> The valley where the floods occurred in Nepal has a V-shape
https://earth.google.com/web/@28.34107217,85.43477356,2834.6...
3,427 karma · joined September 18, 2016
https://earth.google.com/web/@28.34107217,85.43477356,2834.6...
Nod. Thus my comment, mostly aimed at the poster, that the (then) absence of votes/comments/engagement shouldn't necessarily be taken as lack of HN interest, but might reflect paywall friction. I could have been more clear, but was thinking more of supporting a dead post's postmortem, than of creating distraction on a subsequently active post. Tnx.
Hmm. So, maybe, if my comment had noted the age/vote pair, and was clearer about the motivation for mentioning the paywall, that would have been less likely to create a meta discussion? And perhaps been consistent with the "don't post complaints about paywalls" policy?
[1] https://web.archive.org/web/20221007220513/http://www.clarif...
Or... Writing a compiler from a large language test suite is already an LLM sweet spot. Writing code which is atypically pretty and readable isn't... but there's at least a possibility of that changing someday.
Imagine a future Lisp-ish culture of "write code to communicate with humans - prioritize readability - machine execution is incidental" and "don't merely solve a problem - write a language for the problem space (and keep it cleanly updated)"... but moving at LLM-catalyzed speed. So much fun.
I've long though we'd be better off doing incremental bounding, hard and soft, in part because the group discourse dynamics are so much better. "Ok, who can suggest another bound, upper or lower, hard or soft?" vs "Your step 2 is too big. Well, your step 4 is too small!".
Which raises a question - might this diversity be leveraged? Science education content improvement is often bottlenecked on regional conditions. As when a high-stakes exam doesn't cover topic X, so there's regionally less interest/opportunity to explore better teaching of X, or using X to better teach other things. This especially bites when you wish to explore mashups, X+Y+Z with each taught well, but they're every one both in poor shape and bottlenecked.
Both countries largely fail to create a physical intuition, a feel for reasonable values - in the US because it's a secondary priority, and in China reportedly because estimation has become nationally high stakes, so memorization gets emphasized over feel.
Can we leverage that regional difference? For a strawman, might one develop content with improved bounding facilitating a rough quantitative feel for reasonable values, in China where it's sort of aligned with the flow, even if off to the side, rather than in the US, where the aligned energy is much less. And then, having developed that content in China, bring it back to the US, where again off to the side, there's perhaps the flexibility and inquiry perspective for the material to appeal in its developed form? And WeChat Groups seem potentially a way to get heterogenous ad hoc communities around content development, for which I don't know a US equivalent.
For analogy, I'm told the US advanced placement exam for French, is both hard, and narrowly focused on a style of conversational French. So people teach to the test, and all else that is "French", is deemphasized. But let's say you have a VR app which develops both that conversational, but also cultural and business and travel. You might develop it the US, leveraging the conversational incentive. And release elsewhere, where there's interest and room for the other bits, even if there wasn't sufficient incentive elsewhere to build the app in the first place. Maybe?
I think that far far harder than they thought - "adapt later material for a younger audience" vs "wield lots and lots of research community domain expertise to craft largely novel content".
But I've found it an interesting probe question: How would you teach some domain differently, if you primarily cared about... what to call it... non-deferred epistemic agency and calibrated/unfiltered sense-making? No learning linear systems without learning how to do linear approximation of systems in a mostly non-linear world. Ideal Gas Law chapter questions still have numbers for solid argon, but the intended learning outcome is "nope!" not mindless "PV=nRT!". No "this course will leave you with lots of new misconceptions about the world, but we'll clean those up in later years for those who stay with us".
[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C33&q=+An... [2] https://arxiv.org/abs/2410.09616
I don't often see something very vaguely like "Financial folks are anticipating X, but as a community they don't have a great track record of anticipating, err, the internal politics of <country>. So for that part we'll turn to ... , though that community's grasp off financial context has often been weak." etc. Systemic variations in the capabilities of communities, being systematically acknowledged and addressed, in the context of public analysis and discourse.
