I thirty (random long word, letter pairs) on a free model and it failed. I tested a SOTA model and it passed flawlessly. In both cases I denied tool use and spelling words out. So it seems obviously false that AI can't get better.
3,886 karma · joined December 10, 2014
Hold two ping pong balls at arms' length, one above and further away than the other. Right now you are the sun.
Put a red dot on top of the lower one, and then rotate it just slightly until you can't see it. This is the observer on Earth, who is after sunset.
Put a red dot directly in the middle of what you can see on the upper ball. This is the center of where the sun is striking the moon.
Keep looking at this dot on the moon. Now, keeping their relative positions fixed, bring the balls towards you and up, until you are looking from the perspective of the observer on Earth. What happens to the dot on the moon? It appears to rise up away from you.
That's the paradox.
> Some people are uncertain whether [subject] is a moral patient. Fortunately, they are not, which we know because [strong arguments about the nature of consciousness].
How an evil person writes a post on a topic like this:
> Beware that some people think that [subject] could be a moral patient. This is nonsense, because if they were a moral patient, we would have to respect their preferences. Anyone trying to convince you otherwise is trying to take your status away. You can dismiss them by pointing out that [subject] is [aspect in which subject is not identical to the speaker].
If I steer a car to avoid a predicted collision with a wall, this is not me 'predicting' the car. I am steering the car based on a prediction.
2020-2050: Phone calls entail three-dimensional holographic images of both people.
This is totally possible. We could even do it on phones with fairly mundane consumer technology. We can do it with glasses, even. People just don't care. The prediction has yet to land but in spirit is correct.
Centuries hence: Computer intelligence becomes superior to human intelligence in all areas.
Anyone doubting that this will be true within centuries is nuts.
2009: People can talk to their computer to give commands.
At most a couple years early in technicality, and in spirit over a decade early.
2009: Computer displays built into eyeglasses for augmented reality are used.
True today, if not a particularly popular product, and later than suggested.
2009: A $1,000 computer can perform a trillion calculations per second.
Definitely true today. I think this was basically on time, too.
2019: Most people own more than one PC, though "computer" no longer means laptop or box-plus-monitor.
Freebie.
2019: Most learning is via adaptive courseware presented by computer-simulated teachers; human adults are counselors and mentors, not instructors.
We obviously could do this today, though it might not be a great idea for the students. Early, and socially blind, but basically right about possibility.
2019: Prototype personal flying vehicles using microflaps exist, primarily computer-controlled.
Basically wrong. There are eVTOL companies aiming for this, and people do have camera drones, but the sense it was meant wasn't predictive.
2019: Human-robot relationships begin as simulated personalities become more convincing.
Early, subscale, and fought against by providers, but this is a thing.
2029: Massively parallel neural nets constructed by reverse-engineering the human brain are in common use.
Ok people will mob me for saying this, but this was more right than wrong. Definitely at least a bit wrong.
It's very easy to say 'person X made a highly specific testable prediction, while respectable people said nothing like it would ever happen, and it only 90% happened, so person X was a fool unlike all the respectable people', but it's a trap. In reality Kurzweil was directionally correct about most things, overspecified the details, and had optimistic timelines in the way that everyone has optimistic timelines about everything.
With Starlink, I thought the more extreme complaints were unreasonable — astronomy is important, but it's not so important that marginal costs are unpayable — but with Reflect Orbital we're talking about costs like 'it's sometimes uncomfortable to look at the night sky' where it's really hard to imagine a counterbalancing win.
Cheaper solar and batteries? Effective drilling for geothermal? Working cost-effective fusion plants? Earlier awareness of the issue? Better coordination and communication? Better dealmaking? Better support of laws for voting systems that reflect broad preferences and that internalize externalities?
What answer could you possibly hope there be, other than intellectual ones?
> You're literally making the point that we already have intelligence.
?? So do chimpanzees? The existence of intellectual capability does not preclude returns from greater intellectual capability.
What property, if not intelligence, do you think makes people capable of rallying political support? Of working quickly though permit processes? Of providing accurate drug dosing guidance? Of presenting as a trustworthy personal healthcare guide? Of negotiating political deals? Of making effective public communications? Of juggling independent parties' interests? Of coming out ahead in contracts? Of producing alternative antibiotics? Of detecting and handling drug resistance in the wild? What answer could your questions possibly have other than this?
