Do AI concepts need new ways to promote them?

The time to imagine AI is advancing.

After OpenAI also started the AI ​​revolution with its own o1 consider the example launched in September 2024 – which takes a long time to answer questions but with the benefit of high performance, especially for complex problems, mathematical and scientific problems – the commercial field of AI is full of copies and competitors.

There is DeepSeek is R1, Google Gemini 2 Flash Thinkingand just today, LlamaV-o1both of which aim to provide a similar “idea” to OpenAI’s new o1 and upcoming o3 families. These models are compatible “chain-of-thought” (CoT) activation – or “self-motivation” – forcing them to think about their analysis in the middle, to double-check, to check their work and in the end they will find a better solution than just shooting outside. the settings as fast as you can, just like other major languages ​​(LLMs) do.

However the higher cost of o1 is o1-mini ($15.00/1M access tokens vs. $1.25/1M access tokens for GPT-4o on API for OpenAI) has caused some people to be unhappy with what they claim to benefit from. Is it worth paying 12X as much as a well-known, top-notch LLM?

As it turns out, there are many converts – but the key to unlocking real imaginations may lie in the user making them different.

Shawn Wang (founder of AI news service Resin) displayed on its title Property weekend guest post from Ben Hylak, formerly of Apple Inc., developer of visionOS (which powers the Vision Pro spatial computing headset). The post has been successful because it clearly explains how Hylak promotes the OpenAI o1 model to receive valuable (for him) outputs.

In short, instead of writing what the user writes on the o1 model, they should think about writing a “summary,” or a detailed description that contains a lot of information about what the user wants the model to produce, who the user is. what kind of information do they want the model to produce for them.

As Hylak writes Property:

With many models, we are taught to describe the model as we want it to respond. for example, ‘You are a professional programmer. Think slowly and carefully“

This is different from how I have found success with o1. I don’t advise them how – only what. So let o1 take control and plan and plan his ways. That’s what independent thinking is all about, and it can be a lot faster than you can manually check and chat like a “vulnerable person”.

Hylak also includes a nice screenshot of the o1 model that produced good results for the tour series:

This blog post was so helpful, OpenAI President and co-founder Greg Brockman also shared it on his X account with message: “o1 is a different kind of genre. Success requires using a new way of communicating with regular chats.”

I tried myself on my desire again and again to learn to speak Spanish well this was the resultfor those who want to know. It may not be as exciting as Hylak’s speed and response, but it shows solid potential.

On its own, even when it comes to LLMs as unimaginative as Claude 3.5 Sonnet, there may always be room to improve their motivation for better, more modest results.

As Louis Arge, former Teton.ai engineer and current creator of the neuromodulation device openFUS, wrote on X“One way I’ve found is that LLM relies on their own interests more than what I can say,” and he gave an example of how he convinced Claude not to be “cowardly” by starting to “fight” with him in the first place. on the results.

All of which goes to show that agile engineering remains an important skill as the AI ​​era continues.