As the competition in artificial intelligence continues to intensify, the discussion surrounding DeepMind, OpenAI, and Anthropic has evolved beyond simply "whose model is stronger" to a deeper question: how do top AI labs organize research, choose technical paths, and make long-term trade-offs between computing power, capital, and safety?
In a recent InfoQ interview with Sebastian Mallaby, author of "The Infinity Machine: How Google's DeepMind Mastered the Game of Intelligence" and a renowned historian of technology and finance, we attempted to understand the growth of DeepMind from a perspective closer to an "internal observer," as well as the true structural differences between it, OpenAI, and Anthropic.
As a long-term observer of Hassabis and his team, Mallaby believes that DeepMind's advantage comes not only from the computing power and funding provided by Google but also from a unique organizational approach: it allows scientists long-term freedom to explore while also enabling the rapid formation of strike teams to drive breakthroughs at critical moments. This mechanism allows it to consistently tackle problems requiring a decade or more of investment, like AlphaGo and AlphaFold.
Meanwhile, the differing trade-offs labs make between "safety and speed" are becoming important variables affecting the AGI competitive landscape. And as AI scaling expands from the training phase to the inference phase and even to future agent-level intelligence, the very structure of AI progress is changing.
In this sense, today's AI competition is no longer just a race of model capabilities but a long-term game centered around organizational models, research methods, and systemic resource allocation.
Why Has DeepMind Succeeded?
InfoQ: In your book, you mention that Hassabis's mentor at MIT, Tomaso Poggio, once said that many Nobel laureates he met were both brilliant and incredibly lucky. A small number of laureates were so gifted that they were destined to win no matter what they researched. In 2009, Hinton met Hassabis and thought he was even more competitive than himself. If you had to describe Hassabis in a few words, what would they be? Why?
Mallaby: If I were to use three words to describe him, I would say prodigiously talented, fiercely competitive, and kind and humble.
His genius goes without saying. His competitiveness stems from an innate desire to win, to achieve something remarkable. I think this is related to his participation in international chess tournaments from a young age. By the age of six, he was already immersed in intense competition, which made him extremely driven. This drive became his motivation.
But conversely, if someone were only competitive and talented but not kind, others wouldn't want to work for him, wouldn't like him, and might even try to trip him up or hinder him. His advantage is that people genuinely like him. After interacting with him, people feel he is decent, has good values, and wants AI to benefit humanity. It is precisely because of this that people are willing to support his cause. I find it hard to imagine he could have achieved what he has today without that kindness.
InfoQ: So you believe DeepMind's current success is related to the founder Hassabis's humble and kind personality? Compared to other AI labs, what is unique about DeepMind?
Mallaby: Yes, I think these are closely related to Hassabis's personality.
DeepMind's uniqueness lies first in its founding in 2010, a time when AI couldn't even reliably identify a picture of a cat. The relevant technology was completely immature, yet people were still willing to join this company and commit to a venture that had not yet succeeded. It was this超前 (chāoqián -超前 means 'ahead of its time') investment and foresight that allowed DeepMind to maintain an industry lead for roughly the first decade, until the launch of ChatGPT ushered in fierce competition. For the ten years prior, they were far ahead, all thanks to their early start, which undoubtedly is directly attributable to Hassabis. He believed in the possibility of AI from the age of 18, even before entering Cambridge University, and was steadfastly committed to this vision early on. This is perhaps another trait that defines him: he is not only competitive, a genius, and kind but also a visionary pioneer.
InfoQ: Your book mentions that Hassabis doesn't actually like business management. His ideal organizational form is a "Bell Labs + academic research institute." The DeepMind team has always operated independently in the UK. Many early policies allowed scientists long-term research, did not demand short-term products, and permitted them to publish papers publicly—completely different from many Silicon Valley companies. Don't these conditions, which seem "too free," create management difficulties for Hassabis? Usually, people accustomed to freedom, especially geniuses, are unwilling to be constrained.
