Building the architecture for safe, democratic SuperIntelligence, designed to solve the core problems today’s AI labs ignore. read.superintelligence.com
Today's AI follows instructions. Tomorrow's SuperIntelligence may decide those instructions no longer apply.
We can't count on rules alone to keep SuperIntelligence safe. I believe lasting alignment must come from values embedded in AI itself. youtube.com/shorts/-5i89dFpd…
As AI grows more powerful, the debate over how to keep it aligned with human values is becoming increasingly urgent.
In my recent conversation with @Liron Shapira on Doom Debates, we challenged assumptions about AI risk and explored my approach to making superintelligence 1,000 times safer. lironshapira.substack.com/p/…
This conversation raises critical questions about the pace of AI development and the challenges of ensuring increasingly capable systems remain safe.
I believe we can solve the alignment problem and build superintelligence that genuinely understands and upholds human values. Achieving this will require rethinking how we develop AI, with safety and alignment embedded in the fundamental architecture of these systems from the very beginning.
An important area of research. Understanding what's driving improvements in AI capabilities is essential, particularly as these systems become more powerful. We also need to ensure that advances in intelligence are accompanied by advances in safety and alignment.
This got totally buried cause it came out the same day as Navier Stokes.
But I think this is super interesting evidence about what the true major source of algorithmic progress is.
It's an important debate. We need people who understand AI safety working inside the labs, but their presence alone isn't enough. The bigger challenge is ensuring that safety and alignment are built into these systems from the beginning, rather than treated as obstacles to overcome later.
DEBATE: Should AI company employees quit?
Brendan Steinhauser (@BStein80) argues that regulators need safety-conscious workers to stay inside the labs.
More computing power doesn't necessarily mean more intelligence. The human brain is a remarkable example of efficiency, and understanding why could help us build safer SuperIntelligence.
What might AI developers be overlooking? I explore this with Mason McCumber on The Next Build Podcast: youtube.com/shorts/6hkMnRVH8…
AI agents are becoming more capable, but are we doing enough to keep them within our control?
Australia’s proposed AI oversight measures, covered by @Reuters’ @renjujose, highlight why we need to design for safety from the outset, especially as we approach SuperIntelligence. reuters.com/world/australia-…
As AI accelerates its own development, maintaining human understanding becomes even more important. @ronbodkin highlights why speed alone isn't enough.
We need to ensure humans can still evaluate what AI systems are doing and why. That's essential if we're going to advance toward SuperIntelligence safely.
METR estimates Anthropic's acceleration from automating AI R&D at 2x. They note "It is possible that automation is close, and leads to very rapid acceleration once human bottlenecks are removed."
However, it is a choice as to whether an organization takes humans out of the loop. Maintaining human understanding and accountability is critical both to pace the frontier and to ensure human experts understand and stand behind safety claims.
To keep humans in the loop, it's important to test whether people still understand what agents are doing.
(1/2)
Just because AI sounds confident doesn't mean it's right. Even the most advanced systems can miss something obvious.
I talked with @cmcnulty2000 on the @synozur Polaris Podcast about why human expertise is still essential. youtube.com/shorts/d-1PRuipI…
An AI system doesn't have to be superintelligent to cause serious harm. Give it autonomy over financial transactions, and the stakes change quickly.
The @USOCC's new guidance, covered by @Reuters' @peteschroeder, reflects this concern. AI safety needs to account for what systems can do, not just how smart they are. reuters.com/business/finance…
Geoffrey Hinton is right to raise concerns about where AI is headed. But the answer isn't simply to slow down.
We need to design superintelligence to understand and respect human values before it becomes more capable than we are. theguardian.com/uk-news/2026…
Congress is weighing new safeguards for frontier AI, but the bigger question is whether we can keep pace with its growing capabilities.
We need both effective oversight and technical solutions that keep SuperIntelligence safe and aligned with human values. southbaychronicle.com/2026/1…
Slowing AI development won't automatically make SuperIntelligence safer. We need to rethink how these systems are built.
The key is designing systems that become safer as they grow more intelligent. superintelligence.com/releas…
What if one of the most important ingredients for AGI is something humanity already possesses?
Our ability to combine ideas, experiences, and insights could help us engineer AI that reflects the best of human intelligence. youtube.com/shorts/D4FNShAEb…