2025-07-27 10:31:15
Hi all,
Welcome to our Sunday edition, when we take the time to go over the latest developments, thinking and key questions shaping the exponential economy.
Thanks for reading!
Azeem
The idea of an AI-transformed economy has escaped the margins of niche blogs and podcasts. This week, The Economist made it its cover story: humanity’s next step could be machines that automate the generation of ideas. Not just labour, but thought. If that happens, growth could explode.
But their more immediate concern is the fallout between the haves (the owners of AI capital) and the have-nots (everyone else). The pressure to redistribute, regulate or surveil is set to overwhelm existing institutions.
One such institution is the job. As investor puts it, the job – once a vessel for income, identity, and social legibility – is fracturing. AI is eroding the scaffolding that made careers coherent. Early-employee ladders splinter as firms think through automation.
And those decisions matter. As I wrote just recently, when firms automate high-expertise tasks, wages fall but employment rises. When they automate routine tasks, employment shrinks but pay rises. Either way, the wage ladder bends. The job’s structural role in society weakens.
This will likely have tangible effects before any long-horizon superintelligence arrives. Political commentator Robert Shrimsley warns that the last wave of disruption hit blue-collar workers. This one targets the professional middle class, whose secure jobs underwrite political moderation. If this security disappears, the legitimacy of the system itself starts to crack.
There’s no doubt that superintelligence can provide incredible benefits. But before we can talk seriously about a superintelligent economy, we may need to confront an immediate, unstable one.
See also: Mark Zuckerberg says that Meta is seeing early signs of self-improvement in its AI models.
Antarctic sea-ice has collapsed to levels four standard deviations below the 1991–2020 average. Statistically, that should happen once every 31,600 years. And yet, it’s happened three times in the past 24 months.
Its ripple effects on ocean currents, weather systems, and ice-sheet stability will be enormous. Physics doesn’t negotiate. Nor does biology.
2025-07-26 21:13:51
Hi Azeem here,
The US-China AI rivalry has become the defining geopolitical lens through which many view the future of technology. And for good reason: it’s shaping export controls, industrial policy, and innovation trajectories on both sides. However, when a single narrative becomes dominant, it can overshadow more nuanced analysis.
That’s why I’m thrilled to hand the mic to , one of the sharpest analysts tracking China’s evolving tech landscape. Grace brings a unique perspective to the table that moves beyond great-power competition to explore the structural, commercial, and cultural dynamics shaping China’s AI ecosystem. She’s covered Asian tech for global outlets including Fortune, CNBC and the Economist Intelligence Unit. Grace has also advised top Chinese firms like Alibaba, Tencent and Lenovo – and now writes independently at .
Over to Grace!
By
On Wednesday, the White House published the AI Action Plan, a playbook for building an AI innovation ecosystem. It’s a bet on destiny that intelligence, once summoned, will reorder the world.
China isn’t chasing destiny. It’s deploying fast, frugal, open-weighted models and wiring them into the economy.
That pragmatism unnerves Washington. Chinese labs now ship foundation models faster and cheaper, and – crucially – they publish the weights. Silicon Valley reads this as “open the weights, kill the moats” – a threat to the revenues that depend on keeping the model layer proprietary.
Inside China, the logic is flipped. Once the models are treated as commodities, profit shifts to the application layer. Publishing the weights speeds that shift. Free forks and vertical fine-tunes multiply, each funnelling demand back to the originator.
However, this isn’t some grand strategy being directed by the governments on both sides; it is market-driven, shaped by three structural forces – chips, capital, and distribution – that make open‑weight releases the logical on‑ramp to value.
The real split is over where each country’s tech companies believe the profit will land. China bets on applications; America bets on the model itself. Palo Alto obsesses over model‑led destiny: ever‑bigger parameters, safety benchmarks, and a near‑cult-like obsession in the pursuit of AGI. As recounts in Empire of AI, Sam Altman and his peers see AGI as a world‑changing force capable of solving humanity’s most significant challenges. The wager is long, deep‑pocketed and proprietary: pour venture billions into loss‑making models today; own the platform that reorganizes industries tomorrow.
