news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

last month, openai ceo altman grabbed a lot of headlines with a photo of strawberries. now, maybe the strawberry model is really coming of age.

according to the information, as part of the chatgpt service,openai plans to release the "strawberry" model within two weeks.

of course, the report also points out that the final release date may change, so please treat it with caution.

reliable informant @apples_jimmy believes that the strawberry model is expected to be released this week.

he also said that openai expects to release a new model called gpt-4.x in october, which may be called gpt-4.5.

as for the more powerful gpt-5 model, i've heard that it could be available as early as december, but to be on the safe side, i'd suggest you can expect it in the first or second quarter of 2025.

the information reported that it is not clear what form the strawberry model will take.one possibility is that the strawberry model will be a standalone product.

another possibility is the strawberry modelwill be integrated into the selection menu of the chatgpt model, users can switch between different model services at will.

previous reports also mentioned that the biggest difference between the strawberry model and other models is thatit knows how to "think" before answering questions.

moreover, the strawberry model not only excels in mathematics and programming, but if it is given more "thinking" time, it can also answer questions about subjective topics of customers, such as product marketing strategies.

the information points out thatthe strawberry model's thinking time usually lasts 10 to 20 seconds, which has the benefit of helping to reduce errors.

and because the strawberry model spends more time thinking, it is able to recognize when it needs to ask customers more questions to fully understand the user's needs.

in addition, the strawberry model also has some differences from the gpt-4o model. for example, its initial versiondoes not have the multi-model capabilities of gpt-4o, can only receive and generate text replies, but cannot process images.

the information believes that this may be because competitors are also launching similar products, so even if this product is not perfect in some aspects (such as not being able to process pictures), openai can only speed up the pace of launch.

in this regard, the whistleblower @apples_jimmy also mentioned, anthropic and google are also secretly preparing their new models.it is planned to be launched around the time of the us election.

another thing is the subscription price that repeatedly stirs up users’ emotions.

it is reported that chatgpt has launched a new paid tier, chatgpt pro, and has already pushed it to some users.price: $200/month, ten times more expensive than the current $20/month.

if this is true, it may also echo the above reports about the strawberry model.

the information also pointed out that the number of times the strawberry model can be used per hour may be limited like chatgpt plus, and models with higher subscription prices will respond faster.

as of press time, openai has not yet responded to this matter.

cut the strawberry layer by layer

what does openai's "strawberry" model mean to us?

in fact, the predecessor of the strawberry model is "q*", a mysterious entity that caused a stir at the end of last year.

last november, sam altman was kicked off the board of directors without any warning. he was notified even during the meeting, which shocked the entire company and the industry.

the reason given by the board at the time was that he and the team could not reach a consensus on security and risk management. and this risk was related to the top-secret project "q*" at the time.

the project was originally led by ilya sutskever, who has now left openai to start his own business, working on ai safety-related businesses. considering that musk once said that this project "poses a threat to humanity," it's hard not to be curious about what's going on here.

previously, the information and reuters tried to obtain inside information, but in the end they could only confirm thatmathematical calculation ability is the focus of "q*"。

image source: reuters

the big model is good at "text", and can process language and text skillfully, basically on par with humans. but mathematical operations have always been not so good. even though "q*" once caused a huge change in openai, according to reuters, its performance at that time was probably at the calculation level of elementary school students.

the leaked information currently known shows that there are projects within openai that can achieve 90% accuracy in mathematical calculations, which is an amazing improvement.

image source: reuters

to emphasize: it is impossible to confirm at this time exactly how far "strawberry" has come.

all we can say is that if "strawberry" is an upgraded version of "q*", it is likely to be a project that continues to seek breakthroughs in mathematics and computing.

the relationship between "mathematical calculation" and "reasoning" cannot be directly equated, but it reveals openai's ambition.

what are you talking about when you talk about reasoning?

so, what exactly is "reasoning"?

in reality, these two words obviously have a very broad definition. earlier this year, a team led by the chinese university of hong kong did a comprehensive review of model-based reasoning capabilities. the most fundamental definition of "reasoning" is threefold:

cognitive reasoning: the ability to draw meaningful conclusions from incomplete and inconsistent knowledge

the most common example of this kind of reasoning is putting together a jigsaw puzzle. each small piece is a corner of a huge picture, and if you just pick up two pieces at random, they will definitely not match.

you can only hold these small pieces, piece them together, and slowly assemble a completed picture. in this process, there are no instructions or step-by-step diagrams, and you often have to rely on your hands and intuition.

logical reasoning: based on the premises and the relationship between these premises, the conclusion is systematically drawn, and the conclusion is logically implicit or established.

solving math problems is a typical logical reasoning. there are known conditions and problems to be solved. based on these, you can deduce the results step by step. logical reasoning is currently the "hardest bone" in the development of large models.

