s do it.”

Amazon’s Kiro has caused a minor stir in Seattle media circles, where the KIRO radio and TV stations, pronounced like Cairo, have used the same four letters stretching back into the last century. People at the stations were not exactly thrilled by Amazon’s naming choice. With its core audience of developers, however, the product has struck a nerve in a positive way. During the preview period, Kiro handled more than 300 million requests and processed trillions of tokens as developers explored its capabilities, according to stats provided by the company. Rackspace used Kiro to complete what they estimated as 52 weeks of software modernization in three weeks, according to Amazon executives. SmugMug and Flickr are among other companies espousing the virtues of Kiro’s spec-driven development approach. Early users are posting in glowing terms about the efficiencies they’re seeing from adopting the tool.

Kiro uses a tiered pricing model based on monthly credits: a free plan with 50 credits, a Pro plan at $20 per user per month with 1,000 credits, a Pro+ plan at $40 with 2,000 credits, and a Power tier at $200 with 10,000 credits, each with pay-per-use overages. With the move to general availability, Amazon says teams can now manage Kiro centrally through AWS IAM Identity Center, and startups in most countries can apply for up to 100 free Pro+ seats for a year’s worth of Kiro credits.

New features include property-based testing — a way to verify that generated code actually does what developers specified — and a new command-line interface in the terminal, the text-based workspace many programmers use to run and test their code. A new checkpointing system lets developers roll back changes or retrace an agent’s steps when an idea goes sideways, serving as a practical safeguard for AI-assisted coding.

Amit Patel, director of software engineering for Kiro, said the team itself is deliberately small — a classic Amazon “two-pizza team.” And yes, they’ve been using Kiro to build Kiro, which has allowed them to move much faster. Patel pointed to a complex cross-platform notification feature that had been estimated to take four weeks of research and development. Using Kiro, one engineer prototyped it the next day and shipped the production-ready version in a day and a half.

Patel said this reflects the larger acceleration of software development in recent years. “The amount of change,” he said, “has been more than I’ve experienced in the last three decades.”

Microsoft’s ‘Superfactory’: A New Era of AI Infrastructure

Microsoft has unveiled a groundbreaking approach to data center design and operation, dubbed its ‘superfactory,’ focused on facilitating the training and deployment of advanced artificial intelligence models. This innovative system links massive data centers across vast distances – in this case, Wisconsin and Atlanta, approximately 700 miles apart – via a high-speed fiber-optic network.

The ‘superfactory’ represents a shift from traditional cloud data centers, which cater to numerous separate applications, to a unified architecture specifically engineered for single, massive AI workloads. Each facility incorporates hundreds of thousands of Nvidia GPUs connected through an AI Wide Area Network (AI-WAN) for real-time sharing of computing tasks.

Microsoft’s new two-story data center design maximizes GPU density and minimizes latency, aided by a closed-loop liquid cooling system. By pooling computing capacity across multiple sites and dynamically redirecting workloads, the system distributes power requirements efficiently across the grid.

This interconnected infrastructure will be utilized to train and run next-generation AI models for key partners, including OpenAI, and Microsoft’s own internal models. This development highlights the intense competition among major tech companies to build the necessary infrastructure for the rapidly expanding field of artificial intelligence.

AI Startup BluePill Raises $6 Million to Revolutionize Brand Research with Simulated Focus Groups

Seattle-based AI startup, BluePill, has secured $6 million in seed funding to transform how brands understand consumer behavior. The round was led by Ubiquity Ventures, with participation from Pioneer Square Labs and Flying Fish Partners.

Launched earlier this year, BluePill leverages artificial intelligence to simulate consumer reactions to marketing concepts, products, and designs – offering brands near-instant feedback instead of relying on traditional, time-consuming focus groups.

The company builds tailored AI consumer audiences for each brand, utilizing real-world data like social media conversations, surveys, and customer input, mirroring the brand’s target demographic.

Users upload their ideas to the platform and receive immediate predictions of their audience’s response, effectively running a massive, virtual focus group.

BluePill is also developing pre-built, industry-specific AI audiences – such as “U.S. moms” or “Gen Z snack buyers” – allowing brands to directly query for insights without constructing custom models.

BluePill claims its simulated audiences achieve 93% accuracy compared to human panels. The company currently works with brands like Magic Spoon, Kettle & Fire, and the Seattle Mariners, utilizing the platform to test fan engagement and in-stadium activations.

“Our edge is validated, accurate insights – and the fact that we deliver these in minutes for a fraction of the cost makes it a no-brainer,” stated BluePill founder and CEO Ankit Dhawan.

BluePill is already generating revenue through a fixed annual subscription model.

The startup is challenging established marketing research giants like Ipsos, Qualtrics, and Nielsen, which traditionally depend on lengthy, costly human panels. Dhawan highlighted that newer startups increasingly utilize large language models to mimic consumer responses.

Dhawan’s background includes experience as an entrepreneur-in-residence at the Allen Institute for AI (Ai2) and a previous role as a product leader at Amazon, specializing in AI products.

Key team members include Puneet Bajaj and Andy Zhu. BluePill recently garnered attention in GeekWire’s Startup Radar in June.

Sunil Nagaraj, a founding partner at Ubiquity Ventures, emphasized that “predicting consumer behavior is the holy grail of marketing.” Nagaraj, based in Silicon Valley, is an active participant in the Seattle startup ecosystem and was an early investor in Auth0, which Okta acquired for $6.5 billion.

Helion Energy’s Bold Move: Scaling Up Fusion Power Manufacturing

Everett, Wash. – Helion Energy is pursuing a multifaceted strategy, aiming not just to achieve fusion power, but to manufacture it on a commercial scale. The company is simultaneously developing its seventh-generation fusion prototype while constructing a large-scale manufacturing facility – dubbed Omega – near its headquarters in Central Washington.

Helion’s approach centers around building an assembly line to produce the thousands of capacitors needed to deliver massive electrical surges to its fusion generator. This production will fuel the Orion power plant, a 50-megawatt facility in Malaga, Wash. The facility will utilize both human workers and robotics, incorporating both off-the-shelf and custom technology to accelerate the manufacturing process.

‘Helion is a manufacturing company,’ said Sofia Gizzi, Helion’s senior manager of production. ‘It’s not an R&D company. It’s not a science experiment. It’s very much a manufacturing company.’

The Omega facility is crucial to Helion’s strategy, aiming to mitigate supply chain disruptions and quickly adapt to evolving design needs. This approach aligns with a broader effort to restore American production capacity, supported by the bipartisan Fusion Advanced Manufacturing Parity Act, spearheaded by Washington state congressional leaders. The company recently secured $425 million in investment from prominent firms including OpenAI CEO Sam Altman and Nucor.

Helion’s plan involves initially producing approximately 2,500 capacitor units for the Orion plant, with production slated to begin in late 2026. The scale-up is designed to support the construction of subsequent fusion generators, with the facility capable of operating at 50% of its design capacity or less while producing Orion units. Looking ahead to 2030, Helion anticipates a significant expansion in its manufacturing capabilities.