China’s approach is more pragmatic. Its origins are shaped by its hyper‑competitive consumer internet, which prizes deployment‑led productivity. Neither WeChat nor Douyin had a clear monetization strategy when they first launched. It is the mentality of Chinese internet players to capture market share first. By releasing model weights early, Chinese labs attract more developers and distributors, and if consumers become hooked, switching later becomes more costly.
Entrepreneurs then have the opportunity to utilize these models as free scaffolding. Taking the EV industry as an example, over twenty Chinese automakers, including BYD, Geely, and Great Wall Motors, have integrated DeepSeek into their in-car AI systems to enhance smart assistants and autonomous driving capabilities. In healthcare, it is said that nearly 100 hospitals across the country have now integrated DeepSeek for medical imaging analysis and clinical diagnosis support. Every new integration expands the model’s footprint, tightens switching costs, and shifts margins to the services sitting on top.
China’s so‑called “open‑source strategy” is a corporate pragmatism play, not a grand geopolitical scheme. But to see why pragmatism took this particular form, you need to see the three forces driving it – chips, capital, and distribution – turn openness into the default.
America has had export controls on China since October 2022, preventing Chinese model makers from accessing the latest Nvidia GPUs. Overnight, the brute‑force “scale‑to‑AGI” playbook – whether China wanted it or not – was off the menu. Chinese labs were left with a patchwork of last-gen Nvidia GPUs and local chips, at least one to two years behind the US hardware frontier.
Because chip access is capped, efficiency has become the main event. Chinese researchers have been focusing on extracting the most from their hardware. DeepSeek‑V3 delivered GPT-4o performance at 18x cheaper cost, while Moonshot’s Kimi K2 used MuonClip, which likely halved the FLOPs used during a training run. America’s chip curbs have unintentionally turned China’s models into performance‑per‑yuan champions. As many have said, “necessity is the mother of invention.” Now, Chinese models are so competitive that both local and international companies want to build on them.
The same efficiency drive also aligns with corporate reality. With venture money scarce and a lingering urge to prove they’re innovators – not copy‑cats – Chinese founders must signal value fast.
The more the models are used, the bigger their moats and their name. This is an essential characteristic of Chinese AI, stemming from the fact that China’s funding environment is scarce compared to the US. US venture capital plowed $100 billion into AI in the first half of 2025. Meanwhile, Chinese startups across all sectors raised barely $11 billion from VCs. Since China’s leading ride‑hailing app DiDi delisted from the NYSE in 2022, a regulatory storm has swept across the country’s internet sector. Some American funds pulled out of China, and even Chinese ones became skittish. The US government implemented a rule that disallowed US investors from investing in AI and chips. Chinese start-ups in AI now, compared to ten years ago during the internet era boom, raise capital only after they can demonstrate a live product and real usage figures.
Chinese models may be efficient, but DeepSeek is still estimated to have spent over $500 million in total R&D for R1. Luckily, they were self-funded by hedge fund billionaire Liang Wengfeng. But how do the other startups get funding? If you look at the “four AI tigers” – Baichuan AI, Zhipu AI, Moonshot AI, and MiniMax – most had to secure backing from the BBAT quartet. For Baichuan and Zhipu, releasing strong open-weight checkpoints helped persuade Alibaba/Tencent that the teams could ship and attract a developer ecosystem, unlocking nine-figure cheques only after the code was dropped.
Once a model maker has a reputation, what keeps them open-source?
Not every Chinese model is open‑weight. ByteDance’s Doubao 1.5 Pro keeps its parameters under lock and key – but most still publish. As mentioned earlier, this reflects the Chinese mobile-first digital economy. A typical Chinese digizen opens fewer than ten apps a month, almost all of which are funnelled through WeChat or Alipay; in the US, the figure is closer to thirty. Scale belongs to whoever can seed a model across those few choke‑points fastest. It’s led to some unorthodox decisions. Look at what Tencent did. The WeChat team wanted a good LLM fast. Yuanbao, Tencent’s AI division, wanted its proprietary engine showcased. The WeChat team went ahead and plugged in an external open-weight rival, DeepSeek R1, because it was better than anything Yuanbao had made. This is a typical move by Tencent, where product leaders trump company management in product deployment sway.