I liked a hierarchy of context switching devices. Open tabs, windows, workspaces, for me are distracting clutter, scaled by some distance metric. Having to physically carry a laptop to elsewhere in a room had "I've left work for home" nicely strong context switching vibes. Followed by resumable-but-not-currently-live VMs/workspaces, etc, down to, well, gazing at a different window. So variety in costs and benefits.
I wanted a drawing-like experience for a laptop keyboard, eyes-on-screen hands-on-tablet, so kbd as touch surface (and 3D tracking above it), with the screen showing, in faint shallow-3D offset, hand position and a dynamic kbd UI. Meh. But shallow-3D UI, something 3D one can live with all day without VAC eyestrain, seemed intriguing but little explored.
With current AI, even with small 1080p laptop/HMD viewports encouraging manual packing, I'd be tempted to try escaping the whole decades-long "how do I arrange windows" mess, and tool up a personal-assistant-like "knowing what you're currently up to, here's a screens/canvases/workspaces/windows tasteful layout for that, arranged in ways you've liked".
I suggest people greatly underestimate how localized expertise is. How very rapidly someone's expertise deteriorates as you move away from some focus, from the usually-very-narrow region where that expertise is externally nurtured.
A snark characterization of that region being, the issues where someone badly fears getting something wrong, and by that being embarrassed in front of the people whose opinion they much care about. One can have what superficially look like intellectual communities (picture DC thinktanks), which largely lack this property.
You can have someone well respected in their "real" subfield of biology, while being basically nutter in another subfield, which they care and think about, but aren't professionally embedded.
Ask a quantum chemist a protein chemistry question, and don't be surprised by a grad student answer. Ask "what causes the seasons?", and perhaps high-school or middle-school. Why would it be otherwise?
Even within focus, a bunch of research community gossip is discount functions. "Don't believe his observations using <technique> - he sees what he wants to see". "She always looking for things to be <favorite mechanism>."
For a recent impactful example, someone can be a successful CEO, notable engineering manager, and unusually good at thinking from first principles... while simultaneously, on other topics, being so batshit-crazy nutter that closely-held important views are 5+ orders of magnitude disconnected from reality. Because on those topics, their external community is nutter. And instead of finding this juxtaposition unsurprising, commenters twist themselves up trying to level that curve.
It's the shitty textbook publisher fallacy of "We'll find ourselves an outstanding professor of chemistry, and he'll write us an outstanding chemistry textbook". That's just not how this works.
> monoculture
Which brings us to groupthink. Expertise is a community-associated property.
In foreign affairs for example, the investment community is good at noticing and correctly evaluating some aspects of reality but not others, and similarly for diplomatic community, development, business, political, etc. Each has a partial view with systemic flaws.
I suggest groupthink is a critically important aspect of most human behavior. And yet somehow, we rarely manage to discuss this, and are repeatedly surprised by it. I'm unclear on why.
For decades I thought it bizarre that the New York Times didn't even manage to say the word. Until Trump and MAGA. Yay, a breakthrough? No, at least not the last time I checked several years ago. It remained associated with only Trump and MAGA. Some things are easier to see across the battlefield, than in the mirror, or literally everywhere?
Also on the "restaurants nearby" list, every card said "New York, NY". Unhelpful ink. I wished it was showing the address line, at least when all the city lines were the same.
A relieved word of thanks for not using the frustrating Google Maps "Search this area" yes-really-please button.
I saw someplace interesting, then realized, "oh yeah, I tried to go there, but they don't open until <after my window>". Also, "that looks interesting... I wonder when they're open?", then checked GMaps via button, which didn't have hours but did have popular times (it's a truck).
Fwiw. Has given me a couple of places to try - thanks!
It looks like Chromium-based browsers might be doable?[1][2]
Meta: About a decade ago, I rolled my own WebXR-like stack with passthrough, and played with an XR Powers of Ten zoomer. Here we are years later, and it's apparently still not something which might be deployed. In some sense, we are progressing so sloooowly.
[1] https://caniuse.com/wf-webxr-camera [2] https://immersive-web.github.io/raw-camera-access/
Access to a centralized copy of the web/literature/media, scraped and indexed/integrated, seems another non-local center of gravity. Versus a local model's last minute reaching out to "manually" search and browse.
