Human-Like AI Phone Calls That Convert #buildinpublic

Censorship is not an excuse that Open AI and other LLMs’ prices are higher than DeepSeek. All we want is good product with low price.

DeepSeek Chat is officially live on NewOaks AI!

DeepSeek chat is officially live on NewOaks AI!

Wow

The conclusion of this Google paper is very clear: "SFT is responsible for memory, RL is responsible for generalization" Simple conclusion: Supervised fine-tuning (SFT) is like showing students a large number of examples and answers. Students learn by imitating the examples. Reinforcement learning (RL) is like letting students solve problems by themselves, giving rewards for correct answers and penalties for incorrect answers. Students learn through trial and error and summarizing patterns. The researchers designed two tasks to test the model: General Points card game: This is a card arithmetic game that requires using four cards to make the target number 24. The researchers tested: Rule changes: Change the numerical rules of J, Q, and K to see whether the model has learned the arithmetic rules or just memorized the solutions under specific rules. Visual Variation: Change the color of the cards to see if the model can still recognize the cards despite the visual changes. Virtual Reality Navigation (V-IRL): Navigating in a virtual city using instructions and street view images. The researchers tested: Rule variation: Using different ways of giving directions (e.g., absolute direction "north" vs. relative direction "turn left"). Visual changes: Test in cities the model was not trained on to see if the model can recognize landmarks and navigate in new visual environments. Key findings: Reinforcement Learning (RL) is the generalization champion! Across all tasks, models trained with RL excel at adapting to new rules and visual environments. They learn the underlying principles of arithmetic and navigation and are able to handle situations they have never seen before. Supervised fine-tuning (SFT) tends to memorize. Models trained with supervised fine-tuning tend to memorize the training data. They perform well on tasks similar to the training data, but performance drops dramatically when the rules or visual context change. They are essentially recalling patterns in the training data rather than truly understanding the task. Reinforcement learning (RL) improves the model’s visual recognition capabilities. Interestingly, RL training even improves the model’s ability to recognize objects in images, which helps in the virtual navigation task. This suggests that RL can improve the model’s basic visual understanding capabilities. Supervised fine-tuning (SFT) is still a good helper for reinforcement learning (RL). Although reinforcement learning is better in generalization, supervised fine-tuning is still useful. It can help the model initially understand the instructions and give responses in the right format. This makes it easier for reinforcement learning to further fine-tune the model on this basis to achieve better performance. “Thinking time” is crucial for reinforcement learning (RL). Giving the model more “thinking time” (adding verification reasoning steps in reinforcement learning training) can further improve the model’s generalization ability. Application suggestions: If you want an AI model that can truly understand and adapt to new situations (generalize), reinforcement learning (RL) is a better training method. It teaches the model how to learn and how to solve problems flexibly. If you just need an AI model to perform well on tasks that are very similar to its training data, then supervised fine-tuning (SFT) might be sufficient. But it is likely to struggle when circumstances change even slightly. Think of reinforcement learning (RL) as giving your AI a “brain” that can think for itself, while supervised fine-tuning (SFT) is like giving it a “cheat sheet.” In the long run, the “brain” is obviously more powerful.

FREE TRIAL: Increase Engagement with Conversational AI Voice Agent: via @YouTube

Dario is deeply feared by DeepSeek. That is all I can see. Control=Fear

With all due respect, closed mind does not get you the final win. DeepSeek is complete open source LLM. Its victory will belong to Human Species, not like Close AI or Close Claude

Do you agree?

It's hard to believe, but due to H100 restrictions, DeepSeek was forced to train R1 manually, with thousands of Chinese citizens holding flags to act as logic gates.

Picaso style snake, very cool!

@ylecun Compute is still the upper bound of AI

This is a fake DeepSeek account, scam

Agreed

RT @BarrettYouTube: @bindureddy "probably because it's just harvesting data". It's funny how people like @bindureddy think that because it'…

300% ROI within one month, that is the best customer testimonial for NewOaks AI!

DeepSeek is the Child

Which part is the most challenging part when you do it Greg?

Good article