But that isn’t how Alibaba or ByteDance operates. Top company executives rallied top engineers to get their heads in the game. Beyond what is known as the 996 work culture, seven-day stints in the office to refine consumer-chatbot models are, to be honest, quite normal. In that environment, open-source software acts as growth hacks: zero licensing friction, instant plug-in, and viral forks. The BBATs want market share, and openness is the shortest route.
In truth, though, it’s not all just about corporate strategy either. It’s the chip on the shoulders of Chinese entrepreneurs that is driving their embrace of open-source, too. For decades, Chinese goods were called “copycatters.” Many AI founders and leading researchers are choosing to open-source their research to demonstrate to the world their innovation and capabilities. The psychological element shouldn’t be overlooked, as this open-source vs closed-source debate has long existed in the technology industry. For those who are pro-open, it means more scrutiny and potentially faster advancement of their research, as well as broader adoption. For those who are pro-closed, the argument is based on proprietary knowledge and protectionist thinking.
The Chinese Government doesn’t particularly care about the open-source strategy in itself. All they want is for AI to be in every pillar of its economy through its five-year plans. Instead, it’s the entrepreneurs’ philosophy, combined with structural drivers, that have pushed the open-source flywheel to the centre of the Chinese AI ecosystem.
I want to reiterate that China and the US are not running the same race. Deployment is China’s dividend, and destiny is America’s dream. Each is chasing what it values most. Chinese companies integrate open-source models into daily life because speed-to-market incentives pay off fastest at the application layer; Silicon Valley pours capital into ever-larger proprietary models, hoping to reach AGI first – whatever that means at this point.
Chinese model makers are responding rationally to their environment. Chip constraints reward efficiency, while funding structures and market-share chasing dynamics push them toward open-source. Make no mistake, China’s current open‑source “strategy” is market‑driven, not state‑driven. It is pragmatic; as a Chinese saying goes, “兴趣不可以当饭吃,” literally meaning “interests can’t be eaten as food.” In China, even AI must earn its keep.
2025-07-25 18:07:04
OpenAI is set to release GPT-5 next month. Reports suggest it will be a unified package rather than a typical version upgrade.
OpenAI describes it as a system that incorporates several distinct models, each triggered by the nature of the query. It might use Deep Research for a research task, and 4o-level intelligence to copyedit your email. Effectively, it’s a UX upgrade. That may sound underwhelming, but don’t forget, a UX upgrade drove the ChatGPT boom itself. GPT-3, the model behind the chatbot, had been available for two years before the craze began. What changed was the interface: wrapping the model in a conversational chatbot made it intuitive, playful, and useful to a mass audience.
GPT-5’s new unified interface may have the same effect. Being able to shift between agentic tools and chat may make it feel more collegial, as if you have someone you can bounce ideas off or delegate tasks to. [The Verge]
A 30-second scan of four secondary signals that hint at where the curve is bending.
📦 $1 billion of Nvidia chips evaded sanctions and were smuggled into China since May, suggesting attempts to stymie the trade in advanced chips have been largely ineffectual. Chinese middlemen are flipping GPU racks at around 50% above US prices – highlighting the value of bleeding-edge compute. [FT]
🦾 ServiceNow is saving $100 million a year in staffing costs as generative AI agents now reduce time spent on support tickets by 50%. It isn’t the only company to do so; in an internal memo, Microsoft CEO Satya Nadella acknowledged the “enigma” of laying off thousands of staff as the firm thrives financially. [Axios] [Geekwire]
💥 Generative AI is a transformative dynamo and a revolutionary microscope to society, a new paper from the Federal Reserve claims – but not a burn-fast-and-burn-out lightbulb. The general-purpose nature of the tech, like a dynamo, coupled with its function as an Invention of a Method of Invention (IMI), like a microscope, enhances and automates scientific discovery and research workflows. [Federal Reserve]
📈 Google’s AI systems processed a quadrillion tokens (1,000,000,000,000,000) in June, more than double the volume (480 trillion) in May. [X/Demis Hassabis]
A pulse-check on the ideas shaping our long-term trajectory.
