Paddy Ramanathan

𝓘𝓷𝓿𝓮𝓼𝓽𝓸𝓻/𝓟𝓸𝓭𝓬𝓪𝓼𝓽 𝓗𝓸𝓼𝓽
📣 𝕗𝕚𝕟𝕥𝕖𝕔𝕙𝕥𝕒𝕝𝕜.𝕔𝕠 | 💡 𝕚𝕧𝕒𝕝𝕝𝕖𝕪.𝕔𝕠 |🏃 𝕥𝕓𝕖𝕗𝕚𝕥𝕟𝕖𝕤𝕤.𝕔𝕝𝕦𝕓 |

[Paddy Ramanathan @PaddyRamanathan]

1/ The AI era is no longer about "apps"—it’s about the "Energy-to-Inference" pipeline. I just synthesized the key takeaways from the Cisco AI Summit ft. @jensenhuang, @sama, @pmarca, and more.

Here is the 5-Layer AI Stack for 2026. 🧵👇

2/ ⚡ ENERGY: The New Moat. Power is now the primary determinant of AI velocity. We’re seeing "Compute-Utilities" emerge—hyperscalers moving into nuclear and natural gas to secure gigawatt-scale capacity. No power = no intelligence.

3/ 💾 CHIPS: The Bespoke Era. The "Inference Flip" is here. 80% of workloads are now inference, not training. The result? A move toward custom silicon (OpenAI/Broadcom) and hardware-agnostic software like PyTorch 3.0 to break the software lock-in.

4/ 🌐 INFRASTRUCTURE: The Private AI Factory. Networking is the new wall. East-West traffic is 10x global broadband. Enterprises are moving away from public clouds toward private, terabit-scale factories to protect sovereignty and IP.

5/ 🌍 MODELS: Beyond Language. Language was just the interface. The new frontier is "World Models" (shoutout @drfeifei @worldlabs). We’re moving from software that talks to software that simulates physical reality.

6/ 🤖 APPLICATIONS: The Agentic Workforce. The "App" layer is collapsing into the "Model" layer. Coding is the first breakout workflow, but the goal is "Intelligence Autonomy"—agents pushing beyond "human-in-the-loop" to rebuild entire industries.

7/ 🔮 2026 Prediction: Software won't be a product you buy; it will be a capability you own. The era of "renting intelligence" is closing.

Check out the full breakdown and the "Gigawatt Moat" theory here: https://t.co/Ig9plN6qDG

#AI #TechStrategy #SiliconValley #Energy #Agents #FutureTech @elonmusk @davidsacks47 @DavidSacks

[Paddy Ramanathan @PaddyRamanathan]

@a16z “The app layer is back” is an interesting framing. Put simply, AI is now being leveraged for P&L transformation, with the underlying capital owners also investing directly to create enterprise value.
I wrote about this broader shift here: https://t.co/jSLmtFgust

[Paddy Ramanathan @PaddyRamanathan]

Very interesting. Steve Jobs is said to have answered “everyone” when asked which team at Apple was responsible for innovation.
Most staff functions — HR, marketing, innovation, compliance, finance, and others — were created to standardize, govern, and scale specialized capabilities across multiple business or product lines.
With the advent of AI and agents, that model may start to change. A line of business or product team could become more self-sufficient, with AI agents increasingly performing or augmenting staff-function capabilities in a consistent, scalable way. This could shift organizations from centralized support functions toward more embedded, agent-enabled operating models.

[Paddy Ramanathan @PaddyRamanathan]

This is bigger than “AI adoption.”

What we’re watching is a repricing of enterprise returns — shifting from labor toward compute, energy, and AI-orchestrated execution. In many ways, this is an enterprise P&L transformation cycle disguised as a technology cycle.

That’s why frontier AI labs, PE firms, and alternative asset managers are suddenly aligning around workflow transformation and operating leverage.

But the larger long-term risk to incumbents may not be automation alone.

It is AI-native companies using these new execution layers to disintermediate entire enterprise business models — much like internet-native companies did in the digital era.

I explore this in my latest piece on the AI Execution Economy:
https://t.co/2z0lDBvRTm

[Paddy Ramanathan @PaddyRamanathan]

The first phase of AI was about models.

