HNHacker News
TopNewBestAskShowJobs

RogerL

3,169 karma · joined June 9, 2013

submissionscomments
RogerL··on Getty Images and Shutterstock to Merge
And that is true for the foreseeable future, which unfortunately, if my math is correct, is around 17 hours.

In 2 (time units) we'll be doing computer analysis of lens distortion or something to try to suss out the AI. At which point it won't matter for the stock image use case, of course it matters for legal matters and such. And then in 1-2 more units we're going to need public/private key signing implemented in 'cameras of record', because detection will be practically, if not actually impossible.

Is that 'unit' days or years? Dunno, but I bet it is a lot closer to the former.

RogerL··on Getty Images and Shutterstock to Merge
It's like a person in the 60s saying "computers will eat everything". Were they in every device at the time, in every home, on your wrist, in your pocket? No, capability wasn't quite there yet, but it would be. Today we get uncanny valley, but I mean today, as in there are images coming out that are far better than uncanny valley, and it is obvious to all that this is (at least very very likely)a surmountable problem in the very near future. I think if you can put what you want the photo to be in words, stock photos are largely (not entire) dead.

"give me an elephant under a tree" "make it later in the afternoon". "not that late". "emphasize eyes just a scooch and make it look sad and pensive". "Not quite that much". "Clouds could be a bit whispier" Like you'd talk to a photographer, but with instant updates and no retorts like "great, are you going to pay me to camp out for days waiting for the clouds to move in the sky and then somehow hoping the elephant revisits this tree?"

Beats scouring a huge catalog (which, sure, will have AI powered search, but still), and suddenly, it isn't stock anymore, it is very particular to your specific needs. Custom to your needs, faster than getting a stock photo, and so, so much cheaper.

RogerL··on I am rich and have no idea what to do
> having wealth is a social responsibility in itself

Warren Buffet has said that every dollar he owns is an IOU that he owes to society - somebody worked hard, produced value, he is in custody of it for the time being, but it has to be paid back by him.

RogerL··on Do AI companies work?
It was "cooked up" by people like Joshua Maynez, Shashi Narayan, et al, who I'm going to guess understand what is going on, and adopted by the rest of the field.

https://aclanthology.org/2020.acl-main.173/

RogerL··on A high-performance, zero-overhead, extensible Python compiler using LLVM
return [key, value]
RogerL··on NASA investigation finds Boeing hindering Americans' return to moon
Aerospace welding is a different beast. Vast/tedious documentation requirements, experience with dissimilar metals, welding gold (think about the cost of splatter with that one), extremely tolerance requirements, etc. Some welding jobs are done under a microscope. You don't fix welds, you through everything away and try again. Very experienced welders report a long and steep learning curve. Hiring somebody from a different industry means investing significantly in their skill set, and waiting a decade for their skills to ramp up. There is also a significant tech component, in that a lot is being automated, where the welder is setting up the robot, inspecting results, and so on, and only doing manual welds where the robot cannot achieve what is needed.

Meanwhile bog standard pipe welding often pays better and the jobs are easier to get with far fewer requirements.

RogerL··on What are the Olympics shooting competitors wearing on their faces?
I know you must know this, but for others reading, this also is not a static view, like in a video game, but the sights are 'swimming' around usually in a figure 8 pattern as your heart beats. It is not just a matter of getting a good sight picture, but breaking the trigger when things aren't perfectly lined up, because by the time the mechanism fires and the projectile leaves the barrel you will no longer be lined up. This is worse for air pistol due to the slow velocity of the pellet.

In other words it is not like you are holding a tiny dot in the exact center of a circle, it is a messy blur of trying to hold 3 things in alignment where the front sight is about as big as the entire bullseye (hence why you aim below, so there is white visible, you would otherwise obscure the target entirely.

If you do it well, you are consistently hitting something the size of the period at the end of this sentence from 10 meters away (again, air pistol).

RogerL··on Things I said as a manager part 2: Hiring is emotional
This is the key takeway, not hey you know what, I'm evaluating this person or poor me my job is hard. Unless this is for astronaut training for a manned trip to Mars to build a colony that time line is insane.
RogerL··on We need visual programming. No, not like that
it's pretty basic topology - embedding versus immersion. you cannot embed anything but the simplest software in a 2d plane. you end up having to endlessly try to structure things to minimize line crossings, make multiple diagrams of the same code to capture different aspects of it, or otherwise perform perversions just so it fits on a page.

And I live through this. most of my early career was DoD related in the early '90s, when things like functional analysis was all the rage. endless pages of circles and lines which were more confusing than helpful, and certainly didn't actually capture what the software had to do. Bever again.

