OpenAI introduces ChatGPT Enterprise

Plus: Tesla activates $300m Nvidia AI chip cluster, puts 'Dojo' among world's most powerful supercomputers

Author

The Dispatch
August 29, 2023

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each weekday, we scour the web to aggregate the many stories related to artificial intelligence; we pass along the news, useful resources, tools or services, technical analysis and exciting developments in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on under the hood in AI.

Good morning. Today in AI:

  • OpenAI announces ChatGPT Enterprise

  • Tesla’s supercomputer gets an AI-chip turbocharge

  • The EU’s Digital Services Act is now in effect

  • 62% of code generated by GPT-4 contains API misuses

  • Trending tools & more

The Story: OpenAI just launched ChatGPT Enterprise, the most powerful version of ChatGPT tailored for business use. It offers enhanced security, customization, and productivity features compared to the free consumer version. Early adopters include Zapier, Canva, and PwC.

More Details:

  • ChatGPT Enterprise does not use customer data for model training and conversations are encrypted, addressing enterprise security concerns. It is SOC 2 compliant.

  • There is no usage cap, 2x faster performance, and 4x longer context windows to process more complex prompts. Advanced data analysis features (rolled up from Code Interpreter) help teams quickly extract insights.

  • Administrative controls like SSO, domain verification, and usage analytics enable large-scale deployment across organizations. Shared prompt templates allow teams to collaborate on common workflows.

  • API credits are included for fully custom solutions. Companies like Asana and Klarna claim time savings and productivity gains from early use cases.

Takeaways: OpenAI claims over 80% of Fortune 500 companies are now using ChatGPT, and they’re clearly making an effort to highlight data sovereignty in this release announcement. There are a number of useful features to help teams be more productive and creative by harnessing the power of GPT-4 in a collaborative effort. GPT-4 has never had a 32k context window before, so this will be a game changer for the model.

From Tom’s Hardware: On the heels of Elon Musk’s live stream of a fully self driving Tesla, the company has reportedly just activated a 10,000x Nvidia H100 GPU cluster alongside their supercomputer, ‘Dojo’. This combined supercomputer will train Tesla's fleet of vehicles and process data from them. By next year, Dojo will be among the most powerful supercomputers in the world.

More details:

  • Tesla's new AI cluster will employ 10,000 Nvidia H100 compute GPUs, which will offer a peak performance of 340 FP64 PFLOPS - higher than the 304 FP64 PFLOPS offered by Leonardo, the world’s fourth highest-performing supercomputer.

  • Tesla is using the H100 system alongside its custom Dojo supercomputer to bolster overall training capacity. Nvidia cannot meet current AI chip demands, and Tesla has designed a hybrid system that can be powered by both in-house and other cutting edge tech.

  • Dojo's specialized high-performance chips allow customized optimization for Tesla's needs - the combined systems will provide unprecedented scale for training fully autonomous vehicles.

  • Tesla plans to invest $4 billion by the end of 2024 specifically for autonomy computing.

Takeaways: Details to be verified. This $300m bet on Nvidia’s AI chips shows a commitment to the power of computing that’s akin to gambling: Tesla is betting that they can create error-free autonomous vehicles through brute-force AI training at scale. The live stream highlighted how Tesla’s approach to fully-self-driving vehicles has transitioned from hard-coded engineering to neural networks and massive amounts of training data. This new AI cluster and Dojo are how they plan to leverage this approach.

Google's Creative Lab team has collaborated with rapper Lupe Fiasco to build an AI experiment called TextFX - an AI-powered suite of tools aimed at writers, wordsmiths and rappers. The project explores how large language models (LLMs) like Google's PaLM can be applied to creative writing tasks.

The TextFX web app contains 10 different textual "effects" powered by few-shot learning prompts fine-tuned for PaLM. For example, there are prompts to generate similes, alliterations, and acronyms based on input text/lyrics. The prompts were iteratively developed and optimized using MakerSuite. The code is open source and prompts are available on MakerSuite for replication (users may need to join a waitlist). One of the biggest benefits of/use cases for using language models is creative brainstorming, so we’re excited to see the future of this project.

The European Union's sweeping new Digital Services Act (DSA) is now in effect. Under the new rules, online platforms (including Google, Amazon, Facebook) must implement ways to prevent and remove posts containing illegal goods, services, or content, while simultaneously giving users the means to report this type of content.

The DSA bans targeted advertising based on sensitive categories and restricts ads targeted at children. Violations could result in fines up to 6% of global revenue. Platforms like Google, Meta, TikTok, and Snapchat are already rolling out changes to comply, including expanded ad libraries and chronological feeds. The DSA will likely have global implications as firms adjust policies.

