• Posts
  • META's new OPEN SOURCE Coding AI beats out GPT-4 | Code Llama 70B

META's new OPEN SOURCE Coding AI beats out GPT-4 | Code Llama 70B

PLUS: Privacy Concerns about ChatGPT, Pressure on Tech Giants and more...


META's new OPEN SOURCE Coding AI beats out GPT-4 | Code Llama 70B

Today, Facebook/Meta dropped a bombshell - Code Llama, a massive 70 billion parameter language model, their biggest and best yet. It's the latest in the Code Llama lineup, which includes the base model, a Python-tuned version, and the 'Instruct' model tailored for natural language instructions. 

Code Llama's a game-changer for coders and learners alike. It's designed to speed up workflows, make coding more accessible, and improve code quality. Plus, it's free for both research and commercial use under a community-friendly license. 

OpenAI's ChatGPT breaches privacy rules, says Italian watchdog

Italy's privacy watchdog, Garante, notified OpenAI that ChatGPT may violate data protection rules. This follows their previous move to ban ChatGPT due to EU privacy concerns, although the service resumed after OpenAI made changes. Despite this, the watchdog continued its investigation and believes there may be privacy breaches. 

OpenAI, backed by Microsoft, has 30 days to defend itself. This situation highlights the growing scrutiny AI technologies like ChatGPT face under EU's strict data protection laws, especially with new AI regulations on the horizon.

OpenAI partners with Common Sense Media to collaborate on AI guidelines

OpenAI teamed up with Common Sense Media, a group that judges if media and tech are kid-safe. They'll work on AI rules and stuff to help parents and teachers understand AI like ChatGPT. OpenAI will make some chatbots in their GPT Store that are good for families, using Common Sense's standards. OpenAI's boss, Sam Altman, said this move will make sure families and teens can use their AI tools safely.

Common Sense is also working on a way to rate AI apps, like a "nutrition label," to show how they're used and if they're safe or not. They found that lots of teens use ChatGPT, but their parents don't know much about it. So, they want to teach families how to use AI without any bad surprises.

Essential AI chooses Google's Cloud

Essential AI, a startup founded by ex-Googlers, picked Google Cloud for its AI tech. This is big news because big players like Google, Microsoft, and Amazon are all trying to attract startups to use their clouds instead of doing it themselves. Essential AI will use Google's fancy chips, the TPU v5p, for their AI models and build on Google's global setup. 

The company was started by Ashish Vaswani and Niki Parmar, who are big names in AI research. They've got backing from some heavy hitters like March Capital and Thrive Capital, plus tech giants like AMD, Google, and Nvidia. Google Cloud's big boss, Thomas Kurian, is all for this partnership, seeing it as a boost for digital transformation using AI.

China approves over 40 AI models for public use in past six months

In the last six months, China has greenlit over 40 AI models for public use, a move signaling its push to rival U.S. AI advancements. This surge includes approvals for 14 large language models (LLMs) last week, with big players like Xiaomi and 4Paradigm on the list. Since August, Beijing requires AI models to get regulatory thumbs-up before public launch, showing China's aim to keep a tight leash on AI tech.

Companies like Baidu, Alibaba, and ByteDance were among the first to get the nod. The AI race heated up post-ChatGPT's 2022 success, with China's own versions like Baidu's Ernie Bot hitting 100 million users. China's AI model count stands strong, trailing just behind the U.S. in global share.

Google and Microsoft Are Under Pressure to Disclose AI Earnings

Tech giants Microsoft and Alphabet (Google's parent company) are under the gun to prove that their heavy investments in artificial intelligence (AI) are more than just hype. Investors are eager to see hard numbers showing how AI is boosting their bottom line. This demand mirrors the early days of cloud computing when big players like Amazon and Microsoft eventually had to disclose their cloud sales figures.

 While AI is a big part of their operations, it's not easy to pinpoint exact AI-driven sales. Microsoft's recent $10 billion investment in OpenAI and the launch of AI tools like Copilot for Windows and Office suggest they might be ahead in the game. Google's playing catch-up, with investments in AI initiatives and new products like the Gemini language model. Meanwhile, tech companies are also leaning on layoffs as another financial strategy. The pressure's on for these tech giants to show that their AI bets are paying off big time.

US Wants Cloud Firms to Flag Foreign Users in China AI Race

The US government is pressing big cloud companies like Amazon and Microsoft to keep tabs on foreign clients, especially those from China, who are using their platforms for AI development. This is part of a broader tech tussle with Beijing. Under the new plan, these companies must dish out names and IP addresses of these foreign customers. They're also expected to sniff out and report any fishy business. 

