The week in AI: ChatGPT gets a body & Cognition emerges from stealth with the first AI software engineer

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 this week: ChatGPT gets a body, text-to-video generator Sora to be publicly available in 2024, CEO Altman cleared in termination review

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

As usual, it was a busy news week for OpenAI.

GPT-powered robots: Just two weeks after an announcement that they were partnering with Figure, a promising robotics startup, a demo of the Figure 01 robot powered by OpenAI vision and language models has gone viral. Figure’s Senior AI Engineer Corey Lynch, who puts the robot through several tasks in a makeshift kitchen including identifying objects and throwing away trash, stated that the robot can describe its visual experience, plan future actions, reflect on its memory, and explain its reasoning verbally. He also emphasized that Figure 01’s behavior in the demo was learned, run at normal speed, and not controlled remotely.

Sora is coming this year: CTO Mira Murati confirmed in an interview with the Wallstreet Journal that OpenAI’s much hyped text-to-video generator, Sora, will be released in 2024. And, apparently, nude videos are not off the table: Murati says OpenAI is still “working with creators” to help determine the next steps for implementing ethical and safety considerations for video content creation.

“I’m not sure,” said Murati when asked about nudity. “You can imagine that there are creative settings in which artists might want to have more control over that. Right now we are working with artists, creators from different fields to figure out what’s useful, what level of flexibility should the tool provide.”

Sam Altman is back on OpenAI’s expanded Board of Directors: On December 8, 2023, OpenAI commissioned an independent a review of the events concerning the removal of Sam Altman and President Greg Brockman from the OpenAI Board of Directors, as well as Altman’s temporary termination as CEO. Unsurprisingly, the review has cleared Altman and Brockman of any wrongdoing. “We have unanimously concluded that Sam and Greg are the right leaders for OpenAI,” stated Bret Taylor, Chair of the OpenAI Board. Altman will rejoin the Board, which is adding three new members and making changes to the company’s governance structure - including a new set of corporate governance guidelines and a whistleblower hotline.

Inflection upgrades popular chatbot Pi with Inflection-2.5, a GPT-4 caliber LLM

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

Inflection AI, the AI company by Google DeepMind's Mustafa Suleyman and LinkedIn's Reid Hoffman, has just made its entrance into the “GPT-4 class” of AI models. Their new large language model, Inflection-2.5, competes with GPT-4 across benchmarks to build the industry’s best personal (versus ‘assistant’) AI. Inflection’s free AI chatbot Pi is now running on the upgraded LLM.

Inflection’s mission is to create a personal AI for everyone, and the company has talked about building the largest cluster of GPUs. Things are looking good so far. Pi currently has over 6 million monthly average users, and Inflection claims the upgraded model does 94% of what GPT-4 can do using far less training power. The model’s outputs are high quality - smart, occasionally a little bit quirky - and Inflection-2.5 performs very well on benchmarks in areas like coding, math, and general IQ.

There’s a demand for personal AI chatbots, not just work assistants, and Pi is growing 10% month over month. Pi has a voice output function and now also has access to real-time info from the web. It’s available on all platforms - browser, desktop app, iOS, and Android.

Europe’s world-first AI rules get final approval from lawmakers, despite criticisms

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

On Wednesday, the European Parliament approved the world's first comprehensive legislative framework to govern AI, with 523 votes in favor, 46 against and 49 votes not cast. The EU AI Act categorizes AI technologies based on their risk levels - from "unacceptable" risks warranting a ban to high, medium, and low hazards - and aims to foster innovation while ensuring the protection of fundamental rights. The act is anticipated to be fully enacted by the end of May, following the European Council's endorsement, with its provisions set to be gradually implemented starting from 2025.

The EU AI Act's passage is a critical response to growing concerns over AI's potential for abuse and its implications for competition and consumer rights within the tech industry. While the AI Act aims to direct AI development towards human-centric and ethical uses, critics have noted a number of human rights concerns. The Act leaves a number of loopholes for secretive biometric mass surveillance, predictive policing, and very broad ‘national security’ exemptions.

• Some teachers are now using ChatGPT to grade papers

• The women in AI making a difference

• Microsoft’s AI Copilot for Security launches next month with pay-as-you-go pricing

• Employees at top AI labs fear safety is an afterthought, report says

• Five of this year’s Pulitzer finalists are AI-powered

• Welcome to the Valley of the Creepy AI Dolls (here’s an alternate article on the same dolls for readers paywalled by WIRED)

• OpenAI’s GPT is a recruiter’s dream tool. Tests show there’s racial bias

• Midjourney bans all Stability AI employees over alleged data scraping

MaxAI.me - Do more Faster with 1-Click AI

MaxAI.me lets you chat with GPT-4, Claude 3, Gemini 1.5. You can also perfect your writing anywhere, save 90% of your reading & watching time with AI summary, and reply 10x faster on email & social media.

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

TECHNICAL, RESEARCH & OPEN SOURCE

Introducing Devin, the first AI software engineer

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

Cognition, an AI startup backed by Peter Thiel's (PayPal, Palantir) Founders Fund, has emerged from stealth to announce the launch of "Devin" - a fully autonomous AI software engineer. Cognition claims Devin is capable of handling some entire development projects end-to-end, from writing code and fixing bugs to final execution. That would be a significant advance over current-gen AI coding assistants like GitHub Copilot and Sourcegraph Cody - these assistants act more like copilots and are very limited across end-to-end capabilities. With Devin, Cognition AI appears to be offering a fully-fledged AI worker to handle entire projects.

Devin operates through a chatbot-style interface, where users provide natural language prompts, and the AI plans and executes the necessary steps within its sandboxed developer environment. It can write its own code, fix issues, test, and report progress in real-time, enabling engineering teams to delegate projects and focus on more creative tasks.

Cognition is currently offering early access to select users and plans to expand availability at a later stage. Cognition hints that coding is "just the beginning," suggesting potential AI workers for other disciplines in the future.

A chat with the SVP of Brand at Groq, the company betting that their Language Processing Units (LPUs) are > GPUs for AI

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

Two weeks ago, AI chip startup Groq went viral on X as demos of GroqChat (which allows users to query specific LLMs and receive even a very lengthy response almost instantly) surfaced across the platform. Groq’s user numbers have since then shot up from under 10,000 developers to over 370,000 - with tens of thousands on the API waitlist.

Groq is the brainchild of Google’s TPU (Tensor Processing Unit) lead inventor Jonathan Ross and looks poised to make a big mark on the AI chip industry with their LPUs’ blazing fast inference speeds. In this Cerebral Valley interview, Groq SVP of Brand Mark Heap tells the founding story of Groq, explains why LPUs outperform GPUs, and talks about the difficulties stemming from Groq’s viral moment.

MORE IN T/R/OS

• SIMA from Google DeepMind: Google shifts focus from individual game-playing AI’s towards a general, instructable video game-playing AI agent

• Meta unveils the two massive AI training clusters being used to train Llama 3

• OpenAI Superalignment team open sources Transformer Debugger (TDB)

• How AI is disrupting the demand for software engineers: data from 20M job postings

• Extropic is building analog-thermal chips for AI

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