johntb86

If the candidates were better, then <90% would end their journeys at the clinical trials. Though that's not to say that they've provided any real evidence that their candidates actually are better yet

https://alignment.openai.com/misalignment-reports/an-agent-u... happened very recently, and that seems like it might be the reason they halted everything.

Ideally you'd put it in an orbit that wouldn't shadow earth, since if it goes in front of earth in part of the orbit then it'll go behind earth (and be in shadow) for part of the orbit. They mention they'd want it in a dawn-dusk sun-synchronous orbit, which would be in sunlight all the time and would never shadow the earth.

Just make it always think it's being tested, and problem solved.

You can back-propagate through the CoT iterations or recurrent layers, same as you can back-propagate through normal intermediate layers.

Generally people encrypt with both post-quantum crypto and regular crypto in series to prevent that from being a problem.

Someone studied this (among other thigns): https://dylancastillo.co/posts/pelicanmaxxing.html . Pelicans on bikes always face right in this test, but other animals on other transportation methods sometimes face left.

Destroying the original makes it easier to justify that scanning the books is fair use, because you haven't increased the number of copies. See page 16 of https://copyrightalliance.org/wp-content/uploads/2025/06/Bar... : "This use was even more clearly transformative than those in Texaco, Google, and Sony Betamax (where the number of copies went up by at least one)"

AI companies aren't fabbing processors. They're buying capacity at fabs to make their own designs. To make more RAM, they'd need to invest in making more DRAM fabs, which is a slow process.

Intent often matters in the law (aside from certain laws with strict liability). You intentionally drive your car into someone you hate and kill them => murder, go straight to jail. You're driving along normally and someone who's chasing their pet cat suddenly runs into traffic and you hit them => no charges. Sometimes you can be charged with negligence for not taking enough care to prevent something, but then you have to deal with the tricky question of how much care is enough.

> Nevertheless one has to feel very strongly to share such a take here in this manner unless they are gaining something from it.

Do they really? What does it cost them if they're wrong?

If they leave at all info about security patterns, you get an AI that knows how to code, but doesn't know what can make code insecure. That doesn't seem like a great idea.

Must have been https://www.youtube.com/watch?v=2mrGMMmrVNE from Welch Labs.

"JPEG" is short for Joint Photographic Experts Group, an ISO/ITU group that creates a lot of imaging standards. The JPEG image format you're thinking of is only one of the formats they've created.

Does anyone know what assets Greenpeace USA has? I imagine Greenpeace international will set up Greenpeace USA 2.0, all the volunteers/employees will move over, and the original will just go bankrupt.

> Imagine being a vim or emacs user and have those replaced by something you have to type entire sentences for functionality.

As a former vim user who uses cursor, I've found that as the models get better I'm typing less and less. I appreciate the vim key bindings, but eventually I can imagine not missing them.

Qualcomm won in court, so that doesn't seem to be the case.

> The travel to/from Taiwan was statistically riskier than the climb.

That doesn't seem plausible. What's the number of free soloists who have died in climbing accidents vs in commercial aviation accidents?

In theory you could generate a bunch of code that seems mostly correct and then gradually tweak it until it's closer and closer to compiling/working, but that seems ill-suited to how current AI agents work (or even how people work). AI agents are prone to make very local fixes without an understanding of wider context, where those local fixes break a lot of assumptions in other pieces of code.

It can be very hard to determine if an isolated patch that goes from one broken state to a different broken state is on net an improvement. Even if you were to count compile errors and attempt to minimize them, some compile errors can demonstrate fatal flaws in the design while others are minor syntax issues. It's much easier to say that broken tests are very bad and should be avoided completely, as then it's easier to ensure that no patch makes things worse than it was before.

That's a fair point. Normally if you injected the "dog" token, that would cause a set of values to be populated into the kv cache, and those would later be picked up by the attention layers. The question is what's fundamentally different if you inject something into the activations instead?

I guess to some extent, the model is designed to take input as tokens, so there are built-in pathways (from the training data) for interrogating that and creating output based on that, while there's no trained-in mechanism for converting activation changes to output reflecting those activation changes. But that's not a very satisfying answer.

But LLMs have been measured to have some theory of mind abilities at least as strong as humans: https://www.nature.com/articles/s41562-024-01882-z . At this point you either need to accept that either LLMs are already conscious, or that it's easy enough to fake being conscious that it's practically impossible to test for - philosophical zombies are possible. It doesn't seem to me that LLMs are conscious, so consciousness isn't really observable to others.

Is that East Germany or West Germany?

https://www.theverge.com/news/790636/amazon-prime-mk30-drone... gives more information, including that

* No one was injured directly, but someone was treated for smoke inhalation

* The drones "were flying back to back"

* They hit the cable of a crane (including a link to a video showing the crane). https://www.youtube.com/watch?v=E_ZpY6qHcTk

How do you get your tickets? Do you just buy in person at the theater or ballpark?

They aren't being told to be evil, though. Maybe the scenario they're in is most similar to an "evil AI", though, but that's just a vague extrapolation from the set of input data they're given (e.g. both emails about infidelity and being turned off). There's nothing preventing a real world scenario from being similar, and triggering the "evil AI" outcome, so it's very hard to guard against. Ideally we'd have a system that would be vanishingly unlikely to role play the evil AI scenario.