This move is all about preventing sneaky cyber moves and comes from a recent proposal by the Biden administration. It's a big ask for these tech giants, who now have to figure out how to track and report this info.


Code Llama by Meta is a beefed-up version of Llama 2, specially trained for coding tasks. It can whip up code and chat about it, responding to both code and plain language prompts. You'll find it in three sizes (7b, 13b, 34b) and styles, including a Python-savvy model. It's a whiz with popular programming languages like Python, C++, and Java.

"SliceGPT" by Microsoft offers a smart fix for big language models, cutting their size by chopping rows and columns. This tweak keeps almost all of their smarts while using less memory and computing power. It's like slimming down these brainy models without losing their muscle.

"Qwen-VL" by Alibaba is a cutting-edge AI capable of understanding and describing complex visuals and texts. It outshines previous models in recognizing and interpreting images and texts, even in Chinese. Its abilities include celebrity and landmark identification, poetry creation, and complex reasoning from visual inputs like diagrams and charts. Additionally, it excels in real-world applications like driving advice and Python coding. This tool sets new standards in multimodal AI with its advanced capabilities in both visual perception and cognitive understanding.

"From GPT-4 to Gemini and Beyond" evaluates the performance of Multi-modal Large Language Models (MLLMs), like OpenAI's GPT-4 and Google's Gemini, across four modalities: text, code, image, and video. It focuses on their generalizability, trustworthiness, and causal reasoning capabilities, presenting findings from a qualitative study. This research aims to close the gap between current MLLM capabilities and public expectations, enhancing the reliability of these models in various applications. The detailed analysis covers both proprietary and open-source MLLMs, providing insights into their strengths and limitations. 

 "Generative Expressive Robot Behaviors using Large Language Models," by Google Deepmind explores using large language models (LLMs) to create expressive robot behaviors. It aims to make robots display human-like expressive actions, like nodding or saying "excuse me," in various social situations. Unlike previous rule-based methods, this approach leverages LLMs' rich social context and motion generation ability, translating human language instructions into parametrized robot control code. User studies and simulations show this method results in competent, understandable robot behaviors. 

 "Learning Universal Predictors" by Google explores the limits of meta-learning through the concept of Solomonoff Induction (SI). Using Universal Turing Machines (UTMs) for generating diverse training data, it assesses how neural networks, including LSTMs and Transformers, can be trained for universal prediction strategies. This approach leverages meta-learning to its extremes, aiming to create neural networks with broad pattern recognition and problem-solving abilities. The paper provides both theoretical and experimental insights into this innovative approach to meta-learning.

"TIP-Editor: An Accurate 3D Editor Following Both Text-Prompts And Image-Prompts" by Tencent introduces a 3D scene editing framework called TIP-Editor. This tool uniquely combines text and image prompts with a 3D bounding box for editing, providing precise control over the appearance and location of objects in a 3D scene. It features a stepwise 2D personalization strategy for better scene and reference image representation, and employs a localization loss for accurate object placement. TIP-Editor's 3D Gaussian splatting allows effective local editing while preserving the background, demonstrating superior editing quality in various tests.


Dubbing Studio - lets you nail translations and timing in your videos, tweaking accents and tones, making global sharing a breeze.

Keep It Shot - a Mac app that uses AI to rename screenshots with descriptive titles and creates a fast, private, offline search index for easy retrieval using keywords.

Seeing AI - AI app, designed for the visually impaired, has launched on Android with new features and language options, expanding accessibility and enhancing user experience.

Listening - an app that converts research papers into audio, allowing users to listen on the go. It features lifelike voices, easy pronunciation of technical terms, and lets users select specific sections to hear. 

Descript - an all-in-one tool for easy video and podcast editing. It features AI-driven transcription, voice cloning, and audio improvements, making editing as simple as working with a text document.



Nvidia’s AI partners are also its competition.

