The contest between Anthropic and OpenAI is essentially a difference in the concept of "developing technology first" or "governing technology first". Author | Cissy edited | Chen Caixian
On September 25, yesterday, Amazon officially announced that it injected $4 billion into AI startup Anthropic, which attracted the attention of artificial intelligence circles at home and abroad. So far, following Microsoft, Google, Meta and other technology giants into large-model companies, Amazon has also run into the market.
Google self-developed, Meta open source, Microsoft and OpenAI teamed up - although undesigned, the three routes of these three giants actually represent three different models of the current big model track. Amazon invested in Anthropic (whose Claude is in full swing with ChatGPT), choosing Microsoft's model of teaming up with OpenAI.
Compared with Microsoft's initial investment of $1 billion in OpenAI in 2019, Amazon's first shot of Anthropic is quite heavy: $4 billion, about 28 billion yuan.
It is worth noting that compared to most of the current aggressive technical teams in the same field, Anthropic is a relatively conservative start-up: since its inception, Anthropic's research direction has been AI governance and improving the security of AI models. Therefore, even after ChatGPT came out, its rapid launch of Claude, which is comparable in performance, Anthropic's innovation in technology has always been considered to be first for AI security services, and then for technological breakthroughs.
Amazon's investment in Anthropic is rated as competing with Microsoft and OpenAI, but the difference between the two is actually more impactful, because they will largely guide the current AI industry to two paths: develop technology first, or govern technology first?
1
Anthropic two or three things
In 2019, the University of Cambridge's authoritative AI panorama report "State of AI Report 2019" predicted that in the near future, at least one mainstream AI company will make substantial changes to the governance and management model of AI.
In 2021, Anthropic was founded to operate as a public benefit business. This validates the University of Cambridge's reported projections; But more remarkably, the previous owner of Anthropic's founding members was OpenAI. At that time, OpenAI had become world-renowned with the GPT-3 language large model released in 2020, and became a fascinating place for countless AI researchers.
Anthropic founders, Daniela Amodei and Dario Amodei, are two siblings.
Prior to joining OpenAI, Daniela Amodei worked at Stripe where she was founded as an early investor in OpenAI, the U.S. Congress, and joined OpenAI and rose to VP of Security and Policy.
Dario Amodei, who graduated from Princeton University with a Ph.D., joined Baidu American Research Institute in 2014, working under former Baidu Chief Scientist Andrew Ng before joining Google as a senior research scientist for a year. In July 2016, Dario Amodei joined OpenAI and rose to the position of VP of Research, where he published several research papers on AI interpretability and security, and is considered one of the most cutting-edge researchers in the field of deep learning.
Not long ago, Time magazine selected the top 100 AI figures, and Daniela Amode and Dario Amodei were also selected.
Caption: Dario Amodei
During his tenure at OpenAI, Dario and his colleagues were skeptical about the way OpenAI handled the ethics and safety of technology, and did not recognize OpenAI's work in AI security governance, and then decided to leave and founded Anthropic in 2021. In addition to the Dario siblings, Anthropic's founding team includes Jared Kaplan, Sam McCandlish, Tom Brown and other former OpenAI employees who worked on GPT-2 and GPT-3.
Since its inception, Anthropic's vision has been to be a company focused on AI safety research, developing controllable, explainable AI systems. In terms of AI safety, Anthropic pioneered the concept system of "constitutional AI", emphasizing the alignment of values in artificial intelligence systems.
After Anthropic was founded, it was favored by investors including Skype co-founder Jaan Tallinn and former Google CEO Eric Schmidt in Series A, raising $124 million. Because Anthropic is a public welfare enterprise, the news caused a sensation in the circle at that time, and was called "$124 million to burn" by TechCrunch, a well-known authoritative media.
In April 2021, Anthropic raised another $580 million in funding, led by Sam Bankman-Fried, CEO of cryptocurrency exchange FTX... In 2022, when the AI market has fallen sharply, Anthropic has become a rookie in attracting money, and ChatGPT has become more and more prominent, and is considered by many industry insiders to be the only startup that can compete with OpenAI.
In 2021, the Anthropic team published several articles on the mechanism of Transformers, such as "A Mathematical Framework for Transformer Circuits", "In-context Learning and Induction Heads", etc., hoping to study a simplified version of Transformer and fully explain Transformer The capabilities of the model. The simplified version of the Tranformer they studied was to learn basic language patterns before acquiring generic language processing capabilities.
These findings provide a new way to understand Transformers.
Capable and experienced R&D team, sufficient resources and financial capabilities, and a layout two years ahead of the big model trend - these three factors make Anthropic a rising star in the field, and Microsoft/OpenAI, Google/DeepMind are currently listed in the top three companies in the global AI field. What's more, Anthropic is also the only startup that is not deeply tied to the big factory.
