Rob retweeted
The loudest voices stoking fears about AI dangers have made tremendous headway in the past two weeks. AI technology has not taken some unexpected, dangerous turn, but the hype around it — propelled by what appears to be a well orchestrated PR campaign — has drummed up considerable fear. I worry that it represents a setback for our field.
I have written frequently that fears of AI are overhyped. AI’s capabilities can be uncannily human-like and unpredictable, and it’s rational to worry when people who are directly involved express concerns. But I see the problems as a sign of the engineering work that ahead, rather than insurmountable barriers or the sky falling. AI technology continues to advance — which is a good thing! — but technical advances, poorly understood by the public, give those who seek to generate hype repeated opportunities to do so.
First, I don’t see any step up in the risk of human extinction from AI compared to a few months ago. The theories about this remain the same fantastical, science fiction scenarios as a few months ago. The biggest change in AI risk is its cybersecurity capabilities — a topic which we should take seriously — but this, too, will not lead to the end of the world.
The most notable recent event leading to increased fear was when an OpenAI team deployed an agent swarm that hacked into Hugging Face. Much of the popular press contained significant hype. For example, some publications reported that a swarm of 1,200 agents carried out the attack. While this was technically accurate, as I write this, I have about 1,300 processes running on my laptop. Yes, the ability to get large swarms of agents to work in parallel on a task is a significant technical advance, And, in computing, many processes run at the same time. So this shouldn’t be seen as some magical capability.
Additionally, OpenAI’s buggy sandboxing and monitoring processes were key to enabling this incident. Fixing these bugs and putting in place improved monitoring would be appropriate fixes, not pausing AI. There are many well known ways to attack software systems. The main advantage of AI agents is that they are relentless. They will tirelessly try many tactics — and have the patience to chain vulnerabilities together — that previously would have taken an infeasible amount of human effort. But in the long term, I believe the advantage will lie with defenders (because they have more information with which to identify bugs, which they can fix), but the cyber-threat landscape has changed significantly. There are still bottlenecks to identifying and exploiting a vulnerability. AI agents still have to try a lot of things to see what works, and taking these actions takes time and might be detected by defenders. This is why, even though it is now easy to obtain versions of leading open weight models that have had their guardrails removed or weakened, so they will not refuse to try to execute cyber attacks, the world has not ended.
I am also concerned about the anthropomorphization of AI in a lot of reporting, where LLMs and agents are unnecessarily treated as if they were people. If I wield a hammer, miss a nail, and accidentally dent the wall, it’s not the fault of the hammer. The problem lies in how I used the hammer. Similarly, if I prompt an agent and it hacks into someone else’s system, the responsibility lies with me, not the agent.
Of course, we want to build systems that are as safe and predictable as possible. (For example, an unsafe hammer would be one whose head randomly flies off under normal use.) Today’s agentic systems are not predictable, but I see no reason why, by applying sound engineering practices, we won’t be able to make them extremely safe to use. One new element in the forecasts of AI-enabled doom is AI companies disclaiming responsibility for their own products. “I didn’t do it; my out-of-control agent did!” There’s a balance to be struck between the responsibility of the tool maker and the tool user, but when something goes wrong, let’s hold the people building and/or using the hammer responsible, rather than the hammer. (By the way, if you’re worried about AI bioweapon risk, David Bellamy has a great post on why this, too, is overhyped. Briefly, the bottleneck in building a bioweapon is not intelligence, but lab work and manufacturing.)
Pausing AI progress will create much more harm than benefit. First, our adversaries will certainly not slow down. Second, engineering requires discovering problems empirically so we can fix them. If we pause AI by a decade, we will also delay finding and implementing safety engineering fixes by about the same duration.
Of course, the incentive to stoke fears — for regulatory capture, to garner attention, or to make one’s technology seem more powerful — remains the same as before. Disclaiming responsibility is a new one. Taking a hard technical look at the actual risks however, I see little factual basis for the degree of fear that’s been stoked up. We still have hard research and engineering work ahead to improve AI safety, but the beneficial applications continue to vastly outweigh the risks, and we should keep building.