Mallaby: I think the most crucial point about DeepMind is that they gave these top talents enormous freedom. Researchers could decide their own research directions, publish academic papers, and enjoy research freedom similar to that in a university.
But if DeepMind had only freedom, it would just be a group of people doing their own research, ultimately failing to produce any finished product. Hassabis combined free exploration with focused攻坚 (gōngjiān - 攻坚 means 'assaulting strongholds' or 'tackling hard problems')—when he judged that a certain research direction was ripe for a major breakthrough, he would order: "Alright, this path can be productized. We will form a dedicated team, push forward at full speed, and implement top-down, centralized management."
He calls such teams "strike teams." From the earliest development of the Atari game agent to the later AlphaGo, then AlphaZero, and up to the AlphaFold series of models, such strike teams were behind them. Within a strike team, there is a leader giving unified commands, everyone sprints toward the same goal, there is time pressure and deadlines, and the work intensity is extremely high.
So Hassabis's genius lies in combining free exploration with product transformation—leveraging breakthroughs in theoretical science to land products like AlphaGo.
He understands academic freedom because he did a PhD and knows the university research model; his thinking about forming strike teams comes from his experience in the gaming industry—he worked at a game company and founded one himself, so he很清楚 (hěn qīngchǔ -很清楚 means 'very clear') how game companies rush deadlines and deliver products on time. He perfectly fused these two models together.
InfoQ: How does a top AI lab like DeepMind organize and manage talent? For example, how do they recruit scientists? How do they manage top researchers? Do they offer high salaries or more resources?
Mallaby: In the very early stages, the company didn't offer high salaries because their funding was very limited. Although they offered stock options, most people who joined didn't believe these options would have future value—because AI at the time had neither mature products nor a clear profit model.
Therefore, attracting excellent talent in the early startup phase was quite difficult. I believe Demis (Hassabis) managed this largely because he and his co-founder Shane Legg were themselves highly respected scientists, which made other researchers willing to join their team.
Additionally, some people joined because they didn't want to deal with the "non-research" tasks in a startup, like dealing with lawyers, interfacing with investors, or considering office space. These things seemed both tedious and boring to them. If Demis was willing to handle these organizational and management tasks, they could focus on their expertise: scientific research. Therefore, some people (like Mnih, who worked on the Atari project) mentioned this was precisely why they joined.
However, after 2014, the situation changed significantly. After being acquired by Google, DeepMind gained ample financial support, allowing it to offer more competitive compensation and establish a more stable,完善 (wánshàn -完善 means 'complete' or 'well-rounded') work environment, including better office conditions and high-quality free meals. Overall, the work experience improved markedly after the acquisition.
Furthermore, as AI technology advanced, research became increasingly dependent on computing power. Achieving breakthrough results with models requires vast amounts of advanced semiconductors and computing resources. Therefore, for scientists, access to sufficient computing power became a key factor. Working at DeepMind provided direct access to Google's computing power and chip resources, which also became a major reason for attracting top talent.
InfoQ: When DeepMind was acquired by Google, it wasn't completely in a "passive survival" state. You mentioned in a previous talk that Meta CEO Mark Zuckerberg had invited Hassabis to dinner to discuss acquisition, and Elon Musk had also extended an olive branch. That is to say, DeepMind already had its own光环 (guānghuán -光环 means 'halo' or 'aura') before entering the Google system. Can this be interpreted as DeepMind having a high degree of autonomy within the Alphabet system? Is this one of the factors in DeepMind's success?
Mallaby: Yes, I believe that having other companies interested in acquiring them at the time actually strengthened their negotiating position when they ultimately sold to Alphabet.
However, I also think a more critical point is that Alphabet's leadership—especially Larry Page, and later Sundar Pichai who succeeded him—quickly realized upon meeting Demis that he was a very special person. They very much wanted to retain him and were willing to let things proceed according to his ideas because they considered him extremely valuable.
