Assessing Beijing Municipality’s push for General AI
Two years on, limited impact
In May 2023, Beijing’s municipal government published two documents setting out measures to promote the development of “general AI”.1
These measures were significant. Fynn Heide wrote at the time that they were evidence of growing Chinese policymaker interest in general AI, and of Beijing municipality leading the implementation of this shift.
(To avoid ambiguity: These were measures from the Beijing municipal government, which has a status similar to a province. I’m not using “Beijing” as a metonymy for the national-level government or China as a whole.)
In this post, two years later, I assess how much impact the measures have had. I compare the apparent results to the four policy priorities that Heide identified in the documents. Limitations in public information, in particular from the most recent year, make it difficult to get a complete picture. But, overall, the results seem modest.
In the following section, I give my bottom lines about how much progress the Beijing government has made on each of the four priorities. The subsequent sections go into more detail, explaining how I approached my research, and then giving more detailed findings on each of the priorities.
But before that: Why does it matter how much the Beijing municipal government has achieved?
Two reasons. First, the 2023 measures are one datapoint for a key question in China AI policy more broadly: When a Chinese government body announces grand ambitions in AI, how likely are these to turn into something meaningful? Second, Beijing municipality is a key player in the Chinese AI landscape. For people watching AI in China, it’s worth paying attention to Beijing municipality for its own sake.
Summary
Priority 1: Increasing the availability of advanced computing power. In particular, by subsidising and aggregating compute within the city, as well as by drawing on compute resources from neighboring provinces.
The 2023 measures led to the creation of a “partnership program,” connecting providers of AI inputs (including compute) with companies training AI models. In the first year, officials reported 8.5e18 FLOP/s being provided through this program. (FLOP/s is a unit of how much computation is happening each second.) This is an amount of computing power comparable to the cluster used for GPT-4, but significantly smaller than the largest planned clusters. However, these figures may be inflated, and it's unclear whether this compute would have been available regardless of the program.
Beijing launched a platform in September 2024 that reportedly aggregates 5.6e19 FLOP/s of compute from 29 providers. The platform currently appears more focused on matching users to compute clusters rather than combining clusters together, such as to enable very large training runs.
I struggled to find data about Beijing using compute from neighboring provinces. This continues to be a stated priority so Beijing might be doing so in a significant way.
Priority 2: Increasing the supply of high-quality training data. For example, with measures to share training data between large-model developers and by supporting the production of new training data.
Government-backed Beijing Academy of AI (BAAI) has published three iterations of a “Chinese Corpora Internet” dataset of Mandarin text, scraped from the internet in mainland China. The latest version is around 1 terabyte, with a higher-quality 500GB subset. I expect this is the most significant example of the government contributing to the availability of more training data.
Officials reported that 101 datasets totaling 1,150 terabytes were shared through the partnership program in its first year. However, the low financial value (only about $1.3 million in transactions) suggests this data might not be particularly valuable or unique for AI developers.
Despite being mentioned in the 2023 measures, I could not find evidence of a government-backed platform for crowdsourcing data labeling. Its absence in later policy initiatives suggests it may have been abandoned.
Priority 3: Supporting algorithmic research. Support for research institutions doing algorithmic research, both on general efficiency improvements, and on more fundamental research, such as alternative paradigms for developing general AI.
The Beijing government continues to fund cutting-edge AI research institutions, with BAAI being the most notable. It is difficult to assess whether the 2023 measures have had any impact here. BAAI predates the measures, and the research priorities mentioned in the documents were so wide-ranging that research groups would likely have worked on them in any case.
Priority 4: Increasing safety and oversight for large model development. For example, by creating independent model evaluation benchmarks, requiring security assessments for models with “social mobilization capabilities,” and promoting research on “intent alignment.”
Beijing-backed institutions have made some progress on model evaluation, with BAAI developing the FlagEval platform shortly after the 2023 measures were announced, and the newer Beijing-AISI also working on evaluations.
The municipality still refers to these kinds of assessments as a priority. I’m not aware of Beijing-specific regulations about them, but this is likely because these regulations were implemented at the national level soon after the 2023 measures were published.