The second phase is about execution.

Programming was the proving ground.
Enterprise operations are the real market.
The winners won’t simply deploy copilots on top of legacy software.

They will redesign workflows around agents.

The rise of the AI Execution Economy:

FINTECHTALK: The AI Execution Economy https://t.co/2z0lDBvRTm
@JTLonsdale @beyondthearc @rshevlin @charlesepotts @karlmehta @davidsacks47 @chamath @sama @garrytan

[Paddy Ramanathan @PaddyRamanathan]

Important signal.
The AI market is shifting from:“Who has the best model?”to“Who can transform enterprise workflows fastest?”
Programming was Wave 1.Enterprise execution is Wave 2.
This is why deployment, orchestration, and workflow redesign are becoming strategic assets — not just implementation services.
I recently wrote about this transition as the rise of the AI Execution Economy. https://t.co/2z0lDBvRTm

[Paddy Ramanathan @PaddyRamanathan]

Great segment, Joe.
One way to frame this is through checks and balances.
The founders designed government with internal checks — three branches — and external checks, where citizens retain overarching rights. What is often missing in industry and regulation is an equivalent system of competition and choice.
The contrast is obvious when you look across sectors.
In less regulated or more competitively contested markets — Uber, Airbnb, Big Tech, e-commerce, cloud, mobile apps — value creation happened much faster. The rate of development was extraordinary because disruptors had room to challenge incumbents first, prove consumer demand, and then force the system to adapt.
Uber challenged taxi medallion monopolies. Airbnb challenged hotel and zoning incumbents. Amazon and Shopify changed commerce. Apple and Google opened mobile distribution. Cloud platforms let startups build without buying their own data centers. In each case, the market moved faster than the regulatory apparatus, and consumers benefited from speed, choice, lower friction, and new business models.
Now compare that with pharma and banking.
In pharma, regulation is obviously necessary because safety matters. But the problem is when incumbents use regulatory complexity as a moat. Smaller biotechs face approval timelines, compliance burdens, trial costs, and gatekeeping structures that large pharma can absorb far more easily. The result is slower innovation and fewer paths for challengers.
Banking has the same pattern. Regulation protects consumers and the system, but it also entrenches large institutions. New banks, fintechs, and payment innovators face licensing, capital, compliance, and supervisory burdens that incumbents helped shape and can better afford. This creates a system where “safety” can become a euphemism for protecting the existing order.
The issue is not regulation versus no regulation. The issue is regulatory capture — or more precisely, regulatory grab. Incumbents learn how to turn regulation into a barrier against competition.
That is the danger now with AI. Anthropic and its allies are trying their level best to define the rules of the game in a way that sounds like safety, but could easily become incumbent protection. Once the largest model companies help write the regulatory framework, the risk is that future startups are forced to ask permission from the very architecture their competitors helped design.
The deeper principle is this: society needs to balance the rights of incumbents with the rights of disruptors.
In politics, Trump was a force of nature and could break through the institutional barriers. But most disruptors are not like that. In business, especially in regulated markets, most challengers can be quietly suffocated long before consumers ever get a choice.
That is why competitive regulatory choice matters. A biotech should not be captive to one approval pathway forever. A fintech should not need to become a bank before it can challenge a bank. And an AI startup should not have to comply with a safety regime designed by Anthropic, OpenAI, Google, and Microsoft before it can even enter the market.
The free press, markets, courts, consumers, investors, and competing jurisdictions should all be part of the check-and-balance system.
Without that, regulation becomes less about protecting the public and more about protecting the powerful.

[Paddy Ramanathan @PaddyRamanathan]

Could Sierra become the AI-native company that disrupts Salesforce — the original SaaS trailblazer?

Salesforce defined the SaaS era. It took Marc Benioff @Benioff more than a decade after founding the company — and years beyond its IPO — to reach roughly $15 billion in market value.

Sierra got there in about three years as a private company with this raise announced today.

That contrast may tell the bigger story. Salesforce was built for the cloud-software age. Sierra is being built for the AI-agent age. And while Salesforce became a horizontal enterprise platform, Sierra is starting with a narrower wedge: customer service. But that wedge sits right in the heart of Salesforce’s empire.