When acquiring DeepMind, deep learning pioneer Geoffrey Hinton from Canada joked: "Even buying just Demis alone would be worth £100 to £150 million." This actually reflects their high evaluation of Demis—he is not just an entrepreneur but a person with genius-level insight.
Because of this, Google gave DeepMind extremely high regard and full resource support after the acquisition, including substantial research funding. This is also one of the important reasons for DeepMind's subsequent continued success—essentially, because of Google's recognition and trust in Demis and its willingness to support his research direction long-term.
InfoQ: Even with a high degree of autonomy, Google is, after all, a commercial company. How did DeepMind balance long-term research with short-term results?
Mallaby: For a considerable period, DeepMind's focus was always on long-term research, not productization. It basically did not launch truly commercial products, although it did attempt some medical applications, but that was more due to the push from his co-founder Mustafa Suleyman rather than a core demand from Google.
Google at the time permitted, even supported, DeepMind's focus on basic research. It invested nearly $1 billion annually to support these research projects—a very large investment, but one that cash-rich Google could afford.
This situation changed fundamentally in 2022. With the launch of ChatGPT, Google suddenly realized that traditional search engines could be颠覆 (diānfù -颠覆 means 'overturned' or 'disrupted') by new search methods based on large language models. This threat made Google nervous, so they quickly pivoted, hoping to compete with OpenAI, and began demanding that DeepMind shift from being primarily research-focused to a more product-oriented direction, especially in developing large language models.
At first, I also wondered if Demis would be uncomfortable with this, given that he is essentially a scientist.
But later, when I spoke with him, his view was interesting. He said: "You have to remember, before founding DeepMind, I ran my own startup, Elixir Studios, which was a game company—we made products. I'm not averse to making products, and I'm also a very competitive person; I want to win this competition too."
He also mentioned that at this stage, AI development has entered a new state: building large language models is, on the one hand, creating products, and on the other hand, itself advancing the scientific frontier.
In fact, since the emergence of ChatGPT, we have seen a flood of new technological advancements, such as:
- Models with longer context windows (can remember more information)
- Multimodal models that can handle images, video, and audio
- Models with complex reasoning capabilities
- And agent models capable of performing tasks
These frontier explorations are themselves happening within "product forms." Therefore, at the current stage, scientific research and product development are not an either-or choice but two paths that can be advanced simultaneously.
Earning Billions of Dollars Can't Compare to Winning a Nobel Prize
InfoQ: In the first few years after the acquisition, what was the biggest cultural conflict between DeepMind and Google? To what extent did Google allow DeepMind to maintain research independence?
Mallaby: These cultural differences primarily centered on "safety" issues. DeepMind wanted to establish a more independent, special governance mechanism to ensure AI wouldn't be used merely to serve the commercial interests of one company. They envisioned forming an institution类似 (lèisì -类似 means 'similar to') an "ethics and safety review committee," whose members might include figures like former US President Obama, to decide how AI should be used.
Google sometimes seemed to support this idea superficially but had no real intention of implementing it. Therefore, the two sides engaged in three years of反复博弈 (fǎnfù bóyì -反复博弈 means 'repeated game theory' or 'back-and-forth struggle') over this issue, involving lawyers, investment banking advisors, etc.
During this period, Demis even tried to find other出路 (chūlù -出路 means 'way out'), such as contacting Alibaba founder Jack Ma, hoping to secure funding to make DeepMind independent again. This was arguably the core "cultural conflict" between them.
However, at the research level, the differences were not very apparent. Google was very supportive of DeepMind's frontier research, like the AlphaGo project.
In fact, during AlphaGo's matches in South Korea, Google co-founder Sergey Brin, then-CEO Eric Schmidt, and other executives personally attended to watch. They were very invested in and enjoyed this historic moment.
InfoQ: You mentioned that intense cultural conflicts erupted between Google and DeepMind, and you also wrote in your book that Demis almost left Google with his team. But why did he ultimately decide to stay? Did Google make some compromises?
