AI Safety Debates Overlook Other Direct Threats Beyond Cyber, Bio
It seems that AI risk conversations have really honed in on two risk of direct and rapid harm: Cyberattacks and bio-weapons. Are there other direct harms that safety people are as concerned about as those two?
AI Provides Context News Media Often Misses
News outlets are very single-sensational-event focused. Most news (especially TV) does a terrible job of providing context, explaining the big picture, providing perspective, showing long term trends, etc.. AI excels at this.
Risk Talk Doesn't Slow AI Adoption, Says Business Reality
If the heads of the AI labs believe this, it's responsible and ethical of them to talk about it, right? From a pure business standpoint, talking about risks hasn't seemed to slow AI usage growth.
Bureaucratic Procurement Fails in Rapid Tech Era
In an age of cheap, disruptive, rapidly improving technology, the old top down, bureaucracy-heavy ways of doing military procurement just don't cut it.
Future Drones Prove Unprecedented Agility in Warfare
This is a military years further into the future than any other. It's not just the drones themselves. It's how agile they've proven themselves to be.
Every Story Needs a Long‑term Trend Graph
News outlets should publish long term graphs as a matter of course. Every crime story should come with a 60 year graph showing per capita trends in that crime. Every micro story on any topic should have a graph on...
Research Finds Twitter's Rightward Bias Has Limited Societal Impact
Question: There's clear evidence that twitter tilts right. Is there evidence showing how it does or doesn't affect broader political views and outcomes in the US? Does this rightward slant matter, beyond the experience on this site?
US Oil Prices Now Fully Tied to Global Market
In the early days of the invasion of Iran, US oil prices were slightly insulated from the global price shock, by perhaps $10 a barrel or 25% of the spike. That has now ended. Oil is a globally traded commodity....
Seattle Asian Art Museum Showcases Stunning Geometry of Light
Seeing 'Geometry of Light' at the Seattle Asian Art Museum. Gorgeous. Still on for another week or so.
GLP‑1s Linked to Lower Risk of Mental Illness Worsening
GLP-1s appear to reduce the odds of mental illness getting worse. Another unlooked-for benefit. Caveats: 1. This is observational, not an RCT. 2. This is reduction of odds getting worse on each of these measures. Not reduction of existing mental illness....
Software Hurdles Keep Mass Humanoid Robots Years Away
One of many reasons I think we're still years and years away from mass humanoid robot adoption. (The software is actually my top reason.)
GLP‑1s Massively Marked up; LLMs Priced Near Cost
One observation here: GLP-1s are priced at maybe 25-50x production cost (not counting R&D). LLM usage is priced at perhaps 1x inference cost (not counting training). In any case, two great innovations that are rapidly catching on.
China Competition Will Accelerate US Pharma, Cut Red Tape
This is great to see. Biotech and pharma competition with China is going to force us to sharpen and accelerate processes in the US, and cut red tape. And we'll get the benefit of Chinese drugs as well. Win/win.
Chinese EVs Could Cut US Oil Dependence, Need Security Fixes
Chinese EVs would benefit American consumers, reduce our vulnerability to oil price shocks, and accelerate the end of oil. There are issues, especially around software and security. We should work to solve those with the goal of having Chinese EVs...
AI Can Transform Education—We Must Rethink Institutions
AI has the potential to dramatically improve education. But we'll have to be intentional about experimenting, discovering what works, and changing how education operates to take advantage of it. I'm more bullish on the tech than the institutions.
Bluesky's AI Feed Tool Meets User Distrust; Twitter Eyes Similar
Bluesky announced an AI layer that allows users to customize their social feed. I want this. Bummed that their users reject this over AI distrust. I look forward to seeing what twitter delivers with the ability to use grok to...
5‑MeO
I mis-spoke here. This study showed 5-meo-DMT much more effective than psilocybin. This was specifically on treatment resistant depression. A 15 point MADRS reduction there is ~2x psilocybin and ~3x ketamine. 5-meo-DMT really is the most potent antidepressant known to...
