Arguably the introduction of RL a few years ago is already that better paradigm. RL training trains AI to meet goals, not to match the text of training data as the classic pre-training step does.
Whatever advancements we make are probably going to continue to be glued to the existing LLM process.
It's been commonly reported as being over AI but it had less to do with AI inherently and more to do with that the maintainer regularly made changes without ever discussing with the much more active contributors, including reverting some people's work and putting in untested broken AI code once.
sad that this image is already deeply aged by sam's more recent actions
It's crazy to portray the recent Anthropic events as having anything to do with LLM welfare concerns
aside from the AI angle, what makes the Anthropic case important is that it is -- unless there is something I am forgetting -- the first time a major tech company has stood up to Trump instead of accommodating him
For RimWorld, there's RIMAPI (an HTTP REST API mod), which I've seen multiple discussions about people connecting LLMs to. I've been meaning to try that out.
tbh the Simulators post was frequently misunderstood by people so I'm pretty happy at seeing a new write-up of the idea that's easier to refer people to
It's a little annoying how the comparison sets up a lot of unnecessary arguments about the relative value of AI and humans.
The easier argument is to point out the energy usage of training divided by the number of users. The energy use of microwaves and home computers look bad compared to that.
and all of humanity in 2 weeks uses more energy than a single person's use of AI. really makes you think
-okay I don't really expect that to happen but I just want to keep everyone's idea of the spectrum of slow to fast takeoff wide. I saw someone say we're in fast takeoff the other day and it peeves me deeply
There won't be any Dyson spheres because they're a caveman's idea of how to use the Sun. A superintelligence with a true understanding of reality and why anything exists will embody a perspective where physics causes the Sun to spontaneously reorganize into computronium, siphoning measure from our-
I was won over from realizing that you have to translate your expectations into numbers if you want to bet / participate in a prediction market on them, and that any belief you're not willing in principle to bet on (assuming a perfect judge / resolution criteria) isn't a real belief.
- the parts of the hamster responsible for consciousness go untouched in this process, so the modified hamster is still conscious (and if the original LLM is computationally conscious, it still is, separately)
- the process of adapting an LLM to work in a hamster brain necessitates changes that make it conscious. Running the adapted version on a computer would be equally as conscious as running it in a hamster.
It's fun to imagine how exactly this impacts the question of consciousness in LLMs. Some possibilities:
- the modified hamster is now just equally conscious or not as the original LLM imprinted on it.
theories of consciousness that deeply prize continuity are overrated imo. There are practical benefits to certain forms of continuity of course (lack of time gaps in awareness and ability to react, consistency in behavior), but any theory of mind that doesn't handle uploading and forking is flawed.
I thought these takes would've had that effect here when I first saw them but they landed without too much fanfare:
- "stop shaming people for not using alt text on images, as people can already use AI to help understand images"
- "please use alt text on images, it helps people *and* helps train AI"
not bad. i think our "should swap" logic is wrong somehow. it feels like it "cares" about where you start dragging relative to the dragged thing. but it shouldn't. it sohuld only care when center of one thing cross center of another thing. ⏺ Right. Currently it compares the cursor position against the target's midpoint. But the cursor could be at the top or bottom edge of the dragged element depending on where you grabbed. Need to compare the dragged element's center against the target's center.
ngl. to me it's a miracle that i can talk to a computer like this, and it knows what i'm talking about and does the right fix. many engineers would struggle with this (including myself)
this is like the time I discovered that watermelon doesn't make other people's tongues itchy. I think these product designers have something different about them
started noticing that it's common for music apps with sleep timers to have the options max out at 1 hour, and it has me questioning why multiple companies would pick a max time that is so obviously lower than the usual time it takes to fall asleep
If abiogenesis happened any later, we wouldn't be here, so its earliness doesn't tell us much. We came into existence scarily close to the end of the Earth's habitability: the sun's expansion will make it much less inhabitable in a few hundred million years.
x.com/anderssandbe...
anyone looking to replace their use of Discord with something more private shouldn't just jump ship to another proprietary system that's run the same way, and should instead look for something open source, end-to-end encrypted, and distributed/federated, like Element(Matrix).
Error correction is much easier for intentionally-designed systems than natural life since the evolution of natural life directly depended the ability to have mutations but we can design things to check their work with checksums etc. So grey goo that doesn't evolve or drift in purpose is conceivable
"large language model the size of a small language model" might be the ideal compromise
It also makes no sense when you consider how Waymo's driving safety is much better than regular drivers. Unless you think the Philippines universally has much better drivers than exist in the US.
I bet my manifold market points on Sonnet 5 coming out this week, not this. I'm afraid they need to put this back and do the right thing
for years I considered my Bluesky usage as something I pushed myself to do to hedge against Twitter, but in the last month or two it finally passed some threshold of activity and interestingness. I forget to check Twitter as much now and when I think to check it I'm wondering more often why I bother
It's some kind of email format conversion error
In real life, people can easily visibly self-organize into groups for their combination of opinions/interests. In social media, we're grouped by algorithmic forces and it's not easily apparent how a sub-group can walk away, so instead we get more aggressive behavior policing to try to shape groups.
though if they made the bot woke and the first time it said "whoops" was when *it* happens then I would finally be convinced we're on a good timeline again