Mallaby: Ultimately, what Demis most wanted to do during that time was still research. Because if he really had to go out and start a new company to become independent, he would not only have to run around raising funds but also hire a bunch of lawyers to "tear up the agreement" with Google's legal department, which would certainly make Google furious. For him, rather than getting bogged down in this kind of endless商业纠葛 (shāngyè jiūgé -商业纠葛 means 'commercial entanglements'), he clearly preferred to devote all his energy to攻克 (gōngkè -攻克 means 'overcoming' or 'cracking') AI scientific challenges.
So I think that's the reason he ultimately chose to stay.
Another interesting point: Demis's base is in London, not Silicon Valley. In Silicon Valley, the entrepreneurial fervor of "you must own your own company" is almost a creed; but in London, the atmosphere is less intense. Demis显然 (xiǎnrán -显然 means 'clearly') preferred to stick to the original intention he had when he sold the company to Google.
Before him lay two截然不同 (jiérán bùtóng -截然不同 means 'completely different') paths: one was to build a vast商业帝国 (shāngyè dìguó -商业帝国 means 'commercial empire') independently, becoming a super-rich billionaire; the other was to delve deeply into science, invent true AI, and ultimately win the Nobel Prize.
Clearly, compared to those billions of dollars, he craved that Nobel medal more.
InfoQ: Can we discuss a hypothetical: if DeepMind had not been acquired by Google, what would it be like today?
Mallaby: In the UK where I live, one often hears the sentiment: "Ah, if only DeepMind hadn't been sold to Google, we would have an independent British AI giant now."
But I completely disagree.
I think DeepMind was really too short on money at the time. You can完全可以 (wánquán kěyǐ -完全可以 means 'can completely') look at this acquisition from another angle: this was not a "loss for the UK AI industry"; on the contrary, it was a major victory for the UK. A天才的 (tiāncái de -天才的 means 'genius') British entrepreneur like Demis convinced an American giant to willingly invest nearly $1 billion annually into an AI lab in London.
Think about it: Americans bringing large sums of money to invest in the UK—isn't that a great thing?
If Demis hadn't sold the company to Google then, given his talent and competitive drive, he certainly would have succeeded. He likely would have remained independent or turned to investment from big names like Elon Musk. But that path would have been much more difficult, because the available funds would be smaller, and he would have to face all sorts of internal friction—after all, Musk is an extremely combative person.
Without Google's financial support, a miracle like AlphaGo might have been delayed for a long time. But regardless, Demis would still have eventually become a leading figure in AI. That is his mission and his天赋所在 (tiānfù suǒzài -天赋所在 means 'where his talent lies')—if he wants to achieve something, probably nothing in this world can stop him.
Behind AlphaGo and AlphaFold Lies the Ability to Choose Problems
InfoQ: AlphaGo可以说 (kěyǐ shuō -可以说 means 'one could say') made DeepMind famous overnight. Internally at Google,伴随着 (bànsuízhe -伴随着 means 'accompanying') AlphaGo's success, did Google's top management reassess the value of AGI? Do you think this victory changed DeepMind's level of autonomy within the group? How was this specifically reflected?
Mallaby: For everyone at Google, DeepMind's victory with AlphaGo was nothing less than a stunning wake-up call.
Through that match, everyone saw the terrifying power contained in AI. Earlier, Sergey Brin had asserted: "Making a system that can play Go must be incredibly difficult." Demis proved him wrong with facts. Google's executives had to admit that AI's evolution speed had far exceeded their expectations.
From then on, R&D budgets began to skyrocket. DeepMind's funding was already high, but after AlphaGo, it roughly doubled again.
One could say this victory became Demis's most powerful bargaining chip, allowing him to说服 (shuōfú -说服 means 'persuade') Google to keep pouring money in. And the emergence of AlphaZero a year later (2017) further consolidated his position, proving that this investment was completely worth it.
