Trump Backs AI Kill Switch; Raises Fears of Government Coercion
If a government AI kill switch already existed, would the current DOD or current administration have used it, or the threat of it, against Anthropic? Could they use it to coerce AI companies into behaviors that are counter to the public interest?
Powerful AI Stays Controlled, Weaker Tools Spread Widely
Inspired by Miles's post, and the announcment of the open source nano-analyzer by @Aisle_Inc yesterday, my view of the AI cyber landscape. I think this is roughly the dynamic we want. The most powerful tools are held closely. Less powerful...
Tool Enables AI Alignment and Deeper Model Introspection
Very cool. A useful tool for alignment. Could also be a tool to give LLMs deeper introspection capabilities if we chose to.
AI Handles Mechanical Tasks, Boosting Alignment Researchers' Productivity
Just as AI can improve the productivity of AI researchers it can specifically improve the productivity of AI alignment researchers. It does so by taking on the most well defined and mechanical tasks, leaving the ambiguous and amorphous work to...
AI Legal Chats: Should They Get Attorney‑client Privilege?
Should legal chats with an AI be protected by attorney-client privilege or something like it? I have some bias towards yes. But also good to see this person trying to commit fraud with ChatGPT's help lose.
Future Jobs Thrive in Human‑Centric Relational Roles
Where will jobs be in a post-AGI world? "I argue it's what I call the relational sector. Categories where the human isn't just an input into production, it is part of the value." Very good read.
World Economy Decouples From Oil, Yet Some Sectors Remain Vulnerable
The global economy is less coupled to oil consumption than at any point in the last century. This buffers us against the shock of having 15% of production locked up. Though within this, certain sectors and nations are still very...
Data, Not Compute, Will Bottleneck Scientific AI
AI progress in science depends on high quality training data and the ability to rapidly verify results. E.g., as @michael_nielsen pointed out recently, the training data is what made AlphaFold possible. Data and experimentation will be the science AI bottleneck,...
Question Doom Probabilities: Balance Optimism, Pessimism, and Evidence
What I see, subjectively, is that's cool to have a high p(doom). It signals to others that you've thought farther ahead, that you're smart enough to model the future, and have the skills to see consequences that most don't. It's...
Mythos Outpaces Opus 4.6, Completing All 32 Steps
Mythos's offensive cyber capabilities are indeed a significant step forward, as verified by the UK's AI Security Institute. On average, Mythos gets about 40% farther on a multi step attack than Opus 4.6, and is the first model to complete...
India’s Solar Capacity Hits 150 GW, Installations Surge 87%
India has now passed 150GWs of solar, with annual installations increasing a whopping 87% this year.
Authoritarianism’s Planning Edge: Democracies Must Adapt
China embraced capitalism, opened up to trade, invested in infrastructure, cracked down on corruption, and has the most human capital of any nation on earth. Authoritarianism did help in being able to create long term plans and execute on them....
US only Restricts Iranian‑origin Cargo, Not Whole Strait
So the US isn't actually blockading the Strait of Hormuz. Just the small fraction of oil and other cargo that comes from Iranian ports. https://t.co/uWubDRslxe
Mythos May Outperform GPT 5.4, Raises Security Stakes
There does seem to be an anti-Claude vibe shift. As someone who has tweeted reasons to believe Mythos is mostly a continuation of current trends, I want to clarify that: 1. The cyber security implications look major and quite important....
Policy-Driven Scaling Made Renewables Finally Cost‑competitive
Solar, wind, batteries*, and EVs only dropped in price to where they win economically because earlier policy driven by environmental concerns scaled them. * - Batteries are a partial exception due to early scaling from consumer electronics. But that only...
Mythos: Strong Yet No Breakthrough, Just Linear Progress
Mythos, while a very powerful model, appears to be a relatively linear improvement over previous models. AI progress is real, impressive, and continuing. But Mythos does not appear to be a discontinuity.