RogerL··on I have no constructor, and I must initialize
And it is not true (for any reasonable reading of the quote). There are very popular languages that don't get the deserved hate that C++ does. Sure, Python is slow, packaging/versioning is painful, but it is nothing like C++ complaints.

I mean, a standard (and stupid IMO) interview question is rate your C++ expertise from 1-10, and if you answer more than about 5-6 you get bounced for lying or not recognizing your limitations, while they gleefully point out Stroustrup wouldn't answer 9-10.

RogerL··on Mouth-based touchpad enables people living with paralysis to use computers
Lift your foot up and try to balance in your seat and see how it feels. Try using a standing desk with it. Put a pen between your toes and try to draw a square and feel major muscles in your upper thighs struggling to make these fine control motions. Put on high heels and try to do anything that isn't gross movement. Sit in a train/plane/car and think of trying to use a laptop with a foot controller. Heck, put your mouse on the ground right now, turn pointer speed down as slow as possible to accommodate the gross movements of your legs, and try to use it.
RogerL··on Mouth-based touchpad enables people living with paralysis to use computers
I went to grad school with a guy w/o arms. Sure, he could write with his feet. It was extraordinarily hard and difficult. Nothing changes the physical facts - we neither have the nerves or fine control in our legs/feet that are present in our hands.

We also build fine motor control in very young childhood; children can't color within the lines (for example) because the nerves just aren't there yet. I would expect someone born without arms to have better fine motor control of their legs than an adult that tries to develop the ability. And it takes them years - 7 or 8 years to get pretty decent control. Even if I had the neuroplasticity of a young child (I surely don't) I wouldn't want to spend years trying to develop that control.

I face this every day on the piano. You distribute motion between fingers, wrist, forearm, elbow, and shoulder based on the amount of fine control needed. Sure, you can use your shoulder to trill, but it is always going to be very slow and clumsy. finges are best, but you face fatigue, so you usually use wrist rotation to get muscles with more endurance suppling the gross motions, and then fingers for the fine control of dynamics. You can't change this physical reality, at the most you can compensate (e.g. someone with a fused wrist would have to find alternatives).

RogerL··on Microsoft's official Minesweeper app has ads, pay-to-win, and is hundreds of MBs
Net profit margin has been going up since 2015 or so, release year of Windows 10. Gross profit has also been increasing YoY over the same time period. I know it isn't quite that simple - 2/3 of the business is cloud and productivity (Office, etc), and cloud is relatively new, adding a lot of growth, but there is nothing that forces them to eeek out nickels from 13yo via minesweeper ads.
RogerL··on AI in software engineering at Google: Progress and the path ahead
I look at is as asking an intern to do some work that I don't have time for. Do I have to check their work? Yes. Might I have to correct and guide the outcome? Again, yes. Am I going to ask them to implement something novel and groundbreaking? Not really, that'd be a disaster unless they are a prodigy. None of that removes my capacity, or any kind of danger.
RogerL··on Anyone got a contact at OpenAI. They have a spider problem
Is it (I am not a worker in this space, so genuine question)?

My thoughts - I teach myself all the time. Self reflection with a loss function can lead to better results. Why can't the LLMs do the same (I grasp that they may not be programmed that way currently)? Top engines already do it with chess, go, etc. They exceed human abilities without human gameplay. To me that seems like the obvious and perhaps only route to general intelligence.

We as humans can recognize botnets. Why wouldn't the LLM? Sort of in a hierarchal boost - learn the language, learn about bots and botnets (by reading things like this discussion), learn to identify them, learn that their content doesn't help the loss function much, etc. I mean sure, if the main input is "as a language model I cannot..." and that is treated as 'gospel' that would lead to a poor LLM, but i don't think that is the future. LLMs are interacting with humans - how many times do they have to re-ask a question - that should be part of the learning/loss function. How often do they copy the text into their clipboard (weak evidence that the reply was good)? do you see that text in the wild, showing it was used? If so, in what context "Witness this horrible output of chatGPT: <blah>" should result in lower scores and suppression of that kind of thing.

I dream of the day where I have a local LLM (ie individualized, I don't care where the hardware is) as a filter on my internet. Never see a botnet again, or a stack overflow q/a that is just "this has already been answered" (just show me where it was answered), rewrite things to fix grammar, etc. We already have that with automatic translation of languages in your browser, but now we have the tools for something more intelligent than that. That sort of thing. Of course there will be an arms race, but in one sense who cares. If a bot is entirely indistinguishable from a person, is that a difference that matters? I can think of scenerios where the answer is an emphatic YES, but overall it seems like a net improvement.

