Devin, the new cool AI developer

Along with : GPT 4.5 Turbo leaked & Grok going open-source

Hey there, 

With everything happening in the AI world - Musk's legal issues with OpenAI, EU regulations, and constant updates- Devin stands out. This autonomous AI software engineer builds and deploys apps without help!

What kinds of self-reliant AI could we see in the future? Stay till the end to read some of the thoughts I am having :)

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection 

Table of Contents

Devin is an AI Autonomous Software Engineer created by Cognition AI. It is equipped with a shell, code editor, and browser within a sandboxed computed environment, just like a human developer would have.

Devin provides real-time progress reports, accepts feedback, and collaborates on design choices, enabling a collaborative work experience for the user. 

Here are some of Devin's impressive capabilities:

  • It learns new technologies by reading resources like blogs.

  • It builds and deploys apps from start to finish.

  • It autonomously finds and fixes bugs in codebases.

  • It can fine-tune its own AI models.

  • It addresses bugs and feature requests in open-source repositories.

  • It has even completed some real jobs on Upwork.

The results? As per Cognition - Devin correctly resolves 13.86% of issues end-to-end, significantly surpassing the previous state-of-the-art of 1.96% on the SWE bench. (source)

A post recently sparked interest among the AI community regarding the upcoming GPT-4.5 Turbo model by OpenAI. The company inadvertently published a blog post about the GPT-4.5 Turbo model, which was indexed by Bing and DuckDuckGo.

This incident piqued the curiosity of many, leading them to attempt accessing the blog. However, the page was de-indexed, resulting in a 404 error, and was removed from search engine listings.

Despite this, details of the leak spread throughout the AI community, especially on Reddit and X . It appears that OpenAI may release the model by August, before the next-generation GPT-5 model, due to this accidental leak.

The forthcoming GPT-4.5 Turbo model is anticipated to support 256K tokens, doubling the capability of the current GPT-4 model, which supports only 128K tokens. (source)

Elon Musk has announced that the AI chatbot, GrokAI, will soon be going open source. By joining the ranks of models like Llama, Mistral, and Claude, Grok will become part of a rapidly expanding group of Open-source AI models. Previously, it was exclusive to X users for a subscription fee of $16.

The specifications, final release date, and components of the chatbot are yet to be announced. Let's see what innovation and capabilities GrokAI will unveil with its open launch. (source)

Several months after being removed from his position as CEO and losing his seat on the board of directors, OpenAI announced that Altman would be rejoining the company's board. 

Three new members will also join the board: Sue Desmond-Hellmann, the former CEO of the Bill and Melinda Gates Foundation; Nicole Seligman, the former president of Sony Entertainment; and Fidji Simo, the CEO of Instacart. This will bring the total number of board members to eight.

Additionally, according to a report by the New York Times, Mira Murati, OpenAI's chief technology officer, might have played a significant role in Altman's dismissal. Murati is said to have written a private memo to Altman expressing concerns about his management and sharing these concerns with the board. (source)

Microsoft now offers its Copilot GPT-4 Turbo model for free, shifting from its previous subscription model. Announced by Mikhail Parakhin on Twitter, users can activate GPT-4 Turbo in Copilot by selecting Creative or Precise mode.

This move is speculated to clear the path for the upcoming GPT-4.5 Turbo model, illustrating a strategy where Copilot Pro members first receive the latest tech, which is then made available to the broader public, balancing free access with premium benefits. (source)

Recently, Figure AI released a video on its YouTube channel showcasing its humanoid robot not only responding to the conversation but also performing tasks according to instructions which is impressive.

Figure AI has recently incorporated a voice feature into its robots, named Figure 01; this voice capability is powered by OpenAI's ChatGPT.

The neural networks in Figure AI enable fast, precise, and dexterous robot actions. The company has stated that integrating these with OpenAI models offers advanced visual and language intelligence.

The European Parliament has given final approval to the “AI Act” that will govern the use of AI in Europe.

The AI Act imposes significant penalties for non-compliance with the prohibited systems provisions, with fines of up to €35 million or 7% of global turnover.

The rules encompass high-impact, general-purpose AI models and high-risk AI systems, requiring compliance with specific transparency obligations and adhering to EU copyright laws. (source)

Union cabinet greenlights IndiaAI initiative

IndiaAI, under the Digital India Corporation, is set to lead the "India AI Mission" following the Union Cabinet's approval of a Rs 10,371.92 crore budget for five years.

