☕️ Aim For The (AI Coding) Stars

Nvidia, ServiceNow, and Hugging Face unveil StarCoder 2, a code-generating LLM

Co-developed as a family of models, not a single model, by Nvidia, ServiceNow, and Hugging Face, the trio unveiled StarCoder 2, an open-source code-generating LLM. With an emphasis on significantly improved performance at lower costs to operate compared to its older sibling, the trio aims to meet the demand for alternative code assistants that are free, without licenses that exclude commercial use, and can get apps built quickly.

​

What do we know about it?

Trained on 67.5 terabytes of data, compared to its predecessor's measly 6.4 terabytes, StarCoder 2 comes in 3 variants: A 3B model trained by ServiceNow, a 7B model by Hugging Face, and the flagship 15B trained by Nvidia, all free and available in a ​GitHub repository​. StarCoder 2 can be fine-tuned "in a few hours," with the two weaker models being able to be run on any modern GPU while the 15B one demands a bit more power, such as Nvidia's A100. However, although the RAIL-M 1.0 license used for StarCoder 2 is less restrictive compared to others, some still see it as a potential roadblock due to its ambiguity.

Lightricks opens the waitlist for its LTX Studio, an AI-powered filmmaking platform

Up next to add to Lightricks' impressive range of creative editing tools, like Videoleap and Facetune, is the company's new AI-powered filmmaking platform, dubbed LTX Studio​, which has had its waitlist opened to the public. Aimed at professionals like filmmakers and advertisement agencies, LTX Studio provides easy-to-generate storyboards and scripts with extensive customization, as well as a shot editor that allows users to fine-tune individual shots, from frame settings to camera motion.

​

Brave brings its privacy-focused Leo AI assistant to Android users

Brave, an open-source browser, recently announced that it will launch its popular Leo AI assistant to all Android users with iOS only a few weeks later. Centered around privacy and anonymized queries between the user and itself, Leo leverages popular models Mixtral 8x7B, Claude Instant, and LLaMa 2 13B, with the default set to Mixtral. The Leo assistant can perform tasks such as summarizing content, answering questions, generating written content, translating or rewriting pages, writing code, and transcribing audio or video in real time.

​

Apple cancels its decade-long EV project in favor of AI innovation

After being in development for almost a decade, Apple's autonomous EV project, Project Titan, was recently announced to be canceled due to unsteady progress and a drop in demand - but it's not the end of the road. According to Bloomberg News, Apple transferred many of the talents working on the project to the company's GenAI team instead, an industry that Apple previously hadn't been too involved in. "If it is true, Apple will put more focus on GenAI, and that should give investors more optimism about the company's efforts and ability to compete at a platform level on AI," said Ben Bajarin, CEO of consulting firm Creative Strategies.

​

Adobe unveils a new prototype AI-powered music-making tool

Developed by researchers from the University of California and Carnegie Mellon, Adobe announced the aptly named Project Music GenAI Control, a generative AI platform capable of creating audio pieces from reference tracks or simple text prompts. With it, users get a lot of editing power - adjusting tempo, intensity, repeating patterns, structure, remixing audio, and extending tracks to an arbitrary length are just some of the features the platform offers. However, the platform is still deep in the research stage with no public release date in mind, with one official stating that it doesn't even have a proper UI.

​

Copilot For Finance, Microsoft's time-saving answer for finance departments

On Thursday, tech giant Microsoft announced the release of Copilot For Finance, a variant of the company's Copilot chatbot tailor-made for handling tasks in finance departments. The tool aims to help users review data sets for risks, produce reports from unprocessed data, and generally handle tedious, time-consuming tasks. Microsoft adds that its finance department has seen early results already, cutting 1-2 hours of reconciliation down to 10-20 minutes with the help of Copilot For Finance. However, pricing and wider availability beyond a public preview have not been disclosed for the tool yet.

​

Photoroom rises to a $500 million valuation after a lucrative round of funding

Paris-based AI-powered image editing app Photoroom recently confirmed its most recent round of funding at $43 million, which puts the business at a total valuation of $500 million, according to CEO and co-founder Matthieu Rouif. Names of big backers include Balderton Capital, Aglaé, and Y Combinator, the funds of which will be spent on Photoroom's R&D and infrastructure departments. The company also intends to pour a lot of funds into computing power and image rights for training its ground-up models.

​Nebius - Nebius is an AI-centric cloud platform built for the heaviest workloads with Nvidia's flagship H100 GPUs, whether it's for training large-scale machine learning workloads or generative AI models of your own.

​Synthesia - Turn text into studio-quality videos in minutes with Synthesia's AI-powered platform bundled with AI avatars and voiceovers in 130+ languages.

​Junia - Elevate your website's SEO with Junia AI by driving more organic traffic, improving visibility, and climbing rankings with quality, SEO-optimized, and ready-to-rank content.

​Flipner - Capture your ideas with the help of Flipner, a pocket-sized AI writing assistant and content hub that emphasizes your writing style.

Behind The Deepfake Mask (3-min read)

Meet the men behind the recent AI robocall that impersonated Biden, which urged Democrats not to vote in the New Hampshire primary, and uncover the different takes between the two.

​

AGI This Decade (61-min watch)

Join Dwarkesh Patel together with the CEO of Google Deepmind, Demis Hassabis, in a podcast as the two discuss the nature of intelligence, its scalability in artificial form, and the long-term repercussions behind it all.