Today, Sierra is still only a tenth of Salesforce’s valuation. But the speed of value creation is stunning. It raises an uncomfortable question: is Salesforce the incumbent platform that adds AI, or is Sierra the AI-native architecture that eventually absorbs the category?

There is also an irony in the leadership arc. Benioff brought Bret Taylor @btaylor into Salesforce as co-CEO in what looked, at the time, like both a business and political hedge. Bret Taylor’s appointment was not incidental but strategic—reflecting his deep entrenchment within Silicon Valley’s power networks and positioning him, ultimately, to assume the role of Chairman at Twitter prior to 2020 election cycle.(I wrote about this https://t.co/JvYkbcwJw0).

Later reporting around “Twitter Files” fueled claims—disputed by some—that the company under Taylor's and the then CEO leadership worked closely with government actors during the 2020 election cycle, with critics alleging bias against Donald Trump and the MAGA movement. Since then, both Benioff and Taylor appear to have adapted to the new political climate, positioning themselves closer to the Trump/MAGA center of gravity rather than fighting it.

But strategic tradeoffs have consequences.
The deal that once looked like succession planning may now look like the seed of disruption. Taylor left Salesforce and built Sierra. And if Sierra continues on this trajectory, the ultimate twist may be that Benioff’s chosen heir doesn’t inherit Salesforce.
He disrupts it.
Or perhaps, one day, acquires it.
@elonmusk @davidsacks47 @realDonaldTrump @scottmcnealy @rabois @shaunmmaguire @mtaibbi @shellenberger @bariweiss

[Bret Taylor @btaylor]

Sierra is raising $950 million from new and existing investors, led by Tiger Global and GV, at a valuation of over $15 billion. We now have more than $1 billion to invest in becoming the global standard for companies wanting to transform their customer experiences with AI. 

We’ve never had such conviction in the opportunity for Sierra and our customers. Just a couple of years ago, we had four design partners. Now, Sierra is serving over 40% of the Fortune 50, and agents built on our platform are powering billions of customer interactions — everything from refinancing homes to processing insurance claims, returning orders, and helping people raise millions in fundraisers.

We’re deeply grateful to our customers for helping show what’s possible. If you’re not yet using Sierra, we’d love to partner with you. https://t.co/tT4tBSeSoR

[Paddy Ramanathan @PaddyRamanathan]

@Jason I'm in. https://t.co/pOBn4oBUZg

[Paddy Ramanathan @PaddyRamanathan]

I got a few minutes with the great Dušan Stojanović, Founding Partner at True Global Ventures — a one-of-a-kind investor with a rare ability to spot major shifts before the rest of the market catches on.
He gave a sneak preview of the upcoming TGV Conference, where the conversation will center on the 𝐧𝐞𝐰 𝐬𝐭𝐚𝐜𝐤: 𝐀𝐠𝐞𝐧𝐭𝐬, 𝐌𝐨𝐝𝐞𝐥𝐬, 𝐚𝐧𝐝 𝐒𝐭𝐚𝐛𝐥𝐞𝐜𝐨𝐢𝐧 𝐑𝐚𝐢𝐥𝐬.
The timing could not be better. It comes right on the heels of Animoca Brands, a TGV portfolio company, announcing stablecoin approval in Hong Kong alongside Standard Chartered and Hong Kong Telecom - Anchorpoint, the new co, will issue the HKDAP
TL;DR from the conversation:
𝐓𝐡𝐞 𝐄𝐦𝐞𝐫𝐠𝐞𝐧𝐜𝐞 𝐨𝐟 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐎𝐧-𝐂𝐡𝐚𝐢𝐧 𝐅𝐢𝐧𝐚𝐧𝐜𝐞 (𝐀𝐎𝐅): 𝐓𝐡𝐞 𝐍𝐞𝐰 𝐅𝐢𝐧𝐚𝐧𝐜𝐢𝐚𝐥 𝐋𝐚𝐲𝐞𝐫
𝐒𝐚𝐚𝐒 𝐈𝐬 𝐭𝐡𝐞 𝐍𝐞𝐰 𝐌𝐚𝐢𝐧𝐟𝐫𝐚𝐦𝐞: 𝐌𝐨𝐝𝐞𝐥𝐬 𝐚𝐧𝐝 𝐀𝐠𝐞𝐧𝐭𝐬 𝐓𝐚𝐤𝐞 𝐂𝐞𝐧𝐭𝐞𝐫 𝐒𝐭𝐚𝐠𝐞
𝐓𝐡𝐞 𝐔𝐈 𝐈𝐬 𝐃𝐞𝐚𝐝: 𝐀𝐠𝐞𝐧𝐭𝐬 𝐁𝐞𝐜𝐨𝐦𝐞 𝐭𝐡𝐞 𝐈𝐧𝐭𝐞𝐫𝐟𝐚𝐜𝐞, 𝐌𝐨𝐝𝐞𝐥𝐬 𝐇𝐨𝐥𝐝 𝐭𝐡𝐞 𝐋𝐨𝐠𝐢𝐜 𝐚𝐧𝐝 𝐃𝐚𝐭𝐚
𝐓𝐡𝐞 𝐑𝐢𝐬𝐞 𝐨𝐟 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐕𝐞𝐧𝐭𝐮𝐫𝐞 𝐂𝐚𝐩𝐢𝐭𝐚𝐥
More here 👉 https://t.co/4yBvCJ7izj

