Microsoft is releasing their latest Small LLM (SLM) Phi-4, which now comes with advanced reasoning capabilities (think o1). Based on maths AMC benchmarks, it seems to beat much larger models including Gemini 1.5 Pro. What is particularly interesting with this model release, is the technical report and lots of details on how they generated synthetic data.
Not a lot of information nor the paper is available yet, but this is joint research from Sony and University of Illinois on join audio-video synthesis - think Sora with audio. Super interesting.
Meta rolled out new AI-powered features for its Ray-Ban smart glasses, including live virtual assistance, instant language translation, and Shazam integration for identifying music hands-free. These upgrades position the glasses as a versatile everyday tool for real-time assistance. Source: The Verge
YouTube now lets content creators decide which AI companies can use their videos for model training. The program initially includes partnerships with 18 major tech firms, such as OpenAI, Microsoft, and Meta, signaling a new era of transparency in AI development. Source: The Verge
Google Labs introduced “Whisk,” an experimental tool that leverages Imagen 3 and Gemini AI to empower users to remix, refine, and transform images seamlessly. This tool expands creative possibilities with image-to-image AI functionality. Source: The Verge
In a recent interview, former Google CEO Eric Schmidt expressed growing concern over advanced AI systems. He argued that intervention, including potentially “shutting down” self-improving AI, may be required to prevent unforeseen consequences. Source: The Sun
SoftBank’s Masayoshi Son committed to investing $100 billion in the United States’ AI industry, with plans to generate 100,000 new jobs over the next four years. The announcement followed discussions with President-elect Donald Trump. Source: New York Post
Lockheed Martin established a new subsidiary, Astris AI, to accelerate artificial intelligence adoption in defense and commercial sectors. The initiative aims to enhance operational capabilities and drive innovation across industries. Source: Reuters
Oklo, a nuclear startup chaired by Sam Altman, has secured a 20-year agreement to supply 12 gigawatts of electricity to data centers. This deal highlights the tech industry’s growing interest in sustainable nuclear energy to meet increasing AI and cloud computing demands. However, Oklo still requires regulatory approvals and must construct its small modular reactors, with plans to have its first reactor operational by 2027.
Microsoft has surpassed its competitors by adding 485,000 new subscribers to its cloud services in the past quarter, bringing its total to over 5 million. This growth is attributed to the company’s strategic investments in artificial intelligence and cloud infrastructure, which have enhanced service offerings and attracted a broader customer base. In contrast, rivals like Google and Amazon have reported slower subscriber growth during the same period. Microsoft’s focus on integrating AI capabilities into its cloud platform has positioned it favorably in the competitive tech industry landscape.
I know a bit about nuclear (both in Sweden and Finland there is plenty of nuclear power), but I don’t know much about these micro nuclear reactor designs. The 2027 timeframe seems incredibly aggressive, but maybe that’s normal for these types of reactors?
Midjourney has introduced ‘Moodboards,’ a feature that enables users to upload curated image collections to inspire new AI-generated art. This addition allows for the creation of multiple personalization profiles, facilitating seamless organization and deployment of diverse styles. Source: VentureBeat
Google has enhanced its Gemini Code Assist tools by integrating the Gemini 2.0 language model and expanding connections to external code repositories and cloud-based databases. These improvements aim to boost developer productivity within integrated development environments (IDEs). Source: TechZine
The United Arab Emirates’ Technology Innovation Institute (TII) has launched Falcon 3, an open-source family of language models available in four sizes—1B, 3B, 7B, and 10B parameters. These models are designed to democratize access to advanced AI capabilities, offering high performance while being efficient enough to run on lightweight hardware, including laptops. Source: Zawya
Databricks has secured a $10 billion Series J funding round, elevating its valuation to $62 billion. The investment, led by Thrive Capital, Andreessen Horowitz, and Insight Partners, is intended to support the company’s AI product expansion, potential acquisitions, and international go-to-market operations. Source: YourStory
@platypus opens the first-ever monthly AI Pulse News Roundup Thread, inviting users to share and discuss breaking news, explore key developments in real-time, and archive highlights of the month. The new format encourages community engagement around AI advancements.
@platypus shares insights on a paper from University of Maryland and Adobe Research titled DynaSaur, which presents an unconstrained method for creating LLM agent actions.
@PaulBellow discusses Amazon’s new generative AI model, “Olympus,” highlighting its multi-modal capabilities and strategic implications for Amazon in the competitive AI space.
@sps mentions Amazon becoming the training compute provider for Anthropic, leading to discussions about both companies’ interdependencies.
Various users comment on Amazon’s developments, including @PaulBellow’s observation of potential competitive shifts in the AI market and whimsical discussions around creative names like Trainium.
User participation is active, with users posting diverse content, such as breaking news about OpenAI filing for a trademark on its reasoning models, Amazon’s AI enhancements for its services, and updates on AI projects globally from different tech firms.
@mitchell_d00 suggests creating a “proud parent thread” to involve families in AI projects, emphasizing the community aspect.
Reports on significant developments include:
Soracom’s venture into mixed-reality systems with Android XR.
Meta’s new AI features for Ray-Ban smart glasses.
YouTube allowing creators to control which AI companies use their content.
Future trends include Google’s Gemini Code Assist, Falcon 3’s launch by the UAE, and significant funding rounds for Databricks to enhance AI product expansion.
@vb notes the increasing role of nuclear energy in tech, particularly regarding Sam Altman’s nuclear startup and its 20-year power agreement for data centers. Various models, including Microsoft’s new small language model Phi-4, are discussed, with emphasis on competition among major players like Google and Microsoft in AI infrastructure.
This thread reflects vibrant discussions on ongoing advancements in AI, showcasing community engagement, humor, and diverse insights into the ever-evolving technological landscape.
Two things: (1) “Series J”??? Wow!!! and (2) super exciting to see SLMs getting better and better, Phi-4 and now Falcon 3 - does anyone have any experience with small LLMs and any takeaways from actual applications?
I kind of envision it. Having the generic language logic and rules that can be applied to local data sources. it don’t need data just instructions on how to read data that is generic and modular. More like a smart word processor.
This paper from Meta presents Large Concept Models (LCMs), a novel approach to natural language modeling that operates on sentence-level embeddings instead of token-level representations, as used in traditional Large Language Models (LLMs). There are lots of interesting goodies here, such as its sampling mechanism - instead of sampling tokens sequentially, LCMs generate semantic level outputs by predicting sentence embeddings; they also show very impressive zero-shot generalization capabilities.