🧵 A breakthrough in biofabrication could edge us closer to a plastic alternative. Scientists have developed a simple, scalable method to grow bacterial cellulose sheets with aligned nano-fibers, finally rivaling plastic’s tensile strength. Strong, flexible, transparent – and compostable. [Nature]
💸 Copper-doped β-NaMnO₂ could drive battery costs lower than today’s sodium-ion cells. Until recently, though, it barely survived 30 charge cycles — compared to 100–500 for typical sodium-ion chemistries. A Tokyo University breakthrough just pushed that to 150 with zero capacity loss. Still early, but a big leap for cheap, scalable chemistry. [Robotics and Automation News]
UBTECH’s Walker S2 has become the world’s first humanoid robot capable of autonomously swapping its batteries. This means it can operate 24/7, offering a major advancement for continuous autonomous operation in factories and customer service roles. [LiveScience]
Was this email useful?
Thanks! Want to share this with someone like you? Forward the email.
2025-07-24 18:53:14
The White House’s new AI Action Plan reads like an industrial‑scale mobilization order. Its 90 recommendations slash permitting delays, sideline DEI and climate hurdles, and fast‑track gigawatt data‑center grids and fabs. They also hard‑wire “objective, non‑ideological” procurement rules that privilege open‑weight US models.
The aim is simple: unquestioned technological dominance. Full‑stack export diplomacy — bundling US chips, open‑weight models and standards — plus tighter chip controls keep allies close and China at arm’s length. The document is even titled ‘Winning the Race’.
This is AI as statecraft. Compute, energy and standards bodies are the new strategic high ground. Private industry is invited to write the rulebook, recommending which regulations to scrap in exchange for momentum. The upside is speed. The downside (mainly for the rest of the world) is a regulatory vacuum primed for social, environmental and geopolitical blowback. Expect a trans-Atlantic clash pitting Washington’s ‘objective’ AI against Europe’s DEI-anchored regime. [White House]
A 30-second scan of four secondary signals that hint at where the curve is bending.
📈 Alphabet has hiked its 2025 AI capex to a record $85 billion, around 22% of all its revenue. The figure, announced in Google’s latest earnings call, eclipses many sovereign budgets, indicating that scale itself is becoming a moat. The firm also announced its AI Overviews in search have 2 billion monthly users, while Gemini has hit 450 million monthly active users. Search revenues remain robust, climbing 12% year‑on‑year to $54.2 billion. [FT, WSJ]
👨🏻💻 China has 2.2 million open-source developers, compared with 1.7 million in the US. This eastward shift in open-source talent undermines Western model licensing and fuels a community-driven AI stack outside US influence. [SCMP]
🏛️ More than 500 organizations lobbied regulators in Washington DC on AI in the first six months of 2025. OpenAI alone spent $1.8 million on lobbyists, showing that regulatory capture is scaling with the model race. [FT]
🔋 The average battery size in Chinese battery electric vehicles (BEVs) rose around 4% month-on-month in June to 67.5kWh. At this pace, range anxiety doesn’t stand a chance. [X/EVCurveFuturist]
Last week, I chatted with , CEO of The Atlantic. He shared with me some useful prompts he uses to pop his filter bubble and sharpen his thinking:
“Imagine you’re totally skeptical of this person. Pretend you want to find every error. Now read the essay and write the most obnoxious but intelligent critique you can.”
That gives you a guide to harden the argument or rephrase it. It’s especially useful in politics.
“Read this story and identify the six embedded assumptions you don’t even notice — because you’re in a filter bubble — that will trigger readers on the other side.”