[Paddy Ramanathan @PaddyRamanathan]

@mazemoore @NateSilver538 isn’t acting as a data scientist here (nor he is one in the true sense)—he’s shaping narratives to fit a predetermined agenda, with the analysis bending to match the story.

[Paddy Ramanathan @PaddyRamanathan]

Balaji, I think your assessment misses the central point: this conflict is fundamentally being used as leverage in a broader U.S.–China trade negotiation. My view is that when Trump visits China in May, the message will be something akin to “tear down this wall, Mr. Gorbachev” — a symbolic demand for a rebalancing of the economic relationship. China, in turn, may agree to a more balanced trade deal, giving the U.S. room to accelerate reindustrialization. If that happens, the MAGA movement will likely celebrate it on July 4th as a defining victory, casting Trump as the figure who helped save America — and, in their eyes, the West — from a new form of communism cloaked in globalism.

[Paddy Ramanathan @PaddyRamanathan]

This is great. Godspeed, David Sacks.
My only caveat is that AI is bigger than science and technology alone. It is closer to a new industrial revolution. Yes, the opportunities across the AI stack are enormous — from energy, chips, and infrastructure to models and applications. But AI will also reshape education, the workforce, and the overall pace of human development in profound ways.
That is why PCAST would be better served by broader representation, not just from technology companies, but also from leaders in education, labor, industry, and public policy. I write/podcast about these things at https://t.co/FRf5PxlKez

[Paddy Ramanathan @PaddyRamanathan]

In my latest FINTECHTALK with Bill Capuzzi, we unpack why the bigger story is not just brokerage infrastructure, but the new financial stack forming around:
Alts
tokenization
stablecoins
TradFi/DeFi convergence
prediction markets
AI as a force multiplier
Read/Listen - https://t.co/ISSlUnw6YR

[David Sacks @DavidSacks]

In December, President Trump signed an Executive Order tasking us with the development of a national framework for AI, what he called “One Rulebook.” This was in response to a growing patchwork of 50 different state regulatory regimes that threaten to stifle innovation and jeopardize America’s lead in the AI race.

Today we are releasing that framework. It will help parents safeguard their children from online harm, shield communities from higher electric bills, protect our First Amendment rights from AI censorship, and ensure that all Americans benefit from this transformative technology.

We look forward to working with our colleagues in Congress to turn the principles we are announcing today into legislation.

https://t.co/9fwatPYP5M

[Paddy Ramanathan @PaddyRamanathan]

If you Missed Jensen Huang’s GTC Keynote, Here Are the Big Takeaways

It was not just another product launch. The deeper story is that NVIDIA is no longer simply selling chips into an AI boom. It is helping define the operating logic of the next industrial era.

1. Physical AI is moving from sci-fi to platform shift

Jensen closed with an Olaf-like Disney robot — a signal that humanoids are moving out of the lab and into the commercial imagination. Today that may show up in theme parks. Tomorrow it could extend much further: humanoids in businesses, and eventually in many homes.