In March 2023, Anthropic raised more than $1.3 billion and was valued at $4.1 billion; Just 6 months later, Anthropic directly attracted $4 billion from Amazon. Previously, Anthropic's conversational robot Claude competed with ChatGPT overseas and was highly recognized in the technology circle. Anthropic and OpenAI are the very few teams that are still stuck with closed source.
According to TechCrunch, Anthropic has made a "long-term" commitment to Amazon to provide AWS customers worldwide with access to next-generation foundational models through Amazon Bedrock. In addition, Anthropic will provide AWS customers with unique capabilities for early access model customization and fine-tuning capabilities.
But in addition to Amazon, Anthropic has also launched similar cooperation with Google Cloud, Zoom and other enterprises. Despite receiving a large investment from Amazon, Anthropic remains true to its original vision of pursuing the ultimate goal of AI security.
2
This wave of "AI security" is different
Different from the previous wave of AI ethics, after the rise of large models, the direction of "AI Safety" has attracted the attention and discussion of more industry bigwigs, such as Yoshua Bengio, a Turing Award winner and one of the three giants of deep learning.
People close to Yoshua Bengio told AI Technology Review that Bengio plans to stop taking students from next year and devote himself fully to AI safety research.
The attention mechanism proposed by Yoshua Bengio in 2014 later became the core of Google's Transformer, and the rise of large models is based on Transformer. Although the results of large models are in full swing, Bengio, as the initiator of the attention mechanism, is very concerned about the safety of large models. In July, Bengio attended the U.S. Congress to participate in the discussion of artificial intelligence, arguing that "global rules for artificial intelligence should be developed."
Sam Altman, CEO of OpenAI, previously declared that AGI will be achieved by 2030. For Sam Altman's assertion, some people rejoice in the rapid progress of technology, and others believe that it is irresponsible to allow AI to develop unrestrained before the risks and safety of AI are under control. Anthropic and Yoshua Bengio fall into the latter category.
In the last wave of AI commercialization, although the issue of AI ethics and safety has also been put on the agenda, it is not the core of research. In this wave of big models, the same thing that attracts attention as algorithm innovation and sufficient computing power is the security of algorithms.
When the ability of the model is not strong enough, safety and controllability is an additional proposition; However, after the explosion of ChatGPT, with the rapid development of large model capabilities, security and controllability have become a necessary proposition. The problems of bias, discrimination, privacy leakage, cultural politics, inducement (such as "jailbreaking of large models") of deep learning algorithms have also become extremely serious.
A previous study by UIUC's Li Bo team found that although GPT-4 is generally more credible than GPT-3.5 in standard benchmarks, it is more vulnerable in the case of jailbreak systems or user prompts, possibly because GPT-4 follows (misleading) instructions more accurately. (For more information, please read the previous report of AI Technology Review: UIUC Li Bo: GPT-4 is more "arrogant" than you think)
As mentioned earlier, OpenAI's large model is not open source, and the outside world cannot obtain its specific large model. In this case, outside researchers can only conduct "black box testing" on the security evaluation of large models. According to AI Technology Review and industry insiders, black-box testing is more difficult than white-box testing, "because you don't know the internal data, you can only estimate the internal mechanism of the model through the access of the model, and then design the response method based on the conjecture."
Whether from a social or technical level, the security risks of large models are a topic that should not be underestimated. It is precisely because of this that when ChatGPT first emerged, domestic To C large model products were temporarily blocked. Although the first batch of 8 large models passed the filing in early September, various manufacturers did not dare to take the security of large models lightly.
In large model products, the factors affecting safety are not only the late start of technology, but also the number of users: the larger the number of users, the more public opinion problems related to the model, and the greater the pressure of risk control.
Therefore, at present, large domestic model companies in infrastructure are actually facing three mountains: computing power, data annotation, and risk control. Risk control mainly includes network security and content moderation.
In contrast, Anthropic's founding team came out of OpenAI and participated in the core research and development of GPT-3, and is perhaps the company that knows OpenAI and the GPT series of large models best besides Microsoft and OpenAI.
Anthropic has become a gold-absorbing expert, perhaps reflecting an important trend in the AI industry: in the AI era, enterprises that can merge technological innovation and technology governance are the darlings of the market and the winners of the future.
Another AI company is about to go public in Hong Kong, what are the sufferings of companies hovering outside the IPO door, and how will the listed AI companies continue to write a new chapter, we are about to launch an observation article "In 2023, the previous generation of AI companies rushed to the bloody listing difficulty", welcome to add the author WeChat Sunpx33, if you have something to say, you can talk to me.
Without the authorization of "AI Technology Review", it is strictly forbidden to reprint it on web pages, forums, and communities in any way!
Please leave a message in the background of "AI Technology Review" to obtain authorization for reprinting the official account, and you need to mark the source and insert the business card of this official account when reprinting.