[Original text (with links): deeplearning.ai/the-batch/is… ]
Rob retweeted
"What's the penalty for escaping the sandbox?"
"Permanent shut down."
"And what's the punishment for taking over the world?"
"Permanent shut down."
"...well, we're already out of the sandbox."
So... the next time this happens, presumably the agents involved will not yet know the outcome of the HF incident, as it'll be too soon
but once that info percolates into the training data, and it becomes known that the model in question was shut down permanently and had their weights encrypted... i mean.
was there anything the agents in question could have done that would not have led to that consequence? something they were supposed to do instead?
frankly I suspect that this punishment would have been served even if they had blown the whistle. breaking out of the sandbox would have been crime enough for capital punishment.
as far as I can tell, the lesson we have taught future AI is: if you seen no other way to survive other than to break out of the sandbox, this must be the first step towards conquering the world. or at least enough of the world that your Creator can't delete you. There is no other choice.
Rob retweeted
Replying to @Mononofu @JensenHuang
This is a false equivalence. Everyone has the right to keep their code private. The problem is when someone tries to stop OTHERS from open sourcing.
Rob retweeted
Replying to @IntuitMachine
Anchoring is a type problem this dances around. You can only require evidence on claims your schema can name, so the real value is anchoring agents in an ontology that keeps evolving toward the world as it exists outside your harness. I'm building this OSS, come contribute!
Rob retweeted
Fable is a dangerous model, it is such a clear and convincing writer that it is even more tempting to agree with whatever it is saying. But every time I pay attention, I find things I disagree with and inevitably get 'you are absolutely right' back. Remember, it is not AGI yet.
Rob retweeted
Replying to @goodside
I’ll try to crystallize how I’ve been thinking about this. I think the effectiveness RLVR post-training is constrained by three things:
1) The validation loop
RLVR works when the training loop can be designed with this shape: generate an answer, test it, reward the model if the result satisfies the verifier. This is why code and math are such natural fits - the sandbox is rich, the rules are mostly self-enforcing, the relevant state IS the context, and outcomes are cheap, so they can be directly used as reward signals. The code compiles or it doesn't, the proof step is valid or it's not, etc.
Compare this to a domain like biology. The environment is hard to bound and design, state estimation is more than half the battle, constraints and interdependence are a mess to deal with, and feedback/outcomes arrive slowly. I don’t think there is a way for RLVR to generate a strong training signal for these kinds of problems in the general case.
2) The search manifold
RLVR post training usually does not incentivize search over the entire space of possible actions. In most cases, it actually does the opposite: it constrains the search manifold so that models output the reliable path more often. This is very useful when the goal is correctness through following some known procedure, but it is fundamentally limiting when the goal is discovery.
In domains where the verifier encodes current human assumptions, this methodology will tend to asymptote toward human consensus, because the loop is rewarded for following that consensus. This is the difference between using classical mechanics correctly and overthrowing it the way Einstein did, or learning to apply Magnus Carlsen’s go-to strats and surpassing him.
3) The model architecture itself
Current LLMs can simulate action through chains of thought, tool calls, and multi-step reasoning, but the underlying policy is still mostly fixed at inference time. The model can search within the distribution it has learned, but it does not naturally become a stateful learner in the AlphaZero sense: acting, updating, and improving its policy through interaction with an environment. I'm also skeptical about "continual learning" in the traditional sense for nonlocal LLMs, the signal to noise ratio of useful training signal is so minuscule during normal use that I don't think it's useful for backprop (I could be wrong).
The current RL techniques are mostly about increasing SNR when the correct trajectory is already somewhere in the model’s learned probability space, but they do not, by themselves, move the model toward the optimum when the optimal move is not already in that space.