The right mental model is not a niche robotics market. It is something closer to a Cambrian explosion, or the early smartphone era, when a new device class rapidly became a new platform. The physical world is becoming the next compute canvas.

That is why the next chapter of AI is not just digital AI — it is Physical AI. World models, robotics, autonomy, and embodied systems are no longer side stories. They are one of the biggest vectors pushing AI from software assistance into real-world economic transformation. AI is moving from generating answers to generating action.

2. The inference inflection is here — and the stack is being rebuilt around production

One of the most important themes Jensen reinforced is the inference inflection. The center of gravity in AI is shifting from training to inference.

That matters because the market is moving from “build the model” to “run the model continuously, efficiently, and at scale.” This is where silicon design, orchestration, latency, deployment economics, and real-time responsiveness start to matter as much as raw model capability.

As inference becomes the dominant workload, the hardware and software stack must increasingly be optimized for high-volume, low-latency production use cases — especially agents, real-time decisioning, and world models operating in dynamic environments. The implication is bigger than faster chips. The entire ecosystem is being redesigned for production AI. I had covered this an earlier piece here

3. The AI factory is the new industrial model

Jensen’s framing of the AI factory may be the most important idea from GTC.

The old factory of the Industrial Revolution took in raw materials, labor, and energy, and turned them into finished goods. The AI factory takes in data, compute, and energy, and turns them into tokens, predictions, software, decisions, automation, and increasingly physical outcomes.

This is not just a metaphor. It is the correct mental model for where AI is going. See AI Factory image below.

4. Compute becomes the new labor — and token-per-watt becomes the new productivity metric

In the AI factory model, compute begins to function like a new form of labor. Data is the raw material. Energy is the critical input. And output is no longer just content — it is intelligence operationalized into products, workflows, services, and machines.

Software was about digitizing workflows. AI is about industrializing cognition.

That is why Jensen’s emphasis on token per watt matters so much. It is an AI-era productivity metric — the equivalent of energy efficiency and output optimization in the traditional factory model. The long-term bottleneck in AI may not be capability. It may be the ability to deliver intelligence efficiently. The constraint is shifting from “can the model do it?” to “can the system do it economically at scale?”

That means energy, compute efficiency, and deployment economics will become strategic differentiators.

5. This is bigger than a software cycle

I increasingly think AI should be understood not as just another software wave, but as an Industrial Revolution-scale shift.

Software transformed workflows. AI transforms the production function itself.

This is the move from chatbot to factory, from software layer to production layer, from digital intelligence to physical intelligence.

6. We are entering the era of the AI-native company

Just as the internet era produced internet-native companies like Amazon, Google, Uber, Airbnb, and Facebook, this era will produce AI-native companies designed from the ground up around AI factories, inference economics, autonomous systems, and machine-mediated decision loops.

The winners will not be the firms that merely add AI features. They will be the firms that redesign their business models and operating systems around AI as a production architecture.

The sectors where this transformation is already becoming visible include autonomous vehicles, customer support, software development, engineering, healthcare, robotics, and search. These are early indicators of a broader shift in which intelligence becomes scalable, programmable, and embedded into core operations. See screen shot from Jensen’s presentation showing the players in each of the category.

7. The AI application stack is getting clearer — and OpenClaw could be a platform-shift moment

Another useful lens from GTC is the emerging application stack. The market is taking shape across several important layers: Also in screenshot below

Frontier model builders

Model-to-production platforms

Agent frameworks and protocols

Inference frameworks

This matters because the AI economy will not be won by models alone. It will be won by the systems that connect models to production, production to workflows, and workflows to outcomes.

That is also why OpenClaw could become one of those moments people later compare to the PC, Linux, the mobile phone, or ChatGPT. Agentic frameworks are not just developer tools. They shape how personal agents and enterprise agents are built, coordinated, and deployed. That has implications not only for productivity, but for the future human interface to computing itself.

8. Conclusion: GTC was a blueprint for the new industrial age

My core takeaway from GTC is this:

NVIDIA is not just supplying picks and shovels to an AI boom. It is helping define the blueprint for a new industrial age — one where data, compute, energy, inference economics, and embodied intelligence become foundational inputs of economic power.