RogerL··on Why are most sofas so bad?
Survivorship bias. The cheap stuff doesn't last. "Take from the dresser of deal, Lacking the three glass knobs, that sheet..." - Wallace Stevens, 1922. "Deal" being cheap pine or other soft heartwood (in the poem he is trying to evoke a scene of poverty and maybe a bit of gaudiness). Not that buying old can't be an excellent strategy, it having survived either a long life, or perhaps it was just never used in which case all bets are off.
RogerL··on LilyPond: Music notation for everyone
LilyPond rules when you want to manipulate music algorithmically. Like make a music reading training app that creates examples based on your performance so far.
RogerL··on Beyond A*: Better Planning with Transformers
It is in the abstract, very first line. "While Transformers have enabled tremendous progress in various application settings, such architectures still lag behind traditional symbolic planners for solving complex decision making tasks. In this work, we demonstrate how to train Transformers to solve complex planning tasks ..."

This paper is interesting (to me) as an example of how to use transformers for decision making. I don't particularly care if it is up to A* standards at the moment.

RogerL··on The killer app of Gemini Pro 1.5 is using video as an input
Don't we have that? My browser offers to translate pages that aren't in English, youtube creates auto generated closed captions, which you can then have it translate to English (or whatever), we have text to speech models for the major languages if you want to hear it verbally (I have no idea if the youtube CC are accessible via an api, but it is certainly something google could do if they wanted to).

I'll probably get pushback on the quality of things like auto-generated subtitles, but I did the above to watch and understand a long interview I was interested in but don't possess skill in the language they were using. That was to turn the content into something I already know, but I could do the reverse and turn English content into French or whatever I'm trying to learn.

RogerL··on How to hire low experience, high potential people
If you ask for a story, and the person needs/wants the job, that story is not likely to be accurate. You get some version of "my greatest weakness is I just work too darn hard and give to the company", pitched to whatever the question actually is. Sure, some people won't, but how do you know? Proxies and stories are usually a terrible way to interview somebody.
RogerL··on How to hire low experience, high potential people
Oh, but you just rely on your 'intuition' when interviewing. I'm sure that'll lead to no bias.
RogerL··on Ask HN: Why aren't all heaters computers?
That device&company seems like a joke. Costs around $1000, multiyear payback (~3years) if you use it every day, you only get 400w from the computer(the rest just from a regular heater also built in), and the payback is "up to" $28, which I would assume would be the payback for 24/7 use.

At my rate of using space heaters, I'd expect probably a 10-20 year payback period (pay schedule is not clear, so it is a guess, and generous). Drop $1000 on a computer I don't use, hoping that the company stays in business for decades? No thanks. I'll run a pihole or a server or something and at least get my own compute out of it if I want heat from compute.

But a great answer to the question. Computers are around $1000 for anything approaching heavy compute, you are probably running around a 400 to 750 watt power supply, so you get a few hundred watts of heat out it (depending on load, obviously). Vs the $10 or so you'd pay for that much space heater (you get 1500 for ~20-30). Loooonnnng payback period.

RogerL··on Ask HN: Why aren't all heaters computers?
Asian countries are already struggling to keep up with energy demands of semiconductor factories.

And then there are the other environmental impacts - waste chemicals, waster water, and greenhouse emissions.

RogerL··on Three American climbers solve the 'last great problem in the Himalayas'
I replied separately, but this is being written about because up to now, every last elite climber said 'nope, I'll 100% die on this one'. Well, perhaps there were aborted attempts (I'm not aware of any), but you abort when it becomes impossible, ie death is looming.

People can quote stats for Himalayan climbs and the such, but a climb like this lacks meaningful statistics. It is so far from a trudge up Everest, and unequalled, that I don't see how to assign a stat. Right now it stands at 0%, which is probably not the right number.

RogerL··on Three American climbers solve the 'last great problem in the Himalayas'
It's 100%, for everyone. This is being written about because until now, despite a community of hyper competitive individuals vying for first ascents in extreme environments, no one has done this route until now; it was unimaginable, basically, as something you could possibly survive. But these mad lads had the imagination to see it was possible, and the skill sets to achieve it.

As for 'dice rolls', it depends. In climbing there is "objective danger" - danger you cannot really control, like your route passes under a serac that collapses frequently, and "subjective danger" that you can control - you step on a loose hold and fall. They spend months planning, trying to reduce the objective dangers as much as possible, and of course years of training to reduce the subjective dangers. Summarizing all of that as a 'dice roll', while it might capture the percentages of particular high risk climbs, seems to miss the mark a bit. Spend a night in a climbing hut, or hanging in a tent/portaledge off a mountain side and it's endless talk about this sort of thing. How to master your body, the terrain, and often most importantly and hardest, your mind.