Comically, in advance of the DSA going into effect, Amazon filed a petition asking the EU to reevaluate its classification as a “very large online platform”. Amazon is the largest e-commerce platform in the world several times over.

From our sponsors:

Get smarter in 5 minutes with Morning Brew (it's free)

There's a reason over 4 million people start their day with Morning Brew - the daily email that delivers the latest news from Wall Street to Silicon Valley. Business news doesn't have to be boring...make your mornings more enjoyable, for free.

The ROBUSTAPI benchmark covers Java APIs

The Story: Researchers from University of California San Diego have found that large language models might generate code that compiles and runs, but still contains API misuse that could lead to bugs or vulnerabilities when deployed. This is a dimension of code quality that has been overlooked in existing benchmarks that focus on functional correctness.

More details:

  • The researchers created a new benchmark, ROBUSTAPI, to evaluate API misuse in LLM code generation. It has 1208 real Stack Overflow questions on 24 Java APIs that are commonly misused.

  • The benchmark checks generated code snippets against 40 validated API usage rules using static analysis on the abstract syntax tree.

  • Experiments on GPT-3.5, GPT-4, LaMDA, PaLM show high rates of API misuse, even with relevant examples. Highest misuse rate was 70% in GPT-4.

  • GPT-4 had the most compilable responses, but a higher misuse rate than GPT-3.5, showing advanced coding ability ≠ robust code. Relevant examples reduced misuse rates for some models, showing they can learn correct API usage.

Takeaways: This might not be surprising to some, but even state-of-the-art LLMs generate ‘sound code’ that could lead to bugs and issues when deployed. The benchmark methodology here is good; the analysis of different conditions like zero-shot vs relevant examples is insightful. What the research truly highlights is that we need our benchmarking tools to develop and improve, concurrently with LLMs, as coding assistants.

Reinforcement learning from human feedback (RLHF) is a primary method of training AI models, but is computationally expensive. Researchers from Google DeepMind have proposed Reinforced Self-Training (ReST) as a more efficient approach. ReST iterates between a ‘Grow’ phase where the model generates sample data, and an ‘Improve’ phase where a subset of samples is used to fine-tune the model offline.

ReST improved translation quality over baseline models in tests. It’s more efficient than standard reinforcement learning, and allows researchers to inspect data to diagnose issues. ReST works by sampling and scoring model outputs, providing a simple way to align large language models. The DeepMind researchers present ReST as a general technique for aligning models using offline reinforcement learning. ReST does not require complex training - just sampling and scoring.

Trending AI Tools & Services:

  • Code Llama: integrated with HuggingFace spaces - there are three sizes (7b, 13b, 34b) as well as three flavors (base model, Python fine-tuned, and instruction tuned)

  • QR Code AI: Customized, functional QR codes with AI generated imagery

  • Granica: A developer-first efficiency platform - optimizes AI pipelines by reducing data storage costs, ensuring privacy, and enhancing performance

  • AIModels.fyi: Describe your problem, get an AI that can solve it

  • FineShare: Explore a vast library of 100+ AI voice models and create song covers with your favorite AI vocals in just one click

  • Debunkd: Verify statements and check for AI generated images

Guides/lists/non-AI:

Social media/video/podcast:

  • Chris Fernandez, founding CEO of EnsoData, solves your sleep problems... with AI [Podcast]

  • This week in data: The real cost of generative AI [VentureBeat Video]

  • DeepMind-like gaming AI: incredible driving skills! [YouTube]

  • What's wrong with LLMs and what we should be building instead - Tom Dietterich - VSCF2023 [YouTube]

  • (Discussion) Hot take: people who find ChatGPT not useful either haven't used it enough or have an ego problem [Reddit]

  • Counting cars on a CPU with GPU-level power. With just a few lines of Python code, you can now run powerhouse AI models like YOLOv5 [X]

Did you know? 

A mysterious new ‘city’ near San Francisco backed by major Silicon Valley investors is facing growing scrutiny, according to a New York Times report. Led by former Goldman Sachs trader Jan Sramek, the secretive company Flannery Associates has spent over $800 million acquiring thousands of acres of farmland in Solano County to build an experimental metropolis. While details remain scarce, backers include LinkedIn's Reid Hoffman, venture capitalist Marc Andreessen, and billionaire VC Michael Moritz.

According to the report, the firm wants to use the land to ‘rethink government and how buildings are constructed’. But Flannery Associates now faces lawsuits from local residents and zoning laws as its land grab comes to light. Despite big tech funding and the AI boom, the futuristic city's development faces substantial hurdles. This controversial project’s viability remains uncertain.

❝

Elon desperately wants the world to be saved. But only if he can be the one to save it.

Sam Altman, CEO of OpenAI on Elon Musk, August 2023