Fear-Driven TV News Skews Public Perception of Crime
Fear porn also dominates local TV news. Past studies have found a correlation: The more local TV news you watch, the more you believe crime is rising. Social media amplifies this, bu the problem is old and is all about media incentives. If it bleeds, it leads.
Profit-Driven Crime Coverage Undermines Democratic Discourse
It's not just that people are bad at this. It's that news media have consistently highlighted crime stories because they're profitable. That affects people. The news incentives are broken and toxic to democracy.
AI Policy Debate: Substitute or Augment Human Workers?
An extremely important topic. I'm more optimistic than Erik, but none of us know for sure. And as he says, "A key policy question is whether we use AI mainly to substitute for people or to augment them."
AI Pushes Back on Your BS: Benchmark Results
I love this. A benchmark of how often an AI model will push back on your BS.
Cool Buildings by Sending Waste Heat Into Space
Air conditioning is an energy hog, right? Right. But there are also incredibly low-energy ways to cool buildings - by literally sending their waste heat into space.
Warrantless Data Searches Violate 4th Amendment, Need Reform
Massie is substantially correct here and this issue is more and more important as powerful AI and tools like Palantir's gets deployed. It's not 100% as simple as he's saying, as the warrantless collection of American communications happens when a...
AI Hackers Nearing Proficiency; Defense Requires Collective AI
AI models are getting better at hacking. Surprisingly (to me) they're still not world class hackers. But on this pace they could be within a year or two. As I've said recently, the way to defend against this is at...
Real AI: Competition, Not Single-Entity Domination
Sci-fi stories of AI almost always have a single dominant AI or AI corporation that achieves a monopoly. That's a true danger to the world. The real world is hyper-competition between leading companies and open weight projects that democratize AI....
Four to Five Near‑Failure Sets Capture Most Strength Gains
Indeed. Very parallel. 4-5 intense sets of resistance training per week close to failure bring ~all the strength gains of weight lifting. After that, you get bigger muscles but no strength gains.
Humanity Overcomes Doomers: Innovation Beats Pessimism
Concur. History shows that: 1. There's a human psychological bent towards doomerism, particularly about technology. 2. In actuality humanity is remarkably innovative and resilient. We rise to challenges. Often late, but we rise. The burden of proof is on the...
Electric Fleets Will Outpace Consumer EVs on Cost per Mile
Electric trucks and other electric commercial vehicles will take over their markets faster than personal electric cars. Fleet owners and operators care about cost per mile, not just up front cost. And electric drive trains are cheaper in energy and...
AI $10T Annual CapEx by 2030 Unfeasible
This upper bound is an implied AI CapEx of ~$10 Trillion a year by 2030. That's quite implausible. Hyperscaler free cash flow is on the order of $1T / y. Even the very very rosiest AI scenarios would have AI...
Global Shift Toward China Signals US Decline
The second map (most of the world moving towards China and away from the US) is post-tariffs but pre-invasion-of-Iran. Will be interesting to see how this evolves. Probably not in a good direction for the US.
AI's Biggest Beneficiaries Are Users, Not Just Investors
The overwhelming majority of AI value accrues to users of AI and society as a whole. Founders, employees, and investors get rich, yes. And that's tiny compared to the customer value and positive externalities.
AI Self‑improvement Hinges on Breakthrough Discovery, Which Seems Unlikely Now
This is the core question in AI recursive self-improvement to me. If AI RSI is mostly about writing code, or running experiments with tweaks over current algorithms, we shouldn't expect a takeoff to ASI. But if AI models can truly...
LLMs' Massive Memory Inflates Perceived Reasoning Abilities
One superpower of LLMs is that they have memorized an enormous amount of all human knowledge. At a practical level that's awesome. It leads to incredible capabilities. It may, however, lead us to exaggerate their actual reasoning / novel production...
Exercise Volume Overestimated; Intensity and Consistency Matter
We massively overestimate how much exercise is needed. Study after study show the benefits for cardio fitness, longevity, and strength come from: 1. Intensity (at very low volume). 2. Consistency. Volume basically doesn't matter beyond a low amount, except for getting...
AI CapEx Hits $650B Ceiling Before Revenue Surge
My expectation on AI CapEx: There's a natural wall at annual CapEx reaching 100% of hyperscaler free cash flow [including future AI free cashflow]. That's about 50% above current spending levels. Beyond that, you need to take out a lot...
LLMs Aid Top Mathematicians but Remain Unreliable
LLMs are amazing. In math, they're clearly becoming useful to top mathematicians. They also struggle with reliability and not knowing what they don't know.
Northern Europe's Winter Power Gap Calls for Nuclear
This is a mistake. Northern Europe is one of the hardest places on earth to power with renewables. German winter is long and dark. And as we electrify heat, the greatest electricity demand will be in winter. Nuclear should be...
AI Targeting Must Remain Human‑Supervised to Prevent Tragedy
My view of the US bombing of the school in Iran being possibly (probably?) due to selection by Claude: 1. It's a tragedy. I hate seeing my country cause civilian deaths, especially kids. 2. LLMs are not 100% accurate or anything close...
Claude Auto‑creates Graphics, Outpacing ChatGPT
It's really something. My favorite new AI feature in some time. I used to deliberately prompt ChatGPT to make informative graphics in answers. Claude's are better, faster, and just happen when the answer warrants one.
Claude’s Diagram Feature Simplifies Complex Answers Instantly
The new Claude feature of creating helpful explanatory diagrams in its answers to explain things to you is just fantastic. It's fast, clear, and is great as a way to rapidly get the gist of a complex answer.
Rising Nuclear Plant Costs Threaten US Energy Future
Nick is a nuclear expert who wrestles with that most important yet underemphasized issue of nuclear in the US: Continually rising construction costs.