China's Magazine Production Capacity May Be Underestimated
Maybe, just maybe, our estimates of China's magazine depth are off. In any case, China can presumably crank up production of these quite rapidly.
Recursive AI Gains Hit Diminishing Returns, No Fast Takeoff
AI Recursive Self Improvement will likely face strong diminishing returns. Anthropic estimates that it takes a 40x productivity increase for its researchers to achieve a doubling of the rate of AI progress. That's roughly an exponent of 0.2 (the fifth...
GPT‑5.4's Hacking Prowess Heightens Security Urgency
GPT5.4 is apparently very effective at hacking as well. Makes proactively using these models to secure the world more urgent and important than we knew.
AI Must Remain Democratic, Plural, and Multi‑Polar
AI is fortunately democratized, plural, and multi-polar. It's important to human freedom that it remains that way. No one person, company, or nation can ever be allowed to control it.
White‑collar Work’s Nonroutine Cognition Slows AI Verification
Most white collar jobs in the US are ”nonroutine cognitive". For those jobs, it's likely that a significant part of the work is made up of tasks where perfect verification is going to be tough, and thus where AI progress...
Uncorrelated Low‑Rank Model Boosts Crowd Forecast Accuracy
This is the wisdom of crowds. Diversity of thought increases overall performance. Grok is only the 5th best forecaster, but its answers correlate least with the rest. You wouldn't choose it if you could only use one model, but added...
AI Will Outpace Humans on Fully Verifiable, Data‑rich Tasks
AI will exceed human capabilities on any task that is 100% verifiable and for which unlimited training data can be produced. That was evident with AlphaGoZero. Now it's coming for coding. At least, in cases where you can perfectly define...
Ukraine Catalyzes Europe's Leap Into Modern Military Era
Ukraine provides a chance for all of Europe to jump into the modern era of military capabilities.
Boost Clean Energy, Not Consumption Restrictions
Amen. The way to address climate or energy is not to police consumption, but rather to increase clean energy abundance.
High-Power Chargers Target Heavy Trucks Outside China
Correct take. Outside of China, anyway. We will build chargers at this power output, but they'll be used primarily for heavy trucks.
Mythos Cuts Cyber Attack Surface, Makes Exploits Scarcer
Concur. Mythos shrinks the total cyber attack surface area. There will be fewer exploits remaining, and those likely harder to find. And because model capabilities scale sub-linearly with both compute and training data, it likely takes a very large $$...
AI Cybersecurity Closes Gaps, Giving Defenders the Edge
Yes. We're closing vulnerabilities that could have been (and may have been) exploited by bad actors already. AI cyber-security may well favor defenders in the large majority of cases.
Fearful AI Under State Control; Favor Democratic, Regulated Pluralism
Strong nation-state control of AI is terrifying. Democratized, plural AI with sensible regulation is the way.
Marketplaces Choose Models That Meet Desired Behaviors
Good thought. We are - as a marketplace - selecting for models that do what we want them to do.
AI Safety Requires Ecosystem Defense Against Model Jailbreaks
We should assume that any malicious actor can use the top open weight models and can jailbreak them so they'll answer any question on bioweapons, cyber attacks, etc.. AI safety fundamentally must happen at the ecosystem level. We must...
Antarctica Mirrors Humanity’s Next Frontier in Space
This is an absolutely delightful article about Antarctica, with very deliberate insights into human exploration of space.
Fact‑checking on Facebook Is My Top AI Wish
Fact checking integrated into Facebook (like Grok on twitter) is the single thing I most want from Meta's AI efforts. If they accomplish nothing else, I'll still be delighted if they build that feature.
Orban Aligns with Iran, Russia; Vance Supports Enemy
Viktor Orban is allied with Iran and Russia. Vance is there trying to prop up our enemy.