I couldn't find research by BAAI—an obvious candidate to do it—that explicitly focuses on “intent alignment.”
In the rest of this post, I explain how I approached this research, and present more detailed findings on each of the priorities.
My approach
As a way of focusing on the priorities that seemed most significant at the time, I assess Beijing’s progress towards the priorities of the documents, as articulated in Heide’s write-up. I follow Heide’s original structure and compare against his descriptions of the priorities, which I write here in pull quotes.
The evidence that I use is often specific to the particular objectives. For example, the 2023 measures included the intention to create a crowdsourcing platform to boost data labelling. To assess Beijing’s progress here, I looked online for a new Beijing-backed crowdsourcing platform. An important weakness of this relatively ad hoc approach is that I might have overlooked some important data points.
There are also three key indicators that I use for my assessments in many cases:
People’s Daily article (here) from July 2024 about the implementation of the 2023 measures. This reports on an event where officials presented the initial results of the measures. The article is the only source I could find giving such an overview. That said, it has a couple of important limitations. First, it dates from only a year after the measures were announced, whereas almost two years have now passed. Second, we should be careful about taking the claims of Beijing officials (as reported in the CCP’s newspaper) at face value.
Beijing’s AI+ plan, published in July 2024 (CSET translation here). This is Beijing municipality’s version of the national-level “AI+ Initiative,” which aims to integrate AI into various sectors. Beijing’s AI+ plan explicitly describes itself as “implementing the spirit” of the 2023 measures. Comparing the 2023 measures and the Beijing AI+ plan gives a sense of what changed in the year between the two of them being published. As with the People’s Daily article, a limitation here is that this information is already a year old.
Activities of BAAI. The Beijing Academy of AI (also known as the Zhiyuan Institute, more detail here) is a prestigious AI research institution, backed by the Beijing municipal government. As such, we might expect it to play an outsized role in implementing plans from the municipal government about AI development. Several times, I looked for what BAAI in particular had done that might be relevant to a given priority.
1. Increasing the availability of advanced computing power
The city’s science and economic policy bodies will seek to create a “compute partnership” (算力伙伴) between Aliyun — Alibaba Group’s cloud compute subsidiary — and the Beijing Supercomputing Cloud Center to subsidise and aggregate compute. The documents suggest that large model teams based in the city would then have priority access.
(In this post, quotations in the above format are always taken from the Heide summary of the 2023 measures.)
“Compute partnership” refers to the “General AI Industry Innovation Partner Program” (通用AI产业创新伙伴计划) established by the 2023 measures.
Companies can join the program as one of several kinds of partner, including “compute partner” (for companies that provide compute) or “model partner” (for companies training models). I presume that the intention of the partner program is to more effectively match together companies working on different parts of the AI supply chain.
The Partner Program has grown significantly. There are at least 260 partners now, of which at least 33 are compute partners: I compiled data on this here. The program seems still to be active, with a Beijing official referring to it in a speech in April 2025.
Amount of compute provided in the partnership: We have some numbers about this from the July 2024 People’s Daily article. The newspaper reported that compute partners in the partnership program have provided approximately “8,500P” of computing power. “P” refers to petaFLOP/s, i.e. how much computation is happening per second. So the headline figure is 8,500 petaFLOP/s, or, in different notation, 8.5e18 FLOP/s. This is not trivial but would not put someone at the very frontier of AI development—especially when that computing capacity is split between many different companies.
As a comparison, GPT-4, which was released two years ago, was reportedly trained on a cluster with a peak throughput of 7.8e18.2 And, according to the same source, there are AI clusters being planned today that are more than 100 times larger than the one used for GPT-4.
I’m also not sure what we should make of official claims, such as the claim in the People’s Daily about 8,500P of compute. My guess is that officials would tend to overstate the provision of compute, to make their measures sound more successful. They are particularly likely to overstate counterfactual provision of compute. I expect compute companies would have provided a lot of computing power, even without the 2023 measures creating a partnership program, but officials have an incentive to link the provision to the initiative. On the other hand, there might be some incentives to downplay compute provision, e.g. to avoid triggering further U.S. controls on Chinese access to compute.
