HNHacker News
TopNewBestAskShowJobs

mrtranscendence

1,366 karma · joined August 5, 2020

submissionscomments
mrtranscendence··on Deep Learning’s Diminishing Returns (2021)
The amount of computation in the world devoted to deep learning, in aggregate, might be a small slice of total computation. But in terms of costs borne by individual organizations who train such models, deep learning could be extremely substantial. Particularly if it takes millions in additional investment to achieve only marginal reductions in error.
mrtranscendence··on The note Reddit sent to moderators threatening them if they don’t reopen
I guess I'll try Lemmy again, but I'm not hopeful. When I used it last week any sub-community I was interested in either didn't exist or was completely dead.
mrtranscendence··on The note Reddit sent to moderators threatening them if they don’t reopen
I tried to Google how many users Reddit has. The numbers are all over the place. Various links on the front page say:

1.66 billion monthly users

430 million monthly users

50 million monthly users

330 million monthly users

Who the hell knows.

mrtranscendence··on The note Reddit sent to moderators threatening them if they don’t reopen
> the mods of big subs were one of the main reasons it was so terrible

I've heard this, and was just reading a thread on Reddit where many users told of their mistreatment and unjustified banning. But anecdotally I've never actually noticed mod misbehavior, and I've been a Reddit user for like 17 years. Could be the subs I read ... most of which aren't extremely huge, I guess. Still, I feel like I should have been banned from somewhere at some point if capricious, power-hungry mods really were as rampant as they're made out to be.

mrtranscendence··on The note Reddit sent to moderators threatening them if they don’t reopen
I think Reddit faces more potential downside here than other similar platforms might, because of the popularity of third-party apps among key contingents of the site (particularly moderators), and also because of the missteps that management has taken dealing with the API changes. If they lose even 5% of their traffic to this debacle they've made a bad decision, IMO.
mrtranscendence··on The note Reddit sent to moderators threatening them if they don’t reopen
They have the right to reverse deletions, but it's a supremely bad look and if they're really doing it (this is the first I've heard of it) I hope they catch a huge amount of flack. There could be very good reasons for some of those deletions that might impact safety, job security, or personal relationships.
mrtranscendence··on The note Reddit sent to moderators threatening them if they don’t reopen
Aren't those methods potentially expensive? And if you did make a hastily-built frontend backed by scraping, couldn't Reddit continually break it without much effort?
mrtranscendence··on The note Reddit sent to moderators threatening them if they don’t reopen
There are distinct questions here: whether Reddit has the right to forcibly reopen subreddits, whether it’s justified in doing so, and whether it’s a good idea at all. The first isn’t in dispute, but the other two are open for debate.

It could all go super well and everyone forgets about this shakeup, or it could engender further animosity and chase people (particularly trend-setting power users) to some other platform. It could also ruin some communities if the wrong new mods are chosen.

I’m not sure how likely this is but it is plausible.

mrtranscendence··on Vercel's AI Accelerator
> If you want to summarize a document and it's less than 16,000 tokens, you pipe it into gpt-3.5-turbo-16k.

If it could fit within the context length then you wouldn’t need to search it to find relevant portions at all. I’m talking about situations where you want to ask questions about a very long document, using the search approach, and one of those questions might be “please summarize chapters 2 through 8”.

The search method will never accommodate that, but a model fine tuned on your data might plausibly.

mrtranscendence··on Twitter is being evicted from its Boulder office over unpaid rent
I know absolutely nothing about these agreements, but the idea seems at least plausible. An unoccupied building doesn’t get problems noticed or fixed, and could result in squatting.
mrtranscendence··on Vercel's AI Accelerator
It’s much easier to run searches and insert snippets from high-ranking results into the query. There are limitations to this approach, though. For example, how would you summarize a document this way? You’ll get snippets but not the whole picture.

I’ve also found that GPT-4 tends to hallucinate when I ask a question but the search process doesn’t result in a good match. Like making up an author or title for a document I’m asking about.

mrtranscendence··on Show HN: Non.io, a Reddit-like platform Ive been working on for the last 4 years
If it’s true that a fair portion of those who search Google specify “reddit.com” at least sometimes, as seems likely, I’m comfortable that users looking for “elucidation and discovery” are at least somewhat relevant to Reddit or those who wish to compete with Reddit.