The mission is designed to build a comprehensive AI ecosystem -  promoting innovation, improving computing access and data quality, developing local AI skills, and attracting talent.

A key feature of the mission includes setting up a large-scale computing facility with over 10,000 GPUs, in collaboration with the private sector.

Additionally, it plans to launch an AI marketplace to provide services and pre-trained models to innovators, serving as a pivotal resource for AI development. (source)

Midjourney has launched a new feature titled "Character Reference," designed to create consistent characters across multiple reference images, addressing a challenge many users have faced.

Now, users can utilize prompts and adjustments to generate character images, controlling the degree to which the new image resembles the original character.

Moreover, In the advanced options of Midjourney's web alpha version, users can drag or paste an image, selecting it as a prompt, style reference, or character reference. (source)

Inflection has launched its latest model, Inflection 2.5, which the company claims is competitive with other leading large language models (LLMs) such as GPT-4 and Gemini.

Here are some of the notable updates in Inflection 2.5:

  • It approaches the performance of GPT-4 while using only 40% of the compute for training.

  • It can now perform real-time web searches to provide users with up-to-date information.

  • It incorporates IQ capabilities comparable to other leading LLMs.

Inflection 2.5 merges raw capability with a user's distinct personality traits and an empathetic touch. It is available on pi.ai. (source)

Command-R is a scalable generative model that excels in Retrieval-Augmented Generation (RAG) and Tool Use tasks. It is specifically designed for large-scale production workloads in enterprise settings.

  • Command-R has several key advantages:

  • High accuracy in RAG and Tool Use tasks.

  • Low latency and high throughput for efficient processing.

  • Extended context length of 128k tokens and competitive pricing.

  • Strong capabilities across 10 major languages.

  • Availability of model weights on HuggingFace for research and evaluation.

Cohere, the company behind Command-R, has taken an enterprise-centric approach, working closely with businesses to customize its models to meet their specific requirements. (source)

  • OpenAI has launched a new tool to help developers debug and understand small language models. The transformer debugger is an open-source tool that allows users to explore the inner workings of these models and investigate their behavior. 

  • An experimental "meta" prompt to help users create effective prompts for Claude, streamlining task-specific AI interactions. Accessible via a Google Colab notebook with an API key, this tool aids in generating custom prompts, serving as a beginner's guide or a method to test various prompt variations easily.

Tool: OpusClip

Problem: It's time-consuming to turn long videos or podcasts into engaging social media snippets.

How to use it:

  1. Visit OpusClip and create an account.

  2. Paste the URL of your video (from YouTube, Zoom, Streamyard, etc.).

  3. Tell OpusClip the content type, your preferred clip length, and the video's main topic.

  4. Click "Generate video" (takes about 15 minutes).

  5. OpusClip creates multiple scored snippets for you to choose from.

We used OpusClip on our 'Leading with Data' content. It generates shareable video snippets, but we found it can have difficulty isolating very specific topics, even when we mention them during the creation process.

  • For those interested in image processing and AI model training, this article from Stability AI explores and benchmarks compute solutions, particularly focusing on comparing the performance of Gaudi 2 accelerators and Nvidia A100 GPUs in image processing and running AI models.

  • In a recent Leading with Data episode, I spoke with Sarabjot Singh from Tatras Data about his early work on recommendation systems and the origins of Tatras Data. We also explored his company's blend of traditional tech with Generative AI and his key advice for aspiring Generative AI professionals.

  • I discovered an engaging post by Rajat Monga, the lead developer of TensorFlow at Google, that excellently summarizes the computing advances of the past decade and possible future directions. I plan to feature him in an upcoming episode of "Leading with Data" - if you have any questions for him, please share.

  • The course by DeepLearning.AI offers participants how to leverage knowledge graphs within RAG applications. Suitable for anyone looking to deepen their understanding of knowledge graphs and their applications in AI, and for those who wish to enhance the performance of LLMs with structured, contextual data.

It’s not that the promise of AI automatically delivering software has not been discussed before - it is the pace of development that could be remarkable. Imagine bots like Devin releasing software updates 24 x 7. Today - most apps on mobile make a 15-day release; software product releases would run in months. What happens when this shifts to daily updates?

Stretching the idea a little more - can we have real-time bug fixes on applications as soon as a user shares the crash report? Something to ponder over…

How do you rate this issue of AI Emergence?

Would love to hear your thoughts

Login or Subscribe to participate in polls.

Reply

or to participate.