The week in AI: A smarter, more direct ChatGPT

In partnership with

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence; we pass along the news, useful resources, tools and services, and highlight the top research in the field as well as exciting developments in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on in AI.

NEWS & OPINION

OpenAI makes ChatGPT smarter, more direct and less verbose for subscribers

-------------------------

It’s not GPT-5 yet, but OpenAI quietly announced upgrades to the GPT-4 Turbo model powering ChatGPT for subscribers on Twitter/X at the end of last week. The new model brings with it improvements in writing, math, logical reasoning and coding - in addition to a more up-to-date knowledge base. It was trained on publicly available data up to December 2023, in contrast to the previous edition of GPT-4 Turbo available in ChatGPT, which had an April 2023 cut-off.

“When writing with ChatGPT [with the new GPT-4 Turbo], responses will be more direct, less verbose and use more conversational language,” OpenAI wrote. Based on our own usage and experimentation with the updated model, it is indeed much more ‘straight to the point’, a welcome improvement.

The update - which follows the GA launch of new models in OpenAI’s API (notably GPT-4 Turbo with Vision, which adds image understanding capabilities to the normally-text-only GPT-4 Turbo) - arrives shortly after the model lost its crown as the best AI chatbot to Anthropic’s Claude 3 Opus on the LMSYS Chatbot Arena, a crowdsourced platform where users can evaluate LLM’s and rank their outputs. The new model is back to #1.

Apple to overhaul entire Mac line with AI-focused M4 chips

-------------------------

After peaking in 2022, Macintosh sales fell 27% in the last fiscal year. In the holiday period, revenue from the computer line was flat - and Apple is now preparing a major overhaul of its entire Mac computer line with a new family of processors. The M4 chips will have a strong focus on artificial intelligence capabilities to boost performance and enable new AI-driven features across Apple's products. The M4 lineup will include at least three main varieties - an entry-level version called Donan, more powerful models dubbed Brava, and a high-end chip codenamed Hidra.

Apple plans to release the updated M4-powered Macs beginning late this year and continuing into early 2025. The revamp will span the iMac, MacBook Pro, Mac mini, MacBook Air, Mac Studio, and eventually the Mac Pro. The M4 chips and associated software updates are part of Apple's big AI push to catch up with rivals and reinvigorate sluggish computer sales.

Microsoft invests $1.5B in UAE AI firm as US and China battle to exert AI influence in the Middle East

-------------------------

In a move attached to the escalating tech cold war between the United States and China, Microsoft has announced a $1.5 billion investment in G42, an influential AI firm based in the United Arab Emirates (UAE). The investment is not just financial but deeply strategic, aimed at securing a foothold in the Persian Gulf - a region where both superpowers are vying to expand their technology influence.

G42, known for its expansive AI and biotechnology capabilities, has had previous entanglements with Chinese technology and personnel, so the partnership raised eyebrows in Washington. Gina Raimondo, the commerce secretary, traveled twice to the Emirates to talk about security arrangements for the partnership. “When it comes to emerging technology, you cannot be both in China’s camp and our camp,” she said.

Microsoft's partnership will allow G42 access to its cutting-edge AI chips and cloud services. In exchange, G42 will phase out Chinese technology, including Huawei’s telecom equipment, from its infrastructure and commit to stringent security protocols vetted by the U.S. government. This agreement underscores a broader American strategy to curb Chinese technological incursion while bolstering its own tech diplomacy. Chinese leader Xi Jinping’s sees Washington as leading an all-out campaign of “containment, encirclement and suppression” of his country. But there’s not much he can do about it.