InfoQ: Both AlphaGo and AlphaFold have attracted significant attention in the industry. In your view, what methodological commonalities do AlphaGo and AlphaFold share?
Mallaby: They are both examples of an "Infinity Machine."
Take Go as an example. The first move can be placed on any of the 361 intersections, then the second player has 360 choices. Next, 359, 358... If you multiply these possibilities, you quickly get an astronomically large number—every move, every response, and the response to that response creates an explosively growing possibility space.
So, it's a huge search space, an almost infinite set of possibilities. I call it the "Infinity Machine." That is, a system like AlphaGo is essentially a machine capable of extracting "meaning" from almost infinite possible moves.
If you look at protein structure—an image of a protein, you see how it folds. There are countless tiny bends, and each bend can change in different directions. Therefore, the number of possible structural combinations a protein can form is even larger than in Go; in a sense, it's very close to "infinite."
But DeepMind still invented a machine capable of predicting the correct protein structure from all these possible combinations—the number of atoms involved in these combinations even exceeds the number of atoms in the universe. So, from this perspective, it is also an "Infinity Machine." I think this is the core connection between the two: how to extract meaningful results from such vast amounts of data and possibilities.
AlphaFold is indeed a very important, milestone achievement, globally. So, why did DeepMind choose to open-source it?
On the one hand, it was out of consideration for helping the world and advancing science; on the other hand, it was also because DeepMind believed that their ability to create AlphaFold was largely thanks to a long-accumulating scientific community.
For example, the CASP competition—initiated by academia, held every two years, where different research teams compete in protein structure prediction. Before DeepMind won in 2020, this competition had been running for about 18 years.
That is to say, before DeepMind solved this problem, the entire field had a large amount of foundational research work. Therefore, DeepMind also wanted to express gratitude by giving back to academia—they opened up the AlphaFold system, allowing the entire field to use the prediction results.
If they had just utilized everyone's open research成果 (chéngguǒ -成果 means 'results' or 'findings') without公开 (gōngkāi -公开 means 'making public') their own research, it would have seemed morally inappropriate. So, this was also an important reason for their choice to open-source.
Of course, there is also a practical factor: in commercial applications, like drug development, just AlphaFold 2 (the version that later won the Nobel Prize) is actually not enough.
You not only need to know the structure of a protein but also understand how proteins interact with each other and with other molecules. This is precisely what AlphaFold 3 and 4 aim to solve.
And AlphaFold 4 was not open-sourced; it is a proprietary system. This also means it is gradually moving towards the productization stage, so DeepMind and Google chose to keep it internal.
InfoQ: Both AlphaGo and AlphaFold are very great achievements. So how does DeepMind choose "topics worth investing a decade in"? What判断依据 (pànduàn yījù -判断依据 means 'criteria for judgment') are replicable?
Mallaby: Demis often mentions a concept called "Scientific Taste."
He has a keen intuition: he can "smell out" which unresolved problems in the AI field, though extremely challenging, are只要 (zhǐyào -只要 means 'as long as')拼命攻关 (pīnmìng gōngguān -拼命攻关 means 'working desperately to tackle') for two years, will definitely break through. He is extremely adept at making these strategic big bets. This ability stems not only from his deep understanding of the scientific frontier but also from seeing through the operating logic of the scientist community.
He once told me a fascinating story about AlphaFold.
In 2018, the team had been researching for two years. The system they made was the strongest in the world but far from the ultimate goal of predicting "all protein shapes." It was just better than other AIs, but still far from solving the problem. At the time, the team leader, Andrew Senior, even said somewhat discouraged to Demis: "We can't solve this; it's too difficult. We've tried our best. The 2018 system performs well and is领先全人类 (lǐngxiān quán rénlèi -领先全人类 means 'leading all of humanity'), but please don't force us to predict all proteins in nature; it's simply impossible."
