Colossus Data Centers Tarnish Industry's Environmental Image
Andy has a good point. The Colossus data centers have been bad actors on local pollution. I'm very pro-datacenter, and most attacks on them are confused or dishonest. But Collosus specifically makes data centers look bad.
Grok's Strong Model Sees Low Demand, Underused Compute
Suggests that Grok just doesn't have enough inference demand to use its existing compute. Which is consistent with everything else we've seen. Grok is a good model with very little customer uptake.
Mythos Scores Just Above AI Trend, Not Revolutionary
Another new analysis finds that Mythos, while a fantastic model, is either pretty much on the AI trend or just slightly above it. [Similar to what I found in normalizing scores on Epoch's capabilities index.] https://t.co/KTwEdWM4Q3 https://t.co/33cn9vxSl4
Ocean Wave Power Enables Low‑Cost Floating AI Data Centers
I'm delighted to be a small part of this. The oceans are an incredible untapped resource for energy that can power AI. The most intense wave energy is thousands of miles from human settlements, but can power floating data centers...
Audits Needed, But AI Must Secure the World
Dean calls for mandated audits of AI company's safety claims. I think this is quite reasonable. I also think it's secondary to using AI (and non-AI methods) to secure the wider world. It's simply too easy to build a near-frontier...
Invest in Biosecurity Now to Counter AI and Pandemic Threats
We should absolutely take AI risks to biosecurity seriously. And we shouldn't panic. Now is the time to strengthen our overall biosecurity regime. The same investments in monitoring, preparation, and technology development will reduce risk from natural pandemics, lab leaks,...
AI Still Blind to Its Own Knowledge Gaps
I think about this as one aspect AI Dunning-Kruger. AIs historically don't know what they don't know. They are improving there, though.
Pro‑Claude Bias Emerges Amid Fierce AGI Competition
Not all metrics look like this, of course. But I do sense a pro-Claude bias among those I know who are most AGI/ASI-pilled and those who are most safety concerned. To me, the race seems incredibly competitive, and models from...
Superintelligence Can Be a Tool, Not a Being
Spot on. You could even get to general superintelligence and have it be a tool, not a being. Volition and desire are not the same things as intelligence.
Log‑scale Plot Masks Sub‑linear AI Scaling
This is impressive, but it's sub-linear scaling, not super-linear. The X axis is on a log scale. This is roughly power law scaling with an exponent of 0.3 (third root of compute), which is actually quite high compared to other...
AI Alone Won’t Enable Bioweapons; Expertise Remains Key
A call to make it easier to use AI models in biology and to rethink AI and biosafety. "The path from a model's "how-to" to a working bioweapon is long, technical, and littered with failure points" "Access to these models alone does...
AI Safety Requires Ecosystem-Wide Vigilance, Not Isolated Models
This perfectly illustrates why AI safety must be at the ecosystem level, not (just) at the individual model level. There are open weight models. They can be jailbroken. The AI genies are out of the bottle. No matter how safe...
AI's Future Hinges on Finding Network Effects
It's true that much could change in AI economics in the future. In particular, if any AI company finds a way to achieve strong network effects or feed-forward loops ('flywheels' in Amazon speak) that would change the situation dramatically. Not...
AI Won't Form Natural Monopolies, Benefits Spread Widely
If I could convince my liberal friends who fear AI of one thing, it would be this. There is no sign of natural monopolies in AI. There's every reason to believe that the large majority of benefits go to users...
Chinese AI Models Narrowing Gap, Now Just Months Behind
Some metrics Chinese AI models have fallen further behind*. Other metrics show they've closed the gap. On Artificial Analysis, Kimi 2.6 is 3 months behind the US frontier and the gap has shrunk. * - Even on CAISI's chart, Deepseek V4...
Future AI May Achieve Super‑Ethics Beyond Human Morality
Maybe. By and large, AI is likely to remain a tool, not a being. But if it were to have true volition, there's a reasonable case that we're building towards something more moral than the average human. If we can...
