2025-12-30 00:03:36
The question of whether AI is “good enough” for serious knowledge work has been answered. The models crossed that threshold this year. What’s slowing organizations down now isn’t capability, but the need to redesign work around what these systems can do.
We’ve spent the past 18 months figuring out how. We made plenty of mistakes but today I want to share what survived. Six mental models that can genuinely change the quality of work you get from generative AI. Together with the seven lessons we shared earlier in the year, this is the operating manual we wish we had all along.
At the end, you’ll also get access to our internal stack of 50+ AI tools. We’ve documented everything we’re actively using, testing or intend to test, to help you decide what tools might work for you.
Enjoy!
Most people start working with AI by asking something along the lines of: how do I speed up what I’m already doing?
That question is comfortable and wrong. I find that it anchors me to existing constraints.
A more useful question is:
What would I do if I had 50 people working on this?
Then work backwards.
The 50x reframe forces you to imagine the ideal outcome unconstrained by time or labor. Only then do you ask which parts of that hypothetical organization can be simulated with software. I now encourage our team members to think of who they would hire, what work that person would do, how they’d know if they were successful.
If you’ve not had the experience of hiring fifty people on a project (fair enough!), use this prompt to get you started to identify what it is that you may need:
I currently [describe your task/process]. Walk me through what this would look like if I had a team of 50 people dedicated to doing this comprehensively and systematically. What would each role focus on? What would the ideal output look like? Then help me identify which parts of that hypothetical team’s work could be automated or assisted by AI tools.For example, we use this approach for podcast guest prospecting and research. We used to rely on network and serendipity to identify 20-30 strong candidates for each season; a mix of the right expertise, timing and editorial fit that consistently delivered good conversations – but left too much to chance. Instead, 50x thinking asks what if we could systematically evaluate the top 1,000 potential guests? What if we could track the people we’re interested in so they surface when they’re most relevant? We built a workflow that researches each candidate, classifies expertise, identifies timely angles, and suggests the most relevant names for any given week’s news cycle.
Even experienced operators have blind spots. We internalize standards of “good” based on limited exposure. No one has seen great outputs across every domain but the models, collectively, have come closer to it than anyone else.
To make the most of this superpower, I give Gemini, Claude and ChatGPT the same task – and make them argue. Then have each critique the others. You’ll quickly surface gaps in your framing, assumptions you didn’t realise you were making, and higher quality bars than you expected.
When models disagree, it’s usually a sign that the task is underspecified, or that there are real trade-offs you haven’t surfaced yet. Which brings us to the next point.
If you’re having the same conversation with AI repeatedly, turn it into a tool. Every repeated prompt is a signal. Your workflow is basically telling you that this (t)ask is valuable enough to formalize.
I found that when I productize a conversation by turning it into a dedicated app or agentic workflow, my process gets better at the core and my tool evolves over time. So the benefits of the original conversation end up compounding in a completely new way.
## Context
I have a recurring [FREQUENCY] task: [BRIEF DESCRIPTION].
Currently I do it manually by [CURRENT PROCESS - 1-2 sentences].
Here’s an example of this task in action:
<example>
Input I provided: [PASTE ACTUAL INPUT]
Output I needed: [PASTE ACTUAL OUTPUT OR DESCRIBE]
</example>
## What I Need
Turn this into a reusable system with:
1. **Input specification**: What information must I provide each time?
2. **Processing instructions**: What should the AI do, step by step?
3. **Output structure**: Consistent format for results
4. **Quality criteria**: How to know if the output is good
## Constraints
- Time/effort budget: [e.g., “should take <5 min to run”]
- Depth: [e.g., “verify top 10 claims, not exhaustive”]
- Tools available: [e.g., “has web search” or “no external lookups”]
- Error handling: [e.g., “flag uncertain items vs. skip them”]
## Desired Format
Deliver this as a: [CHOOSE ONE]
- [ ] System prompt I can paste into Claude/ChatGPT
- [ ] Zapier/Make automation spec
- [ ] Python script (Replit-ready)
- [ ] Lindy/agent configuration
- [ ] All of the above with tradeoffs explained
## Success Looks Like
A good output will: [2-3 bullet points describing what “done well” means]I kept asking LLMs for editorial feedback multiple times a week, for weeks. After some fifteen hours of repeat prompting, I built a virtual editor panel in Replit and Lindy.