Under
Climate folks: Who are the experts on how undersea cable data is used to measure North Atlantic salinity and currents, and detect potential AMOC weakening?
AI Alignment Should Prioritize Pro‑social Values over Owners
Fascinating and important research. As I've said lately, we don't want to align AI to its creators or its owners. The ideal alignment of AI is to a pro-social set of values. This suggests that such alignment is happening and...
Humans Will Continue Overseeing AI for Years
This is a pretty good way of thinking about this, and leads to the conclusion that humans will remain as overseers of AI for quite a long time to come.
Low AI Doom Probability; Scaling Claims Overlook Core Limits
1. From a bayesian perspective, low p(doom) should be the default. The burden of proof is on those with a high p(doom). They have not made a case that convinces me. 2. AI recursive self-improvement models ignore so much hard stuff....
Benchmark Flaw Slashes AI Progress Estimates, Undermines METR Predictions
Correcting for this AI benchmark error: - Cuts the length of task the AI can accomplish by a factor of 7. - Cuts the historical AI improvement rate by a factor of 3. This is one of many reasons I don't think...
AlphaFold Enhances Experiments, Sparks Research, Yet Misses Early Drug Impact
Nice to see empirical study of how our most impressive breakthrough in AI for biology is impacting science. Point 1 is important for those who dream that AGI or ASI will just solve all of science by thinking. They won't....
LLMs Out‑Persuade Human Campaign Ads, Raising AI Risks
Hyperpersuasive AIs - used for political reasons, criminal reasons, or commercial ones - are a plausible AI risk. A number of studies have found that LLMs are more persuasive than humans. This one finds that 4 US LLMs are more...
Benchmarks Overstate AI Progress Despite Ongoing Improvements
AI is amazing and constantly improving. And there's reason to believe that the benchmarks exaggerate progress in various ways. Another example today.
Submarine Reactors Prove SMRs Work, but Cost Skyrockets
Great perspective on the argument that naval submarine nuclear reactors prove that SMRs can work. Turns out they are crazy expensive. (Though this probably exaggerates cost somewhat.)
Home Robots Face Precision, Utilization, and Privacy Hurdles
I want to elaborate on this. A couple different concerns about the household humanoid robot scenarios: 1. Extremely high precision requirements. The tolerance for error in home environments is going to be low. AI systems and robotic systems don't start out...
Low AI Doom Boosted by OpenClaw Optimism Despite Weak Security
I have a very low AI p(doom). But it's lifted a bit by the enthusiasm about OpenClaw, and willingness to embrace it with a terrible security model.
China Emerges as War’s Biggest Winner, US Falters
No. China is the big winner. Putin will still lose in Ukraine. But Xi looks increasingly stable and sane to the world, as the US looks worse and worse, and weakens and distracts itself. China's EV industry stands to benefit...
AI Progress Is Collaborative, Not a US‑China Race
I think highly of Vinod but am skeptical of this very common AI viewpoint. First, it's proving difficult for anyone to really break out of the AI pack. AI models are much more similar than they are different. There will...
Now's the Second Best Time to Buy an EV
The best time to buy an EV was before the oil crisis. The second best time to buy an EV...
Israel Targets Iran’s Power, Not Its People
There is little evidence that the Israeli government cares about the Iranian people. They care about destroying and suppressing Iran's military and economy and ability to influence the region. Those are not the same thing.
Renewables, Not Nuclear, Are Displacing China's Coal Fastest
That impacts coal and gas demand but not oil demand. It's also wrong in terms of what's happening. I love nuclear as part of the solution. But look at the country building the most nuclear, China, and you see more...
Norway Leads, Many Nations Surge Toward EV Dominance
These are the countries building resilience against future oil shocks. And just getting cooler cheaper cars.
China's EV Surge Cuts Fuel Demand, Drives Oil Decoupling
China is working on decoupling itself from dependence on the global oil market. All while making cars that are quieter, cleaner, cheaper, and just plain cooler than the gasoline cars of the past. Trucks now too.
Aggressive EV Adopters Cushion Oil Price Shocks; US Lags
Countries that have adopted electric vehicles the most aggressively will suffer the least from oil price spikes. The US, of course, is not one of those. [Corrected graph.] https://t.co/jl1blEELO0
Renewables and EVs: Natural Hedges Against Fossil Price Swings
Renewables are a hedge against volatile coal and gas prices. Electric vehicles are a hedge against volatile oil / gasoline prices.
Solar Remains Top‑rated, yet GOP Vilifies It
Interestingly, solar never got polarized among the US populace. For 20 years, it's polled as the most popular energy source in the US, by wide margins, across the country. It's just that certain members of the GOP political class hated...
LevelTen Index Misreads Solar Prices, Not Actual Costs
Okay. I've been seeing this make the rounds. It is just a large misunderstanding of the second graph, to the extent that it's drawing the wrong conclusions. The price of electricity from solar and wind in the US *has* flatlined...
Hedge Funds Care About Clean Energy's Cost Advantage
Trivia: The first time I gave a talk that said clean energy was just going to win on cost, with hardly a mention of climate, was at a @Scaramucci SALT conference. I just figured hedge fund managers would care a...
Clean Energy Now Cheaper Than Fossil, Outpacing Military Budgets
It's always been the case that the way we win on climate is just to make clean energy the cheapest option. That trend has been clear for 15+ years. Now that crossover is happening. Policy still matters, but the most...
Electric Fleets Win on Cost per Mile, Outpacing Personal EVs
Commercial electric vehicles will take over faster than personal vehicles. Fleets care about cost per mile. And electric drive trains already win there.
Space Data Centers Demand Unprecedented Scaling
Nothing impossible here. But many things need to scale in unprecedented ways for space based data centers to happen.