While many industry leaders focus on cramming complexity into their reasoning models, Alibaba's Qwen Team has condensed its efforts into just 32 billion parameters with the release of QwQ. Other key takeaways include:

• AI pioneers Andrew Barto and Richard Sutton win the A.M Turing Award for their work on "hedonistic" machines
• The Pentagon brings in AI power from data startup Scale AI to help with its strategic decision-making
• Google goes all-in on AI Overviews with an update and an experimental feature to replace conventional search results

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

THE BIG AI STORY

Shares of Chinese tech giant Alibaba soared by as much as 8.4% on the Hong Kong stock exchange, breaking the company's previous 52-week high after it released its latest reasoning AI model - QwQ-32B (Qwen-with-Questions). Available on Hugging Face under the Apache 2.0 license for use in commercial and research applications, the Qwen Team's latest addition to the industry is free for all.

​

What's the verdict, and where is AI headed from here?

As the name suggests, QwQ-32B is a 32-billion parameter model, much smaller compared to DeepSeek-R1's 671 billion, even if only 37 billion of those are in active use for any given query. Alibaba claims QwQ-32B achieved "impressive results" and, with the release of public benchmarks, has shown that it can rival many of the industry's best reasoning models, such as DeepSeek-R1 and o1-mini.

QwQ-32B's size and others like it showcase how effective models have become - less than two years ago, ​OpenAI's GPT-4​ totaled a mind-boggling parameter count of 1.8 trillion. Unsurprisingly, as a result, Alibaba and China as a whole continue to remain highly optimistic about the upward curve of AI development, with the latter ​having already bet $53 billion on the technology over the next three years​.

6 QUICK HITS

Andrew Barto and Richard Sutton, pioneers of reinforcement learning, have won the A.M. Turing Award, the computer science field's medal of honor, for their work in training "hedonistic" machines. Starting back in the late 1970s, their research taught machines to learn from rewards and punishments, known as "reinforcement learning," which set the foundation for AI today - be it for things such as Google's champion-beating Go-playing program or tools like ChatGPT. Despite early setbacks, their work is now celebrated for transforming AI research and driving billions in investments.

​

The Pentagon has inked a contract with data startup Scale AI, as announced by it in a blog post, to integrate AI agents into the military's workflows under a program called Thunderforge. Using Microsoft's LLM systems and Anduril's simulation infrastructure, the goal is to modernize outdated planning processes and thus allow for faster strategic decision-making. Initially deployed with the US Indo-Pacific and European Commands, Thunderforge will be scaled across combatant commands to give operational planning in emerging conflicts a makeover.

​

After the bumpy launch of AI Overviews, Google is expanding its features with the introduction of AI Mode. Leveraging its Gemini 2.0 line, in this experimental mode, the usual standard blue-link results will disappear, replaced by a conversational AI that builds responses using "advanced reasoning, thinking, and multimodal capabilities." Currently, AI Mode is available to just Google One AI Premium subscribers and is being tested via Search Labs, with feedback expected to refine the experience before a broader rollout.

​

OpenAI reportedly plans to launch a suite of specialized AI agents with tiers that could reach up to $20,000/month to leverage highly proficient AI agents for a variety of fields. According to The Information, for example, one "high-income knowledge worker" agent will reportedly be priced at $2,000/month, while another for developers is estimated at $10,000/month, and another for "PhD-level research" at a staggering $20,000/month. SoftBank has already committed to spending $3 billion on OpenAI's agent products this year as it takes bolder and bolder steps to recoup billions in previous losses.

​

According to an internal email shared with Reuters, Amazon's cloud computing arm, AWS, has recently formed a new team focused on developing agentic AI, an emerging technology aimed at automating tasks. Led by AWS executive Swami Sivasubramanian and reporting directly to CEO Matt Garman, company officials see agentic AI as a potential multi-billion-dollar business opportunity and seem to be all hands on deck. The new group follows Amazon's continued efforts to update its AI-powered services, including an upcoming version of Alexa.

​

Google co-founder Larry Page is reportedly launching a new startup, named Dynatomics, to take advantage of AI to transform product manufacturing, as first reported by The Information. Together with a select group of engineers, Page is working to develop AI systems that create "highly optimized" product designs, which are then prototyped in factories. According to the article, Chris Anderson, former CTO of electric airplane startup Kittyhawk, is at the helm of this stealth operation.

3 AI TOOLS

​Lindy - Lindy helps businesses build AI agents in minutes to automate workflows such as email management, customer service calls, sales outreach, and more.

​Hunch - Chain together AI tasks in a visual, no-code workspace with Hunch, the platform that allows users to share their work as a tool for the entire team.

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

AI READ & WATCH

Disinformation Runs Deep in The Data (2-min read)

A new NewsGuard report reveals that a Moscow-based disinformation network called "Pravda" is distorting AI chatbot outputs by flooding the internet with hundreds of false claims. The network may have influenced major chatbots - recycling a third of its arguments - and published millions of articles, but how severe are the effects in reality?