natural language reasoning: this is a process of integrating multiple kinds of knowledge, either explicit or implicit, to draw new conclusions about the world

friends who like to read detective stories and mystery novels should be able to understand this easily. this kind of reasoning is like encountering a murder story. there are some vague hints and some unclear information in the book. you must combine various clues to deduce who the murderer is and what the crime process is.

if you only look at the internal openai documents obtained by reuters, the goal of "strawberry" is to plan, access the internet, and perform deep research.

these all seem more like the last type of natural language reasoning, which is nothing more than a strengthening of it. it is hard to say whether it can be regarded as an improvement in reasoning ability.

but,openai is not so rigid about "reasoning", but has a more ambitious vision。

two months ago, john schulman, one of the founders of openai, said on the podcast dwarkensh that the progress of gpt-4 was largely due to post-training technology.

“it is very complicated to create a model that has the functions that people care about through post-training,” said john schulman. “it requires a lot of investment and the accumulation of a lot of research and development work, which to a certain extent forms a barrier.”

john schulman defines reasoning as follows:

“reasoning means that some calculation is required, or some deduction is required. from this definition, it is necessary to be able to calculate and calculate step by step while processing the task.”

it can be seen that in his definition, reasoning and computing behaviors are highly bound together, and he hopes that machine reasoning can be carried out in real time - just like humans, who can analyze and judge while receiving information.

however, even if a person is not good at math, it does not prevent him from thinking logically and completing various types of reasoning. why is the mathematical ability of a machine so important?

it can be understood like this:mathematics is never just about calculations; it is also a way of expressing information.

mathematics is a language that relies more on symbolic form and precise meaning. 1 is 1 and 0 is 0. when using computational symbols and formulas to present information, it is actually lower-dimensional than natural language.

in other words, the reason why the big model is "capable of writing" is that it is based on "being able to calculate", converting natural language into computer language.

this was laid down as early as the 19th century by one of the most important mathematicians in history, george boole (the same boole who created the boolean variable).

George Boole

boole was a man with devout religious beliefs, and he even wanted to explain the existence of god through mathematical reasoning.

regardless of what his final conclusion was, the wealth he ultimately left to the world was his book “an inquiry into the laws of thought,” in the opening chapter of which he explained his grand goal: to use the symbolic language of calculus to express the basic laws of reasoning, a type of thinking activity.

this also explains why, once people talk about ai's performance in mathematical operations, they are more nervous in their expectations:

if we can crack the language of mathematics, we may be close to cracking the thinking process.

openai's technical veteran has left again

there is a very strange phenomenon: it seems that every time openai has big news about technical insiders, it is always accompanied by drastic personnel changes.

coincidentally, several internal employees of openai also officially announced their resignation today.

for example, alexis conneau, former research director of openai audio agi, announced his resignation to start his own business, while he is also an important technical member of the gpt-4o research team.

before the release of gpt-4o, he excitedly predicted that it could usher in a new era of human-computer interaction.

he has extensive work experience in major companies such as google and meta, and joined openai in april 2023. in his words, his main job is to equip the gpt model with a "talking mouth."

prafulla dhariwal, head of the research team behind gpt-4o, praised conneau:

alexis conneau came up with the vision for her before anyone at openai, and worked tirelessly to bring it to life!

or, arvind neelakantan, who worked at openai for four and a half years, also switched to the "enemy camp" meta ai research team today.

he has participated in the development of many important openai projects, including embeddings, gpt-3 and gpt-4, api, and chatgpt.

neelakantan said that his work experience at openai was the highlight of his career. he will focus on the development of the next generation of llama models at meta ai.

in response, logan kilpatrick, former openai developer relations director, also sent his farewell wishes.

since the beginning of this year, openai has seen a wave of resignations, and its founding team has been "falling apart."

former chief scientist ilya sutskever had just announced his withdrawal from openai, and jan leike, one of the inventors of rlhf, followed in his footsteps and left.

the reasons for leaving are basically different. in addition to the aftermath of last year's "palace drama", it may also be due to personal career planning, etc.

openai’s drastic personnel changes are unlikely to have an impact on the ai ​​competition landscape in the short term. at the same time, amid a chorus of pessimism, the current immature ai industry can no longer tolerate a one-year window period.

as news about the model continues to come to light, we are looking forward to seeing another magnificent ai age of discovery in the second half of the year.

at the very least, it will be more interesting than the boring new ai technologies in the first half of the year.

it is foreseeable that the progress of the underlying models of ai technology will be like a powerful driving force, leading to an explosion in the entire application end, just like the gpt-4 that came out that year, bringing us long-awaited surprises.

at that time, we as users will always be the biggest beneficiaries.