news

last year, openai held a developer conference (devday 2023) that caused a stir in the industry in san francisco, launching a series of new products and tools, including gpt-4 turbo that supports 128k context, api price reduction, new assistants api, and gpt-4 turbo for visual features, dall·e 3 api, and a vastly improved json model, as well as the ill-fated gpts and app store-like platform gpt store.

ceo sam altman said with confidence: "we hope you can come again next year. compared with the new results we are carefully building, what we released today will seem very outdated," raising expectations again.

after just experiencing a new round of resignations of senior management teams, openai’s annual event devday 2024 is approaching, and everyone has a question in their minds: what will happen next?

however, this year’s openai devday has quietly changed its tone. first, it was announced that the format of the conference would be adjusted, from a large-scale event to a series of roadshow-style developer interactive sessions.

what’s even more surprising is that, according to official spoilers, openai will not release its next-generation major flagship model during devday, but will focus on updates to apis and developer services.

in short: the next generation model is still roasting in the oven, so let’s not mention it at all.

devday is no longer a model conference, this time it is really a "developer" conference.

although openai announced that it will not release a new generation of flagship models on devday this year, the industry still has many expectations for this developer conference. according to speculation, openai may focus on launching new api capabilities such as long-term reasoning and function calling, while enhancing the agent capabilities of the model to provide developers with more autonomy and the possibility of integration with other apis.

the api of 4o voice mode is also highly anticipated

silicon star also communicated with several developers to understand their real needs for apis. multi-modality and performance became the focus of attention.

talkme is a language learning toc product for overseas markets. ceo jia zijian told silicon star that ai developers who go overseas usually choose openai’s api, and a small number of them use gemini and claude. the main api used by talkeme is openai, plus gemini assistance.

"gpt-3.5 turbo has basically been eliminated, and everyone will switch to 4o-mini or 4o to improve business logic. for more complex tob saas products, you may consider gpt-4, which has stronger capabilities and is relatively more expensive."

o1 is also currently under consideration by talkme, but due to its high cost, it is currently limited to pre-research on experimental functions and cannot be put into formal production of toc applications.

"o1's capabilities are indeed very strong, but from an engineering perspective, using o1 for some, 4o for some, and 4omini for some can basically cover the development scenario."

regarding the expectations for the openai api, jia zijian said there are three main points:

the first is low latency.

the second is multi-modal capabilities under low latency. “part of our multi-modal capabilities are currently using 4o and some are using gemini.”

the third point is more stable json output. "the publicization of openai is not yet so rigorous. it is more likely to be dominated by scientists' algorithms. as a result, the json output by openai is not very stable. in many scenarios, it is better to write your own regular rules to parse the output text content. in this dimension, google is worse than openai if done well, it will be more beneficial to developers if there is stable json output.”

in terms of price, "except for o1, other prices are relatively cheap."

chen yang, director of cocos engine operations and manager of lao cai miao, conducted an analysis from the perspective of domestic ai products. he told silicon stars that in terms of current capabilities and response speed, openai is still the strongest. on the other hand, domestic developers consider more about compliance and roi, while domestic qwen, doubao, minimax, models such as wenxin are also catching up, and the price/performance ratio of some products after discounts is not weaker than openai.

he said, "future demand must be multi-modal. traditional http post requests will be gradually eliminated. domestic doubao has taken the lead in using webrtc, and openai webrtc is about to be opened. therefore, in addition to the quality of the model, the future market must consider the comprehensive effect, the combination of asr, tts, webrtc, etc., from this perspective, domestic products have greater advantages in integration, and there is a chance to gain an advantage in the price war.”

yuhe technology targets enterprise scenario applications. technical partner chi guangyao hopes that the update of the openai api can provide improvements in two aspects: built-in engineering capabilities to reduce or detect hallucinations; for o1, provide more flexible usage options, allowing users to specify whether use complex reasoning capabilities in this request.

devday’s shift in focus aligns with openai’s recent strategy of taking more incremental steps in the development of generative ai.

the company is currently optimizing its flagship models gpt-4o and gpt-4o mini, facing the challenge of staying ahead of the technology in the race for generative ai. according to some benchmarks, openai appears to have lost its absolute technological lead.

one of the reasons is the difficulty in obtaining high-quality training data. according to data from originality.ai, more than 35% of the world's top 1,000 websites have banned openai's web crawler.

a study by mit's data provenance project found that about 25% of data from "high-quality" sources has been restricted from use in training ai models.

to address these data acquisition challenges, openai has signed a series of high-dollar licensing agreements with publishers and data brokers.

another aspect of the pressure comes from the cost of operating the company. according to reuters, openai expects revenue this year to reach us$3.7 billion and a loss of us$5 billion, which will mainly be used to pay for operating service costs, employee salaries and office rent. it does not include other large expenditures such as employee equity incentives.

in order to solve the funding problem, openai is promoting a new round of financing. in early september, financing news broke that openai may receive an investment of us$6.5 billion at a valuation of us$150 billion. according to sources, this valuation depends on whether openai can transform its corporate structure into a profit model similar to its competitors anthropic and xai.

however, the latest news is that apple has decided not to participate in the latest round of financing of us$6.5 billion. in addition, the departure of the chief technology officer, chief research officer and vice president of research last week has also added uncertainty to openai.

although openai chief financial officer sarah friar revealed to investors on september 27 that this round of financing has been oversubscribed and is scheduled to close before the end of the month, financing negotiations have not yet been finalized, and the participants and specific amounts may still change.

the focus on apis may be related to the financial pressure openai is facing.

according to research reports and related data published by futuresearch, openai’s annual revenue is us$3.4 billion, 76% of which comes from chatgpt paid users (chatgpt plus, chatgpt enterprise and chatgpt team subscriptions), and api interface services contribute 15% of the company’s revenue , approximately us$510 million.

the api-driven service model once helped platforms like aws grow rapidly. as more and more companies use amazon's apis to develop new products or services, thereby generating network effects, the value of its ecosystem continues to grow.

although openai's current direct revenue from apis is not high, in the context of chatgpt's slowdown and openai taking more incremental steps to develop generative ai, it may be possible to provide better apis for the company opens up new market opportunities and promotes collaboration with other businesses.

without gpt-5, executives left and transformed into a for-profit company. this series of events indicates that commercialization has become an urgent problem for sam altman to solve. openai will consider profits more than realizing agi.

for a long time, sam altman has used superb skills to manipulate public sentiment while running his own business, making him the leader among ceos in the technology industry.

but don’t be surprised if you hear him talk about openai’s plans to move to a “profit model” at openai devday in san francisco.