• Boston Dynamics unveils fully electric Atlas robot for real-world applications

• A 73 y/o congressman wanted to understand AI - so he went back to a college classroom to learn

• Introducing OpenAI Japan

• Stanford University’s AI Index Report 2024 with essential trends

• An interview with Google Cloud CEO Thomas Kurian about Google’s Enterprise AI strategy

• TikTok may add AI avatars that can make ads

• OpenAI fires two researchers over information leaks

• Perplexity AI: ‘Like Wikipedia and ChatGPT had a kid’ - inside the buzzy AI startup coming for Google’s lunch

• Generative AI is coming for healthcare, and not everyone’s thrilled

MaxAI.me - Outsmart Most People with 1-Click AI

MaxAI.me best AI features:

• Chat with GPT-4, Claude 3, Gemini 1.5.

• Perfect your writing anywhere.

• Save 90% of your reading & watching time with AI summary.

• Reply 10x faster on email & social media.

Join 1M+ users on MaxAI.me to do more faster!

TECHNICAL, RESEARCH & OPEN SOURCE

Google and Meta innovate to overcome context window limitations in LLMs

-------------------------

Researchers from Google and Meta this week have proposed major innovations to address context window limitations in large language models.

Context window limitations: Transformer models, widely used in language models, suffer from a critical limitation known as quadratic inefficiency. This means the computational resources needed grow excessively with the increase in input sequence length. This inefficiency makes it impractical to process long texts, such as entire books or lengthy conversations, in a single step, limiting the models' ability to understand and generate context-attuned responses.

Google's Infini-attention Model: Google’s team introduced "Infini-attention", which is integrated into Transformer models to efficiently manage infinitely long input sequences without overwhelming memory and computational resources. The model cleverly combines the traditional attention mechanism with a new technique called compressive memory. It stores previous computations in a compressed format, allowing the model to "remember" and "refer back" to much older data without the need to keep it all actively in memory. This mechanism supports both local and long-range dependencies within the text, crucial for tasks like summarizing lengthy documents.

By allowing continuous streaming of input data with minimal memory footprint, Infini-attention can handle extensive data sequences more efficiently, vastly improving on traditional models in both scalability and processing speed.

Meta's MEGALODON Architecture: The MEGALODON architecture includes technical enhancements like complex exponential moving averages and timestep normalization to optimize data flow. It processes data in chunks, applying attention mechanisms individually to each segment. This method not only mitigates the traditional quadratic complexity of full attention mechanisms but also facilitates efficient parallel processing, enhancing speed and scalability. Moreover, MEGALODON's versatility extends beyond text. Its architecture makes it suitable for a range of data types, including audio, video, and sensor data.

It’s amazing to see how much work has been done in the last year to expand context windows. It wasn’t long ago that ChatGPT, Bard, etc. were limited to 4k tokens and you couldn’t input more than about 5 pages worth of words without receiving an error message. Now, some of the most powerful models are in the hundreds of thousands of tokens per context window and there are no signs of slowing down.

Elon Musk’s xAI introduces Grok-1.5 Vision

-------------------------

xAI just introduced Grok-1.5 Vision, a multimodal upgrade to the open-source model that allows for processing visual information. Grok 1.5V can process documents, charts, screenshots and photos, with a focus on real-world understanding.

xAI created a new ‘RealWorldQA’ benchmark to evaluate spatial understanding, with Grok-1.5V outperforming GPT-4V and Gemini. Grok-1.5V will roll out to testers and existing users soon, with significant improvements across images, audio, and video expected in the coming months.

Perhaps in part because it was launched as just a ‘spicy’ chatbot alternative to ChatGPT, Grok feels a bit under-appreciated in the broader LLM discussion, and this impressive vision upgrade shows the open-source model is here to compete. With Elon’s arsenal of data across X and Tesla and a chip on his shoulder to compete with OpenAI, it might be time for the industry to start paying attention.

MORE IN T/R/OS

• OpenAI makes major updates to Assistants API

• Microsoft’s VASA-1: Bringing portraits to life with AI

• Holodeck: Language Guided Generation of 3D Embodied AI Environments

• ChatGPT can predict the future … when it tells stories set in the future, about the past

That’s all for this week! We’ll see you next Thursday.’