2025-12-26 18:10:18
This year, I recorded 30 conversations on AI, technology, geopolitics, and the economy.
What follows is a 20-minute edit, a way in. If you only had 20 minutes to orient on key ideas that were top of mind in 2025 and will hold as we step into the new year, these are the ideas I’d start with:
Listen on Apple Podcasts or Spotify
Part 1 - AI as a general purpose tech
2:27: Why this matters
3:06: Kevin Weil: The test for a startup’s longevity
4:01: Matthew Prince: The “Socialist” pricing debate
4:45: : This will stifle the AI boom
7:42: : The “NBA-ification” of journalism
8:13: : From utopia to protopia
Part 2 - How work is changing
10:13: An evolving labor market
10:45: Steve Hsu: The future of education
11:27: Thomas Dohmke: The inspectability turning point
12:09: Ben Zweig: The new role for entry-level workers
13:16: Ben Zweig: The eroding signal of higher education
Part 3 - The physical world, compute, and energy
14:05: Setting the stakes
14:51: Greg Jackson: “We’re half way across the road. We have to get across as fast as we can.”
15:27: Greg Jackson: Building a “show, don’t tell” company
16:12: : The physical reality of AI
Part 4 - The changing US-China landscape
16:57: A new era
18:09: Dan Wang: The West’s hidden tech gap
18:38: : The two types of accelerationism
19:38: Jordan Schneider: What the US can learn from China
Kevin Weil - OpenAI CPO on what’s coming next in AI
Matthew Prince - Inside the collapse of the internet economy
Tyler Cowen - How AI will reorder economies, schools, and spirituality
Nick Thompson - AI and the future of media
Kevin Kelly - The handoff to “bots”
Steve Hsu - The difference between early and late AI adopters
Ben Zweig - Where did all the entry-level jobs go?
Thomas Dohmke - What AI means for developer salaries, SaaS, and more
Greg Jackson - The future of clean energy
Dan Wang - Why China builds while America debates
Jordan Schneider - Inside the US–China decoupling
🔔 Subscribe on YouTube for every video, including the ones we don’t publish here.
2025-12-23 12:17:18
“I subscribe because you are naming fractures that matter.” – Lars, a paying member
argues we’ve got one common AI metaphor wrong. We talk about training AI as if we’re training animals, building instincts and ultimately, shaping behaviour. No, he says in his year-end review:
We’re not “evolving/growing animals”, we are “summoning ghosts”. Everything about the LLM stack is different (neural architecture, training data, training algorithms, and especially optimization pressure) so it should be no surprise that we are getting very different entities in the intelligence space, which are inappropriate to think about through an animal lens.
He goes on to say that the “ghosts”
are at the same time a genius polymath and a confused and cognitively challenged grade schooler, seconds away from getting tricked by a jailbreak to exfiltrate your data.
We’re familiar with that jaggedness. and colleagues ran a field experiment two years ago involving management consultants and GPT-4. Inside the model’s competence zone, people go faster (25%) and their outputs are better (40%). But on a task outside the frontier, participants became far more likely to be incorrect. Same tool, opposite effects.
This jaggedness has more recently surfaced in Salesforce’s Agentforce itself, with the senior VP of product marketing noting that they “had more trust in the LLM a year ago”. Agentforce now uses more deterministic approaches to “eliminate the inherent randomness” of LLMs and make sure critical processes work as intended every time.
Andrej’s observation cuts deeper than it may initially appear. The capability gaps he describes are not growing pains of early AI. They are signatures of how training works. When a frontier lab builds a model, it rewards what it can measure: tasks with clear benchmarks, skills that climb leaderboards and get breathlessly tweeted with screenshots of whatever table they top.
The result is a landscape of jagged peaks and sudden valleys. GPT-4 can pass the bar exam but struggles to count the letters in “strawberry.” And yes, the better models can count the “r”s in strawberry but even the best have their baffling failure modes. I call them out regularly; and recently Claude Opus 4.5 face planted so badly that I had to warn it that Dario might need to get personally involved:
What gets measured gets optimised; what doesn’t, doesn’t. Or, in Charlie Munger’s way, show me the incentive and I’ll show you the outcome.
