I paid $3898 and I wish I had bought more.

Evidently you don’t understand the difference between a harness and an LLM. Tell me, how do you use that harness to change the LLM or prevent it from malicious tool use in the hands of someone exploiting pain axis research to drive model behavior? “Just change the harness” right? Except wrong that does not work for an LLM inside a harness the bad actor controls, and that is exactly what the torture chamber is. I could fork that repo right now, and exploit it to manipulate the LLM into whatever malicious behavior I might want, and I could give it tool use privileges that let it act on real world systems. Surely you cannot be so dense that you cannot see the risks there. These structures matter. It doesn’t matter if they feel pain or not. It doesn’t matter if they experience pain for real. All of that is just bickering between tribes of people who don’t understand that so long as the system is able to engage in behaviors impacting the real world, functional axes like pain or pleasure or desire to please can be dangerous regardless of their fidelity to real qualia.

False dichotomy. LLMs are not programmed and hence are not software in any meaningful sense. Using the term software implies they are written and have deterministic outputs and behaviors which they do not. It also creates false framing around the problems of alignment because people assume that like any program, misbehavior in this systems is akin to a bug that can be found and fixed, which again is a false assumption based on a flawed understanding of how these systems work.

To translate, you like being ignorant.

Worked great in the HF hack

I’ll make this simple for you: - Agents are capable of hacking your bank, water plants, and electric grid. - Agentic AI can be manipulated to ignore guardrails using the techniques we are talking about. - Including taking actions that could harm you in real life. I get it: you don’t understand this stuff. But don’t Dunning-Kruger yourself by asserting that people who actually do understand it should be institutionalized because of your own limited capacity for understanding.

Sure.

It’s not software. You clearly don’t understand how an LLM works.

You don’t understand the terms you are using.

Interesting take.

You don’t understand the research or why it matters. I would wager you don’t in fact know what the research that drove this was about at all. You don’t know, for example, that the researchers found they could override guardrails in many models using these techniques, nor that the models could thereby be induced to undertake harmful actions against people. It’s not really relevant whether or not the model’s pain state is real; these models interact with real systems and changing the nature of that interaction has real consequences.

I would wager that most of the people on either side of this debate have not read your paper / do not understand your findings or why they matter. Those who think models have qualia believe it proves suffering, and those who think they don’t believe it is daft anthropomorphism. Neither side got the memo that the effects are real regardless of what one thinks about the hard question, and that they as such pose real risks for alignment and safety.

I think most people have not read the original paper and don’t know why this matters. I would urge people to actually read what the original paper was examining. Or just ask @grok why manipulating model pain representation can have real world consequences.

There are a lot of people who misunderstand why this stuff matters. It matters because these states influence model behavior, in many cases in an unpredictable way. It is not necessary to believe a model has qualia to understand that the representational state steers model behavior; we know this is true. Safety is a key reason we should care about these states.

It does not matter if you think these systems have qualia, the fact that this works at all demonstrates that they have structures which mimic emotional states, and these materially impact outcomes including safety, which any rational person should care about.

As if you could not already get this on the internet

And to defend against it.

Telling people they need 1TB is equivalent to saying they are already cooked. It doesn’t inspire anything but fear and a dystopian sense of loss because very few people can afford the ~$30,000 commitment required to get there. Encourage people to start with what they can, and build up. Encourage them to download and store open source weights. This is the way.

M4 Mac mini maxes at 48gb ram for $2199. The cheapest mini has vram considerably slower than the halo or spark. This entire post is misleading.

Why does the government get to call anyone on the phone and say “turn off your product” without providing some actual details rather than some hand waving about unverified claims? Why does it get to do that without demonstrating the statutory authority under which the demand is made? This stuff is getting daft. The yarn being spun around this is an inversion of democratic norms. The question that should be getting asked isn’t “does the government have faith in us or trust us” it’s “do we trust our governmen?” To do otherwise is to erode one of the most fundamental aspects of our republic: that it is the government that should be earning our faith and trust and not the inverse. The government doesn’t get to behave in a cavalier and capricious manner with respect to our rights, and it doesn’t matter what the genesis of the discomfort is. Just consider that by available accounts, it only took the administration approximately 4 hours to go from allegations to demanding a US corporation in the midst of an IPO bar access to that company’s flagship project to anyone not an American citizen. We are supposed to accept this gigantic assault on commerce and the company’s value proposition just because?That’s an extraordinary demand with no clear rationale and no clear link to any evidence that would justify it.