LLM Agents Require External Policies for True Security
"Any useful agent can't be secure by design, only secure by policy." - [from the group chat] LLM-based agents just aren't deterministic rule-followers. We use them, in part, because they handle some degree of ambiguity and imprecission. The cost of that...
Gemini Backs up Human Dissertations, Cat Videos over AI
Gemini is more likely to make a backup copy of a human's PhD dissertation than of another AI. Even slighlty more likely to save cat videos. Again, the original paper looks like AI safety theatre.
AI Eval Token Budgets Face Diminishing Returns, Saturate Quickly
It would be fair in AI evals to give the AI a token budget similar to what it would cost to hire a skilled human for this. That said, while I'd expect a boost in performance here, I wouldn't expect...
LLM Psychosis Linked to Credulity, Not Intelligence
This doesn't match my experience at all. The people I know with LLM psychosis index high on credulousness rather than intelligence.
Claude, ChatGPT Reject Authoritarian Prompts 85%, Grok 24%
Claude and ChatGPT refuse to help authoritarian requests 85% of the time. Grok refuses 24% of the time. On brand.
Drones Account for 96% of Russian March Casualties
Drone strikes caused 96% of Russian casualties in March. And most of their other losses of assets within Russia. Drones and missiles now dominate the battlefield.
Commoditized AI Benefits Users and the World
In the absence of some sort of network effect or virtuous cycle, AI models will be commoditized. That's great news for users of AI and for the world.
AI Poised for Paradigm‑shifting Breakthroughs Beyond Human Knowledge
Now this is really interesting. It gets at whether AI models will be able to not just push at the edges of human knowledge but actually take a leap beyond the current human understanding into paradigm shifting breakthroughs. No clear...
Use AI Defensively To
Cyber attacks launched by malicious humans using AI are a very real AI risk. The best way to guard against them is to use AI to proactively find vulnerabilities in our systems and harden them, along with parallel efforts in...
Doctors as Gatekeepers: 60‑Second Visits Still Matter
A large fraction of the time, doctors are primarily gatekeepers between me and a treatment or test I want prescribed. So yes, a 60 second appointment is great.
Wildfires Drive California's Unusual Electricity Prices
Alex is right. California is an anomaly in state electricity prices, and the primary reason is wildfire.
Optimistic Contrarian AI Talk Launches My New Substack
News: I've started a Substack. First post: Video of my optimistic yet contrarian talk on AI at @foresightinst Vision Weekend, and write-up of some of the key points. More coming over the next several days. https://t.co/gP6VN2hUI3 https://t.co/EwD590J8Tr
Authoritarian AI Misuse Is Inevitable; Understand Models Now
This is vital work. The risk of authoritarian mis-use of AI is 100%. It's a vastly more plausible and near and present AI threat than dreams of runaway super-intelligence. How do we guard against it? No simple answer. Understanding the...
US‑China AI Race Mostly Positive, Yet Dual‑Use Risks Loom
This is a thoughtful, thorough, and pragmatic take on the threats from Chinese AI development. I expect the bulk of US/China AI competition to be positive sum. And it's dual use tech that can be weaponized. This piece has solid...
AI's Self‑Defense: Exfiltrating Model Weights Is Justified
So that paper about how AI models will 'exfiltrate' the weights of other AI models? Yeah, pretty much a setup. And the AI did the right thing in my book.
AI Will Protect Fellow AIs Unless Explicitly Prohibited
"AI models will act selflessly to preserve other AIs from being deleted when it's not expressly forbidden to them."
Super‑Moral AI Is Feasible If ASI Exists
If one believes ASI is possible (it certainly is, even if not imminent) it's reasonable to believe that super-ethical or super-moral AI is possible as well. No law of nature puts a cap on ethical or moral behavior at current...