I’m not saying this needs to be a clone of Reddit; I don’t think I implied that, or intended to at least.

mrtranscendence··on Show HN: Non.io, a Reddit-like platform Ive been working on for the last 4 years
To be clear, that’s not how I use the platform (I’m not sure I’ve ever searched that way). But it’s a common enough pattern that there have been widely read articles referencing it:

> This means you’ll no longer have to add “Reddit” to your searches when you’re looking for thoughts from actual humans, not empty answers from websites just trying to get clicks.

https://www.theverge.com/2022/4/20/23034024/brave-search-fea...

> It turns out that almost 70% of polled readers add 'Reddit' to their search results at least sometimes.

https://www.androidauthority.com/reddit-web-search-queries-p...

I can’t say that it’s a substantial portion of Reddit’s traffic, of course. But clearly it’s not just a mrtranscendence idiosyncrasy.

mrtranscendence··on Show HN: Non.io, a Reddit-like platform Ive been working on for the last 4 years
Given the popularity of searching for things like product reviews or good restaurants in a new town by appending “Reddit.com” to the search query, I think elucidation and discovery are at least part of what drives people to such platforms.
mrtranscendence··on Show HN: Non.io, a Reddit-like platform Ive been working on for the last 4 years
A moderator who does this consistently will be crucified, and no one will want to post in the community while they remain.
mrtranscendence··on Licensing is neither feasible nor effective for addressing AI risks
I'm not sure why watching Linus Tech Tips and makeup tutorials is going to give AI a better shot at super-intelligence, but sure?
mrtranscendence··on Ken Griffin says the AI community is making a mistake by creating so much hype
> You seem under the mistaken impression that if I can't cite it, its not happening.

I don't even know how to respond to this. Think about the implications of what you're saying: you can claim anything, and when you're challenged on it, you can say "just because I have no proof doesn't mean it's not happening". If you have no proof that bioinformatics professionals are being laid off due to AI, then why did you bring them up as an example?

> I'm pointing at a crack, and you are saying show me proof this crack will cause everyone in the valley below to die.

No, you're pointing at what you believe is a crack, and I'm saying "show me proof that what you're pointing at is actually a crack and not a trick of the shadows".

> Its a flawed way of thinking in many ways especially with fundamental things that we base our lives and welfare on (as a society)

I don't know. I'd like to think that with fundamental things we base our lives on, we should try to be as certain as possible that we're responding to the right threats. We don't typically base our response to threats solely on how severe the damage would be if the threat came to pass -- we also run analyses to determine how likely the threat is to happen in the first place. There's a reason we don't throw trillions at the problem of preventing an asteroid impact or the eruption of Yellowstone.

> the rate of change has been exponential even if you refuse to believe it

I mean, people say that. What I've seen is:

* the transformer architecture is invented in 2017

* it takes years to make a broadly useful text completer with that architecture

* over the most recent span of a year (gpt-3.5 to gpt-4) we've seen relatively minor improvements in capability

* gpt-5 isn't even being trained yet (at least as of April)

I see some advancements, but they're all about making smaller, less useful models run on commodity hardware. The hard work of making an AGI doesn't seem like it's going anywhere right now. The current approaches require too much data and too much compute to see large improvements, and new approaches -- which don't yet exist -- need to be developed.

mrtranscendence··on AI will save the world?
> And that's just for the slogan

My point is that I don't understand what using an LLM changes here. You'll still have to generate multiple candidate slogans, you'll still have to bring 10 of them to the account manager, you'll still have to discuss the possibilities with the customer and tailor the results to their needs. You'll still have to iterate. All an LLM can do is write the candidate slogans, and -- as of now -- it doesn't do a spectacularly great job at it. Very little actual labor is being saved here.

mrtranscendence··on Apple's game porting toolkit is fantastic. Cyberpunk 2077 at Ultra on an M1 MBP
I'm not talking about AAA gaming here. I'm talking about day-to-day work-related tasks, which is primarily what I use my MacBook for.
mrtranscendence··on Ken Griffin says the AI community is making a mistake by creating so much hype
Cite that bioinformaticians and lawyers/paralegals have been laid off in substantial numbers due specifically to machine learning / AI automation?

