AI Week in Review 25.06.14
Astra video upscaling, o3-pro, o3 price cut, Magistral Small & Medium, DIA browser, Krea 1, Cartesia Ink, Copilot Vision, Meta buys Scale AI, V-JEPA2, Apple's AI-enhanced Siri is delayed.
Top Tools
While we have pegged OpenAI’s newly-released o3-pro as “the AI that thinks too much” for its slow, over-thinking behavior, yet o3-pro deserves the top tool billing this week for being, at least for now, the top SOTA AI reasoning model, achieving 93% on AIME 2024, 84% on GPQA Diamond, and nearly 2800 ELO on Codeforces. The o3-pro model is highly capable for complex analysis tasks.
In addition, OpenAI dropped the price of its o3 reasoning model by 80%, making it more cost-competitive with other AI models. At $2 input and $8 output per million tokens, o3’s pricing is in line with Gemini 2.5 pro and cheaper than Claude 4 Sonnet.
AI Tech and Product Releases
Apple's Worldwide Developer Conference (WWDC) was focused more on liquid glass interface than AI this year, yet Apple made several AI-related announcements:
Live Translation between languages is coming to messages and calls, and it will run directly on device. For calls, it will be an audio translation.
Visual Intelligence allows users to search for objects within images, similar to Google's "circle to search". Apple will add this to iOS26 later this year.
Developer Access: Developers will be able to build apps that leverage Apple's on-device AI models.
Image Playground for AI-driven image editing gets a boost with ChatGPT integration.
The biggest AI news though was a set-back: The long-awaited AI-powered Siri update won’t arrive until next year. Apple has slipped into an AI laggard.
Mistral AI introduced Magistral, its first reasoning model, in two variants: Magistral Small, a 24B open source AI reasoning model, and a larger Magistral Medium for enterprise use. Magistral is optimized for real-world applications across many domains (such as finance, healthcare, law) requiring traceable, interpretable reasoning, and it supports multilingual output for their European users. Magistral Medium achieves 73.6% on AIME 24, on par with DeepSeek R1, while the Magistral Small reaches 70.7%.
While not state-of-the-art, Magistral is claiming to be 10 times faster than other AI reasoning models. Magistral Small is available on Hugging Face, while Medium is accessible via Le Chat and cloud API providers.
The Browser Company, maker of the Arc browser, has launched the Dia Browser, a new AI-native browser that allows users to interact with their tabs conversationally. Dia leverages AI with customization:
A notable aspect of Dia's design is its emphasis on user customization and integration of AI, a step beyond typical browsing capabilities. Dia allows users to toggle between chat and search modes, providing context-aware responses that enhance productivity by utilizing data from open tabs.
Midjourney will soon launch an AI video generation product, and Midjourney videos are popping up on X. Midjourney is inviting subscribers to rate AI-generated videos to help train the model.
Moving in the other direction from Midjourney, Krea AI has released their first AI image generation model, KREA-1:
Krea 1 ensures highly realistic, crisp textures, a wide variety of styles, and deep artistic knowledge – making AI images not look less AI anymore.
Cartesia has introduced Ink, speech-to-text models optimized for real-time chat. The Cartesia Ink model is both more accurate in its speech-to-text transcription and also the fastest at time-to-complete-transcript.
Microsoft's Copilot Vision can now "see" a user's screen and provide interactive guidance for applications. Microsoft is also introducing an AI agent to Windows 11 Settings, allowing users to describe issues for suggested solutions or automated fixes.
Google has rolled out Audio Overviews in Search via Google Labs.
Lemony launched an “AI in a Box” device providing secure, on-premise AI for private, cloud-independent generative AI workflows. The Lemony solution simplifies enterprise AI for highly regulated industries, addressing privacy and compliance concerns.
AI Research News
A paper from Apple research called “The Illusion of Thinking” created a big stir with claims that AI reasoning models only pattern-match and don’t truly reason. This sparked heated debate, with critics arguing that Apple used flawed methodology, and that token limits, not reasoning limits, caused apparent failures. The rebuttal is correct; the paper’s test setup was flawed.
Meta released the V-JEPA 2 model, which learns a “world model” from video, being trained on a dataset of over 1 million hours of internet video. The paper “V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning” explains the architecture and approach; they claim state-of-the-art performance for their 8B size on visual QA. Through post-training for an action-conditioned world model, V-JEPA2 can enable more adaptable robot planning and reduce training costs for robots and autonomous systems.
