VCG
·
gettyimages.comChatGPT New Models Are Out: They’re Not Strawberry But o1
Curated by
cdteliot
4 min read
593
3
OpenAI has unveiled its latest AI models, o1-preview and o1-mini, designed to enhance complex reasoning capabilities in areas like coding, science, and mathematics. As reported by TheWrap, these new models, part of OpenAI's "Strawberry" AI reasoning project, represent a significant leap forward in AI technology, with the ability to break down and solve intricate problems step-by-step.
Not One But Two
Contrary to prior reports, OpenAI o1, formerly codenamed "Strawberry", is not a single model but actually a family of models
1
2
. The first two have been available for Pro and Team subscribers on September 12, 20242
. They're called o1-preview and o1-mini1
. The former works like the typical GPT models before it, though it has more advanced self-checking abilities, while the latter is a more compact version aimed at coding3
4
. These will also be open to Enterprise and Educational subscribers sometime next week2
. Will there be more models added under the o1 umbrella? Possibly, but we still have to wait and find out.4 sources
Not Quite as Expected
openai.com
Everyone was looking forward to Strawberry because it's said to have an autonomous searching feature, meaning it can surf the web on its own and come up with accurate responses based on the latest and most relevant sources it scraped. Sorry to break your bubble, but you're about to be disappointed. OpenAI o1 can neither browse the Internet for information nor process the files you provide
1
. Talk about a letdown! It seemed to be able to analyze photos at the very least, but even that feature was disabled since OpenAI deemed it needed further testing2
. What is true, however, is that it's so much slower than GPT-4o but also more organized and smarter in a sense3
.3 sources
The Redeeming Point
OpenAI o1 may not have met most expectations, but it certainly satisfied the anticipation of its self fact-checking feature and logical algorithm that is better at problem-solving and breaking down complex tasks.
1
2
The Vice President at Thomson Reuters, Pablo Arredondo, confirmed that o1 is more capable of examining legal briefs and solving LSAT logic games than its predecessors. Ethan Mollick, a Wharton management professor, praised o1 for getting all the answers right to a difficult crossword puzzle. Moreover, it got 83% correct answers on the International Mathematical Olympiad qualifying exam, which is a great boost compared to the mere 13% GPT-4o score.1
3
OpenAI claims that their new o1 models are more competent in areas such as science, coding, language, and data analysis than GPT-4o, GPT-4, and GPT-3.5. OpenAI research scientist Noam Brown explains that o1 runs on "reinforcement learning" with "reasoning data", so it can think sharper albeit slower. With the way it works, you have to sacrifice time for accuracy because the longer it thinks, the more precise its answers are.4
5
5 sources
Double-Edged Sword
techxplore.com
You may think that o1's advanced thinking skills are great, but it has downsides too. First is the slow loading times that we were all informed beforehand. Like a real human, it needs time to ponder things and consider the next course of action. If you want a good reply from it, you have to sit and wait until it arrives at an answer.
1
2
Next, it has the tendency to overthink. Because it's designed to reason everything out, it can trip over itself trying to find the best solution to the simplest problem. Not everything has to be complicated, but o1 won't just let it go until you're frustrated. It refuses to settle on the easy way and insists on the twisty path.3
Lastly, it may give you detailed responses, but "detailed" can only go too far. OpenAI o1 may just add extra, unnecessary steps and insert information you didn't ask for. Its suggestions may become excessive, cluttered, or confusing, so always look out for that.4
5
5 sources
Is o1 Really the Best GPT Has to Offer?
The straightforward answer is no. Yes, it has its perks, but the cons far outweigh the pros. It is by no means that much better than GPT-4o. In fact, GPT-4o has more useful features than o1 and it's significantly cheaper too
1
2
. OpenAI o1 doesn't have GPT-4o's multimodal processing, ability to scour the Internet, and instant speed when creating replies3
. Even OpenAI confided on their website that they still recommend GPT-4o over o1 on most AI prompts3
. The pricing is vastly different as well, with o1 being very expensive. If you're playing around with it, you should keep an eye on your tokens. OpenAI charges $15 for every million input tokens and $60 for a million output tokens2
4
. If you ask us, it's not worth the cost unless you have specific uses for o1 in mind.4 sources
Closing Thoughts on OpenAI o1
The buzz was that Strawberry would be the closest thing to Artificial General Intelligence (AGI), but with o1's release, the collective realization is that it's still far from the ideal. The features people were looking forward to, such as autonomous Internet navigation and completely accurate answers, were basically nonexistent.
1
It can't scrape the web and it still commits errors and succumbs to hallucinations. GPT-4o is still the better choice for AI tasks since it has more capabilities than the latest models.2
Despite its shortcomings, we must also recognize o1's superiority when it comes to intelligence and reasoning. If you want an AI to guide you past logical problems, o1 may just pull you out of the mire and push back the headaches. Just make sure that you can afford it because o1 doesn't come for cheap.3
We recommend you only use it minimally while development is still ongoing. OpenAI o1 is still newly born and it will certainly improve in the months to come. Who knows? Those rumored features we're all excited about may be released in future updates. Let's not give a judgment yet and see where OpenAI plans to take o1.2
3 sources
Related
How does o1's reasoning capability compare to other AI models
What are the limitations of o1's current capabilities
How does o1's training data differ from that of GPT-4o
What specific safety protocols does o1 implement to prevent harmful responses
How does o1's performance on benchmarks like jailbreaks impact its reliability
Keep Reading
OpenAI is Training Next Model
OpenAI, a leading artificial intelligence company, has announced that it has begun training its next flagship AI model, which is set to succeed the groundbreaking GPT-4 technology powering ChatGPT. This development comes alongside the formation of a new Safety and Security Committee tasked with evaluating and improving OpenAI's processes and safeguards.
78,811
OpenAI's Strawberry: Everything You Need to Know
OpenAI is poised to release its latest AI model, codenamed "Strawberry," which promises to revolutionize reasoning capabilities in artificial intelligence. According to reports from The Information, this new model is designed to take a more deliberate approach to problem-solving, spending 10-20 seconds processing complex queries before delivering more accurate and thoughtful responses.
570
OpenAI Unveils o1 Model
OpenAI has unveiled its latest AI model, o1, previously code named "Strawberry." This model is designed to enhance reasoning capabilities in artificial intelligence. As reported by multiple sources, this new model series aims to tackle complex problems in science, coding, and mathematics by spending more time "thinking" before responding, mimicking human-like reasoning processes.
81,247
OpenAI Bans Reasoning Probes
According to reports from Ars Technica, OpenAI is threatening to ban users who attempt to uncover how its latest AI model reasons through problems. This move has raised concerns about transparency and the company's commitment to open-source AI principles.
12,681