Altman’s simple words made AI startups tremble:"Sheltering" OpenAI is destined to die!This is the speech delivered by OpenAI CEO Altman at the latest YC alumni sharing meeting. To this end, he also specifically added: Don’t spend too much energy on the UI interface.


It resonated with countless people at once. Many people in the industry said:

The reality is, I see a lot of this happening.


In addition to this sentence, Ultraman's speech contained many golden sentences, and many of them were "details" that were excerpted and forwarded by everyone——

Including the progress of GPT-5 and 6, views on the future trend of AGI, and personal entrepreneurship.


For example, "AI girlfriend" is just a beautiful trap, don't try it easily.

I deliberately chose a name like ChatGPT that is very machine-like so that humans will not have feelings for it.


Netizens after listening to the speech even described it as a "gold-like harvest."


So what exactly did Ultraman share in this speech? Let’s take a look.

01

"If you simply follow me, you will die."

At YCAlumniReunion2023, Altman had some entrepreneurial advice. In the summary of some enthusiastic netizens, it can be roughly divided into three points:

Companies that simply package OpenAI will not survive long.

We are optimistic about AI medical consultants and AI personalized one-on-one tutoring.

It is very unwise to blindly raise funds.

The first point is about simple packaging. Altman gives an example of a company that is solving the small flaws of the GPT model, especially those that only focus on the interface.

First, OpenAI is already solving most of the work; secondly, the competitive advantage of these companies that only compete with OpenAI in terms of cost and speed will not be long-lasting and sustainable.Businesses must provide truly unique value.

Some netizens thought that some time ago Sequoia Capital proposed that generative AI has entered the second stage——

Hype and quick demos are being replaced by truly valuable and complete product experiences.

Various views actually coincide with each other.

Many people in the industry expressed their affirmation:

Don't put yourself on the chopping board to be killed by the next version of OpenAI.


(Well, there is an official fandom version)

But someone else made a different point: There are already too many of these and it seems silly to me. It revealed several pieces of information: monopoly is a good thing; first-mover advantage will win... but competition will promote progress, and ChatGPT does have many shortcomings that can be improved.


That being the case, what directions are worth pursuing?

The second point is that Ultraman is optimistic about the empowerment of AI in medicine and education——

AI medical consultant and AI personalized one-on-one tutoring. Both of these have great social value.

For example, in the education scene, OpenAI itself is also very active.

It is reported that OpenAI is preparing to buildOpenAI Academy, expected to be launched by the end of 2023. It may be a free online teaching system accessible to everyone, and teachers can interact with GPT-5 and receive feedback and guidance during the course.

Prior to this, OpenAI also demonstrated two educational scenarios in GPT-4 customer cases. GPT-4 is transformed into an AI learning assistant, which can serve as both a virtual tutor for students and a courseware assistant for teachers.

The third point is that it is very unwise to raise funds blindly without a plan. Don’t think that everything will be fine once you have money.

The netizen who concluded the summary also said:Just because Ultraman can do it, doesn't mean you can do it too.


02

What does GPT-5 look like?

Another hot topic is OpenAI’s next generation large model——GPT-5andGPT-6.

In this regard, Ultraman also revealed the appearance of the next generation large model in his speech, but not much:

GPT-5 and GPT-6 will haveMultimodal outputcapabilities, compared to existing GPTreliabilityhigher,personalizationCustomized experience is better.

Regarding specific multi-modal capabilities, some netizens predict that it can better complete NLP and visual scene alignment tasks such as grounding, such as image annotation, visual question answering, etc.


Combined with the fact that Ultraman also hinted that GPT-5 will haveSpeech recognition, synthesis and emotion detectionLooking at the new functions, GPT-5 should have the ability to align various modalities such as "pictures, text, listening and speaking".

Taking advantage of the topics of GPT-5 and GPT-6, Ultraman also "gave a message" to the large model:

Now, "Moore's Law" (ScalingLaws) in the field of large models has begun to take effect. The cost of training large models is becoming lower, and the price of calling the GPT interface will become cheaper.

ScalingLaws is a law proposed by OpenAI in 2020. Simply put, as the size of the model, the size of the data set, and the calculation floating point numbers used for training increase, the performance of the model will improve.

Altman believes that as energy and computing costs drop rapidly in the future, more powerful AI capabilities will emerge, and "many things that were previously unimaginable will be made."

But even if the future of large models is so bright, Ultraman said that "the road to AGI is still far away"——

Whether it is GPT-5 or GPT-6, bothFar inferior to AGI.

Even in terms of the "human-like" standard, no AIChatbot can achieve it at this stage:

Even with the cutting-edge AI technology behind it, I didn’t feel like I was chatting with a “person.”

Altman believes that true AGI will be able to master "autonomous reasoning," that is, developing new knowledge over time. For example, only AI that can write papers and conduct experiments based on existing knowledge of physics can meet the threshold of AGI.

However, the output of AI should be treated with caution:

People will forgive humans for making mistakes, but they won't forgive computers. The standards are different.

Regarding the arrival of the AGI moment, Ultraman also gave an idea:

At that point, people may experience some degree of identity crisis, but nothing too bad and confusing. We've been through this moment many times, where technology will eventually become ubiquitous.

03

"The result of an undergraduate's persistence"

Finally, regarding ChatGPT itself, Ultraman also found two small "tits".

First, the reason why ChatGPT is called ChatGPT and does not have a nickname like a person’s name is because Ultraman does not want to go there."AI girlfriend" productgo in the direction.


Second, ChatGPT is not a project that OpenAI initially “invested in”.

From beginning to end, OpenAI evenOnly one undergraduateInsisting on doing research in the direction of large models, others have more or less turned to other fields such as robotics or game AI to "publish papers".

This undergraduate who only works on large models isAlec Radford, and now the number of citations of papers on Google Scholar has reached 9.6w+.


In 2016, Alec Radford obtained a bachelor's degree from Franklin W. Olin College of Engineering and then joined OpenAI, where he worked for 7 years.

It's what he insists onlarge language modelThe route has changed the development direction of OpenAI company and even the entire AI field.

In this regard, Ultraman said half-jokingly and half-seriously:

When recruiting, remember to look for people who “don’t look like they’re going to join the company.”

OpenAI did not determine the research direction of large models at the beginning, but we still recruited AI researchers, and after we made breakthroughs, more people joined us, which eventually led to ChatGPT.

Combined with OpenAI's "industry-free" rule for recruiting talent, breaking convention seems to have become a new driving force for the company to gain inspiration and vitality.