Just now, ChatGPT officially announced that its mathematical capabilities have been upgraded again. Netizens: I am finally proficient in addition and subtraction within ten.

Since the release of ChatGPT, its capabilities have been continuously unlocked by people, such as writing neural networks and making smart speakers. During the trial, people gradually discovered that mathematical ability is a major shortcoming of ChatGPT, and even simple "chicken and rabbit in the same cage" problems can be calculated incorrectly.

Probably with this in mind, ChatGPT has just announced an important update: improving "authenticity" and "mathematical capabilities."

This is the third update of ChatGPT since its launch in November last year. However, because the "update instructions" are too vague, people still need to go through it. A process of exploration of new capabilities.

A few days ago, Stephen Wolfram, a computer scientist and the father of the Wolfram Language, combined the science and engineering artifact Wolfram|Alpha with ChatGPT, injecting super computing knowledge into the latter to complement each other, and the effect is quite good .

So, can ChatGPT’s math ability after this update compete with it?

It seems... The comparison result is not satisfactory:

"Only It can be said that neural networks are not used for this." Sebastian Raschka felt helpless.

Some people also found that the upgraded ChatGPT "grew increasingly grumpy":

##"Which teacher taught you math?" Faced with a question about addition and subtraction within ten, his tone was very much like a parent helping his child with homework.

#Maybe this is an "accidental phenomenon"? It seems that mathematics is really difficult.

In any case, we can look forward to a subsequent wave of interesting demos.

So Volume: ChatGPT and its Competitors

“The next 6 to 12 months will bring an explosion of experiments once companies can use OpenAI 's API is built on top of ChatGPT. The killer use case that emerges may be around the impact of generative AI on knowledge management."

Nicola Morini Bianzino.

At a recent public event, Nicola Morini Bianzino, global chief technology officer of Ernst & Young, said that there is currently no "killer" use case for using ChatGPT in enterprises. But that could change soon, and he predicts the next six to 12 months will bring a lot of experimentation, especially once companies are able to build on ChatGPT using OpenAI’s APIs.

Bianzino describes the impact of generative AI on knowledge management as the “dialectics of AI.” "Knowledge companies tend to store knowledge in a very flat, two-dimensional way, which makes access, interaction and conversation difficult. We tried to build expert systems 20, 30, 40 years ago. It didn't go very well because they Too rigid. I think this technology has the potential to overcome many of the problems of expert systems," said Nicola Morini Bianzino.

At the same time, ChatGPT competitors are also emerging, and the track is becoming more and more "volume". From Anthropic’s Claude, DeepMind’s Sparrow, Google’s LaMDA to Character AI, there seem to be new competitors entering the arena every day.

Anthropic is a San Francisco startup founded in 2021 by several researchers who left OpenAI. Less than a year after its founding, the company announced a whopping $580 million in funding, and on Friday was reported to be on the verge of an additional $300 million.

The company has developed an AI chatbot called "Claude", which is currently available in closed beta via Slack integration and is reportedly similar to ChatGPT and even has some improvements . Anthropic describes its mission as "committed to building reliable, explainable, and controllable AI systems."

DeepMind is also a force that cannot be ignored on this track. The company introduced "Sparrow" in a paper in September, which was hailed as "an important step toward creating more secure and less biased machine learning systems." Sparrow is "a useful conversational agent that reduces the risk of unsafe and inappropriate answers" and is designed to "talk to users, answer questions and help find evidence."

However, Geoffrey Irving, a security researcher at DeepMind and the lead author of the Sparrow paper, said that DeepMind considers Sparrow to be a research-based proof-of-concept model that is not yet ready for deployment.

In a Time article two weeks ago, the company's CEO and co-founder Demis Hassabis said that DeepMind was considering releasing its technology sometime in 2023. "Private beta" of the chatbot Sparrow. This would allow companies to develop reinforcement learning-based features such as citing sources — a capability that ChatGPT does not have.

Let’s talk about Google’s LaMDA, this model sparked heated discussion last summer—— Google Engineer Blake Lemoine was fired for claiming LaMDA was sentient. Even if not as much as Lemoine believes, LaMDA is still considered one of ChatGPT’s biggest competitors. Google said in a 2021 blog post that LaMDA's conversational skills have been "years in the making." Like ChatGPT, LaMDA is built on the Transformer architecture and is also trained on conversations.

According to Google, “During training, LaMDA discovered some of the subtle differences that distinguish open conversations from other forms of language.”

The New York Times mentioned in a report on January 20 that Google founders Larry Page and Sergey Brin met with company executives last month to discuss ChatGPT’s possible search for Google’s $149 billion threats to the business. A Google spokesperson said in a statement: "We continue to test our AI technology internally to ensure it is useful and safe, and we look forward to sharing more experiences with the outside world soon."