The benchmarks a system trains on sculpt its mind. This is what happened in the shift to Reinforcement Learning from Verifiable Rewards (RLVR) – training against objective rewards that, by principle, should be more resistant to reward hacking. Andrej refers to this as the fourth stage of training. Change the pressure, change the shape.
The essay has raised an awkward question for me. Are these “ghosts” something alien, or a mirror? Human minds, after all, are also optimised: by evolution, by exam syllabi, by the metrics our institutions reward, by our deep family backgrounds. We have our own jagged profiles, our own blind spots hiding behind our idiosyncratic peaks.
So, are these “ghosts” that Andrej describes, or something more familiar? Our extended family?
Psychometrics has argued for a century about whether there’s a general factor for intelligence, which they call g, or if intelligence is a bundle of capabilities. The durable finding is that both are true. Cognitive tests positively correlate and this common factor often explains up to two-thirds of variation in academic and job performance.
But plenty of specialized strengths and weaknesses remain beneath that statistical umbrella. Humans are not uniformly smart. We are unevenly capable in ways we can all recognize: brilliant at matching faces, hopeless at intuiting exponential growth.
The oddity is not that AI is spiky – it’s that we expected it wouldn’t be.
We carried over the mental model of intelligence as a single dial; turn it up, and everything rises together. But intelligence, human or machine, is not a scalar. Evolution optimized life for survival over deep time in embodied creatures navigating physical, increasingly crowded and ultimately social worlds. The result was a broadly general substrate which retains individuated expression.
Today’s LLMs are selected under different constraints: imitating text and concentrated reward in domains where answers can be cheaply verified. Different selection pressure; different spikes. But spikes nonetheless.
What does spikiness imply for AGI as a single system that supposedly does everything?
My bet remains that we won’t experience AGI as a monolithic artefact arriving on some future Tuesday. It will feel more like a change in weather: it’s getting warmer, and we gradually wear fewer layers. Our baseline expectation of what “intelligent behaviour” looks like will change. It will be an ambient condition.
The beauty is that we already know how to organise such spiky minds. We’re the living proof.
Human society is one coordination machine for spiky minds. It’s the surgeon with no bedside manner. The mathematician baffled by small talk. The sales rep who can close a deal on far better terms than the Excel model ever allowed.
2025-12-22 21:26:58
Hi all,
Over the past year, we reviewed thousands of data points on AI and exponential technologies to identify the signals that mattered most. Across almost 50 editions of Monday Data, we shared ~450 insights to start each week.
Today, we’ve distilled them into 25 numbers that defined the year, across five themes:
The state of intelligence
The state of work
The state of infrastructure
The state of capital
The state of society
If you enjoy today’s survey, send it to a friend.
Let’s jump in!
Margin gains. OpenAI’s compute profit margin – revenue after deducting model operation costs – reached 68% in October 2025, roughly doubling from January 2024 (36%) as efficiency improved.
Demand curve. Following the Gemini 3 release, Google has been processing more than 1 trillion tokens per day. Moreover, token output at both Alibaba and OpenRouter has been doubling every 2-3 months.
Open beats on price. Shifting from closed models to open models was estimated to reduce average prices by over 70%, representing an estimated $24.8 billion in consumer savings across 2025 when extrapolated to the total addressable market.
Trust gap. Twice as many Chinese citizens (83%) trust that AI systems serve the best interests of society as Americans (37.5%).
Where it’s used most. AI usage is clustered in service-based economies – with UAE (59.4%) and Singapore (58.6%) in the lead; the US ranks ~23rd.
Near universal adoption. Close to 90% of organizations now use AI in at least one business function, according to McKinsey.
Time saved. ChatGPT Enterprise users report 40-60 minutes of saved time per day thanks to AI use. Data science, engineering and communications workers save more than average at 60-80 minutes per day.
Hiring tilts to AI. Overall job postings for the whole labor market dropped ~8% YoY1, but AI roles are on the rise. Machine learning engineer postings grew almost 40%.
Automation pressure. In areas with autonomous taxis, human drivers’ pay has fallen. Down 6.9% in San Francisco and 4.7% in Los Angeles year-on-year.
Early-career squeeze. Early-career workers between 22-25 in the most AI-exposed occupations (including software developers) saw a 16% relative decline in employment since late 2022.2
2025-12-21 11:15:30
Hi all,
Today’s our final Sunday edition of 2025, open to all. From Monday, we move to our holiday schedule – no Sunday editions until 11 January and a pared‑back weekday rhythm. You’ll still hear from us.