AI Business and Policy
Meta secured a 49% stake in Scale AI for about $14.3 billion, valuing the data-labeling startup at around $29 billion. As part of the agreement, Scale’s CEO Alexandr Wang will join Meta to spearhead a newly formed “superintelligence” lab focusing on artificial general intelligence (AGI). Scale remains operationally independent while deepening its commercial relationship with Meta.
Meta’s Scale AI purchase is part of a bigger story: Meta is betting big on AI by bidding up AI talent. CEO Mark Zuckerberg is personally focused on rebuilding Meta’s AI team, including its Llama model team, with aggressive talent recruitment efforts. Is Scale AI the largest Aqui-hire in history?
At its Advancing AI event, AMD unveiled its new Instinct MI350 Series AI chips that are 4x faster on AI compute and 35x faster on inferencing than their prior AI chips. AMD also introduced open, scalable rack-scale AI infrastructure and previewed the next-gen MI400 Series and Helios Rack system. They are still far behind Nvidia in market share.
Tesla sued a former engineer for allegedly stealing trade secrets from its humanoid robotics program, Optimus, and using them to launch a rival startup. The lawsuit accuses Zhongjie “Jay” Li of downloading confidential “advanced robotic hand sensors” data to personal devices before founding Proception. Proception then claimed to have built similar humanoid robotic hands, resembling Tesla's designs.
Barbie-maker Mattel and OpenAI have teamed up to bring generative AI to toy-making and IP repackaging. This marks OpenAI's first licensing deal with a toymaker, aligning with its strategy to integrate AI across diverse industries.
Disney and Universal sued Midjourney for unlicensed use of 'Star Wars,' 'The Simpsons' and more in their AI image generation models. The movie studios said:
“Midjourney is the quintessential copyright free-rider and a bottomless pit of plagiarism.”
Spanish AI startup Multiverse Computing secured $215 million in funding for its CompactifAI technology. This innovation compresses large language models up to 95% without performance impact. Their "slim" models offer 4x-12x faster inference, 50-80% cost reduction, and can run on devices like PCs and phones.
Wikipedia paused an experiment using AI to summarize articles after editors pushed back. Editors raised concerns that AI "hallucinations" could damage credibility, despite summaries being labeled "unverified." Wikipedia remains interested in AI summaries for accessibility despite the pause.
Concurrent with their Magistral AI reasoning model release, Mistral AI launched Mistral Compute, an AI infrastructure platform in partnership with Nvidia, positioning itself as Europe's alternative to US cloud providers.
Outset, an AI-moderated research platform, secured $17 million in funding, bringing its total to $21 million. Their AI-powered platform conducts video interviews at scale, offering results 8 times faster and 81% cheaper than traditional market research.
Wandercraft secured $75 million in funding to accelerate its AI-powered robotics development. This capital will propel the launch of Eve, a personal exoskeleton, expand clinical use of Atalante X, and deploy the industrial humanoid robot Calvin-40.
The new Meta AI app is causing a major privacy concern, with users unknowingly publishing their private conversations, audio, and images. People are sharing sensitive details like tax evasion inquiries, legal issues, and personal addresses, often unaware of the public nature of their posts. Be aware!
Senator Cynthia Lummis introduced a new AI regulation bill called the Responsible Innovation and Safe Expertise (RISE) Act. It pairs a conditional liability shield for AI developers with mandates for transparency on model training and specifications. Sen Lummis wrote on X:
If we want America to lead and prosper in AI, we can’t let labs write the rules in the shadows. We need public, enforceable standards that balance innovation with trust.
New York passed the RAISE Act, which aims to prevent frontier AI models from causing major harm by requiring large labs to publish safety reports and report incidents. It targets models trained with over $100 million in computing resources, imposing civil penalties up to $30 million for non-compliance, and now awaits the Governor's signature.
AI Opinions and Articles
Sam Altman’s “The Gentle Singularity” has a dose of self-promotion in his predictions, but its worth listening to nonetheless, because he gets fundamentals right. For example, he notes that:
From here on, the tools we have already built will help us find further scientific insights and aid us in creating better AI systems. … The economic value creation has started a flywheel of compounding infrastructure buildout to run these increasingly-powerful AI systems.
This ‘flywheel’ of competition, investment, and AI-drive algorithmic improvement will combine to propel AI to new heights. The result will be abundant intelligence on tap.
In the 2030s, intelligence and energy—ideas, and the ability to make ideas happen—are going to become wildly abundant. These two have been the fundamental limiters on human progress for a long time; with abundant intelligence and energy (and good governance), we can theoretically have anything else.