Superman III was powerful metaphor for our relationship with technology. The true danger today isn't the machines actively taking over, but rather us being "sucked into the machine" gradually losing our authenticity and agency to become components of a larger technological

RT @IAmTimNguyen: The problem with Eric Weinstein is that his theory has already been discredited, per my article hosted on Sabine's blog.….

RT @kenneth0stanley: I worry that so much discussion of AI risks and alignment overlooks the rather large elephant in the room: creativity….

So, true intelligence is more than just having capabilities. Intelligence is "more with less"! . An important feature in nature is the ontogenetic capacity to acquire new capacities efficiently. The brain "bootstrapped" itself and evolved by phylogeny to overcome its own slow.

To overcome the phylogenetic "speed limit", life evolves "extragenomic" systems i.e. brains. These are *inferential* organs designed to acquire and process high-frequency information. They create their own compact, causal models of the world.

This new system is what David's recent LLM paper calls "Knowledge-In" (KI) emergence, where complexity arises from extracting information from a complex environment. "In KI cases, the word emergence is often substituted with the words engineered, developed, evolved, trained, or.

But the world is full of high-frequency information: a predator's location, a fleeting opportunity. This creates a selective pressure for a new kind of learning system, which is one that can operate much faster than evolution. This is the origin of *ontogenetic intelligence*.

Evolution (phylogeny) is a powerful learner, but it faces an information bottleneck. As Krakauer explains, its learning rate is limited by *generation time*, which is roughly "one bit per selective death." It's incredibly slow, optimised for stable, low-frequency environmental.

What is intelligence? . Many definitions focus on capabilities, but David C. Krakauer from @sfiscience argues for a deeper, evolutionary view as a starting point. We must understand two timescales of information acquisition: . - the slow march of evolution (phylogeny) and - .-

RT @max_nlp: Really enjoyed discussing the state of AI benchmarking alongside Prof Mark Bishop, @IAmTimNguyen, Enzo Blindow & @ecsquendor a….

RT @keyonV: Can an AI model predict perfectly and still have a terrible world model?. What would that even mean?. Our new ICML paper formal….

RT @hardmaru: Every ML Engineer’s dream loss curve:. “Kimi K2 was pre-trained on 15.5T tokens using MuonClip with zero training spike, demo….

RT @karpathy: Scaling up RL is all the rage right now, I had a chat with a friend about it yesterday. I'm fairly certain RL will continue t….

The "Superintelligence Strategy" paper offers an interesting path forward (in Dan’s opinion):. Deterrence (MAIM): Acknowledging mutual vulnerability to prevent a destabilizing race. Nonproliferation: Keeping WMD-level AI capabilities out of the hands of rogue actors.

Beyond state-vs-state conflict, Hendrycks identifies a more subtle threat: the gradual "erosion of control." Economic & military pressures create powerful incentives to cede authority to AI systems without clear limits. As the paper notes, this is not a sudden takeover: "This

So if a race to monopoly is a path to escalation, what's the real strategic dynamic? The paper introduces Mutual Assured AI Malfunction (MAIM). This is the default reality where any state making an aggressive bid for a strategic monopoly on AI can expect its project to be

Hendrycks critiques the call for a "Manhattan Project for AGI" as a flawed "take over the world strategy." His reasoning is based on simple game theory: it would be "extremely escalatory.". A rival nation would not sit idly by while another builds a monopoly on intelligence. The

Discussing his new paper "Superintelligence Strategy" @DanHendrycks said: . "I think it's harder to make cutting-edge GPUs. than nukes.". Unlike enriching uranium, which is a known process, GPU fabrication is an incredibly complex and concentrated industry. This physical

@BasedBeffJezos The result is their "p-bit" i.e. a physical probabilistic bit. Implemented as a tuneable double-well potential in silicon, it acts as a fractional bit whose state "dances" between 0 and 1. The payoff is unprecedented energy efficiency, generating controllable entropy at just a.

The major breakthrough from @BasedBeffJezos's team is porting this entire paradigm from exotic superconductors to standard silicon. Engineering programmable stochastic physics in silicon was a monumental challenge, requiring them to innovate on the control of electron mechanics.