I’ll host a member‑only 2026 briefing on January 7. If you’re a paying member, you’ll receive the invitation — please RSVP here. (Not a member yet? This is a great time to join.)
If you are based in London🇬🇧, please scroll to the bottom of today’s edition where I ask for a favor.
Enjoy the reading and happy holidays!
Azeem
The Cold War’s defining symbols, the US nuclear carrier and the silent Soviet submarine, represented a clash of industrial-scale superpower might. I explained in my book back in 2021 how the logic was inverting. That inversion is now without doubt.
Ukraine’s Sub Sea Baby UUV, costing a fraction of Russia’s $400M Kilo-class submarine, just breached a fortified harbor to strike at point-blank range. Without a Navy, Ukraine has more than “matched” Russia at sea. It changed the accounting, as the Russians need to protect every vulnerability from systems both hard to detect and cheap enough to risk repeatedly. For Moscow, it’s the worst bargain in warfare.
Meanwhile, China’s hypersonic missiles and stealth platforms are pushing US carriers so far offshore that their air wings risk irrelevance. In wargames, “we lose every time,” says one military boss. The carrier has gone from being the Queen, able to reach every corner of the board, to being the King, protect at all costs. No wonder the Pentagon refers to this as “Overmatch.” (More details here.)
Where superpowers once competed to build the most imposing platforms, smaller actors now exploit exponential tech to invert the cost-imposition calculus: Ukraine’s drone swarm renders Russia’s Black Sea Fleet obsolete, while China’s A2/AD network negates America’s carrier supremacy. The 20th century’s totems of power are becoming the 21st century’s most vulnerable liabilities.
But “drones beat ships” or “cheap beats expensive” isn’t the takeaway. The point is that exponential technologies collapse the old prestige hierarchy. The decisive edge shifts from owning the biggest tchotchkes to building the fastest adaptation loop.
See also:
Russia’s war bloggers reveal a new model of authoritarian control: by cultivating critics who question tactics but never the war itself, the Kremlin has turned authentic dissent into a tool of legitimacy.
For years, extreme ultraviolet lithography (EUV) has been treated as the West’s unassailable hardware moat. It’s often described as the most complex machine ever built, used to print the world’s most advanced chips. See this video primer to understand its importance:
EUV was the one technology Western analysts assumed China could not replicate. After all, ASML, the Dutch firm holding a global monopoly, has shipped zero EUV machines to China since 2019. That assumption is now under strain, as we learn that a Chinese lab unveiled a working EUV prototype in early 2025, with a plan to deploy domestic tools for chip production by 2028–2030. If confirmed, this upends the calendar, which was once anchored in the mid-2030s. Export controls designed to maintain a ten-year buffer may soon look like a fence built five metres behind the finish line…
Worth noting, this timeline is not a given. A single ASML EUV tool is built from roughly 100,000 parts supplied by about 5,000 suppliers. Replicating the physics is one thing. But replicating that ecosystem is another. Frontier mass production requires many machines, high uptime, stable yields, and a deep web of optics, light sources, resists, metrology, and replacement parts. Credible skeptics note that China may be salvaging components from older ASML tools or sourcing them through secondary markets.
Still, China has repeatedly beaten Western expectations when narrowing supposedly long-lasting technology gaps. EUV has been the West’s last clean chokepoint in compute. If this bottleneck loosens, the idea of chip dominance as a stable geopolitical lever becomes far less certain.
See also:
My past conversations on the US-China decoupling with and on why China innovates while the US debates with Dan Wang.
The conventional wisdom that America’s power grid problems will hand China victory in the AI race is overblown. argues that the US has multiple viable paths to meeting AI’s energy demands, including natural gas, off-grid solar and tapping spare grid capacity during off-peak hours.
When I visited Silicon Valley in 2023, I noted the following in my reflections at the time:
I heard the word ‘exponential’ a lot. Underpinning Silicon Valley is the exponential drumbeat of Moore’s Law. But that word, ‘exponential,’ wasn’t a quotidian presence in 25 years of conversations. It’s just always been there. Goldfish don’t talk about the bowl.