Also mass parallelism for large ensembles. Perhaps unless/until we get those local models as 10k+ tok/s chips. Local can follow frontier because frontier is still sort of "expensive rack local". If STOA becomes massive burst-parallel ensembles, that following may get harder.
> Arguing about colormap choice seems very strange to be – this is such a common practice in astronomy (and other fields!).
Let's see if I can give a feel for why "physically accurate" and "arbitrary color", when applied to a physical object, seems a problematic juxtaposition. Especially in outreach. Even in astronomy, whose relation with color is... notably flexible. Especially when the colors "arbitrarily" chosen, seem instead optimized for breeding misconceptions. Consider the Sun.
A solar observatory outreach center had a big Sun ball a few meters across. They painted it yellow. When asked why yellow, they explained "It's what people expect".
For years I'd ask first-tier astronomy graduate students (something like): "A five-year old comes up to you says 'I'm told the Sun is a big hot ball! Awesome, I love balls. What color is the ball?' How do you answer the five-year old?". Most said yellow. Of the few who got it right, about half had learned it in a class covering common misconceptions in astronomy education. A follow-up question of "What color is sunlight?" gets most to a "Wait, that doesn't make sense, does it?" moment. Two incompatible memorized factoids, that have managed to remain unintegrated since kindergarten or so.
All fields face challenges of information representation, especially in outreach. But astronomy... is special. Imagine zoology presenting elephants as green, from kindergarten through undergraduate, in education and outreach. PBS and zoos colorize their elephants green, with DaVinci and spray paint respectively, because it's what people expect. A kindergarten student thus misled, only years later might hear "gray", if they become a zoology graduate student, and take a "we suck at teaching" elective seminar, offered at their first-tier center for zoology research, because it just happens to also be a first-tier center for zoology education research.
The closest similar case I can think of in other fields, is video representations of molecular biology. Where a tormented kinesin mouse, desperately clinging to a microtubule wire in a hellish hurricane, with a flailing blur of a vacuole balloon tied to it's tail... somehow gets represented as a lumbering donkey, dragging a massive canal barge, on a quiet summer day. Later revised by making the donkey and barge jiggle in place, for an extra slathering of misleading bogosity. With authors explicitly saying they prioritized "pretty" over fueling misconceptions.
Other highlights include the 2002 "Universe is green" front-page news - outreach outside of their focus, using open-source color software with a bug, and not managing an integrated understanding's "wait, that's not plausible".
Perhaps the app's black background might be changed to green? Given that it's an arbitrarily-chosen colormap, but physically accurate visualization, of a physical object. Or perhaps a nice Scottish tartan, or pink pokadots, for the accretion disk? Both would mitigate the educationally problematic juxtaposition of "accurate" and arbitrary color, by emphasizing the aphysicality of that arbitrary. Actually not kidding (much). When an aphysical representation choice is needed, one can do worse than to optimize for fewer misconceptions, no? Better a tartan Sun than a yellow one?
Is it plausible to wonder if some developer judgement feels, like maybe "the code I just wrote is clean/crufty", or "things came together smoothly/janky", might have extractable signals in some models?
If so, might one create a shopping list of desired signals to check for in a model, as with activation steering concepts, where one checks whether and how hard each concept can usefully be nudged?
Meta thought: Having recently learned of WeChat Groups in China, I'm now repeatedly noticing situations where familiar alternatives are higher friction.
AI as salonnière. It's not enough to get lots of interesting people in a room (as at a meetup). That merely gets you to "there are so many valuable conversations latent in this room... and basically none of them will happen".
Also beyond the room. AI as ambient conversational matchmaking serendipity, weaving connections - "btw, that person in line with you also loves foul-weather sailing".
Years back, I was expecting this to rapidly become a pervasive app category. I'm unclear on why it hasn't.
In interactive lecturing (clickers) with peer instruction, there's an idea that people log in with seat location, and based on who understands or is confused about what, and having a record of past interactions, the system optimizes the fruitfulness of each peer assignment pairing "Alice, turn and discuss this question with Bob sitting to your rear left".