But Demis intuitively felt that predicting all proteins was possible. To verify his judgment, he personally attended every technical seminar of the AlphaFold team.
He didn't speak much at the meetings but冷静地观察 (lěngjìng de guānchá -冷静地观察 means 'calmly observed') whether the team's discussion was "fluent."
所谓 (suǒwèi -所谓 means 'so-called') "fluent" meant whether people were continuously and quickly generating new research灵感 (línggǎn -灵感 means 'inspirations'). Demis believed that as long as inspiration was still flowing—even if some ideas didn't seem right at the moment—as long as they still had the "ability to generate new ideas," it meant the space for scientific progress was far from exhausted.
Based on this observation, Demis decided to withstand the pressure and refuse to shut down the project. He not only didn't listen to the old leader's suggestion to retreat but also replaced the commander, promoting the young John Jumper.
Jumper firmly believed success was within reach, and Demis gave him unreserved support. Two years later, they delivered research results worthy of a Nobel Prize.
This story tells us: as a leader, you not only need that scientific taste to judge "what can succeed," but you must also learn to sit in the conference room,捕捉 (bǔzhuō -捕捉 means 'capture') those sparks of碰撞 (pèngzhuàng -碰撞 means 'collision' or 'clashing') thoughts, and listen to the sound of inspiration flowing.
InfoQ: AlphaGo and AlphaFold are just阶段性实验 (jiēduàn xìng shíyàn -阶段性实验 means 'stage experiments') on the path to AGI. Do you agree with this statement?
Mallaby: Well, you could say that. I think they确实 (quèshí -确实 means 'indeed') advanced the science of artificial intelligence. And at the end of this long path will be Artificial General Intelligence (AGI).
However, from another perspective, you could also argue that the success of large language models comes from another technical route in AI research, one that does not include reinforcement learning, so it is different from the AlphaGo path.
And it is also somewhat separate from AlphaFold. AlphaFold did use Transformer models, but a special form of Transformer. Research related to AlphaFold and Transformers大致集中 (dàzhì jízhōng -大致集中 means 'roughly concentrated') between 2018 and 2020, more precisely in 2019-2020.
Meanwhile, OpenAI was already building large language models based on the Transformer architecture. Therefore, I tend to see this as another independent technical route in the progress of AI.
So,宏观上看 (hóngguān shàng kàn -宏观上看 means 'from a macro perspective'), any major breakthrough in the AI field will推动 (tuīdòng -推动 means 'push') future development; but from a more specific technical origin perspective, the birth path of large language models is relatively independent.
The True Landscape of AGI Competition
InfoQ: Based on your understanding of the AI industry, what do you think are the core differences between DeepMind, OpenAI, and Anthropic?
Mallaby: First, a big difference lies in its relationship with Google. As I said before, DeepMind has substantial financial support from Google, which is a huge advantage. Because even today, the cost of training these models remains extremely high, and the revenue they generate is actually not much.
So at this stage, you must have ample financial backing, and Demis is lucky to have Google behind him. In comparison, Anthropic and OpenAI need to constantly find investors for funding, which is itself difficult—this is a key difference.
But in other aspects, they are actually quite similar. For example, they all have their own large language models, and these models are generally of high quality. At different times, one might lead the other two, but this lead is dynamic.
If you look at their differences from another angle, you can start with "safety." Anthropic places great emphasis on safety issues, even recently having a controversy with the Pentagon over the use of AI in military systems.
DeepMind's performance in this aspect is somewhat different. I think Demis himself also values safety, including in military applications, but he hasn't had a direct confrontation with the government. So you could say Anthropic is the company most willing to take risks to promote safety issues; DeepMind (and Google) are in the middle.
As for OpenAI, it's interesting. From external appearances, it似乎 (sìhū -似乎 means 'seems') doesn't emphasize safety as much—for example, when Anthropic had its dispute with the Pentagon, OpenAI's attitude was more like: "No problem, we can provide AI."