Another powerful player is

Character AI, a company founded by one of the authors of the Transformer paper ​​Noam Shazeer founded ​​​ and gradually became well known. The company’s AI chatbot technology allows users to chat or role-play with anyone, imitating historical figures such as Queen Elizabeth and Shakespeare. The technology is currently free to use, and Character is "studying how users interact with it before formulating specific revenue generation plans."

It is rumored that Baidu will release a chatbot similar to ChatGPT

What has attracted more attention from domestic AI practitioners is that according to reports from Reuters, Bloomberg and other foreign media, Baidu plans to launch an artificial intelligence chatbot service similar to OpenAI’s ChatGPT in March.

Baidu plans to integrate chatbot-generated results, rather than just links, when users make search requests, sources said. "The tool, which has yet to be named, will be embedded in the main search service and will return conversational-style search results to users."

In an internal discussion in December last year, Baidu CEO Robin Li once shared his views on ChatGPT: "Turning such a cool technology into a product that everyone needs" is the most difficult. I hope Baidu will new "At least we can have a high-growth, innovative business, truly above and beyond our expectations" in the next year.

According to a report from Science and Technology Innovation Board Daily on January 30, Baidu does have internal plans to launch a chatbot similar to ChatGPT, but the specific time is not precise. Baidu CEO Robin Li positions the project as "leading an intergenerational change in search experience." He pointed out internally that related technologies have reached a critical point, and Baidu has greater opportunities in it.

Detection tool: Let the text generated by large language models have no place to hide

Although ChatGPT has powerful capabilities, it is also abused in areas such as school assignments and paper publishing. has caused widespread concern. Therefore, the academic community began to explore methods and tools for detecting text generated by large language models (LLM) such as ChatGPT.

Several researchers at the University of Maryland have studied the watermarks output by language models such as ChatGPT. In the paper "A Watermark for Large Language Models", they proposed an efficient watermark framework. The embedding of watermarks has negligible impact on text quality and can be detected using efficient open source algorithms without accessing the API or parameters of the language model.

Our method can detect relatively short synthetic text (as few as 25 tokens) while making it statistically impossible for human text to be labeled as machine-generated.

##Paper address: https://arxiv.org/pdf/2301.10226v1.pdf

Stanford UniversityIn the paper "DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature", several researchers proved that sampling from LLM Text tends to occupy the region of negative curvature of the model's log-probability function. Exploiting this observation, they defined a new curvature-based criterion to determine whether a passage was generated by a given LLM.

The researchers call their method DetectGPT, which does not require training a separate classifier, collecting a dataset of real or generated passages, and explicitly watermarking the generated text. DetectGPT generates random perturbations of passages using only log probabilities calculated by the model of interest and another general-purpose pre-trained language model (such as T5).

The results show that DetectGPT is more discriminative than the zero-sample method of current model sample detection, especially the detection of fake news reports generated by 20B parameters GPT-NeoX from the strongest The 0.81 AUROC of the zero-sample baseline has been improved to 0.95 AUROC. Code and data will be released in the future.

DetectGPT Illustration of detecting GPT-3 generated text.

Paper address: https://arxiv.org/abs/2301.11305

In addition to detection solutions presented in the form of papers, some individuals have also launched powerful detection tools. For example, an ML engineer from Hive AI who is working on the ChatGPT detector has a solution that can recognize text generated by ChatGPT, GPT-3 and other popular AI engines.

Judging from the internal benchmark test results, this solution is significantly better than similar methods such as GPTZero and OpenAI GPT2 Output Detector . On the in-house dataset, model balance accuracy is >99%, compared to ~60% accuracy for GPTZero and 84% accuracy for OpenAI GPT2 Output Detector.

Demo address: https://hivemoderation.com/ai-generated-content-detection

Finally, GPTZero has also received an update - GPTZe##roX, a new AI detection model specially created for educators. The model can process a mix of AI-generated and human text and highlight the parts of text most likely to have been generated by AI. In addition, a pipeline was built to handle batch uploads of files in PDF, Word, and .txt formats to easily run multiple files.

##Demo address: https://gptzero.substack.com/p/gptzerox

In short, as AI-generated text detection tools become increasingly rich and perfect, large-scale language models such as ChatGPT are bound to become more and more formal in their application, helping people unleash the power of AI more efficiently.

The above is the detailed content of Just now, ChatGPT officially announced that its mathematical capabilities have been upgraded again. Netizens: I am finally proficient in addition and subtraction within ten.. For more information, please follow other related articles on the PHP Chinese website!