☕️ Meta's Llama 4 Launch Gets Off To a Rocky Start

The first of the Llama 4 herd arrived earlier this week with the debut of Llama 4 Scout and Maverick, albeit with doubts and concerns as some raised an eyebrow at Meta. Other key highlights include:

• OpenAI's Sam Altman shares an update on the standalone consumer release of o3 and GPT-5 delays

• IBM unveils its next-generation made-for-AI mainframe built to handle the workloads of tomorrow

• Microsoft's AI-generated Quake II demo gives a glimpse of how AI could be integrated into gaming

Join us at AI Tangle as we untangle this week's happenings in AI!

Meta's Llama 4 Scout and Maverick meet the public as the company debuts its next-gen family of AI models

On Saturday, Meta unveiled the first pair of Llama 4 AI models from its latest open-source AI family as it scrambles to catch up with the industry's best. The two new additions, Llama 4 Scout, with a uniquely massive context window of 10 million tokens, and Llama 4 Maverick, are already available to users via Meta AI on WhatsApp, Messenger, Instagram Direct, and the Meta AI website. The company's biggest rendition, the staggering 2 trillion parameter Llama 4 Behemoth, will continue its training for the time being to better "serve as a teacher" for the company's future AI models.

​

​But real-world use doesn't seem to add up.​

However, despite ​touting a successful release with all the benchmark scores​ and Llama 4 Maverick ranking second on LM Arena, some had reason to believe that Meta may have gamed the system. Though LM Arena isn't the most accurate metric for a variety of reasons, companies haven't generally at least admitted to fine-tuning their models for these. Independent researchers and testing users across social media, like X/Twitter and Reddit, have been disproportionally critical and suspicious of the release in comparison to other mainstream debuts, be it Anthropic's, OpenAI's, or Google's, though Meta executive Ahmad Al-Dahle has staunchly denied any such unsubstantiated claim or rumor.

After nearly a year of absence from the scene, one of the leaders in the GenAI image generation race, Midjourney, has resurfaced to release its first new AI image model in ages. Sticking to its usual naming scheme, Midjourney V7 began rolling out as part of an early alpha-test phase starting on Thursday, a week after OpenAI's image generator in ChatGPT went viral and three weeks after Google began experimenting with its own.

OpenAI shifts gears as it prepares for an o3 consumer launch after canceling it in February

In spite of practically canceling the standalone consumer launch of its o3 reasoning model earlier in February, OpenAI's plans have seemed to change course. In a post on X/Twitter on Friday, company CEO Sam Altman stated that both o3 and its successor, o4-mini, will be released in just "a couple of weeks," though it comes at the cost of delaying GPT-5 by a few months. GPT-5 is still on track to unify a variety of OpenAI's offerings, such as voice, Canvas, search, and deep research, all at varying "intelligence settings," depending on whether a user is subscribed to the company's Plus or Pro plans.

​

IBM unveils its new z17 mainframe, purpose-built to handle AI workloads and agents on-site

After nearly five years in the works, IBM has announced the release of its latest mainframe, the IBM z17, designed to speed up AI adoption with "over 250 use cases" ranging from AI agents to generative AI apps. Powered by the new IBM Telum II processor, the z17 can process 450 billion inference operations daily, a stark 50% improvement over its predecessor, and supports 48 IBM Spyre AI accelerator chips to accommodate where the AI market is heading, with plans to double that within 12 months. Enterprise customers can look forward to the system shortly, as it is set to become generally available on June 8.

​

Microsoft shows off an AI-generated Quake II demo to showcase the potential of AI in the future of gaming

Microsoft has released a browser-based demo of the classic Quake II arena shooter powered by its Copilot AI platform in an effort to showcase the future potential of gaming via its Muse family of AI models. The demo lets users play through a Quake II level at a "playable" framerate and a resolution of 640 x 360 pixels by using either a keyboard or controller, with the AI generating instant visual responses. However, Microsoft realizes it's not without its problems it's not without its problems, ranging from fuzzy enemies to inaccurate damage counters to struggles with object permanence.

​

DeepSeek unearths a new method to give its AI reasoning models a leg-up over the competition

​In a paper recently published on arXiv, Chinese AI startup DeepSeek, in collaboration with researchers from Tsinghua University, uncovered a new approach to boost the reasoning capabilities of its LLMs. This new technique involves combining two methods - generative reward modelling (GRM) and self-principled critique tuning. The report details that the resulting DeepSeek-GRM models have "achieved competitive performance" compared to leading public reward models as the industry waits in anticipation for the debut of DeepSeek-R2 as early as this month, according to a speculative article by Reuters.

​

OpenAI is reportedly in talks to acquire Jony Ive and Sam Altman's AI hardware startup for roughly $500 million

OpenAI has reportedly considered buying io Products, an AI device startup connected to OpenAI's own Sam Altman and former Apple designer Jony Ive's co-launched studio, in a deal rumored to be valued at at least $500 million. The deal could broaden OpenAI's product portfolio by adding hardware innovations, such as a screenless phone and AI-enabled household devices, to its already rather expansive lineup of software and robotics solutions. The move would also position it to compete more directly with Apple, even as the company faces struggles with restructuring and funding challenges.

​Speechify - Speechify turns long, dense reading into a pastime with high-quality voiceovers powered by over 200 human voices, sparing you the time and helping you get through long walls of text faster.

​Needle - Meet Needle, a seamlessly integrating platform that simplifies how teams access and manage knowledge across their favorite tools by leveraging AI to do organization-wide searches and save time.

​Dynium - Minimize or completely prevent data leaks with Dymium, a platform that boosts security posture across workflows and helps comply with regulatory requirements.

​Clarity - Clarity is your personal AI-powered assistant for working with documents. Talk to your documents, get instant insights from them, and collaborate in a team.

AI's Place in Delivering Mental Health Therapy (4-min read)

With many countries struggling to cope with the shortage of mental health clinicians, some have attempted to pivot to AI in hopes of easing the burden. Provided sufficient training, new research suggests that it may be more than just a shot in the dark.