. . AI: Startup Says It's Built an Automated Software Engineer (3.12.24)

Your daily digest of the top stories in AI.

Friends, today's edition covers breakthroughs, strong moves, and big fears. These are the stories the AI community is talking about most.

If you find these summaries valuable, we hope you’ll think of someone you can forward them to.

Now here’s today’s news,

-Marshall Kirkpatrick, Editor

Cognition Labs Secures $21M, Launches AI Engineer 'Devin'

First impacted: Software engineers, AI developers
Time to impact: Short

Cognition Labs launched its AI software engineer, Devin. Devin's performance on the 'SWE-bench' benchmark (a platform that simulates real-world GitHub issues) was reported as resolving 13.86% of the issues end-to-end, a significant leap from the previous state-of-the-art of 1.96%. The model is also capable of executing complex engineering tasks and learning from its mistakes and has shown advancements in problem-solving. Devin has been tested in real-world scenarios, successfully completing tasks on Upwork, including writing and debugging code for a computer vision model and compiling a report on the resulting data. We highly recommend you check out the video in the link for more info! [Introducing Devin, the first AI software engineer] Explore more of our coverage of: AI Development, Series A Funding, Software Engineering. Share this story by email

Musk to Open Source AI Tool Grok

First impacted: AI developers, Business owners
Time to impact: Short

Elon Musk says that Grok, an AI tool run by X, will be open sourced this week. The decision is potentially influenced by Elon's critique of OpenAI for not being open sourced as well. [via @elonmusk] Explore more of our coverage of: Elon Musk, AI Tools, Grok Launch. Share this story by email

Government Report Flags AI as Potential 'Extinction-Level' Threat

First impacted: Policy makers, AI researchers
Time to impact: Medium

A report by Gladstone AI, commissioned by the U.S. government, warns of potential national security risks posed by AI, even suggesting it could be an extinction-level threat to humanity. The report suggests policy changes such as limiting AI training power, requiring government approval for advanced AI use, banning the sharing of AI model specifics, tightening control over AI chip production and export, and directing federal funds towards making advanced AI safer. Many people were upset about this report; Eric Hartford of Abacus.AI for example tweeted, "Wow @TIME, what a bunch of clowns! 🤡 What's next, an AI prohibition? AI speak-easies? Bring it on!" [U.S. Must Act Quickly to Avoid Risks From AI, Report Says] Explore more of our coverage of: AI Security Risks, Government AI Regulation, AI Threat Assessment. Share this story by email

Meta Invests Heavily in AI Infrastructure

First impacted: AI engineers, AI researchers
Time to impact: Medium

Meta plans to bolster its AI infrastructure with the addition of two 24k GPU clusters, designed to support existing and future AI models, including Llama 3. By the end of 2024, the company aims to expand its infrastructure to include 350,000 NVIDIA H100 GPUs, which it says will bring its total computing power to nearly the equivalent to 600,000 H100s. Former Meta AI team member Natalia Burina shared the article and added, "What’s mindblowing about @AIatMeta is that everything is built in house down to the hardware and data centers. This means an unprecedented amount of control over AI products and a major competitive advantage." [Building Meta’s GenAI Infrastructure] Explore more of our coverage of: Meta AI Infrastructure, GPU Clusters, AI Storage Solutions. Share this story by email

New Attack Extracts Information from AI Models

First impacted: AI researchers, Data scientists
Time to impact: Short

Researchers have reportedly created a new technique that can pull out detailed data from AI language models like OpenAI's ChatGPT and Google's PaLM-2, even managing to recover the embedding projection layer of a transformer model for less than $20. The study's authors (mostly from Google DeepMind but including one co-author from OpenAI) say they've also extracted the full projection matrix of OpenAI's Ada and Babbage language models, exposing their hidden dimensions, and they believe it would take fewer than 2,000 queries to recover the entire projection matrix of the gpt-3.5-turbo model. [Paper page - Stealing Part of a Production Language Model] Explore more of our coverage of: AI Language Models, Data Extraction, OpenAI Research. Share this story by email

That’s it! More AI news tomorrow!