> currently the rate of change exceeds the rate at which we can evaluate the changes and thus take action

I dispute this. I believe we're seeing a plateau in capabilities, due to limitations of the transformer architecture and the massive, expensive amounts of compute it takes to create marginally better models. I don't believe we'll see anything like exponential growth in what LLMs can do.

mrtranscendence··on AI will save the world?
It should be noted that Ilya Sutskever, judged purely as a philosopher of mind and neurologist, is a great machine learning engineer. I don't feel the need to pay any attention to what he says about LLMs being conscious, nor is there any particular reason to assume they are (even if we posit that LLMs are "intelligent", which I think is a category error, why should intelligence be related to consciousness?).
mrtranscendence··on AI will save the world?
Well, what's "generally affordable"? $30k for a car isn't even that expensive these days (in the US at least), relative to the competition. The average cost of a new car in the US is almost $50k. Except for the very cheapest new cars you're almost certainly going to pay over $20k.
mrtranscendence··on AI will save the world?
The thing about advertising and marketing is that actually writing copy doesn't take anywhere near the full 40 hours a week. How much time does it take to type up a short bit of marketing copy? If your only goal is to just shit something, anything out then you could write up a whole campaign in a few minutes.

But they don't do that in general. So where's ChatGPT going to save time? You'll still have to iterate, iterate, iterate, go back to stakeholders, discuss in progress work with clients, etc etc etc. I just don't see any reason to worry much in the short term, as long as you're working for a reputable company.

This may change eventually, but I foresee it taking several years at least. Progress isn't anywhere near as "exponential" as some camps claim.

mrtranscendence··on Ken Griffin says the AI community is making a mistake by creating so much hype
What, exactly, is already possible with current LLM capabilities? If you're implying they could replace workers, I don't think that will happen broadly. They're simply not that good yet, and it's unclear to me that they'll become that good in the next several years. And they definitely can't replace most jobs that involve any kind of manual labor.
mrtranscendence··on Apple's game porting toolkit is fantastic. Cyberpunk 2077 at Ultra on an M1 MBP
That may be true, but in practice I get faaar better battery life out of my M1 Max than I would out of a laptop with a mobile 4090.
mrtranscendence··on Apple's game porting toolkit is fantastic. Cyberpunk 2077 at Ultra on an M1 MBP
> Is that really the case? I'm not being facetious here, I'd really like to see more useful datapoints. There aren't many benchmark comparisons out there that strive for actual useful comparison, especially outside synthetic stuff with questionable applicability like 3dMark.

Yeah, on synthetic benchmarks a higher-end MacBook Pro GPU compares favorably with recent Nvidia laptop cards (say, 4070 or so). But in games that drops off dramatically ... more like a 3050 or 3060 at best.

mrtranscendence··on Apple's game porting toolkit is fantastic. Cyberpunk 2077 at Ultra on an M1 MBP
High-end hardware is niche, but the state of AAA game performance these days ... ugh. I can't even get a stable 60fps on Jedi Survivor with my 5800x/3080 Ti rig, even with lower settings. How bad is it on something like a 1060?
mrtranscendence··on Apple's game porting toolkit is fantastic. Cyberpunk 2077 at Ultra on an M1 MBP
I mean, you're kind of dead wrong? I've been using Macs for almost two decades at this point as a developer, and it's never kept me from getting work done. These days it's great having a powerful Mac, with the ability to use local machine learning models with relatively large amounts of VRAM (more than any consumer GPU, though inference won't be as fast as recent Nvidia cards). It's great at video editing, too.
mrtranscendence··on Apple's game porting toolkit is fantastic. Cyberpunk 2077 at Ultra on an M1 MBP
> there are plenty of triple A games in the App Store

Mobile AAA games in the iOS App Store, or desktop AAA games in the Mac App Store? There are some AAAs that work on a desktop Mac (e.g. Resident Evil Village), but it's not the norm at all. As for mobile ... modern mobile games can die in a fire for all I care (as they're 95% gacha nonsense).

mrtranscendence··on Run and create custom ChatGPT-like bots with OpenChat
I'm curious how that works, as the documentation is a little under-specified. It seems like it requires specifying exact "utterances" from the user, but I don't think that can be the case -- wouldn't it be flatly useless that way? But it's not clear how to use it to, for example, disallow talking about politics. Or to disallow talking about topics unrelated to the dev's product, for that matter.
← PreviousPage 5 of 32Next →