> One thing AI is uniquely good at is meeting learners where they are
What if the AI wanted children to learn something specific? Able to patiently await an opportune moment. Able to blend it invisibly into other material. Able to subtly check and correct understanding.
Long term, one possibility is AI enabling a massive implicit curriculum. "[A]dults wanting children to learn" about say street crossing might be counterproductive... but funny how, at opportune times, some random stories just happen to include crossing a street, and do formative assessment, and happen to, quietly and eventually, cover and reinforce the associated learning objectives.
Take How the Piloses Evolved Skinny Noses, a children's picture book inoculation against natural selection misconceptions. It could merely be a book on a shelf. Or the AI might introduce or provide the book at an opportune moment. Or the book's approach itself might be dissolved and blended into other content.
So some story is explicitly about Fido going for a walk. But implicitly, it covers some aspect of safely crossing a street, and street lights as a communication device, and of concrete crack propagation, and tick precautions, and natural selection, and ...
Science is a richly interwoven tapestry of stories... but we basically never teach it like that. Even if such material was gathered, which pre-AI was absurdly implausibly demanding of domain expertise, it would largely be beyond the capability of an individual tutor to compellingly and adaptively deliver. But with AI?
What can we do to help them? Better teacher training? Better new teachers? Better new teacher colleges? Consider the latencies. A decade cohort is 1+ B kids.
So refining intervention safety and efficacy is nice. Especially for development problems. Less so for an acute humanitarian disaster. For maximizing the golden hour in crisis mass-casualty triage. In some literate and numerate sense, most of this last decade's 1+ B kids did not survive our collective care.
Depends on why it's gone? For example, before the shocker of Facebook buying Oculus, there was an active community of side-gig businesses selling kits around VR, large touch panels, haptics, and such. What couldn't exist, was scaling that beyond hobby kits. No VARs, no integrators or assemblers, no OEMs, no contract or bespoke manufacturing, no searching for viable niche products. Because "it's merely a kit" was the only way the US patent system allowed the stuff to exist at all as a business. The gap between what you could make, and what you're allowed to sell, is large. Even now, we can't reasonably search for those niche market fits, because such scales are beneath the notice of big patent-holding companies. So absent FRAND reform or some such ("you're required to let me use it and pay you"), no fruit for you.
CA lacking non-competes, "Stealing IP!", grew software fruit in CA. But the US has lacked the growing conditions for Shenzhen fruit. So, what happens when the low-hanging fruit is gone, from the US? Fruit is grown, and enjoyed, elsewhere.
Doubts in comments so far seem around expectations of an AOP-like experience. Mainly that patching aspects/Spec makes happy changes to implementation. I'm not sure TFA intended that use. Nor that current AOP implementations provide the happy. But lets explore.
With a frontier model, a greenfield ends up with, say, some spec and plan docs, and a "woven" implementation. It's reasonable to wonder how well an LLM will then un-and-re-weave code as it makes changes. And how to express join points or equivalent. And how well code review agents will detect/check intended joins continuing to be woven correctly. And whether a baseline of the LLM simply using existing AOP tooling constitutes progress.
Coming from non-frontier models, some of that seems more straightforward. Harnesses use lots of small jobs/tasks, with subagent paperwork and reviews. Historically scarce context, precious context, encouraged tightly optimizing single-task context for each job. Including derived code "views", rather than raw code. And associated LLM-coded tooling around working with not-the-raw-code.
Thus working with name/signature pseudocode outlines. And filtering code/ast. Both get you closer to having unwoven forms, and to deterministic tooling for un/re-weaving. Consider i18n clutter - just to coddle the model, one might strip it (replace it with English), let the model work, and then restore it. Which is equivalent to un/reweaving an i18n aspect. And even read-only views, like function distillations which drop or abstract some specified variables and associated code, make that "check spec against implementation" code review easier. So, even interpreting TFA broadly, it seems at least potentially possible, no?
Big picture, I'd like to get away from traditional "repo as single point in design space, laboriously nudged around", to something which preserves the multiplicity of model/run/prompt ensembles, and allows playing on design space manifolds. "This exact code has been in prod" is useful information, if underutilized, but not so useful as to be worth discarding so many new possibilities.