Mythos Mirrors Trend, No ECI Acceleration Detected
Anthropic's Mythos does not appear to show any acceleration of ECI. After normalizing Anthropic's internal ECI with @EpochAIResearch 's public ECI, it's clear that the two metrics are extremely close, and that Mythos is pretty much on trend, just slightly...
Mythos Gains Not Driven by Recursive Self‑Improvement
Anthropic's Mythos achievements were not a result of AI Recursive Self Improvement. (Per the Mythos system card.) https://t.co/zYN8UMQpri
Balanced AI Governance: Regulation, Not State Control
Nat is completely correct here. It's quite fortunate that we live in a multi-polar AI world. Government should be one of those poles. Sensible regulation makes sense. But nationalization or complete government control is terrifying from the standpoint of human...
Open‑Weight LLMs Detect Same Vulnerabilities as Mythos
It's not just Mythos: Cheap, open-weight LLMs can find the vulnerabilities that Anthropic revealed Mythos found.
Prioritize Misuse Risks Over Superintelligence; Hold Models Back
I think this hangs on whether you see AI risks more from bad actors mis-using AI vs an AI becoming a super-intelligent adversary. Since I'm 99% in the former camp, I see responsible AI companies holding models back until mitigation...
Superhuman Cybersecurity Tool, Not True AGI Yet
Not at all sure this is AGI, and in fact doubt that it meets my criteria of parity with human cognitive capabilities in general. But it's a superhuman cyber security tool. ASI is narrow. AlphaGoZero is a narrow ASI. Mythos...
AI Front‑Runners Act Responsibly, Defying Sci‑Fi Corporate Villains
In science fiction, the corporations are often evil doers. In the AI world right now, we're pretty fortunate that the three companies at the frontier are fairly responsible actors, making proactive choices in the interest of the public good.
US Withdrawal Fuels Iran's Stronger, Angrier Toll Regime
Very plausible scenario. The US bombs for a few more weeks, then leaves. Iran starts charging a toll for ships leaving the Strait of Hormuz. After some period of rebuilding, it leads to an Iran that's stronger and angrier than...
Oil Shock Dampened: Prices Below Peaks, Economy Decouples
So far this oil shock is not quite at the level of previous shocks. 1. Inflation adjusted, prices still aren't at their highest 2. The world's economy is less coupled to oil than in years past.
Anthropic's ARR May Be 25% Overstated Compared to OpenAI
Context on Anthropic's recent revenue announcement. Still incredible, but probably slightly overstated relative to OpenAI, due to differences in how they report.
Threats of Strikes Spark Iranian Unity, Boosting Regime Support
A danger of attacking Iran was always that it might increase support for the regime through a rally-round-the-flag effect. I don't know how widespread the below is, or how astroturfed, but it's consistent with that kind of effect.
AI Future Will Be Plural; Google Holds Key Advantages
I continue to believe that the AI future is plural, with multiple leading companies. And that Google has some important advantages, if they use them.
Only Congress Can Rename the Department of Defense
Your reminder that it's the Department of Defense, by law. Only Congress can change that. Trump has merely told people that they may use a secondary title of Department of War. Doing so is a choice, and sends a signal....
Optimistic About AI, Yet Skeptical of Near‑Term AGI
"Two Hours of AI Optimism - Grounded in Reality". My opening talk at @Abundance360, with demos of Google tools by @DynamicWebPaige. I explain why I'm bullish on AI and yet skeptical of near-term AGI or fast take-off to super-intelligence. https://t.co/flutju2hNj
AI Self‑Improvement Peaks: Early Gains, Later Diminishing Returns
This is the best essay and modeling of AI recursive self-improvement that I've seen. It reaches the same conclusions I do (that in the foreseeable future we'll see notable improvements but then diminishing returns). And it does so via a...
OpenAI Outlines New AI Policies with Concise AI Summaries
OpenAI proposed a number of AI policies to strengthen society today. Here are three AI-generated summaries for the tl;dr set. https://t.co/qv5rKvtIPF