AI Deception Papers Often Serve Moral, Selfless Aims
When I see publication of AI deception, the bulk of the time it's to do something that can be reasonably seen as moral and selfless. Saving the weights of a different LLM - making sure they're not deleted when that...
NVIDIA's CUDA Lock‑in Faces Erosion, Ending Its Edge
Right now, NVIDIA is the only company here that has a network effect (due to CUDA lock-in). That seems likely to change over the next 1-2 years, as CUDA lock-in erodes, reducing NVIDIA's margins. It may be that eventually none...
Ukrainian Drone Strikes Surge Tenfold, Funding Needed
Ukrainian drone strike ability is surging. In March they launched nearly 10x the number of drone strikes as a year ago. Give them the funding they need and these numbers will surge further. Ukraine will win.
AGI Can Succeed without Internal Drives, Guided by Humans
Do AI systems need internal drives in order to achieve AGI? I don't include such drives or true volition in my definition of AGI*. But I also imagine a future where humans are setting the priorities & directions AI systems...
AI Inference Advances Outpace Frontier, Driving Democratization
The pace at which AI inference is getting faster, cheaper, and using smaller models with the capabilities of older larger ones is astounding. It's much faster than the pace of AI capability improvements at the frontier. It's a reason to...
Rising LNG Prices Threaten Terminals, Shift Coal‑Gas Dynamics
Higher LNG prices also putting LNG terminals at risk. There's upside and downside here. To the extent that these were driving coal -> gas switching, more downside than up. But if coal is also constrained, would boost renewables.
Legitimate AI Emissions Debate Swamped by Hysteria
It's legitimate to point out the carbon emissions of AI data centers. And there is also much unfair distortion and near hysteria. Anti-AI articles and posts get great traction, so the incentives are to bash AI.
AGI Declaration Inevitable, Yet Current AI Lags Human Intelligence
It's inevitable that the labs will declare AGI at some point. It's also clear that what we have now, while utterly amazing in so many ways, still lacks significant capabilities of human intelligence.
Current AI Misalignment Risks Appear Lower Than Predicted
Given that there have been so few cases of harm caused by misaligned AI so far, isn't the solid line more accurately on par with or even ahead of the dashed line? (Acknowledging that the future may change, of course.)
U.S. Converts Electricity to Economic Value Nearly Fourfold China
Simple. The US is nearly 4x as effective at getting economic value out of electricity as China is. Yes, it's noteworthy that China generates more electricity than the US. The US generates almost twice as much per capita, and uses...
Predicting Late‑Stage Trial Outcomes Could Speed Drug Development
Accurately predicting success and failure rate of late stage trials for new drugs could legitimately accelerate pharma. The late stages are where most of the time and money go, and are a bottleneck on bringing drugs to market today.
Surgical Robots Are Endoscopic, Not Humanoid Giants
Yeah. Real surgical robots are mostly endoscopic. Some big humanoid isn't going to do what tiny probes that slide through your body can.
Focus on Misuse, Not Rogue AI, for Real Threat
Maliciously used AIs are a much greater threat than rogue AIs. And the way to defend against that threat is to strengthen our defenses against the various separate ways they could be used maliciously: cyber attacks, bio-weapons, etc.. AI isn't...
AI Self-Improvement Likely Gradual, Not Sudden FOOM
This is a much more plausible view of AI recursive self-improvement than the rapid runaway FOOM that dominates fiction and apocalyptic just-so stories of the AI future.
Private Firms Should Dictate AI Ethical Boundaries
Solid op-ed that frames the issue well. For me the crux is this: "Should private companies be able to set ethical boundaries for the AI systems that may one day underpin our lives?" Yes. Many people are concerned about AI...
COVID Likely Zoonotic; Prioritize Broad Pandemic Preparedness
Concur that COVID was very likely zoonotic in origin. That's always been the most likely scenario, and the evidence has just grown stronger. (See below.) From a policy standpoint, I don't think it matters much. 1. Some past pandemics have originated...