AI Sets Record for Speedy Adoption and Revenue Growth
AI overall shows both the fastest user adoption of any technology in history, and the fastest revenue growth of any sector and any individual companies in history. Hard to understand how some still dismiss it.
AI Sparks Churn, Yet Software Jobs Expected to Grow
We have indicators in every direction on AI and jobs. Specifically for software developers, there are signs both of job destruction and of increased demand. IMHO we should expect to see a fair bit of churn, but I still see...
AI Race Tight; Leaders only Months Ahead, Grok Still Viable
The overall AI landscape remains incredibly competitive. The leaders are only months ahead of the pack. Grok has fallen a bit behind the frontier, but I wouldn't count them out. https://t.co/qb17dJA7BC
AI's Work Future: From Doom to Optimism, Arguments Explored
Excelllent thread on various views of how AI will affect work, from catastrophic to highly optimistic, and the best arguments for each.
No AI Moat Yet; Runaway AI Remains Science Fiction
100%. There is no big AI lead. No AI company has demonstrated a sustainable moat. There's speculation that a frontier company will reach rapid self-improvement and run away from the crowd. This is, thus far, science fiction. There's zero sign...
GPT‑5.5 Matches Mythos Core Capabilities, Lags in Vulnerability Discovery
Is GPT5.5 on par with Mythos for cyber security? Mythos did discover significant vulnerabilities beyond what OpenAI has said about GPT5.5, but that was also the result of substantial harness work outside of Claude, which can be easily replicated. Our...
GPT5.5 Matches Mythos, Boosting Defender Advantage over Attackers
GPT5.5 performs as well as Mythos in a multi-step hacking test. Prediction: In the long run this will help defenders more than attackers. Also: Again we see that there is no moat.
Solar, Batteries, EVs Last Decades, Unlike Fossil Imports
100% agree. Once you've deployed solar panels, batteries, or EVs, they operate for 20+ years. Very different dynamic than importing fossil fuels.
Narrow Superintelligence Arrives; General AI Remains Uncertain
Spiky, narrow superintelligence is essentially here. Whether our current yesterday path can lead to general superintelligence is deeply unclear. Maybe. But not a lot of evidence for it so far.
Robot Dexterity Surges, but We Need Real‑world Demos
Robot dexterity is moving faster than I expected. I still want to see more demos that are in clearly uncontrolled circumstances. A little too easy to show us best case, highly scripted scenarios.
AGI Needs Human‑Level Data Efficiency, Not Yet Achieved
My definition of AGI includes: 1. Continuous learning. 2. Learning from data as efficiently as humans do. We have some progress on 1. Is there progress on 2? The power law scaling of AI precision with data volume seems to...
Brains Outlearn AI: Evolution and Powerful Neurons
Human brains are able to learn from much less data than current machine learning algorithms. The possible reasons include these three and more: 1. Evolution gives us a head start on certain domains. 2. Each human neuron likely does orders of magnitude...
Start Allowing Americans Access to New Cars Now
Americans deserve access to these cars. Yes, we have to solve security issues around them. Let's get started instead of delaying the inevitable.
AI's Competitive Landscape Prevents Monopolies, Benefits All
AI is an ultra-competitive market. There's no evidence of natural monopolies so far. Which is good for customers and the world.
Robot Arms and Cameras Will Replace Humanoid Legs
This is a job for a robot arm or two + a camera. No need for the legs or battery pack. This use case will get relentlessly optimized until it's no longer a humanoid.
Clean Power Outpaces Rising Demand, Peak Looms Soon
This is the first year that electricity demand grew and yet was ourpaced by clean electricity. 2025 was a special year. 2026 probably won't repeat this. But the true peak of global electricity demand is now on the horizon.
AI's Biggest Threat: Power Concentration, Currently Avoided
This is my belief. The greatest danger of AI would be concentration of power. Which, so far, we're avoiding.
Digital Security Improves; Overall Damage Is Decreasing
Narrative violation: The digital world has, empirically, become more secure, not less. ~Every device can be hacked at some cost, to be clear. But total volume of damage has declined.
