Cosine's AI coder shatters expectations

Read time: under 4 minutes

Welcome back, Superhuman

As the AI industry holds its breath for news about OpenAI's "Strawberry" project, a new AI coder has burst onto the scene, threatening to steal the spotlight. Find out why Cosine’s autonomous software engineer could get us one step closer to AGI.

Today’s Insights

• A record-setting coding assistant

• Tutorial: Interact with a PDF using AI

• Decoding ancient epics with AI

• Everything else you should know today

• 5 new AI tools to boost your productivity

• AI-Generated Images: Espresso gadgets

NEXT IN AI

Cosine unveils record-setting AI software engineer

Cognition made waves in March with a coding-focused LLM called Devin that scored a record-setting 13.9% on the industry-leading SWE-Bench. Well, get ready to extend your graph. A Y Combinator-backed startup called Cosine claims to have more than doubled that score by teaching an autonomous software engineer to think just like a human.

Here’s how it works: Most AI models are fine-tuned through trial and error — taking random guesses until they happen to land on the right answer. But the UK’s Cosine, which just raised $2.5M in seed funding, thinks it’s come up with a better approach: “We believe that if you want a model to behave like a software engineer, it has to be shown how a software engineer works,” CEO Alistair Pullen said.

The details: 

• Cosine shows its Genie model real-world examples of coders working through problems

• By understanding the logic behind each decision, Genie can start to figure out how to navigate coding problems on its own

• The experiment is paying off: Genie scored 30% on the SWE-Bench, which assesses how well LLMs perform across different coding tasks; that’s a whopping 10 points ahead of the former leader, Factory AI’s Code Droid 

• GPT-4 doesn’t stand a chance in comparison: Genie performs about 2,196% better than OpenAI’s state-of-the-art model

Why it’s important: Code is the scaffolding behind the websites and apps we use daily. Genie can already fix glitches, build new features, and automate repetitive coding tasks. The next step: software that can essentially create, edit, and improve itself, unlocking the door to runaway growth.

A fully autonomous software engineer might also be the hidden key to achieving AGI. That’s because coders need to constantly work through difficult, multi-step problems — a capability that, if mastered, would be a major step toward reaching human-like intelligence.

PRESENTED BY AE STUDIO

From Idea to AI Solution Instantly

Think about the biggest challenge you’re facing at work right now. Got it? AE Studio’s new AI tool will help you solve it.

Here’s how it works:

1. Answer 3 questions about your business needs.

2. The AI churns out proven solutions.

AE Studio is the quintessential business problem solver. They once taught an AI to brew beer and market it — it sold out. True story.

If you’re fed up with the same problems at work, try AI ideas by AE Studio.

Give it a try here.

THE AI ACADEMY

How to “talk” to your PDFs using AI

• Go to Humata AI’s website and sign up.

• Upload your long PDF document and wait for it to get uploaded.

• Once uploaded, click on the Ask button on the right.

• You’ll be redirected to the chatbot that can answer all queries about your document.

• Enter your query, press enter, and it will give you the answer instantly while highlighting the relevant section in your PDF.

You can use the various features of Humata AI to summarize your PDFs, answer questions about your PDFs, extract key information from your PDFs, and more.

Get more tutorials →

AI & HISTORY

How AI is helping decode an ancient epic

Generative AI isn’t just shaping the future — it’s also revolutionizing how we study the past. Historians are already using machine learning to help with one especially daunting task: Piecing together the Epic of Gilgamesh, an ancient Mesopotamian story that dates back 3,000 years. 

Assyriologists have recovered thousands of clay tablets engraved with excerpts from the poem, but it’s so far been impossible to piece them all together to form a cohesive narrative. It’s estimated that about a third of the narrative remains a mystery, according to the New York Times.

How it works: Since 2018, a team at the University of Munich has used machine learning to match up 1,500 fragments from the epic, which is considered one of the first-known works of literature. They’ve already uncovered 100 lines that had previously been shrouded in mystery. 

That, in turn, gives us a better picture of today’s major religions, which may have been heavily influenced by the story — including one passage that tells of a global flood and a man who survives by building an ark. The same technology is also being used to interpret and decode other pieces of historic texts, like medieval music fragments and a hymn to the ancient city of Babylon.

PRESENTED BY REMOTE

Hire top-class talent from anywhere in the world (easily)

With Remote, you can hire, pay, and manage full-time/contract workers in any country (even where you don’t have a legal entity).

You get the employees you need, Remote handles the payroll, benefits, taxes, stock options, and compliance–it’s that simple.

Create an account and score 15% off service fees for one year.

AI & TECH NEWS

Everything else you need to know today

• Unlikely Partners: Nvidia is teaming up with the state of California to train 100,000 students, developers, and data scientists on how to use a variety of advanced AI tools.

• Fueling the Hype: Perplexity CEO Aravind Srinivas appeared to suggest that the pro version of his platform is already running OpenAI’s hyped “Strawberry” technology.

• Bang for Your Buck: According to Bloomberg, even the stripped-down version of Apple’s smartphone — the iPhone SE — is expected to feature Apple Intelligence.

• Safe Haven: Meta has signed a multi-year agreement to protect Universal Music Group artists from “unauthorized AI-generated content” on its platforms.

• EU Uproar: Nine European countries have issued complaints against Elon Musk’s X for allegedly using posts to train Grok without first receiving users’ permission.

🧠 Brain Food: Researchers at Ontario’s University of Waterloo are working on an AI model that can analyze video footage to determine the portion size of someone’s meal. The tool may one day be used to evaluate the nutritional content and calorie count of food in real-time — guiding users toward healthier lifestyles.

PRODUCTIVITY

5 AI Tools to Supercharge Your Productivity

✅ Scispace: Chat with PDFs, explore new papers, and discover concepts with an all-in-one AI tool for students and researchers.

✅ Omnifact: Give your team access to generative AI while maintaining control over your data.

✅ AICamp: Use internal knowledge, speed up workflows, and build AI assistants tailored to your needs.

✅ Yescribe: Automatically transcribes audio and video into text, helping you focus on what’s really important.

✅ Salesify: Speed up your sales cylce with AI-driven insights and coaching.

PS: Want more? Check out our Top 100 AI Tools.

^{* indicates a promoted tool, if any}

PROMPT OF THE DAY

Act as an Advertiser

Prompt: I want you to act as an advertiser. You will create a campaign to promote a product or service of your choice. You will choose a target audience, develop key messages and slogans, select the media channels for promotion, and decide on any additional activities needed to reach your goals. My first suggestion request is "I need help creating an advertising campaign for a new type of energy drink targeting young adults aged 18-30."

You can adapt the prompt to your specific needs.

_{Source: @devisasari on GitHub}

AI-GENERATED IMAGES

Draw me like one of your French Presses

Midjourney Prompt: gouache painting [insert coffee machine name here, ex: moka, chemex,french press, etc.] stamped with stamped with "Espresso", simple minimal, high-quality,
--ar 2:3 --v 6.1 --stylize 30

Acquire new customers and drive revenue by partnering with us

Superhuman is the world’s biggest AI newsletter for businesses and professionals with 600,000+ readers working at the world’s leading startups and enterprises. Companies like Amazon, Hubspot, and Salesforce feature their products in Superhuman. You can learn more about partnering with us here.

🧞Your wish is my command

Thanks for reading.

Until next time!

Zain & the Superhuman AI team

p.s. If you liked this newsletter, share it with your friends and colleagues here.