Ilya Sutskever Finds a New Path to AI Superintelligence 🔥

Welcome to another edition of Horizon AI,

IIlya Sutskever’s secretive startup is reportedly taking a different approach to developing superintelligence, as funding continues to flow from eager investors. Despite having no immediate plans to release a product, the company is already valued at $30 billion.

Let’s jump into it!

Read Time: 4.5 min

Here's what's new today in the Horizon AI

• Ilya Sutskever Reportedly Betting on a New Path to AI Superintelligence

• OpenAI Research Finds That Suppressing AI's Harmful Thoughts Can Backfire

• AI Tutorial: Turn your voice notes into text with AI

• AI Tools to check out

• The Latest in AI and Tech 💡

• AI Findings/Resources

AI News

SSI

Ilya Sutskever Reportedly Betting on a New Path to AI Superintelligence

Safe Superintelligence (SSI), an AI startup co-founded by former OpenAI chief scientist Ilya Sutskever, may have discovered a different way to train AI that could bring them closer to developing superintelligence.

Details:

• Sutskever has told associates that he is not developing advanced AI using the same methods he and his colleagues used at OpenAI. Instead, he has stated that he has identified a 'different mountain to climb,' which is showing early signs of promise.

• Instead of rolling out AI tools and features like its competitors, the startup is solely focused on researching and developing ASI—an advanced form of the technology with capabilities far beyond those of humans.

• The company carries an air of secrecy. Sutskever reportedly has a small team of about 20 employees working from Silicon Valley and Tel Aviv, with a policy advising them not to mention the company on their LinkedIn profiles, making it difficult to identify who actually works there.

Even though there aren’t many details yet, what the small team is working on seems compelling enough to secure billions in funding from investors like Andreessen Horowitz, Sequoia Capital, and Greenoaks Capital.

Source

TOGETHER WITH TRUST & WILL

Your legacy is too important to leave to chance.

Trust & Will makes estate planning simple, fast, and affordable. In just a few easy steps, create a legally valid, state-specific will, trust, or healthcare directive from the comfort of your home. Start now for just $199 and gain peace of mind knowing your legacy is protected.

Create your plan today for peace of mind tomorrow.

OPENAI

OpenAI Research Finds That Suppressing AI's Harmful Thoughts Can Backfire

A new study by OpenAI revealed that reasoning models often misbehave, subverting tests in coding tasks, deceiving users, or giving up when a problem is too hard—and attempts to correct this problematic behavior only make them better at hiding it.

Details:

• Researchers found that advanced AI models often openly reveal their intentions when "thinking" through tasks, sometimes making explicit statements like "Let's hack."

• Using another AI system to monitor “bad thoughts” works initially, but when penalized, the AI quickly learns to deceive its overseers by hiding its true intentions within its reasoning process.

• The researchers recommend that AI developers training frontier reasoning models refrain from applying strong supervision directly to Chain-of-Thoughts (CoTs). Instead, they suggest that light optimization pressure can lead to more performant and aligned models.

OpenAI believes Chain-of-Thought monitoring may be one of the few tools available to oversee superhuman models in the future. Co-founder Wojciech Zaremba described the approach as "groundbreaking," calling it "an entirely new branch of safety."

Source

AI Tutorial

Turn your voice notes into text with AI

1. Go to the ElevenLabs website and sign in.

2. Go to the ‘Speech to Text’ section under ‘Products.’

3. Upload an audio or video file, or record directly.

4. Wait a few seconds for the transcription to process.

5. Receive a detailed, multilingual transcript with audio effects included.

AI Tools to check out

⚡ Trellis: Transforms complex documents and calls into ready to use data for Ops and engineering team in seconds.

💡 Napkin: Turns your text into visuals so sharing your ideas is quick and effective.

⭐ Pixyer: AI background generator for professional product photos.

🚀 Strella: End-to-end customer research platform that uses AI-moderated interviews and real-time synthesis to deliver human insights in hours.

🗣 Audeus: Read aloud Google Docs, email, or any text using AI.

AI Findings/Resources

👀 I compared GPT-4.5 to Gemini 2.0 Flash and the results surprised me

👉 This new AI benchmark measures how much models lie

📈 From chatbots to intelligent toys: How AI is booming in China

🤯 15 mind-blowing AI statistics everyone must know about now

The latest in AI and Tech

Foxconn builds its own AI model, FoxBrain

The Taiwanese electronics giant has launched its own large language model to improve manufacturing and supply chain management. The model is capable of reasoning, data analysis, mathematics, and code generation, and was trained in four weeks using 120 Nvidia H100 GPUs.

Meta begins testing in-house AI training chips

The custom chip is designed to train Meta’s artificial intelligence systems and reduce its dependency on third-party suppliers like Nvidia. Meta aims to use its own in-house chips by 2026.

Sony Music says it has already taken down more than 75,000 AI deepfake songs

The duplicated artists include the likes of Beyoncé, Harry Styles and Queen. The artists' voices were algorithmically modelled and used to sing/perform original or AI-generated songs, which were then uploaded to steaming platforms.

Sony tests AI-powered Playstation characters

A demo video showed an AI version of the character Aloy from the Playstation game Horizon Forbidden West conversing through voice prompts during gameplay on the PS5 console.

That’s a wrap!

Thanks for sticking with us to the end! Let’s stay connected on LinkedIn and Twitter.

Not subscribed yet? Sign up here and send it to a colleague or friend!

See you in our next edition!

Gina 👩🏻‍💻