This is somewhat similar to their strategy when releasing ChatGPT: they don't mind releasing a product that might have risks, choosing to release first and iterate later. In comparison, both Anthropic and DeepMind have held back from releasing some models due to greater caution.
So a rough ranking would be:
- Most aggressive (least conservative): OpenAI
- Middle: Google DeepMind
- Most emphasizes safety: Anthropic
You mentioned OpenAI recently recruited Peter Steinberger, founder of the开源项目 (kāiyuán xiàngmù -开源项目 means 'open-source project') OpenClaw, and asked if this would pose a threat to Google or DeepMind, especially since OpenClaw is very hot right now.
I think within Google DeepMind itself, there are already many excellent researchers working on agents. Even recruiting someone who made OpenClaw is unlikely to fundamentally change the entire competitive landscape.
Another key point: OpenClaw is indeed impressive, but it also carries certain risks. The real challenge is creating systems that are both powerful and safe—only then do they have real value for large-scale application. And whether Peter can achieve this is actually still uncertain.
InfoQ: Among Google, OpenAI, xAI, Anthropic, and Chinese companies like DeepSeek, Alibaba, Tencent, etc., who do you think is closest to AGI?
Mallaby: The definition of AGI (Artificial General Intelligence) has always been imprecise; people don't actually have a unified, clear standard.
From a certain angle, you could完全可以 understand it this way: these excellent large language models themselves possess "generality," "artificiality," and a certain degree of "intelligence." Therefore, some people might think—a certain sense of AGI has already arrived.
But this essentially depends on how you define AGI.
A more pragmatic way to understand it is: AGI should be a machine that is extremely useful in the real world, especially in commercial environments, capable of performing work originally done by humans. In other words, it's not just "smart" but "can do work."
On this dimension, you can see some very concrete progress. For example, systems like Claude Code can already write code very efficiently, to the point that the demand structure for programmers in companies might change—where 20 engineers were needed in the past, maybe only 10 will be needed in the future, because half the code is done by the model.
Of course, this capability isn't unique to one company. Others are doing similar things: like OpenAI Codex, and Google DeepMind's own code generation models; some Chinese models also perform well in code generation.
So the question becomes: who is best now? Maybe at this moment you could say Claude, but six months from now, the landscape could easily change.
In other words, on the question of "how close to AGI," we have entered a new stage: the key is no longer just "whether it exists," but "who is more practical and has stronger替代性 (tìdàixìng -替代性 means 'substitutability') on what tasks," and whether this capability can持续 (chíxù -持续 means 'sustainably') and稳定地 (wěndìng de -稳定地 means 'stably') improve.
InfoQ: So, do you think the AGI breakthrough is more likely to come from model size or new algorithms?
Mallaby: The progress we've made so far is actually the result of multiple factors working together.
It's important to understand: whether the progress comes from Google, OpenAI, or Anthropic, while they确实 are不断扩大 (búduàn kuòdà -不断扩大 means 'continuously expanding') computing power (scaling compute), this isn't everything. At the same time, they are also continuously improving algorithms, optimizing engineering implementations, and making various technical innovations; these together drive the development of AI science.
So, essentially, this is a combination of "scale + technology" driving the process. And I believe the future will likely continue this combined path.
Of course, many people now discuss whether computing power scaling will hit a bottleneck. My view is that whenever people think "scaling is nearing its limit," new ways of scaling always appear.
For example, early foundational models gained huge improvements by scaling up training, but later the marginal returns might have started to decline. Then, a new scaling path emerged—like scaling "inference models" during the inference phase, which brought new performance improvement space.
Looking further into the future, perhaps new dimensions of scaling will appear, like scaling for agents or embodied intelligence.
Therefore, my judgment is: scaling itself won't disappear; it just constantly reappears in new forms, and will still be a long-term core competitive advantage.
This article is from the WeChat public account "AI前线" (ID: ai-front), author: Dong Mei, editor: Cai Fangfang