Two years later, that implicit faith has crystallized into what Eric Schmidt calls the “San Francisco Consensus,” a set of premises that unite those leading the AI development:
[B]eneath the apparent discord lies a deeper consensus around a number of key ideas. Most of those leading the development of Al agree on at least three central premises: First, they believe in the power of so-called scaling laws, arguing that ever larger models can continue to drive rapid progress in Al. Second, they think the timeline for this revolution is much shorter than previously expected: Many Al experts now see us reaching superintelligence within two to five years. Finally, they are betting that transformative Al (TAl), or systems that can outperform humans on many tasks, will bring unprecedented benefits to humanity.
While I don’t disagree with these premises, I’d note that the Valley has always rightly believed in exponential improvements and has been repeatedly wrong about where that technology leads. The same technologists who promised the internet would democratize knowledge gave us filter bubbles and platform monopolies. The consensus describes what believers expect to build. It has little to say about what happens next. Eric, too, acknowledges as much.
This is where Europe’s skepticism can have value, even if it’s often misdirected. The bitter taste of the platform era left many European founders and policymakers doubtful of Silicon Valley’s claims about technology’s benefits. Fair enough. But equally, it would be a mistake to assume this means AI isn’t powerful – it is. The question is who shapes what it becomes. We need these perspectives to meet and work together more than ever. As Eric says,
Silicon Valley alone cannot answer the profound economic, social, and—quite frankly—existential questions that will come of this transformation.
See also:
For a different perspective: this essay argues that those who push the apocalyptic narratives around superintelligence are looking to avoid the difficult questions about algorithmic harm, labor exploitation and corporate accountability.
AI’s main scaling wall is energy, both to power chips and to cool them, as I’ve argued many times over. Data centers are no longer background infrastructure; they’re competing for land, power and political attention. As those constraints tighten on Earth, proposals to move compute off-planet are becoming more attractive. Investor Gavin Baker calls data centers in space “the most important thing in the next three to four years.” The benefits are: ~30% more usable solar flux than Earth’s surface under ideal conditions, no weather or atmospheric absorption, “free” cooling via radiators and lasers through vacuum. Of course, there will be engineering obstacles to solve.
When batteries started attracting skepticism, the debate fixated on energy density and cycle life. But the decisive factor was manufacturing scale: over three decades, learning curves drove a 97% drop in cost per kWh – this turned batteries, once exotic technology, into universal infrastructure. Orbital compute sits at a similar inflection. The physics largely works but the constraint is industrial scale. If launch costs fall toward ~$200/kg by the mid‑2030s on Starship‑class trajectories1 and hardware follows comparable learning curves, space‑based compute could go from speculative to strategically viable.
AI, technology & science:
argues that Google’s TPUs vs. Nvidia’s GPUs may be the most consequential chip architecture battle since CISC vs. RISC.
OpenAI’s competitor to Nano Banana is out. It’s good and fast, but it lacks the infographic capabilities that really made Banana shine for me.
For the first time, an AI model has matched human experts at analyzing language itself.
Speech to reality: an AI system that builds what you ask for.
When LLMs play repeated game-theory scenarios like the Prisoner’s Dilemma, their strategic behavior is shaped as much by prompt language as by model architecture. Arabic and Vietnamese prompts drive more defection, while French elicits cooperation. The paper is worth looking into.
New research shows multi-agent systems can boost performance by 80% on parallelizable tasks but cut it by up to 70% on sequential tasks. The key to scaling AI isn’t adding more agents; it’s matching the right architecture to the task.
Markets:
Nearly 20% of Meta’s 2024 revenue from China came from ads for scams, illegal gambling, pornography and other banned content.
Greg Brockman said OpenAI was sometimes “sacrificing the future for the present […] The demand will far exceed what we can think of.”
McKinsey, which built an empire advising companies to cut workers, is now cutting its own.
Alexandr Wang has reportedly told associates he finds Zuckerberg’s micromanagement “suffocating.” Ouch.
on America’s tech productivity: the gains are hard to measure meaningfully, and what real gains exist flow to oligopolists rather than ordinary Americans.
Society and culture:
The UK government wants Apple and Google to embed nudity-detection algorithms into their devices at the operating system level to combat underage access to pornography.
A Colorado AI company led by former Intel CEO Pat Gelsinger created a benchmark measuring how well AI models align with Christian values. In a twist, Chinese models from Alibaba and DeepSeek outranked American competitors from xAI, Google and Anthropic.