Once you do that exercise, you start spotting things. Like a throwaway line about Brendan Carr — just to pick a random example. It might seem harmless, but it’s guaranteed to piss off a bunch of people. And you didn’t even notice. So cut that stupid little line about Brendan Carr, right?
Watch the full episode on YouTube. I go live on Substack every Friday at 5 pm UK time | 12 pm Eastern | 9 am Pacific.
Was this email useful?
Thanks! Want to share this with someone like you? Forward the email.
2025-07-23 22:01:48
Many of us feel fluent in prompting AI by now, but still feel frustrated when the results fall short. The issue usually isn’t the model. It’s how we talk to it.
Too often, we fall into what developer Mitchell Hashimoto calls “blind prompting”, treating the AI like a helpful colleague instead of instructing it with purpose and structure. In one 2023 study, participants who got a decent result early on assumed their prompt was “done,” and never refined it further.
If you hired a new team member and gave them only vague tasks with no feedback, you wouldn’t be surprised if they struggled. The same goes for AI. Models need context. They need iteration. And they benefit from review, correction and calibration, just like people do. And as AI becomes one of the most important inputs in modern work, it’s time to start managing it as such.
There are now dozens of system prompts from labs and startups floating around online. AI labs have invested millions of dollars in developing them. They’re all expertly crafted and sophisticated, but Claude 4 is one of the most extensive to date. Claude’s system prompt is at 24,000 tokens, or ~9,700 words, 453 sentences in length. For comparison, the Gemini-2.5-Pro and o3_o4-mini leaks stand at 1,636 words and 2,238 words, respectively.
For all of you who want to improve your skills, studying these leaks might be one of the best routes available. Our team has studied Anthropic’s internal guide to Claude 4 to identify seven key rules, along with examples (included in the leak and our own in italics), that will enhance your prompting game, guaranteed.
You might not realize it, but being specific, with clear, formatted instructions, can dramatically improve your AI results. Even subtle tweaks in wording can improve the accuracy by as much as 76%. The Claude system prompt clearly defines that the standard responses should be in paragraph format, unless the user specifies otherwise:
Claude responds in sentences or paragraphs and should not use lists in chit chat, in casual conversations, or in empathetic or advice-driven conversations. In casual conversation, it's fine for Claude's responses to be short, e.g. just a few sentences long.
...
If Claude provides bullet points in its response, it should use markdown, and each bullet point should be at least 1-2 sentences long unless the human requests otherwise. Claude should not use bullet points or numbered lists for reports, documents, explanations, or unless the user explicitly asks for a list or ranking. There are also numerous stylistic notes instructing Claude precisely how to craft its responses in different contexts, such as:
Claude is able to maintain a conversational tone even in cases where it is unable or unwilling to help the person with all or part of their task.Conversely, Claude remains thorough with more complex and open-ended questions. And for those seeking more affective conversations, the system prompt writes:
For more casual, emotional, empathetic, or advice-driven conversations, Claude keeps its tone natural, warm, and empathetic.How can we make use of this? When prompting your large language model (LLM), clearly define the AI’s role, your exact task and the desired output. This can include the format of the output in terms of the style, length, and formatting, to more precisely get your desired result.
You are a senior copy editor. Task: rewrite the following 200-word blog intro so it’s clear, concise, and in plain English. Output: one paragraph, no bullet points, max 120 words.It is becoming more well-known that providing clear examples of what you want can result in more aligned outcomes with your goal. However, we also see that describing what not to do can be of benefit, too. These crop up multiple times in the Claude system prompt, such as in the following example:
Any calculations involving numbers with up to 5 digits are within your capabilities and do NOT require the analysis tool.Alongside this are some examples to guide Claude on where not to use the analysis tool:
Do NOT use analysis for problems like "4,847 times 3,291?", "what's 15% of 847,293?", "calculate the area of a circle with radius 23.7m" In fact, the system prompt contains a larger frequency of the word “never” compared to “always”, with 39 instances of the former versus 31 for the latter. The word “always” doesn’t even make it into the top 20 words in the prompt, while “never” has the 13th highest frequency.