That is why this moment matters far beyond tech.

It is not just a software innovation cycle.

It is the early architecture of a new Industrial Revolution. https://t.co/Ig9plN6qDG @elonmusk @davidsacks47 @rabois

[Marc Andreessen 🇺🇸 @pmarca]

My information consumption is now 1/4 X, 1/4 podcast interviews of the smartest practitioners, 1/4 talking to the leading AI models, and 1/4 reading old books. The opportunity cost of anything else is far too high, and rising daily.

[Paddy Ramanathan @PaddyRamanathan]

I rebutted Anthropic’s recent AI jobs doomsday report in my X post below. But the bigger story is that they seem to be promoting a dystopian vision of AI in order to shape regulation in a way that favors Big AI incumbents—positioning themselves to control the agenda and reap the lion’s share of the benefits. @DavidSacks https://t.co/4NVCVru0iu

[Paddy Ramanathan @PaddyRamanathan]

Anthropic’s new labor-impact chart (link in comments) is being read as a map of job loss. I think that overstates the case.
Their Figure 2 compares AI’s “𝐭𝐡𝐞𝐨𝐫𝐞𝐭𝐢𝐜𝐚𝐥 𝐜𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐲” with current observed usage across today’s occupational categories and implies that over time usage may expand toward capability. That is a useful near-term adoption lens, but a weak long-term labor-market model, because it assumes today’s job categories will remain the right containers for tomorrow’s work.
𝐇𝐢𝐬𝐭𝐨𝐫𝐲 𝐬𝐮𝐠𝐠𝐞𝐬𝐭𝐬 𝐨𝐭𝐡𝐞𝐫𝐰𝐢𝐬𝐞.
The printing press did not simply automate “scribes” inside a fixed medieval job taxonomy. It reorganized knowledge production, created new firms and specialties, and helped cities with presses grow about 60% faster between 1500 and 1600. Technology changed the categories of work themselves.
That is the real weakness in doomsday AI charts: they measure exposure inside inherited buckets, while the buckets themselves are moving. BLS explicitly notes that occupation definitions change over time and that classification changes can create breaks in time series. 𝐓𝐡𝐚𝐭 𝐢𝐬 𝐧𝐨𝐭 𝐚 𝐟𝐨𝐨𝐭𝐧𝐨𝐭𝐞 — 𝐢𝐭 𝐢𝐬 𝐭𝐡𝐞 𝐩𝐨𝐢𝐧𝐭.
𝐓𝐡𝐞𝐫𝐞 𝐢𝐬 𝐚𝐥𝐬𝐨 𝐚 𝐩𝐨𝐥𝐢𝐭𝐢𝐜𝐚𝐥 𝐞𝐜𝐨𝐧𝐨𝐦𝐲 𝐚𝐧𝐠𝐥𝐞 𝐡𝐞𝐫𝐞. Anthropic is not just publishing research; it is also advocating targeted regulation, transparency rules for the largest AI developers, export controls, and stronger governance frameworks. It has funded AI-policy advocacy and holds a DoD agreement with a $200 million ceiling. None of this proves bad faith. But it does mean the company benefits from a policy climate in which AI is framed as powerful, destabilizing, and in need of compliance-heavy oversight that large incumbents can absorb more easily than smaller firms and open-source communities.
𝐌𝐲 𝐯𝐢𝐞𝐰: AI will disrupt work, but the bigger effect is not that today’s occupations simply get hollowed out in place. The bigger effect is occupational recomposition — new roles, merged roles, split roles, and new layers of supervision, verification, integration, and domain judgment. More importantly, this shift will dramatically increase the velocity of human progress. As AI augments discovery, design, and problem-solving, breakthroughs across medicine, science, engineering, and industry could accelerate exponentially. That means faster progress toward curing diseases, extending healthy lifespan, transforming energy systems, expanding into space, and unlocking breakthroughs we cannot even imagine today.
The printing press did not end knowledge work. It changed who did it, how it was organized, and which skills became valuable. AI will do the same.
𝐁𝐨𝐭𝐭𝐨𝐦 𝐥𝐢𝐧𝐞:  Figure 2 mistakes a transitional occupational taxonomy for a permanent labor market — and that framing conveniently supports a politics of AI regulation that favors incumbents over open innovation. Recent blog that talks about some of these things https://t.co/Ig9plN6qDG @elonmusk @sama @DarioAmodei @DavidSacks @davidsacks47 @chamath @JTLonsdale @rabois @shaunmmaguire @balajis @realDonaldTrump