RogerL··on The Seamless Communication models
So what? Okay, in the case of a ramp, if you need one you probably are going to have difficulty building one. So pay employee Sally to build it instead, absolutely.

But hearing loss does not impair standing up servers and software. They can pay the employee who probably is the expert at this, the guy with the hearing loss, or go task Emil to go do it to ... avoid 'appearances'?

RogerL··on Magic means large search spaces (2005)
Those weren't around in 2005, the time that was written (which is not me defending the claim about OpenBSD in 2005, I have no idea).
RogerL··on Show HN: Learn piano without sheet music
I want to expand on this. Not only do I agree with what you wrote, but this app is trading access to millions of available sources in a very well known writing system to one more or less unique to this app. That's a terrible investment for anyone who wants to do more than learn 1-2 favored songs.

If you need to sight read (and as rock/pop/jazz people point out, you don't have to for many genres), then you need to sight read.

There are so many other virtues to sheet music. Look at the cover image. I can see a few notes. I can see vastly more notes in sheet music. I can easily evaluate if the piece is playable, I can scan and look for broader patterns. I can see that a bass note is being held for 8 measures (and I may choose to repeat it at some point). I can look ahead quite a bit. I can understand the repeat structures - don't gasp, but you don't have to take repeats, or you can repeat more times than written, especially with 20th+ century music, where you are often expected to do things like choose your own ordering of measures or blocks of measures. There are fingerings. I can see if the composer is writing out finger pedaling explicitly (Couperin normally does, Bach normally doesn't). I can see the pedal markings, general contours of dynamics. I can see the trills, etc., which are often just suggestions rather than hard requirements. I can see the meter, meter changes, keys, key changes, accidentals. I can see a big scary chord coming up and spend a bit more time looking at it while I play a few measures behind. I can see that Bach is repeating a phrase a 5th down, or inverting it, or reversing it. I can see the difference between passages meant to be played in time, and fioritura type writing.

I haven't used these piano roll systems so there are undoubtedly some things that are nice about it for an experienced player that I don't know about, so that paragraph is one sided. But that one side is very important - I'd loathe to go without them, and can't imagine I'd ever trade them for whatever advantage the piano roll might bring. After all, a player can take a sheet of paper Chopin wrote, produce that music at a more or less performance level. So it gives you about everything you need. I could imagine a current composer might find something more expressive about the piano roll (maybe expressing note durations not evenly subdivided by 2 or 3).

I suspect there is something neurological happening that stops some people from sight reading well, just like some people struggle with text. I've read accounts of people trying for years, with seemingly good practice techniques, still struggling.

So things like this, synthasia, etc., seem to have a niche. But in general, I suggest, think about someone proposing an app that instead of displaying printed text output it sonically. Great boon for certain situations or people! Undoubtedly someone is using one to read this very post. But a terrible replacement for reading in general.

If a six year old was relying on screen readers because reading is too hard to learn, after testing for dyslexia and vision problems, you'd urge them to make the effort; the advantages of reading text vastly outweighs the 1st grade difficulties of learning to read (yes, that time span will differ by language and writing system, not the point). Literacy is empowering, and arguing that the auto mechanic down the street can't read yet makes a good living is probably not a convincing argument to not teach a child to read.

I learned to sight read at age 4-5 with a plastic brain (I recall my mother having to teach me the letters a-g, and how to write them, for example), so I may underestimate the difficulties of learning later in life. But if you are in a situation where some kind of notation is helpful (again, not all are), learn standard notation!

edit: I thought of a counter-example. Say you play in a band. You can record your output to midi, and then share it with others. You can quantize midi and turn it into sheet music, but chances are you playing is not rhythmically exact. Sight reading that sort of thing is painful (notes carry 1/16th note into the next measure, that sort of thing), and I imagine a piano roll would often be easier.

RogerL··on Show HN: Learn piano without sheet music
Until you reach the B string. A shape that works on strings 4/5/6 does not work on 1/2/3. This is because G->B is 4 half steps, the rest of the strings are 5 half steps apart. So you need 2 forms for closed chords.
RogerL··on A trick to eliminate 2π (sometimes)
But what is the answer? What should I do differently? I get don't use images, but then what? I can't imagine all screen readers have the same capabilities, or that there is a base common ability, so what should we do?

Googling says MathML is the answer (e.g. https://www.washington.edu/doit/how-do-i-create-online-math-... this site uses MathML and your reader isn't handling it. So now what? (alt-tags? something else?)

← PreviousPage 2 of 30Next →