This Week in AI: June 9–16

This week gave us quite a few topics to talk about. Regardless, the undisputed winner is Apple's WWDC24, where, in a display of showmanship only Apple can pull off, the company announced Apple Intelligence under the very catchy slogan "AI for the rest of us". The slogan is spot on, and the features are practical, coming across as boring in a context where all things AI must be hyped as groundbreaking and revolutionary. The most controversial bit about Apple Intelligence is that Apple has chosen OpenAI as its first third-party integration, with Siri being able to suggest that some users' queries be redirected to ChatGPT where the latter would be more useful. Apple claims OpenAI will adhere to its strict privacy standards, that IP addresses will be obscured, and that queries originating from Apple devices will not be persistent.

Still, many have expressed concern about the move, which appears even riskier when one considers the lengths Apple has gone to ensure privacy preservation by hosting its models on proprietary silicon to ensure no data is retained or leaked. Elon Musk made sure to ride the wave of controversy to remind us that his feud with OpenAI is not over even if he ultimately dropped his now infamous lawsuit against the company: Musk stated that a decision to integrate OpenAI's technology at the operating system level would be an unacceptable security violation that would lead to him issue a ban over Apple devices across his companies, requiring the devices to be checked at the door and stored inside a Faraday cage.

Regardless of whether Musk's big claims are plausible or not, OpenAI has recently been under fire precisely for privacy and security concerns, including the much-commented-on incident where OpenAI may or may not have attempted to mimic Scarlett Johansson's voice after the actress declined to lend her voice to ChatGPT, and the departure of Jan Leike and other security researchers from OpenAI.

The last couple of weeks has also seen an increase in reports of tech firms being put under increased pressure by lawmakers and regulators. Two of this week's headlines touch on the subject, exposing an interesting contrast: Microsoft has decided to delay the general availability of the Recall feature for Copilot+ PCs. The feature will launch for the Windows Insider Program members first to enable Microsoft to gather additional feedback and ensure that all the security and privacy features are up to standard. In contrast, Meta didn't hold back its disappointment about pausing AI model training on European users' data following concerns from its primary European regulator.

The remaining headline selection for the week touches upon an interesting topic variety, including the fact that autonomous vehicle companies are struggling, and the AI-powered video generation race is heating up as not only one, but two Sora competitors went public this week. Additionally,

Unify's dynamic routing aims to solve the LLM fragmentation challenge: Unify provides dynamic routing capabilities to help developers choose the optimal LLM for their applications by evaluating cost, speed, and accuracy on a prompt-by-prompt basis. Unify supports custom router training and has developed a toolkit for on-prem enterprise deployments.

Databricks unveils significant Mosaic AI updates and enhancements: One year after acquiring Mosaic ML, Databricks is announcing significant updates to the platform, which include new building and deployment, evaluation, and governance tools to ensure enterprises can efficiently turn state-of-the-art models into compound AI systems.

Stability AI calls Stable Diffusion 3 Medium its 'most advanced text-to-image open model': Stable Diffusion 3 Medium is a high-performing yet compact text-to-image model that delivers photorealistic and stylized images with improved prompt comprehension, typography generation, and flexible licensing options. SD3 Medium is also resource-efficient enough to run on consumer PCs.

Google's June Feature Drop brings Gemini Nano to Pixel 8 and 8a devices: Google's latest Feature Drop brings Gemini Nano to Pixel 8 and 8a phones as a developer option, enabling enhanced transcription capabilities in the Recorder app. The June Feature Drop also introduces Car Crash Detection on Pixel Watch 2 and delivers smart home control upgrades across Pixel devices.

MLCommons releases MLPerf Training 4.0 results and unveils two new benchmarks: MLCommons released MLPerf Training v4.0 showcasing major performance gains. Remarkably, NVIDIA set new records using software optimizations on its Hopper architecture. MLPerf Training 4.0 introduces novel benchmarks for LoRA-based fine-tuning and graph neural network training.

Databricks' newly open-sourced Unity Catalog is now available: Databricks has open-sourced its Unity Catalog, calling it the industry's first open-source catalog for unified data and AI governance across clouds, data formats, and data platforms. The Unity Catalog enables several groundbreaking features while garnering ecosystem support.

Mistral AI's $640 million Series B is finally confirmed: Mistral AI has secured a massive Series B funding round led by General Catalyst with participation from selected top investors. The funding will allow Mistral to expand its computing capacity, grow its team, and strengthen strategies to take on its competitors.

Benchling is tasked with migrating Sanofi's R&D into the digital era: Benchling, a leading provider of R&D cloud software for the biotech industry, has partnered with Sanofi to transition over 1,500 scientists to Benchling's digital R&D Cloud platform, enabling Sanofi to aggregate complex data, accelerate research processes, and harness AI across the R&D lifecycle.

Prominent South Korean fabless chipmakers Rebellions and Sapeon are merging: As NVIDIA's market cap crosses $3T and major tech giants develop their AI chips to reduce reliance on NVIDIA, South Korean fabless AI chip startups Rebellions and Sapeon are merging to better compete against global rivals in the rapidly growing AI hardware market.

RetailReady raised $3.3M in seed funding to prevent shipping errors: RetailReady, a Y Combinator graduate leveraging AI to help shippers and providers reduce fees for non-compliant shipments to major retailers, recently raised $3.3M in seed funding to grow its team and develop its product, which aims to optimize warehouse packaging compliance processes.

Black Semiconductor secured €254.4M for its graphene-based chip-interconnect technology: Backed by €228.7M from the German government and €25.7M private investors as part of a major €254.4 million Series A round, Black Semiconductor aims to commercialize its groundbreaking graphene-based chip connectivity technology.

Enveda Biosciences secured $55M to search the world's biodiversity for new medicines: Enveda Biosciences, a company building an advanced search engine for new drugs in the unexplored chemical space, announced a successful $55M Series B2 round to support its platform which helps researchers find multiple molecule candidates for targeted diseases and organs.

A research team is fine-tuning Gemini to pave the way for truly personalized fitness and health assistants: In two recently published papers, Google researchers presented complementary approaches to fine-tuning LLMs to enable accurate analysis of personal health data and the generation of personalized wellness recommendations, paving the way for AI-powered personal health assistants.

Brave announced its AI assistant Leo is getting Brave Search integration: Brave Software has rolled out an update to its in-browser AI assistant, Leo, enabling it to enhance responses with real-time Brave Search results to provide accurate and up-to-date answers while maintaining robust privacy protections and encouraging users to fact-check sources and explore further.

NVIDIA released Nemotron-4 340B, an LLM family for synthetic data generation: NVIDIA has released Nemotron-4 340B, an open pipeline of models that allows developers to generate synthetic training data for building custom large language models tailored to various industries without prohibitive data costs.