GSV's AI News & Updates (04/25/24)
Meta Llama 3, Microsoft Phi-3 and VASA-1, Adobe Firefly, Perplexity becomes Unicorn, Meta Smart Glasses, 2024 AI Index Report, Powerschool PowerBuddy
We’re back after a brief ASU+GSV Summit and AIR Show hiatus! If we didn’t get a chance to connect live in San Diego this past week, please feel free to email me at claire@gsv.com.
General 🚀
Meta releases Llama 3: these two variants (8B and 70B parameters) blow the earlier models from Meta out of the water. A larger model with 400B+ parameters is still in training.
These models also power the upgraded Meta AI (Zuckerberg’s answer to ChatGPT). The intelligent AI assistant is now on all of Meta’s apps—Instagram, Facebook, Whatsapp and even on the web as “meta.ai”. Zuckerberg claims that Meta AI is the most intelligent free chatbot available publicly.
Zuckerberg talks about these models and Meta AI here
AI Search company Perplexity is now a unicorn: Perplexity just raised another $62.7M and added an enterprise plan. Perplexity’s revenue has grown from $3M to $20M in 6 months (since its last raise). Perplexity has a chance to begin making revenue from selling ads within or around its “answers,” although the startup is still experimenting with different approaches.
Microsoft’s VASA-1 can deepfake a person with one photo and one audio track: in the future, it could power virtual avatars that render locally and don't require video feeds—or allow anyone with similar tools to take a photo of a person found online and make them appear to say whatever they want.
Microsoft launches Phi-3, its smallest AI model yet: “an open LLM that fits on your phone”. Phi-3 represents a breakthrough in AI, designed to operate efficiently on devices with limited resources such as smartphones and laptops. It is designed to deliver powerful, fast, and low-latency AI capabilities locally on consumer hardware, therefore enhancing accessibility and cost-effectiveness. (Technical report here).
Adobe Introduces Firefly Image 3 Foundation Model: aimed at enhancing user control and efficiency in design creation. Key features include the Reference Image tool (utilizes user-uploaded images as a basis for generating similar designs), Generate Background (creating new backgrounds in product photography), Enhance Detail (sharper images), and Generate Similar (replicating a chosen image).
Meta Smart Glasses Styles and Meta AI Updates: the Meta AI assistant on the smart glasses can now do things like:
Generate images based on text prompts in a "reimagine" feature in group chats
Retrieve real-time information like sports scores, local business details, and more powered by Bing search
Describe objects in the user's surroundings using computer vision (in a future update)
Apple acquires Datakalab, startup specializing in AI compression and computer vision technology: Datakalab specializes in developing efficient AI algorithms that consume less power and are optimized for computer vision and AI compression technology. The acquisition of Datakalab is likely aimed at enhancing Apple's generative AI research and development efforts.
DeepMind CEO Says Google Will Spend More Than $100 Billion on AI
Stanford HAI 2024 AI Index Report:
AI beats humans on some tasks, but not on all
Industry continues to dominate frontier AI research
Frontier models get way more expensive
The United States leads China, the EU, and the U.K. as the leading source of top AI models
The data is in: AI makes workers more productive and leads to higher quality work
Education and the Future of Work 📚
Lessons for Ed-Tech Companies in LAUSD’s Far-Reaching AI Experiment: “Ed” was developed through a public-private partnership with AllHere and will provide students with access to information like grades, attendance, bus schedules, and outside resources for learning.
PowerSchool Unveils Contextual AI Solutions to Reduce Teacher Workloads and Accelerate Student Learning: Powerschool announced the general availability of the first two PowerSchool PowerBuddy™ products designed for teachers: PowerBuddy for Learning and PowerBuddy for Assessment.
MetaMetrics launches Advanced AI-Powered Scoring Tool for Student Writing
How to teach in a world run by AI: AI should be a multidisciplinary subject that can be woven into various educational subjects rather than isolated as a standalone course. We need to ensure that students are not only consumers of AI technology but also informed participants in shaping its impact on society.
What Do We Gain and Lose When Students Use AI to Write?: one very real risk - they may start relying on generative AI to the extent that it wholesale replaces their thinking.