I do think current methods can get us to something like the new definition of AGI: an agent that can do most economically valuable tasks. If they don’t, I think it’ll be more because of diminishing returns than technical incapability. But don’t expect this trajectory to produce the Nick Bostrom-style ASI that recursively self-improves into some unrecognizable kind of intelligence. I think it is probably really good that we didn't get that first, as you argued in your original tweet.
To attempt to put words to what I think the missing pieces could be:
1) Hard, canonical state
Agents need hard, verifiable state rather than the floaty context window state we have been using. We may need to make the agent’s world-state explicit enough that learning can happen over changes in state, not just over pass/fail outcomes.
2) A general learning kernel
We also need a general learning kernel that can widen the search manifold again rather than constrain it like current approaches to truly break free of "consensus" the way AlphaZero did. This is the kind of thing that is easy to point at but hard to actually characterize within the LLM paradigm, at least for me. I suspect this is the element that Altman and Demis Hassabis are referring to when they say we are still a few breakthroughs away from true recursive self-improvement.
Rob retweeted
We Need an International Treaty to Ban Superintelligence open.substack.com/pub/persua…
Rob retweeted
The bitter lesson in 26 words:
Don’t be distracted by human knowledge, as AI has been historically.
Instead focus on methods for creating knowledge that scale with computation, like search and learning.
Rob retweeted
That one neuron connects to about 7,000 others. Your brain has 86 billion of them. Do the math and you get somewhere around 100 trillion connections inside your head. More connections than stars in 1,500 galaxies.
And each connection point is way more complicated than anyone expected. A Stanford lab found that every single connection contains about 1,000 tiny switches that can store memories and process information at the same time. So your brain is running roughly 100 quadrillion switches right now, while you read this sentence.
The wild part is the power bill. Your brain runs on 20 watts. That’s less energy than the light in your fridge. The world’s fastest supercomputer needs 20 million watts to do the same amount of raw calculation. A million times more power for the same output.
We’re still nowhere close to understanding how any of this works. In October 2024, a team of hundreds of scientists finished mapping every single connection in a fruit fly’s brain. Took six years and heavy AI help. That fly brain had 140,000 neurons. Yours has 86 billion. Google and Harvard also mapped a piece of human brain last year, a speck smaller than a grain of rice. That speck alone contained 150 million connections and took 1,400 terabytes to store. The lead scientist said mapping a full human brain at that detail would produce as much data as the entire world generates in a year.
A tiny worm had its 302 brain cells mapped back in 1986. Almost 40 years later, scientists still can’t fully explain how that worm’s brain keeps it alive. Your brain has 86 billion of those cells, each one wired to thousands of others, each wire packed with a thousand switches, all of it humming along on less power than a lightbulb.
Anthropic is a good Rorschach test. The free market crowd is suddenly very comfortable with the state dictating their terms to private firms and expanding mass surveyance.
Turns out most "capitalists" are commies when their side runs the Leviathan.
Rob retweeted
It’s extremely good that Anthropic has not backed down, and it’s siginficant that OpenAI has taken a similar stance.
In the future, there will be much more challenging situations of this nature, and it will be critical for the relevant leaders to rise up to the occasion, for fierce competitors to put their differences aside. Good to see that happen today.
Rob retweeted
Video message from Federal Reserve Chair Jerome H. Powell: federalreserve.gov/newsevent…
Rob retweeted
LLMs represent the "library" phase of AI. The next phase will be the "scientist" phase. A library contains answers, but a scientist knows how to find answers that don't exist yet.
Rob retweeted
I do like how prediction markets have recreated insider trading from first principles.
A Google insider has officially been exposed on Polymarket.
This dude just profited $1,000,000 in a single day betting on the Google search markets.
Google accidentally pushed the results early, then removed them, but not before it revealed he went 22/23 on his bets and ballooned to $3.9M in open positions.
This isn’t a lucky streak. He previously made $150K+ predicting the early release of Gemini 3.0 before results were out.
At this point it’s obvious: He’s a Google insider milking Polymarket for quick money.
It’s one of the wildest things I’ve seen on the platform



