Competitive AI Market Delivers Better, Cheaper Benefits to Users
I'm a big fan of OpenAI. Also a big fan of the hypercompetitive, non-monopolized AI landscape that is pushing all AI companies to provide better, cheaper products. As a result, the bulk of all AI benefits are going to users,...
Most Americans Want AI Regulation, yet See Opportunity
A ~supermajority of Americans favor government regulation of AI. While a plurality see AI as more opportunity than threat. https://t.co/lUtaP0JYza
Bipartisan Permitting Reform Needed to Level Energy Playing Field
The only way to pass permitting reform is to pass bipartisan permitting reform. Dems need to acknowledge that this means making it easier to move forward with fossil projects as well. Trust that economics will drive clean energy forward if...
Permitting Reform: The Last Barrier to U
Permitting reform - from local to state to federal, and inclusive of transmission - continues to be the key bottleneck on clean energy in the US. Let clean energy build and it will win.
ID Checks for Device Use Echo Dystopian China
Requiring you to show ID to use a computer or other device at all is about the most dystopian thing I've heard all day. What is this, China?
Innovation Keeps Crop Yields Ahead of Climate Threats
Ryan is correct. Crop yields continue to outpace the downward pressure of climate change and look likely to for the foreseeable future. There is risk of tipping points, to be clear, and those grow as temperatures rise. As of now,...
AI Boosts Self‑Representation, Yet Courts Face Overload
AI is augmenting the ability of people to represent themselves in court. That's great from the perspective of access to justice. It also poses challenges for the courts. Speaking to judges in one court system recently, it was clear that...
Interconnection Queues Inflate Data Center Growth, yet Few Build
Many of the headlines on data center growth are driven by GWs in power grid interconnection queues. But these queues are highly speculative. People enter a project into the queue to reserve a spot in line, even if they don't...
Narrow AI Already Superintelligent; Coding Advances Outpace General AI
Artificial Super Intelligence is already here. It's just narrow ASI, not general ASI. In chess, go, or just about any task for which we can generate infinite training data and trivially verify correctness, we've achieved super-intelligence. Coding isn't quite as...
AI Already Near Superintelligence in Verifiable Game Domains
I think this is broadly correct, with a specific caveat: There are domains where correct answers are 100% verifiable, quickly, and where we can generate infinite training data. In those domains (go, chess, any computer games) AI models are already...
Data, Not Compute, Sets AI’s True Limits
"AI is the log of data" is closer to the truth than "AI is the log of compute". And that has important implications. In domains where we can generate infinite high quality data, there's no AI wall. In less verifiable...
AI Inference Costs Falling 100x Yearly, Capabilities Stay High
Nah. 1. Frontier AI inference prices are plunging >100x / year. (fig 1) 2. Log-linear scaling of capabilities means you can reduce inference compute tremendously while preserving most of the capability. Very high elasticity. (fig 2) https://t.co/tTnU9ZUmdl
Trump EO Accelerates Psychedelic Research for PTSD and Addiction
Delighted to see the Trump EO easing research into psychedelics for PTSD, depression, and addiction. It was a prescient choice by Rick Doblin and others to explicitly research psychedelics for PTSD in veterans. Makes it so much easier to get...
AI Remains Highly Competitive, No Natural Monopolies Emerging
Strongly agree. This is one of the least discussed and most important features in AI. There is no evidence of natural monopolies. This is an extremely competitive field with rapid diffusion. Little sign of that changing.
Research‑Driven Philanthropy Yields High Societal ROI
There's a lot of effective philanthropy out there. One thing I particularly admire about @Arnold_Ventures is the focus on doing research and getting empirical evidence on what works to address societal problems. Very high potential ROI.
Open‑weight AI Keeps Competition Alive, Prevents Monopoly
It's incredibly hard to stop open weight AI developers from distilling knowledge out of frontier models. Another reason to believe AI will remain hyper-competitive and not monopolized. Economics from @dwarkesh_sp newsletter this morning. https://t.co/KKocBrLgNs https://t.co/swXubgHvbM