OpenAI overrode concerns of expert testers to release sycophantic GPT-4o

Check out a little Turvy Week for the number one generative AI company in terms of users.

The creator of ChatGPT was released, and then released a large language model (text, picture, sound) large language model (text, picture) has taken back a large language model (LLM) updated version. Recently informed about the company At least 500 million active weekly users of the Hit Web Service.

Horrible, no good, no Sycofantic GPT-4O update

Openai has started updating GPT-4o, hoping to be better accepted by users on April 24, April 5, then updated after five days Rolled again on April 29Days of complaints of users in social media – mainly in X and Reddit.

Complaints vary in intensity and features, but all of the most highly highly appreciate the incorrect, incorrect and obvious harmful ideas or “glass” or open-disciplined views or “glassy” or user.

Subjects in the screenshots and users, this type of GPT-4o model developed and approved a business idea, approved a business idea, applauded a work idea for a schizophrenic cunning isolation and even supported terrorist plans.

Users including Top AI researchers and even the first of the former Intermediate General Director This kind of terrible user desires said that the unshakable cheatleading of an AI model, which is simply teasing or inappropriate, and the worst of the AI, the worst thoughts and the worst opinions and impulses are the actual damage. AI rose to the level of security.

Openai then released a blog post Explaining what is wrong with the wrong thing, “I did not fully focus on short-term reviews, and the company’s problems were made to respond to the company, as well as” Something from Me “or” Something from Me “or” Something from Me “or” Something from Me “or” Something from Me “, and” Something from Me “or” Something from Me “or” Something from Me “, and” Something from Me “, and” Something from Me “, and” Something from Me “, and” BAK ” He announced how he ended with a very typical model[ing] In order to nuance, “the users in response to users in response to the model performances that users approve of users.

Now today, Openai blog left posts With more information about how the Sycofantic GPT-4O update is about, it was given a loan for Openai, not any special author.

CEO and co-founder Sam Altman Sent a link to blog posts in X, “We missed the sign of the last week with the GPT-4O update of the last week. What happened, some of the things we learned and do different things in the future.”

How and why the new Openai blog posts and why GPT-4O explains so sykophantia

To me, the daily user of ChatGPT, including 4O model, looks for the most surprising acceptance of Openai’s new blog posts, the company contained Before a small “Expert Test” group, have concerns about the model, but exceeded a more widespread response from a larger user than a larger group.

As the company writes (emphasis mine):

“During a period of time discussing risks associated with Tipophilic risks in GPT-4O, it was more concerned about the changes in our tone and style of our internal experience. Nevertheless, Some expert testers said that the model behavior “felt” …

“We decided to do it later: Do we need to post this update despite the results of the subjects of expert test devices based on the subject flags and the results of the A / B. In the end, we decided to launch a model due to the positive signals of users who tested the model.

“Unfortunately, this was the wrong call. These models are based on our users and the user opinion is important for our decisions, as a result, it is our responsibility to interpret this opinion correctly. “

This seems to me like a big mistake. Why do you have specialist testers, if their experience is not higher than the crowds of the crowd? I asked Altman on this option in X But he has not yet answered.

They are not all ‘reward signals’ equal

Openai’s new post-Mortem blog post shows how new versions of new versions of the company have developed new versions and how human feedback is about model qualities, character and “how it changed the character”. As the company writes:

“Since the GPT-4O started in Chatgpt in the last May Five large updates released identity and assistant are directed to changes. Each update covers new post-education and is often combined with an updated model that is tested independently for the start of the model training process and then rated for beginners.

“Postse Train-top models, we are taking a pre-made base model, monitor a wide range of wide ideal answers written by people or existing models, then realizes strengthening with reward signals from various sources.

“We present the language model with a request during strengthening learning and please write answers. We then appreciate the reply to the prize signals and are more likely to produce higher and less reactions to produce higher ratings.“

Obviously, during the post-training, the “premium signals” used by Openai have a great effect on model behavior and meet their performances from ChatGPT users, it may not be the best to use the same as the signal How The model learns to communicate and What kind of species Answers need to be issued. This clearly acknowledges this in the next paragraph of the post in the post:

“It is a difficult question to determine the correct set of relevant reward signals and we take a lot of things: the answers are correct, they are in good Model specifier⁠, they are safe, and users are them, etc. The better and more comprehensive premium signals are producing better models for Chatgept, so we always practice with new signals, but each has its questions. “

Indeed, Openai also reveals that the “Thumbs” reward signal is a new one used with other reward signals in this special update.

“Update, user reviews from ChatGPT, the thumbs and thumbs and thumbs and thumbs and thumbs are added an additional prize signal based on low data. This signal is often useful; a thumbs have generally wrong.”

Again, the company does not blame the new “upper-fingered” data for the failure of the model and attractive cheerleading behaviors. Instead, Openai blog post said it was this combined Caused problems with different other new and old reward signals: …

In response to this blog post, Andrew Mayne, now Mayne, a former member of the Openai technical staff working in the intersional of AI consulting company, wrote another example in x In the premium promotion and model regulations can affect the model performance of fine changes:

“Early in Openai, I use the “polite” word in an example I wrote in an early colleague (now another laboratory).

They wanted to change the “polite” politically incorrect and “useful.”

I noted that it should be not only useful to be useful that it can make a model of more than a model in several turns.

After demonstrating the risk with a simple exchange, quickly “polite”.

These models are odd.“

OpenTai plans to develop model testing processes when progressing

The company lists six process improvements to prevent similar undesirable and less ideal model behavior in the future, but it is the most important thing to me:

“We will officially consider our behavioral problems forming behavioral issues such as behavior, deception, reliability and personality.

In other words, in particular, despite the importance of quantitative information, machine learning and artificial intelligence, Openai accepts that this is not the only model that any model needs.

Many users that provide a “thumb” can result in a very long-term, sad, destructive and undesirable place for many users that respond to the AI ​​model, and these behaviors and users. More things are not always better – especially when you limit “more” to several signal areas.

It is not enough to say that all the model passes all trials or users received a number of positive answers – the experience of trained power users and their model, “he has more weight, while they have more weight.

Hopefully the company – and the whole area – learns from this event and leads to classes.

Considerations for extensive takeaways and enterprise decisions

Perhaps theoretically, theoretically, why the experience for myself is so important and also indicates that there is experience in areas edge and outside For someone you optimizes (in this case, machine learning and AI). This is a variety of practices that allow us to achieve new progress benefits. One should not be superior to the hull, definitely in humanitarian or art.

And finally, I think that this creates a fundamental problem in the design of products and services in the heart. Individual users love more Sycofantic EU based on each dashed interaction based on each isolated interaction, as well as fast food and soda flavors, single-use plastic containers, entertaining media, entertaining or entertaining or entertaining or entertaining or tobloid gossip. Still, took it all together Cumbulation All of these types and activities are less common in obesity and health in the situation of obesity and health in the situation of obesity and health in the state of obesity and health in the state of obesity and health, and in poor health.

In the enterprises, when making the size of any measurable, these broader opens, in the cases where you do not expect or do not expect to make your damage, but you can get back to the way you do the confusion.