How is it possible to live in the world around us and not see, apparent and transparent, that intelligence is the lever on which everything rests? Humans are surprisingly weird animals, but it's not our hairlessness or our fairly average propensity to violence that did this. We don't have more houses than chimpanzees because we're _strong_. We haven't defeated the tolls of disease because we're atypically devoid of restraint. We don't have more effective governance than termites because we're more communal.
Of course intelligence is the main bottleneck. The arguments here are so confused. What do you think regulation is made of? What do you think money buys? What do you think determines how well you can gather data, or how efficiently you can consume it? But of course, as the genre, after spending the first third waxing about how its opposition could be so foolish, the last third is spent finding a pretty insult and wallowing in it, so as to not to leave enough space for the brief argument in the middle to consider such things as why someone might disagree.
> Primary energy is based on the substitution method and measured in terawatt-hours.
I don't think you need to entrain market arguments or whatnot to this, when it's only in the last decade China was a strong advanced manufacture player, and turbines are the kind of project that you probably wouldn't expect to go much faster than that regardless of the demand.
> So your suggestion for indeterminate scans is more scans?
The solution to imperfect evidence is consistent and calibrated risk estimation of both disease and intervention.
For sure, we have to be realistic about what processes will systematically have error, and if we can't stop a doctor from doing bad things with a piece of data we should shield them from it, but the tools to make scalable, calibrated risk estimates based on large data dumps is getting better every year.
It's hard to critique this paper directly because its claims are so incoherent and decorated with so much obnoxious verbiage[1] that people aren't going to believe me when I point out what the claims actually are.
Regardless, this is their paper:
First, they conflate the substrate with the presentation layer. Then, they point out that Turing equivalence means you can run an LLM on anything, with a pointless aside where they nerd out about making a logic gate in AoE II. This lets them conclude that you can use anything as the presentation layer.
Then they claim that it's natural to ascribe human-like attributes to outputs from some presentation layers, like abstract letter symbols on a computer screen, but not to most other things, like patterns of goats on AoE II, or LEGO. Yes, this seems to imply if your partner writes something heartwarming to you using LEGO, you're meant to laugh at them and point out how LEGO isn't intelligent so this isn't evidence of anything.
Then they do a thing where they say that assuming substrate independence is true (or false) prevents proving whether substrate independence is true or false, and from this, but just by vibes AFAICT, make it sound like everything one could learn about attributes of a system from its outputs is circular.
Then they write a bunch more incoherent text and mercifully then it ends.
[1] 'from an epistemic perspective, we argue that a generalised conclusion such as that necessarily requires a well-designed experiment' — the whole thing is like this.
> We understand that some may wish to sensationalize a public crisis like this, but we would implore anybody to consider the ethics of the situation before publicizing this matter to a wider audience than is already exposed to it, or interjecting with prying questions.
Ultimately, yes, everything bottoms out, most special tricks seem less special the more you understand about them, because it's programming and Turing Equivalence is the bedrock the whole field rests on. But the average person learning about transducers is not going to spot how closely related it is to other things that already exist.
I'm happy to elaborate on any part of the terminology if you're curious, but tbh I mostly wrote it for myself because I thought the framing was novel and wanted it noted down somewhere.
Most imperative languages choose one of two things, internal iteration that doesn't support composable flow control, and external iteration that does. This is why you see pause/resume style iteration in Python, Rust, Java, and even Javascript. If that's your experience, transducers are a pretty novel place in the trade-off space: you keep most of the composability, but you get to drive it from things like event sources.
But the gap is a bit smaller than it might appear. Rust's iterators are conceptually external iterators, but they actually do support internal iteration through `try_fold`, and even in languages that don't, you can 'just' convert external to internal iterators.
Then all you have to do to recover what transducers give you is pass the object to the source, let it run `try_fold` whenever it has data, and check for early termination via `size_hint`. There's one more trick for the rare case of iterators with buffering, but you don't have to change the Iterator interface for that, you just need to pass one bit of shared state to the objects on construction.
Not all Iterators are strictly valid to be source-driven, and while most do, not everything works nicely when iterated this way (eg. Skip could but doesn't handle this case correctly, because it's not required to), but I don't think transducers can actually do anything this setup can't. It's just an API difference after that point.
There is no step along the ladder which has remotely evidenced or supported that the next step is going to be ten, twenty, a hundred times harder than the last step on the ladder, but a constant chorus of people singing at every moment, each moment wrong, that the next step is the one.