Analysing Magic’s “Coworker” Breakthrough: Active Reasoning and the Race to AGI

Updated on February 26 2024

Magic AI (, a stealth startup founded in 2021, has made waves with claims that it has achieved a major technical breakthrough in AI, achieving “active reasoning” capabilities – a significant step towards artificial general intelligence (AGI) – a capability similar to OpenAI’s secretive Project Q*(Q-STAR) model that supposedly got OpenAI CEO, Sam Altman fired last year.

Why Nat Friedman Invested In Magic AI

In February 2024, former GitHub CEO Nat Friedman revealed he and his partner invested $100 million in Magic after seeing demos of its AI technology and being very impressed. According to sources, Magic has created a new type of large language model able to process massive amounts of data and context.

Specifically, Magic claims its model can handle 3.5 million words of text input, 5 times more than Google’s LaMDA model. This virtually unlimited context window allows Magic’s AI to process information more like humans do.

Most importantly, sources say Magic’s breakthrough enables “active reasoning” capabilities comparable to OpenAI’s secretive GPT-3.5 model known as Q*.

Why “Active Reasoning” is a Big Deal in AI?

Active reasoning involves an AI system using logical deduction to solve novel problems it hasn’t been explicitly trained on. This allows the system to apply general principles rather than just pattern recognition which is primarily the base of existing Google and OpenAI models.

Current large language models like GPT-3/4 are limited to correlating inputs with training data. They cannot truly reason deductively or adapt dynamically to new situations.

In contrast, active reasoning would allow AI systems to make logical inferences, understand underlying relationships and “think” more like humans. This brings AI significantly closer to AGI.

Magic’s apparent achievement of active reasoning, combined with its massive context window, suggests a potentially game-changing advance in AI capabilities.

Also Read: Why Amazon invested in Anthropic

Can Magic Outpace Google and OpenAI with “Coworker”?

Magic’s technical breakthrough seems surprisingly fast, considering research giants like Google and OpenAI have yet to definitively attain active reasoning.

Magic claims to have been working on a “coworker” than just a “copilot” and active reasoning is the key to this transition where copilots are trained on pattern recognition. working on coworker

Several factors may help explain Magic’s rapid progress:

Funding and Talent Acquisition

The $100 million investment by seasoned experts provided Magic ample resources to pursue ambitious AI research.

Magic also hired several former senior engineers from Big Tech firms like GitHub and Meta, gaining valuable talent.

Novel Neural Architecture

Magic has hinted its AI architecture is more advanced than standard transformers used in GPT-3/4 and similar models.

The startup claims to have developed “something with a multi-million context window” that is not based solely on transformers. This mysterious architecture likely provides advantages.

One possibility is Magic is using Mamba, an architecture that claims to outperforms widely used transformer model on various long-sequence tasks. Mamba was published in a paper just months before Magic’s announcement.

Focus on Active Reasoning

Magic’s founders have extensive experience researching how to make AI models reason better. The startup seems laser focused on enabling active logical reasoning.

With enough resources and talent devoted to this singular objective, Magic may have simply out-researched larger rivals.

Less Public Scrutiny

As a small private company, Magic attracts far less scrutiny of its AI safety practices compared to regulated Big Tech firms like Google and OpenAI.

With fewer public expectations to slow it down, Magic may have progressed quicker by sacrificing safety assurances. This possibility is concerning.

Will Magic Trigger Moloch Trap in AI Development?

Magic’s goal is to build safe superintelligent AI. However, its rapid unilateral progress heightens worries that profit-driven competition in AI could lead to dangerously uncontrolled outcomes.

Magic’s investors were confident enough in its capabilities to provide $100 million in funding. This incentivizes Magic to commercialize its technology quickly rather than methodically.

Once Magic releases products embedding its advanced AI, competitors like OpenAI may feel forced to recklessly unleash their own systems just to keep up.

These dynamics reflect a “Moloch trap” – competitive pressures compelling organizations to sacrifice safety for speed in developing transformative AI.

Without coordination between companies, we may end up in a “race to the bottom” that leads to the uncontrolled emergence of artificial general intelligence with catastrophic consequences.

To avoid this, experts emphasize the need forAI safety standards, oversight frameworks and cooperation within the AI research community.


Magic AI’s remarkable accomplishment of active reasoning, if accurate, represents a watershed moment in the pursuit of AGI. While the latest investments by Nat Friedman is going to add speed to Magic’s effort towards building a “coworker”, mighty rivals like Google and OpenAI are not going to fall behind. With Sora making waves and Google releasing their largest context window model, Gemini 1.5, superintellgence doesn’t seem to be far away. These are both exciting, desperate and watchful times for all tech enthusiasts like us.

Frequently Asked Questions

How does Magic’s technology compare to OpenAI’s Q* model?

OpenAI’s secretive Q* The model is said to be capable of active reasoning, but very few details are known. Based on limited information, Magic seems to have achieved similar functionality to Q*, especially the ability to logically work through novel problems.

Is Magic’s AI really smarter than Google and OpenAI LLMs?

Magic’s system appears to have a substantially larger context window than any other known LLMs. In this aspect, it seems more capable than Google or OpenAI’s publicly known models.

About Appscribed

Appscribed is a comprehensive resource for SaaS tools, providing in-depth reviews, insightful comparisons, and feature analysis. It serves as a knowledge hub, offering access to the latest industry blogs and news, thereby empowering businesses to make informed decisions in their digital transformation journey.

Related Articles