Compute “interconnectivity” platform: This is perhaps the most significant effort by Beijing municipality to pool compute. Launched in September 2024, the platform reportedly aggregates compute from 29 providers, using market mechanisms to match the supply to demand.3
The announcement claims that the total amount of compute totals around 56,000P (5.6e19 FLOP/s). Around only around half of the computing power is from Beijing and the area around it (the Beijing-Tianjin-Hebei region), with the rest from farther afield.
My impression is that this platform is currently more useful for matching users to compute clusters, rather than combining individual clusters together. Combining clusters would be one way to facilitate very large training runs. There is reportedly standardization to enable matching (such as standardized identity authentication). But the same source claims that there is not yet the infrastructure that would enable combining clusters together, such as standardized interconnections between them.
Subsidized compute: There are efforts at the city level to do this, but they seem relatively small.
The People’s Daily article mentions subsidies for companies’ access to compute, but they total only around 60 million yuan ($8 million).
Shangzhuang “public computing platform”: This AI datacenter from a company owned by the Beijing government will provide compute to universities, research groups, and startups in the city. According to the announcements, it started with 500P (5e17 FLOP/s), with a planned maximum capacity of 2000P (2e18 FLOP/s). This would be roughly a sixteenth or a quarter (respectively) of the GPT-4 cluster mentioned above.
In a potential signal that Beijing companies are already struggling to locate sufficient compute for their goals, the city government plans to draw on additional compute resources from adjacent provinces such as Tianjin and Hebei to meet these goals.
I found official sources from 2024 saying that accessing compute from neighboring provinces is a priority. (Though, surprisingly to me, Beijing’s AI+ Plan does not mention it.)
I struggled to find much data about implementation. However, the sources above about the compute platform refer to Beijing-Tianjin-Hebei as one region, and also talk about using computing power from provinces farther afield.
2. Increasing the supply of high-quality training data
Policy measures announced here include a “data partnership” (数据伙伴) with nine initial members including the Beijing Big Data Centre (北京市大数据中心), as well as a trading platform to lower barriers to acquiring high-quality data for large model teams.
Providing data within the Partner Program: The quotation above references the same Partner Program as above; a further way in which companies can be involved is as “data partners”, providing data for others to use for training models. We have some numbers about the scale of this program, but not with enough detail to be particularly informative:
Looking at my figures from before, we see that the program has grown to include at least 40 data partners. I couldn’t find granular information on how much data they have shared.
The People’s Daily article from July 2024 claims that 101 datasets for large models have been shared, with a total size of more than 1,150 terabytes. But, without more information, it’s hard to know how significant this is. How much duplication is there between datasets? What kind of data is there and how helpful would it be for training general models? Is it data that companies could easily have obtained anyway, such as because it is heavily based on publicly available datasets?
As a very upper bound, this could correspond to 300 trillion tokens, equivalent to the entire stock of human-generated public text.4
However, I expect that this dramatically overstates the importance of the data that was provided. The article says that companies in total paid each other around 10 million yuan for access to data. This would only be around $1.3 million—a fairly trivial amount. For comparison, Reddit said in early 2024 that it expected to make around $66 million that year from licensing its data for training. This low financial value makes me guess that the datasets were not that valuable to the AI developers.
Trading platform: Beijing does have a state-backed “Beijing International Big Data Exchange.” However, its relevance to advanced AI models is likely low. Official reporting says that this exchange has recorded transactions worth nearly 10 billion yuan (roughly $1.3 billion) and has around 300 high-quality datasets, but is not framed in terms of large models, or even AI in particular. Additionally, analysis carried out by He and Arcesati in late 2023 found the following: “[H]alf of the listed products are credit inquiry services offered by the same state-owned financial big data company that runs the exchange on behalf of the government.” (That said, this analysis was from only around six months after the 2023 measures; the situation might have changed since then.)
The municipal government seems to also intend to support the building of high-quality training data collections, to explore making more of its own vast data reserves available for large model training, and to create a platform for crowdsourcing data labelling.
Making data available for large model training: My impression is that the Beijing government has supported significant efforts to make more data available, but that this data often isn’t from the government itself.