Although these arbitrary examples of what not to do may seem random, they help to guide the tool behaviour of the LLM, leading to faster outputs and fewer tokens needed to process the task.
How can we make use of this? The context windows (available window of attention for an LLM) are growing bigger and bigger. For complex problems, provide the LLM with examples of what you do and do not want, to clarify the expectations and improve accuracy.
Edit this paragraph for a general audience. DON’T use jargon like ‘transformer’, ‘parameter’, or ‘hallucination’. If any slip through, revise the sentence immediately.By now, we all know that LLMs tend to hallucinate if they don’t know the right answer. Aware of this, Anthropic has opted to include a line to ensure that Claude can send the user to the best place for the latest information.
If the person asks Claude about how many messages they can send, ... , or other product questions related to Claude or Anthropic, Claude should tell them it doesn't know, and point them to 'https://support.anthropic.com'.Providing an “escape hatch” can help reduce hallucinations, allowing the LLM to admit where it is lacking in knowledge instead of attempting to answer (and often being very convincing even if the result is completely incorrect).
How can we make use of this? Explicitly ask the LLM to mention “I don’t know” if it does not have the knowledge to comment. Additionally, you can provide another “hatch” by adding that it can and should ask for clarifying information if needed to provide a more accurate answer.
If you’re < 70 % confident about a style rule, respond: ‘I’m not certain—please verify in theAP Stylebook or Chicago Manual of Style before editing. Otherwise, edit normally.Claude spends 6,471 tokens, nearly a third of its system prompt, just instructing itself how to search.
2025-07-23 19:43:18
Anthropic estimates that the US will need to build 50 gigawatts of new power capacity by 2028 to meet the demand for AI. That’s nearly twice New York State’s peak summer load, or the output of 33 nuclear reactors. The warning comes as OpenAI and Oracle have just announced new Stargate AI data centers that will consume 4.5 gigawatts of power themselves. Leopold Aschenbrenner, who wrote the Situational Awareness paper in 2024, projected that the largest training cluster in 2028 would be about 10 gigawatts. It probably won’t reach quite that level, since most of the capacity will go to inference and training will be split among multiple providers and clusters. Nevertheless, these developments point to the emergence of some enormous clusters. [Anthropic, Tom’s Hardware, Situational Awareness]
A 30-second scan of four secondary signals that hint at where the curve is bending.
🗣️ ChatGPT has more than 500 million weekly active users, who send more than 2.5 billion prompts every day, 330 million of which come from the US. Sam Altman is revealing the rare user data as he tours Washington DC to make the case for hands-off regulation. [Axios]
⛔ Google’s AI overviews cut external clicks by 7 percentage points and raise session endings by 10 points compared with standard search results, a Pew study shows. This is yet more quantitative proof that generative summaries can cannibalize the open web, eroding the traffic economics on which advertising, e-commerce, and media depend. [Pew Research Center]
🐝 Amazon has acquired Bee, which makes a $49 AI-assistant wrist device, for an undisclosed sum. The purchase signals a renewed push for voice-and-vision AI wearables priced for mass adoption, an alternative path to the smartphone. [TechCrunch]
♻️ UN secretary-general António Guterres wants AI data centers to use 100% renewable energy by 2030. Diplomatic pressure turns the carbon intensity of AI compute from a tech‑industry debate to a multilateral climate priority. [Bloomberg]
As generative AI accelerates coding and research, the bottleneck shifts to determining what is worth building. Andrew Ng calls this the ‘product‑management bottleneck.’ He says winning teams have product managers who pair deep user empathy with a fast, data‑refined gut instinct. They make most decisions quickly and slow down only when fresh evidence contradicts their mental model.
Another bottleneck is ensuring that AI-generated content meets expectations. We recently saw the downside when our AI fact-checking process failed. The broader lesson is clear: progress in any AI‑driven workflow hinges on two equally critical abilities – framing high‑value tasks and rigorously validating outputs – now that AI can handle the generation. [Andrew Ng]
Was this email useful?
Thanks! Want to share this with someone like you? Forward the email.