The Opportunities and Drawbacks of AI-Powered Reading Coaches, Assistants and Tutors: Potential drawbacks of AI reading tools include inconsistencies in generated stories that can confuse and reduce young readers' empathy. While these tools mimic some aspects of traditional books, they often lack the rich, multi-modal experiences of high-quality children's literature and may exacerbate the digital divide and diminish valuable human interaction in learning.
HarperCollins Publishers and ElevenLabs to Bring More Stories to Life Through Audio: the agreement will lead to the production of audio versions of select deep backlist series books that would not otherwise have been created.
Innovation through prompting, democratizing educational technology and more: using LLMs to democratize innovation by making advanced technology accessible to non-technical users through simple prompts. AI-driven prompts can be customized and applied in practical settings to produce specific outcomes, particularly in educational simulations and interactive teaching tools.
Startups and Tools 🛠️
Limitless AI: the world’s most wearable AI. Preserve conversations and ask your personalized AI anything.
Panjaya: localize or tailor your video content for any audience, in any language, while preserving the speakers’ natural voice and lip movements.
Autorepurpose: turn your YouTube videos into text content for Twitter thread, LinkedIn posts, Newsletter, and etc.
Coachvox.ai: charge for access to an AI version of you.
France's Mistral AI seeks funding at $5B valuation: Mistral was last valued at 2B euros in a funding round in December. This would be their 3rd round in a year.
AI leads a service-as-software paradigm shift: the push towards services represents a $4.6 trillion opportunity.
Tech 💻
OpenAI’s CEO Says the Age of Giant AI Models Is Already Over: Altman suggested that the future of AI development will not involve increasing model size but will focus on improving models in other ways, due to diminishing returns on model scaling and physical constraints on infrastructure.
AI Gives Enterprise Device Market Something to Be Excited About: some CIOs remain cautious about the higher costs of these AI-enabled PCs and are initially considering them only for specialized roles like data scientists, rather than widespread deployment across the organization.
Boston Dynamics’ Atlas humanoid robot goes electric: the electric Atlas is designed for real-world applications, with enhanced mobility, dexterity, and perception capabilities that allow it to interact with the environment in meaningful ways.
Amazon wants to host companies’ custom generative AI models: AWS announced the launch of Custom Model Import (in preview), a new feature in Bedrock. The feature lets organizations import and access their in-house generative AI models as fully managed APIs.
OpenAI is introducing more enterprise-grade features for API customers: enhanced enterprise-grade security, Private Link, better administrative control, more options for cost management.
How cheap, outsourced labour in Africa is shaping AI English
Regulation and Policy ⚖️
From “AI” to Probabilistic Automation: How Does Anthropomorphization of Technical Systems Descriptions Influence Trust?: this paper examines the impact of anthropomorphizing language (eg giving the product a human name) in AI system descriptions on people's trust levels. They find that participants are no more likely to trust anthropomorphized over de-anthropmorphized product descriptions overall.
Spotting the deepfakes in this year of elections: how AI detection tools work and where they fail: enhancing transparency and trust through techniques like watermarking, digital fingerprinting, and metadata embedding.
Snap plans to add watermarks to images created with its AI-powered tools
Ex-Amazon exec claims she was asked to ignore copyright law in race to AI: According to the complaint, her director allegedly told her to ignore copyright policies to improve the results. Referring to rival AI companies, the filing alleges he said: "Everyone else is doing it."
Other
Marques Brownlee reviews the Humane AI pin: “the worst product I’ve ever reviewed”
OpenAI opens Tokyo hub: OpenAI has chosen Tokyo as the location for its first Asian office, citing Japan's global leadership in technology, culture of service, and embrace of innovation.
Drake Uses AI Tupac and Snoop Dogg Vocals on ‘Taylor Made Freestyle’
Can AI rewrite our human genome? Profulent has designed a gene editor with AI that can successfully edit the human genome: their results point to a future where AI precisely designs what is needed to create a range of bespoke cures for disease.
Resources
The GSV BIG 10: weekly coverage corner for the top 10 stories, insights, and major plays in learning and skilling.