In a conference soon after the 2023 measures were announced, the municipal government announced the release of “18 high-quality datasets totaling nearly 500 [terabytes] across different domains including economic, political, cultural, social, and ecological areas for large model enterprises to use for training.” Not all of the publishing institutions are particularly linked to the Beijing government. For example, TRS Information Technology is a traded company. And a lot of the data is not about government topics. For example, there is a Tibetan language corpus. (Separately, I’m confused about the claim of “500 terabytes.” When I added up the size of the individual datasets mentioned, the total was less than half of that.)
BAAI has published three iterations of a large “Chinese Corpora Internet” dataset. This is composed primarily of Mandarin text, scraped from websites in mainland China. The latest version is around a terabyte, with a higher-quality 500GB subset. I think this was likely motivated, at least in part, by the 2023 measures. The first version was published in November 2023, fairly soon after the measures were published, and the Beijing government is one of BAAI’s main funders. It is the main example given in an English-language press release from the Beijing government in 2024 about the city’s efforts to facilitate usage of public data.
Platform for crowdsourcing data labeling: Despite the best efforts of Deep Research products from OpenAI and Google, I could not find such a platform. Other than in one specific context (medical data), the AI+ Plan also does not mention data labeling, making me think it’s dropped off the agenda.
3. Supporting algorithmic research
Beijing’s government states that it will aim to help its research institutions develop key algorithmic innovations.
I think Heide is referencing points seven and eleven in the plan. Point seven focuses on innovations to large models, such as how to more efficiently do pre-training. Point eleven focuses on developing alternative paradigms for general AI, such as approaches that are more directly inspired by the human brain.
Beijing municipality seems clearly to be supporting both kinds of innovation. Key examples include that it is one of the main funders of:
Beijing Academy of AI (BAAI), discussed above. This is one of China’s leading AI research institutions.
Beijing Institute for General Artificial Intelligence (BIGAI). This is one of the most prominent Chinese organizations working on alternative paradigms to AI development (Chang & Hannas, 2023).
That said, both institutions predate the 2023 measures, and it’s not clear how much of the ongoing support is because of the measures.
One approach to investigate would be to see whether Beijing institutions are working specifically on technical problems highlighted in the 2023 measures. A difficulty here is that the measures cover a very wide range of topics. Given this broad scope, Beijing institutions would likely have researched topics that are described in the measures, regardless of whether the measures had been published.
As examples, point seven calls for a focus on “model construction, training, fine-tuning alignment, and inference deployment stages.” Other parts refer to data collection and to testing/evaluation.
Probably the most specific and distinctive technical problem described in the measures is “edge deployment technologies that support models with tens of billions of parameters.” I’d be interested to see whether Beijing institutions have focused on this topic in particular, but I leave this out of scope for now.
4. Increasing safety and oversight for large model development
The municipal government wants to see independent, non-profit third parties create model evaluation benchmarks and methods.
There are at least two relevant efforts from institutions backed by the Beijing government:
BAAI does work on evaluations, such as developing the FlagEval platform. The first push to GitHub of FlagEval was in April 2023, soon after the documents were published.
My understanding is that Beijing-AISI also does work on evaluations. Beijing-AISI was launched in September 2024. This timing might indicate a response to the 2023 measures, though I doubt this was the main driver, given the wave of “AISI-like” institutions that emerged in China and internationally over 2024. (“AISI” stands for “AI Safety/Security Institute.”)
Models that have “social mobilisation capabilities” (社会动员能力) — i.e. models which can influence public opinion at scale — will need to undergo security assessments by regulators.
Regulations were introduced for this at the national level soon after the measures. I couldn’t find comparable regulations at the level of the Beijing municipal government, but this doesn’t give much evidence of the Beijing government’s commitment to the 2023 measures. It’s not surprising if Beijing didn’t reintroduce something that already existed at the national level.
Two parts of the AI+ Plan indicate that the Beijing government is still committed to this broad direction. First, the municipality expressed the intention to build a “content security large model platform” to improve the city’s ability to automatically review online content. Second, there is an explicit commitment to “supervise and guide large model enterprises implementing national legal requirements for generative AI services.”
