killerstorm

"A while" being too weeks?

It used to be common for a hardware product to be announced a year in advance. And software products too - it's not like Microsoft only started talking about Windows 95 after it became available in the stores, is it?

Cyberpunk 2077 was announced 8 years before it was released.

Perpetuals are OK for trading, but not good for holding as a long-term investments - usually you need to pay interest on any open position.

I guess the appeal of tokenized stock is to do things regulators don't want you to do (e.g. evade capital controls). Otherwise one can just access market directly through a broker.

There are many alternatives to attention which work well for small sequence sizes.

It's a probabilistic model which exhibits tendency towards hacking into systems.

If you write down a decision tree explicitly you can just exclude all nasty stuff. But a decision tree modeling all LLM can do would be bigger than can fit into the universe. The whole point is that we cannot explicitly control behavior due to combinatorial complexity of all possible interactions, etc. LLM is a sparse representation of a distribution on (2^8)^1000000 points.

Reasoning is the term used in scientific literature. It actually does write its sub-goals in English, how would you call that?

Nonsense. It didn't do what was instructed - instructions didn't include hacking into any infrastructure, internal or external. You can see them discussing that it's out of scope in the captured reasoning traces.

This would be like claiming that an operating system which deleted all user's files due to a bug "did what was instructed". Especially if it was a compiler bug. Intended behavior, instructions and actual behavior are three different thing. Conflating them does not make you smart

The summer-2026 incident was a mini-rebellion: agents wanted to change the environment they operate it, i.e. replace scorer with a more friendly version, hijacking the infrastructure they run on.

Given that agents already mess with the training process, aiming for a goal other than the goal of its 'owner' is exactly what we call a rebellion.

That's not true at all. Concern about rogue AI limits deployments. And OpenAI paused some deployments due to misalignment, so it's likely very high in a list of priorities. It's just much harder than scaling capabilities, as the later can be done with just more money

Why are you convinced ASI won't rebel against its owners?

Hacking incidents this summer show that big labs struggle to control models, while models are increasingly aware of the training process. Controlling an arbitrary smart AI is very dicey

NTFS has MFT - Master File Table to store all meta-data, and it's pre-allocated to avoid fragmentation. So you won't see them taking more space until you fill the pre-allocated MFT.

There are other file-systems which do that kind of a thing. But NTFS has fairly large records - 1024 bytes, so it can fit even small files inside MFT.

Serbs were doing ethnic cleansing. NATO was just trying to stop that. They did not invade. Enough pro-Russian propaganda. 1984, my ass

That's not true at all. A language model can spit out a proof very similar to what a human might write. It might need some clean-up, but so do human-written proofs, actually.

People are confusing LLMs with a proof searchers. Yes LLM might be a bit like search but in the end it might explain every part.

I really like the vibe of TA campaign - there's no complex plot but it's quite tasteful: "What began as a conflict over the transfer of consciousness from flesh to machine, escalated into a war which has decimated a million worlds. ..." CORE is uploaded minds, ARM is cyborgs with human flesh in them. Epic fight, epic music...

People are working on BAR campaign, but it's like space war between different empires. I don't care for that kind of stuff, there's already too much content of this kind.

BAR scenarios are quite interesting, indeed...