Nvidia, a big tech player, helps run AI for companies like Microsoft and Amazon. But these companies are trying to not rely only on Nvidia, aiming to easily switch to other options, including their own chips. OpenAI's boss is even thinking of making their own chips. Microsoft and Amazon are launching their own AI chips soon. THE VERGE

Pentagon plans AI-based program to estimate prices for critical minerals

The Pentagon is developing an AI-based program to estimate prices and predict supplies of critical minerals like nickel and cobalt. This initiative, part of a broader effort to boost U.S. production of these minerals, aims to enhance market transparency and challenge the traditional pricing structures set by futures markets and pricing agencies. The program, managed by DARPA, will use AI to determine a metal's "structural price" based on various factors, including production location and labor costs. REUTERS

FBI recruits Amazon Rekognition AI to hunt down 'nudity, weapons, explosives'

The FBI plans to use Amazon's Rekognition AI service to analyze legally obtained images and videos, identifying nudity, weapons, explosives, and other details. Named Project Tyr, this initiative is still in its early stages. While Amazon previously banned police from using Rekognition for facial recognition, this project doesn't violate that ban as it's focused on object detection rather than facial analysis. The FBI's use of Rekognition raises concerns about surveillance and privacy, especially given the lack of federal laws limiting the use of such technology.  THE REGISTER

Shortwave email client will show AI-powered summaries automatically

Shortwave, an email client developed by ex-Google engineers, is enhancing its AI features. It now automatically provides one-sentence summaries at the top of emails or threads, a step up from its previous manual summary generation. For more detail, users can tap the summary to get a longer version, powered by GPT-4 Turbo. TECHCRUNCH

Nvidia’s Big Tech Rivals Put Their Own A.I. Chips on the Table

Amazon, Google, Meta, and Microsoft are building their own AI chips, aiming to reduce their reliance on Nvidia, which dominates the AI chip market. This shift is motivated by the growing importance of generative AI and the high demand for specialized chips, a market where Nvidia has excelled but struggled to meet demand. THE NEW YORK TIMES

The Copyright Office, historically a small part of the Library of Congress, is now at the center of a major debate over the application of copyright law to AI technology. With AI systems feeding off creative content, traditional copyright norms are being upended. The office is conducting a review on how centuries-old copyright laws should apply to AI, with tech giants and content creators presenting their cases. THE NEW YORK TIMES

Cap VC wants to be the AI-powered ‘operating system’ for VCs

Cap VC, a startup emerging from a venture capital (VC) fund, is launching an AI-powered tool aimed at revolutionizing how VC firms operate and make investment decisions. Recognizing inefficiencies in the VC industry, particularly the challenge of handling complex PDF files, Cap VC has developed a system to convert unstructured data from financial documents into structured data. TECHCRUNCH

DOJ’s Healthcare Probes of AI Tools Rooted in Purdue Pharma Case

The U.S. Department of Justice (DOJ) is investigating the healthcare industry’s use of artificial intelligence (AI) in patient records, focusing on potential violations of anti-kickback and false claims laws. This scrutiny stems from the use of AI to recommend treatments to doctors, which may be influenced by pharmaceutical and digital health companies. These investigations are rooted in the case against Purdue Pharma and its contractor, Practice Fusion. They were involved in a scheme to design automated pop-up alerts that encouraged doctors to prescribe addictive painkillers. This case set a precedent for how AI tools in healthcare could lead to problematic outcomes, even as they hold potential for significant medical advances. BLOOMBERG LAW

Tomorrow.io’s radar satellites use machine learning to punch well above their weight

Tomorrow.io, previously known as ClimaCell, has achieved significant advancements in weather forecasting with its two radar satellites, Tomorrow R1 and R2. These satellites, launched in April and June of the previous year, utilize machine learning to analyze weather patterns, demonstrating capabilities that rival much larger and traditional weather forecasting technologies. TECHCRUNCH

AI steps up in healthcare: GPT-3.5 and 4 excel in clinical reasoning

In a recent study published in npj Digital Medicine, researchers explored the capabilities of large language models (LLMs) like GPT-3.5 and GPT-4 in simulating diagnostic clinical reasoning. The study assessed these models' diagnostic reasoning using open-ended clinical questions from the MedQA USMLE dataset and the New England Journal of Medicine (NEJM) case series. The goal was to understand how well these models could mimic the clinical reasoning skills of healthcare professionals, an important step for integrating AI into clinical care. Results showed that GPT-4 was more accurate than GPT-3.5, with GPT-4 demonstrating 76% to 78% accuracy across different reasoning prompts. However, in the NEJM dataset, GPT-4 achieved a 38% accuracy with conventional chain-of-thought (CoT) versus 34% with differential diagnosis reasoning prompts. MEDICAL LIFE SCIENCES


What'd you think of today's edition?

Login or Subscribe to participate in polls.

What are MOST interested in learning about AI?

What stories or resources will be most interesting for you to hear about?

Login or Subscribe to participate in polls.

Join the conversation

or to participate.