The municipal government also seems keen on more work on “intent alignment” (人类意图对齐).
I couldn’t find published research from BAAI that is explicitly about intent alignment.5 This is notable because BAAI would be an obvious institution to implement this part of the plan; it is backed by the Beijing government and does research on large models.
On the other hand: My quick analysis of Concordia AI’s database of Chinese technical safety research does suggest that Beijing institutions put more weight on alignment relative to other safety research, compared to non-Beijing institutions. I found that around 60% of the papers whose primary institutions are in Beijing were about alignment. The same figure is about 40% for papers with primary institutions in other parts of China. This stronger focus on alignment might be an effect of the 2023 measures. However, I don’t think this is strong evidence, for various reasons.6
Conclusion
Two years after Beijing Municipality announced ambitious measures to promote general AI development, the results appear modest.
A significant challenge in this assessment is the limited availability of reliable information. For key initiatives like drawing compute resources from neighboring provinces, I struggled to find concrete data. The most detailed sources—the People's Daily article and the AI+ Plan—are already a year old. Even when numbers are available, such as the claim of 1,150 terabytes of shared training data, the reporting is sometimes too vague to draw meaningful conclusions.
It seems to me that the most important achievements are the compute interconnectivity platform, BAAI's publication of the Chinese Corpora Internet dataset, and the significant growth of the Partner Program. But there are important caveats for each of them. For example, it’s hard to know how much impact the Partner Program really has; companies in Beijing that can provide compute and data may well have worked with companies that train models, even without an official program to match them.
AI development in China is often impressive, and may well continue to be. But looking into the 2023 measures highlights that grand policymaker ambitions for AI do not always turn into reality.
The relevant term, “通用人工智能”, is sometimes also used for “AGI”.
For the GPT-4 figures, see the “Clusters used to train past frontier models” table in the linked IFP article. The theoretical maximum computing power of the GPT-4 computing cluster is three times larger than the figure I wrote in the body text; because of inefficiencies in the process of training large models, chips in practice spend much of their time idle. I’m not sure whether the People’s Daily figure is equivalent to the theoretical maximum, or to the peak throughput in practice. If the People’s Daily figure refers to the theoretical maximum, then it seems less impressive, relative to the GPT-4 cluster.
The name of the platform, “北京算力互联互通和运行服务平台” or “北京市算力互联互通和运行服务平台”, is used in the AI+ Plan from July 2024, but not in the 2023 measures. That said, the 2023 measures do include a similar idea, with section three talking about building a scheduling platform to more efficiently use diverse compute resources.
This upper bound comes from assuming that there is little duplication within the data and that it is primarily text data. In this case, a very rough estimate is that the dataset would correspond to 300 trillion tokens. (Calculations: There are roughly 3 or 4 bytes per token. A terabyte is roughly a trillion bytes. 1,150 trillion divided by 3 is 383 trillion. Divided by 4 it’s 288 trillion.) As a comparison, Epoch researchers estimate the stock of human-generated public text at around 300 trillion tokens.
I did the following two Google searches to look for this:
site:arxiv.org ("intent alignment" OR "人类意图对齐") ("baai" OR "Beijing Academy of Artificial Intelligence" OR "智源") This searches arXiv for anything that includes both the term “intent alignment” or the Mandarin translation, and one of several ways of referring to BAAI. AI research is commonly published on arXiv.
site:baai.ac.cn ("intent alignment" OR "人类意图对齐") This searches the BAAI website for anything that includes the term or its Mandarin translation. Confusingly, the BAAI website has a repository of AI papers and WeChat posts not from BAAI itself, including some that come up in this search query.
First, I don’t have a particular reason to think that the cause is the 2023 measures, rather than something else. Second, the database only goes up to April 2024. Third, the database uses Dan Hendrycks’ definition of “alignment,” which doesn’t perfectly overlap with “intent alignment.”
Hey Oliver,
I only have had a chance to skim this so I might have missed it. Are other municipalities (Shanghai, Shenzhen, Guangzhou etc), rolling out comparable compute-subsidy or data-exchange schemes, or is Beijing just the main driving force here?