[Paddy Ramanathan @PaddyRamanathan]

Anthropic’s recently published paper, which I rebut above, makes what I believe is a flawed argument about labor-market impact. https://t.co/r0GRwIWfPa

[Paddy Ramanathan @PaddyRamanathan]

Anthropic’s new labor-impact chart (link in comments) is being read as a map of job loss. I think that overstates the case.
Their Figure 2 compares AI’s “𝐭𝐡𝐞𝐨𝐫𝐞𝐭𝐢𝐜𝐚𝐥 𝐜𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐲” with current observed usage across today’s occupational categories and implies that over time usage may expand toward capability. That is a useful near-term adoption lens, but a weak long-term labor-market model, because it assumes today’s job categories will remain the right containers for tomorrow’s work.
𝐇𝐢𝐬𝐭𝐨𝐫𝐲 𝐬𝐮𝐠𝐠𝐞𝐬𝐭𝐬 𝐨𝐭𝐡𝐞𝐫𝐰𝐢𝐬𝐞.
The printing press did not simply automate “scribes” inside a fixed medieval job taxonomy. It reorganized knowledge production, created new firms and specialties, and helped cities with presses grow about 60% faster between 1500 and 1600. Technology changed the categories of work themselves.
That is the real weakness in doomsday AI charts: they measure exposure inside inherited buckets, while the buckets themselves are moving. BLS explicitly notes that occupation definitions change over time and that classification changes can create breaks in time series. 𝐓𝐡𝐚𝐭 𝐢𝐬 𝐧𝐨𝐭 𝐚 𝐟𝐨𝐨𝐭𝐧𝐨𝐭𝐞 — 𝐢𝐭 𝐢𝐬 𝐭𝐡𝐞 𝐩𝐨𝐢𝐧𝐭.
𝐓𝐡𝐞𝐫𝐞 𝐢𝐬 𝐚𝐥𝐬𝐨 𝐚 𝐩𝐨𝐥𝐢𝐭𝐢𝐜𝐚𝐥 𝐞𝐜𝐨𝐧𝐨𝐦𝐲 𝐚𝐧𝐠𝐥𝐞 𝐡𝐞𝐫𝐞. Anthropic is not just publishing research; it is also advocating targeted regulation, transparency rules for the largest AI developers, export controls, and stronger governance frameworks. It has funded AI-policy advocacy and holds a DoD agreement with a $200 million ceiling. None of this proves bad faith. But it does mean the company benefits from a policy climate in which AI is framed as powerful, destabilizing, and in need of compliance-heavy oversight that large incumbents can absorb more easily than smaller firms and open-source communities.
𝐌𝐲 𝐯𝐢𝐞𝐰: AI will disrupt work, but the bigger effect is not that today’s occupations simply get hollowed out in place. The bigger effect is occupational recomposition — new roles, merged roles, split roles, and new layers of supervision, verification, integration, and domain judgment. More importantly, this shift will dramatically increase the velocity of human progress. As AI augments discovery, design, and problem-solving, breakthroughs across medicine, science, engineering, and industry could accelerate exponentially. That means faster progress toward curing diseases, extending healthy lifespan, transforming energy systems, expanding into space, and unlocking breakthroughs we cannot even imagine today.
The printing press did not end knowledge work. It changed who did it, how it was organized, and which skills became valuable. AI will do the same.
𝐁𝐨𝐭𝐭𝐨𝐦 𝐥𝐢𝐧𝐞:  Figure 2 mistakes a transitional occupational taxonomy for a permanent labor market — and that framing conveniently supports a politics of AI regulation that favors incumbents over open innovation. Recent blog that talks about some of these things https://t.co/Ig9plN6qDG @elonmusk @sama @DarioAmodei @DavidSacks @davidsacks47 @chamath @JTLonsdale @rabois @shaunmmaguire @balajis @realDonaldTrump