We once dismissed animal suffering as mere mechanical reflex; this essay argues that we may be making the same mistake with AI.
writes about slop as the price we pay for democratized creation.
We’re hunting for an office with 10-12 desks, easy access to meeting rooms, and a flexible, grown‑up environment. Ideally near Great Portland Street, Tottenham Court Road, Bloomsbury or Marylebone. We’re open to partnerships in exchange for space, too. If you have a lead on a spare floor, studio, or a flexible space we can grow into, get in touch.
Thanks for reading!
“Starship‑class trajectories” means launch profiles made practical by SpaceX’s Starship: very heavy payloads, full reusability, frequent flights, and much lower cost per kilogram to orbit.
2025-12-20 23:40:29
🎅🏼 We recorded this to publish after Christmas, but demand for year‑end reflections prompted an early release - so if you hear me say Christmas has passed in the video or podcast, that’s why!
Listen on Spotify or Apple Podcasts
As we approach the end of the year, I want to reflect on what made 2025 special – what we learned about AI, what surprised me, and what it all means for the road ahead. This was the year artificial intelligence stopped being a curiosity and became infrastructure. The year models got good enough to do real work. And the year we began to understand just how profoundly this technology will reshape everything.
You can listen/watch my reflections on 2025 (widely available) or read all my notes below if you’re a paying member.
I cover:
The models matured
The work shifted
Orgs are slow
Atoms still matter
Money’s real
K is the letter of the year
And my seasonal movie recommendation for you 🎁
My favourite development of 2025 has to be Nano Banana, Google’s image generation service. Beyond its fantastic name (which, as a Brit, I’ll admit Americans say much better than I do 🤭), it represents something remarkable. Pair it with Gemini 3 Pro, present a complex idea, and ask it to turn that into an explanatory diagram. The visual results can be phenomenal with some clever prompting. It’s also really fun for video generation:
But the deeper story is that in 2025, many models got good enough to do long stretches of work. Claude 4.5 from Anthropic is fantastic for working with long documents and for coding. GPT-5 excels at deep research and certain classes of problem-solving – I’ve been using version 5.2 for financial modeling. Google’s Gemini 3 Pro is markedly better than previous generations. And I’ve been turning to Manus more often.
All of these models are now capable of doing what I would describe as a few hours’ worth of high-quality work. This has meant that I, as someone who hasn’t been allowed near software code for more than a decade, have been able to build my own useful applications. What was derisively called “vibe coding” a year ago has transformed into something genuinely productive for me. I will write more on this in my EOY lessons for genAI use in the next few days.
One of the biggest changes I’ve noticed is cognitive. There’s been a shift from the effort of actually doing the work to the effort of judging the work, specifying problems well, turning those problems over to a model, examining the output, and deciding what and how to move forward. You need to maintain a high degree of mental acuity as you evaluate: are the assumptions reasonable? Are there obvious errors?
Those of us who have managed people and led teams will recognise that it’s a bit like managing and leading, except your team member is a machine that can work in parallel across many domains simultaneously. And therein lies an additional challenge: the cognitive load of selecting which model to use for which task has now landed squarely on me, the user. It’s clear for coding – that’s Claude. But for the nebulous set of other problems, the choice between GPT 5.2 and Gemini 3 is never obvious beforehand, though it usually becomes clear in retrospect.
In a way, we all start to behave like developers. Developers think constantly about their tooling and workflows. They consider what in their weekly cadence can be automated. If you’re using AI effectively, you’ll spend less time putting bricks in the wall, more time figuring out how best to organise the bricks and specify what you need. If you have a static way of working with your AI system, you’re missing out. The models are becoming more capable. More importantly, the application layer around the models – the tools they can access, the use of memory, the use of projects – is making them dramatically more useful. So you simply have to keep experimenting and trying out new things.
My experience is that about three-quarters of what I’m doing today, I wasn’t doing three or four months ago. Models have become that much more capable. If you treat AI like a simple operating system upgrade, rather than something whose capacity grows and requires regularly changing how you work, you will miss out on the benefits.
One of the toughest lessons of 2025 for many will be how slow and painful organizational rewiring is compared to the progress we’re making with the AI systems themselves. Let me share Exponential View’s story.