OpenAI's Operator is its 1st true AI agent

In partnership with

Hey folks, this time:
1️⃣ World record on text generation with large AI models
2️⃣ AGI by 2025? AI Reasoning Breakthrough (TTT) & OpenAI’s Next Moves, ..

Especially, in part 2️⃣:

• What is TTT?

• What is ARC?

• What is LoRA?

• What is OpenAI’s new app use function and Operator?

(✨ If you don’t want ads like these, Premium is the solution. It is like you are buying me a Starbucks Iced Honey Apple Almondmilk Flat White a month.)

SambaNova Holds Speed World Record on Llama 3.1 405B

200+ tokens (220 words) per second out of the large Meta model with 405B parameters. This is almost a page a second. Why does it matter? 👇

Meet the Reconfigurable Dataflow Unit (RDU)—a groundbreaking AI processor.

It has a unique architecture:

• On-chip SRAM for rapid data access

• High Bandwidth Memory (HBM) for massive throughput

• DDR DRAM for extensive storage

This three-tier memory system is a game-changer.

Effortlessly runs Llama 3.1 - 405B.

The RDU excels in speed, efficiency, precision, and performance.

It outperforms traditional GPU systems by:

↳ Reducing redundant memory access

↳ Enhancing data movement

↳ Most importantly, by achieving high inference speeds

SambaNova’s RDU is not just about speed.

It's about redefining AI capabilities.

The faster you can influence, the more functionality and features you can put into your AI service, meaning search, voice, app access, and more.

See my DEMO how to experiment with it:

I’ve covered SambaNova before, plus other AI tech features and personal stories that don’t make it into the newsletter. If you’re interested, follow me on X for more updates! → https://x.com/musiol_martin

AGI by 2025? AI Reasoning Breakthrough (TTT) & OpenAI’s Next Moves, Others are likely to follow

As I mentioned in the last episode, the timeline for reaching AGI (Artificial General Intelligence) is getting shorter. Here’s where things stand:

• Sam Altman (OpenAI): Predicts AGI by 2025

• Dario Amodei (Anthropic): Predicts AGI around 2026/27

FYI: OpenAI’s team is getting back to peak performance.

Meanwhile, something interesting dropped — a new paper titled “The Surprising Effectiveness of Test-Time Training for Abstract Reasoning” (TTT).

(Source)

It’s a big deal because it introduces a novel method to significantly improve AI’s reasoning skills.

What is Test-Time Training (TTT)?

It’s a method that enhances an AI model’s performance on challenging tasks in real time. Here’s why it’s groundbreaking:

• Performance Boost: TTT achieves up to 6x better accuracy on the ARC dataset (a tough benchmark), reaching 61.9% on public validation sets.

What is ARC? 
The ARC dataset (Abstraction and Reasoning Corpus) is a collection of challenging tasks designed to test an AI’s ability to reason abstractly. Instead of relying on huge amounts of data, it requires the model to understand patterns and solve problems like a human would — without explicit rules or prior examples. It’s considered one of the toughest benchmarks because it tests general intelligence, not just pattern recognition.

• TTT performs:

  • In-Context Learning: Outperforms traditional end-to-end methods.

  • Data Augmentation: Relies heavily on geometric transformations to fine-tune models.

• Training Innovations:

  • Per-task LoRA Updates: Custom updates for each task perform better than using shared updates.

  • Self-Consistency: Applies multiple transformations, uses a voting mechanism, and refines predictions.

What is LoRA? 
LoRA (Low-Rank Adaptation) is a technique for fine-tuning large AI models efficiently. Instead of retraining the entire model, it adds small, lightweight layers that adapt to new tasks without changing the core model. This reduces memory and computation needs, making fine-tuning faster and more efficient.

Why TTT is a Game-Changer:

• Minimal Data Requirements: It uses just geometric transformations for fine-tuning, so there is no need for massive new datasets.

• Model Agnostic: Works well across different models and model sizes.

• Integration: It combines well with other methods for top results on complex tasks.

OpenAI’s Roadmap to AGI

OpenAI has shared a five-level roadmap leading to AGI.

And speaking of upgrades, the ChatGPT app has fantastic features.

It’s not just a chatbot anymore — it can use your apps, too. Here’s what’s new:

• Enhanced Search: It can now browse the web. See the episode on it. 

• Voice Mode: You can speak with it in a realistic voice. → Episode.

• Code Interpreter (Canva): A built-in environment for coding.

• 🆕 App and Terminal Access: It can interact with your apps and even your Terminal, giving it control over your computer.

I will make a hands-on demo/ training video for that you can access as a premium subscriber. But here is already the OpenAI Announcement and demo.

This leads to OpenAI’s next big project: Operator, the first real AI agent from OpenAI.

What We Know About “Operator”?

Launching in early 2025, Operator is OpenAI’s true autonomous AI agent, capable of tasks like coding and booking travel. It aligns with Sam Altman’s vision of agentic systems as the next major leap in AI.

But keep in mind: OpenAI isn’t the only player here—Anthropic and Google are developing similar capabilities.

Demo: Build a secure, offline AI tool in under 35 minutes.

In this mini-series, I show you how to create an offline AI tool to analyze bank statements, using my frameworks. You’ll learn how to build distributable tools that leverage OS LMs.

No coding expertise is needed—just a clear plan and the ability to ask the right questions to guide your AI through the process. (repost)

Four short videos

1/4:

See the other videos of the demo

That’s a wrap! I hope you enjoyed it.

Martin

• Follow me on X.com.

• Do you write newsletters? I use Beehiiv and highly recommend it.

• AI for your org: We build custom AI solutions half the market price, and time (building w/ AI Agents). Contact us to know more.

• Would you like to sponsor a post?

• My book - Generative AI: Navigating the Course to AGI.

• Generativeai.net 

You might like our last episodes: