Ramez Naam (@ramez) header
Ramez Naam (@ramez) profile picture

Climate and clean energy investor. Author of 5 books. Energy & Environment co-chair @SingularityU. Trying to build a better world.

Seattle
Joined May 2007
Photo by Ramez Naam (@ramez)Photo by Ramez Naam (@ramez)Photo by Ramez Naam (@ramez)Photo by Ramez Naam (@ramez)Photo by Ramez Naam (@ramez)Photo by Ramez Naam (@ramez)Photo by Ramez Naam (@ramez)Photo by Ramez Naam (@ramez)Photo by Ramez Naam (@ramez)Photo by Ramez Naam (@ramez)
Despite this election, I remain an optimist about America and the world. Humanity will continue to produce new ideas and new innovations to improve our lives. Good people will continue to come together to improve the world. And the political tide will turn. We'll make it so.
40
23
8
337
170,510
Gary, to his credit, was correct about this growing gap between AI performance on highly verifiable domains vs real world problems.
Three o3 predictions: – People will initially be amazed - Once they dive in, they will see that it is not reliable - It will work best in closed-domains (like math problems) but less reliably in open-ended domains (like everyday reasoning about the real world).
3
2
30
2,971
For the first time in 20 years, production-ready rare earth permanent magnets are being manufactured in America! Made from neodymium mined in Mountain Pass mine in CA and with no Chinese equipment or other dependencies! Huge congrats to @MPMaterials on this impressive feat!
41
179
52
1,321
93,832
I haven't had time to digest all of this, but it looks fascinating and like an excellent update. Great to see @METR_Evals continue to advance our ability to measure AI. Well done @MKinniment @slimshetty_ @mattputz_ @rowanwhuang and everyone else who worked on it. @BethMayBarnes
We've been working on understanding model capabilities beyond what our prior time horizon tasks could capture. Today we’re sharing a research update. We find that AI R&D-relevant capabilities are improving rapidly and are getting harder to measure.
1
2
36
1,737
Ramez Naam retweeted
Excellent analysis and I strongly agree with transparency being crucial at a time when whole economies and people’s livelihoods are at stake: > The most foundational safety policy in AI ought to be transparency. OpenAI has now set the bar for transparency. Google appears to be following. The responsible thing for Anthropic and other AI companies to do is to engage in a transparency “race to the top” – a race to publish the information that the general public needs in order to make independent evaluations of what’s happening in AI and where it’s headed.
New from me: The most important graphs in AI are wrong. Specifically, measures of the length of task AIs can autonomously accomplish are dramatically different than real-world data from inside OpenAI and Google. 🧵 rameznaam.com/p/the-most-imp…
3
4
14
1,864
This is yesterday's news. The more interesting trend is that the world has passed peak gerontocracy, as we argued last week. Old people's incomes are FALLING relative to the young, and state spending is shifting AWAY from the old. Young people are the richest generation ever.
The most important article of 2026 is this piece by @jburnmurdoch, building overwhelming evidence on the unsustainability of the pension systems in many countries and the broken intergenerational contract. To me this is the single underlying problem explaining the inability of many countries (incluiding my two, UK and Spain) to invest and grow and the malaise in their youth. Every future oriented policy, education, innovation, investment in infrastructure, is being sacrificed to increase pensioners living standards. But the problem is political, and some countries have solved it. E.g. in Denmark the retirement age increases (years later) semi automatically with life expectancy. See the result in the last graph. giftarticle.ft.com/giftartic…
14
32
9
265
53,360
AI is naturally plural and hyper-competitive, with no strong network effects or switching costs. That benefits users & society. It's likely to stay this way, unless we politically restrict competition. 'Safety' will be the argument by which such restrictions are sold.
some quick thoughts on ai lab revenue stickiness 1) i'm going to jot down some ideas about the extent to which frontier lab revenue and gross margins are defensible 2) right now, frontier labs generate revenue from selling digital workers like claude code, providing tokens for data pipelines, and being incorporated into apps 3) the switching costs for selling digital workers seem to be very low. i can’t imagine a failed migration from claude code to codex right now. 4) i could imagine not liking it and switching back, but not a failed migration. nothing like a failed salesforce migration or failed erp migration 5) the cost of experimenting with a swap seems close to zero, and this pretty much maps to how people i know switch between openAI and anthropic models. 6) it feels like, without prepayment, it is very easy for digital worker revenue to switch between providers if another model is better / cheaper for a reasonable time. 7) on data pipelines, this revenue feels more sticky than digital worker revenue because you don’t want to mess up your pipeline. it requires some evaluation. 8) but you can incrementally move from one provider to another by swapping part of the pipeline and looking for errors 9) or, by moving the % of the pipeline that goes to a given model over time. i could even imagine someone doing this for model upgrades. 10) customer product integration seems the stickiest, especially if the customer knows what model you are using. there may be some customer loyalty 11) but in practice, it is probably a small part of overall frontier lab revenue. it seems to me there are fewer harvey[s] being founded 12) the digital workers themselves are pretty much interchangeable at a given capability. 13) this suggests that high frontier lab gross margins are mostly down to the monopoly of the best model in a given domain or the oligopoly of high-end models. 14) anthropic, openai, [and] gdm prices are supported by the fact that there are so few frontier model providers and demand for tokens outstrips supply 15) long term, this seems defensible to me. The capital costs of training models are so high, ai r&d is so expensive, and research agents help lock in a lead. 16) and, of course, they will continue increasing capabilities, eating larger tasks, etc... 17) but it is important to be clear-eyed that the oligopoly of frontier providers (maybe even monopoly in some cases) is what supports the gross margins.
1
7
41
2,411
Ramez Naam retweeted
Detaining a PhD student on her way back to Princeton from presenting scientific work at a conference is not how we get the best and brightest scientists in the world to want to come to the US. I hope this is resolved quickly! More details: dailyprincetonian.com/articl…
Our colleague Eva Yi Xie, a Princeton neuroscience PhD student, was detained by ICE at Reagan National Airport on her way home from SfN, where she’d just presented her poster. She’s being held in Chantilly, VA. Please help us bring her home. 📷
369
647
49
6,717
452,898
disinformation
US TSY. SEC. BESSENT: ANY AI-BASED ATTACK YOU SEE COMES FROM CHINESE MODELS
Readers added context they thought people might want to know
The most notorious of recent cyber attacks, was carried out by a US model against hugging face. openai.com/index/hugging-…
6
42
544
20,171
Ramez Naam retweeted
A country of geniuses in a data center can prove the Riemann hypothesis, but they can’t raise your standard of living unless they can build the things you need in the real world. And recursive self-improvement just makes physical chips more valuable. Great to talk with @saranormous and @eladgil about the physical consequences and requirements of superintelligence!
A really fun @NoPriorsPod with Arda Cofounder, former Chief Research Officer at OpenAI, early exec at Palantir @bobmcgrewai 01:05 – Why superintelligence requires physical AI 02:27 – Why the hard takeoff argument is wrong 05:48 – Labs or startups – who will win in robotics and biology? 13:00 – Why America designs products but doesn’t build them 16:33 – Who needs flexible automation: defense, consumer electronics, data centers 18:19 – Humanoid robots vs. robotic work cells 21:07 – Open-world vs. closed-world robotics 25:20 – Building with Anduril: from days to 90 minutes 28:59 – Palantir’s problem-first approach vs. OpenAI’s technology-first approach 33:15 – From dropping out of Stanford to joining OpenAI as a nonprofit 37:13 – Why general robotics models suddenly work 40:28 – Implications for the robotics ecosystem 42:46 – On-device models, speed, and safety 44:18 – How Forward Deployed Engineering really works 50:57 – Superintelligence out of the data center 53:41 – Self-replicating factories and Von Neumann probes 56:46 – Bob’s answer to the Fermi paradox 59:33 – Conclusion
16
18
1
144
14,851
The below is false, but it serves the administration's purpose. The most severe AI security incidents have come from proprietary US models under testing. But that wouldn't give them an excuse to crack down on open weight models.
JUST IN: 🇺🇸🇨🇳 U.S. Treasury Secretary Bessent says “Any AI-based attack you see comes from Chinese models.”
2
7
62
3,490
The President has no legal or constitutional power to do this. Just as he has no legal or constitutional power to impose a pre-release testing regime on AI. But AI companies have let him get away with the latter, without challenging it.
BREAKING: President Trump says he will impose a “tremendous tax” on any US company that attempts to build data centers outside of the US. Trump says he will not allow data center construction outside the US because “China will take all of them.” The Nasdaq 100 extends losses to -4% on the day after the statement.
4
1
23
1,764
This is indeed the most likely AI future. As long as we don't succumb to regulatory capture or government-created AI moats.
Value created ≠ value captured. AI will commoditize like every general-purpose tech before it. Just ask the ghosts of railway, airline, and fiber equity. $ANT
4
3
1
35
6,028
Yep. I'm an AI optimist. It's the most powerful technology humanity has invented, and we're not even close to maximizing its capabilities or its economic value. My analysis is specifically about the case for RSI and the reasons to update our models based on empirical data.
I think trying to track how much work AI can do autonomously is critical, and this is a useful counter to some of the most audacious claims about the state of autonomous AI research. But I do want to note that the focus on research tasks at the most advanced AI research organizations on the planet also runs the risk of substantially underestimating what AI can already do in terms of economically valuable tasks. Its worth separating the two (and I don't think Ramez is claiming they are not seperated). Most work is not frontier AI research, and there AI can be tremendously successful. Epoch's MirrorCode results, for example, show AI autonomously reimplementing software that the researchers estimate would take a humans 2-17 weeks. This is well-specified work, so it has very different boundary conditions. Plus, in the real world, some human intervention is fine (in my mind, vastly preferrable to no human role for a variety of reasons). Occasional guidance can still leave you with enormous productivity gains, even if that doesn’t satisfy the requirements for fully autonomous AI research. And even in the case of AI research, the same OpenAI report includes researchers estimating that a month of their team’s work would have taken three months without coding agents, on average. In short, RSI/AI-driven research progress may not tightly couple with progress on economically valuable tasks, and the variance and range on those tasks are less well understood.
1
6
48
3,179
Ramez Naam retweeted
ANTHROPIC ANNOUNCES FAST TAKEOFF AS PROBES EXIT THE HELIOPAUSE ($ANT -7%)
65
94
15
2,602
103,619
The whole "intelligence explosion" debate is fun and interesting, but ultimately useless. If it's going to happen there's nothing we can do to stop it. Meanwhile, the AI we'll have in 2 or 3 years will be godlike WITHOUT any intelligence explosion. noahpinion.blog/p/does-it-re…
16
21
199
22,673
Ramez Naam retweeted
Replying to @ramez
all this has happened before, and it will happen again...
1
5
13
974
Ramez Naam retweeted
When AIs do something impressive sooner than we expected, it's both evidence that AIs are improving faster than expected and evidence that the problem was easier than expected. These are hard to separate. I elaborate here (warning: weapons-grade copium): lesswrong.com/posts/WdTKkd5g…
6
7
141
7,014
AI authoritarianism will arrived draped in the flag, holding promises of safety.
Cursor has told Homeland Security they are considering phasing out *all* China developed open-weight models because of the 'geopolitical risk.' The report will be published by multiple House committees later today. This is a huge indicator of future regulatory direction!
3
6
53
2,733
Ramez Naam retweeted
If this is happening we need better American open source models; or else we are stuck inside an implicit tariff wall paying rents to OpenAI/Ant/Google
Cursor has told Homeland Security they are considering phasing out *all* China developed open-weight models because of the 'geopolitical risk.' The report will be published by multiple House committees later today. This is a huge indicator of future regulatory direction!
9
5
56
4,359