Is there any calculation/projection/cost-benefit done on the effects of these measures? I did not renew my Google AI sub because Gemini started to reject transcribing song lyrics due to IP. This seems like a similar thing
maybe this is what will get people to use grok or chinese models. At some point they're too restrictive. Claude refused to help me scrape a public website they themselves have gotten their info from..
I also see Google winning this in the end. They have absolutely crazy behavioral data and can exploit much more effectively. Meta is also up there, but Google has production hardware as well as clouds, etc.
Its either AGI, or a robbery beyond behavioral data and the whole internet that can make the AI labs win. I dont even know what such a robbery would look like, would perhaps be just making the most advanced hacking tool and then holding the rest of the world and banks hostage.
1) Have a single person vibe together a computationally inefficient solution in 1 month that runs transformers on some backend.
2) Hire a team to work for 2 years to design, perfect, optimize, plan, revise, and architect a solution.
The risk/reward for 1) is extremely better than 2). Given the turnover for entertainment in this day and age it doesn't make sense to do all this complex stuff.
It would be much better to use LLMs to create ordinary programs than to use an LLM over and over to emulate something that's thousands or millions of times more efficient.
You can't generalize and say that 1 is better than 2, it depends entirely on what you are developing, for whom and for what purpose. For a one-off dumb thing for a few people 1 is fine, for anything at scale, 2 is likely better.
How do you vibe a transformer solution without the existing real solution, which would be needed to generate the training data.
These are just machines for ripping off existing games/programs with mostly accurate simulations. You can't vibe together a transformer of something that doesn't exist with meaningful control over the results.
This might hold up for theoretical intellectual work without any economical value, but if you are researching anything that can be applied to the industry, you will instead have 1 year of market ownership and first-mover advantage.
I think it says more about what the value of math as intellectual work is, than about intellectual work in general.
Like if an Egyptian historian would come and say "let's wait one year until the methods are good enough to decrypt this language", no one would be bothered or surprised.
You imply "worth" being equal to some known economic value.
You even imply, economic value was determined correctly in advance for arbitrarily long future intervals. And the applicability was known in advance to begin with.
Obviously, you're completely wrong.
Applying simplistic models to real life incurs a penalty. The affected scale multiplies with that penalty. When dealing with things affecting humanity as a whole, that makes careless over-generalizations a serious hazard.
“if you are researching anything that can be applied to the industry, you will instead have 1 year of market ownership and first-mover advantage.”
I would like to disagree with you on that point. Given the very popular rise of https://rea.tools/, it appears the time to reverse engineer anything software based (up to maybe 80% or more?) could be down to minutes. With a recent example, some indie video game developers are not going to show off their video game until release as a few have been flipped and made it to market before the original developers have. Point being, a few moths of market ownership could be possible if llms can’t reverse it in any way.
What a waste. All the anticipation that used to be built up with previews and looks into the dev process, now flushed down the toilet so billionaires can flood the web with crap and make more billions.
Someone is going to make the case against using LLMs to reverse engineer software solely for commercial gain being as illegal as using guns for extortion.
Unless you’re able to get a patent on day one, once an idea is public potentially within the hour someone can make more money off of it and it’s not just your idea. Seems like a valid case to lock it down.
First movers often lose though in the market. Relevant to today's news about DoorDash acquiring Instacart for a few billion dollars: WebVan came out a decade earlier (i.e. at the wrong time) and went bankrupt after 3 years.
A perhaps more relevant personal anecdote: About a year ago I told a non-technical friend that I'd help him migrate his small business website that was originally set up in WordPress (by a developer who is since long gone) to a format that would be much easier for him to manage with no dev support. I planned to move him over to one of the usual suspects like Wix or Squarespace, but this is one instance where my procrastination was a virtue. In the ~9 months since he originally had an issue with his site, AI capabilities had exploded. Turned out it was a lot easier (and cheaper) to use Claude to fully migrate his site and get it up on Cloudflare workers, and I set up everything so he could manage all the content himself. Plus, it took me less than a day to do the migration off WordPress, and most of that time was just me wanting to learn the ins and outs of Claude Code because I was new to it at the time.
The newest submissions to Hutter Prize made crazy jumps by just replacing the LSTM with a transformer and training it.
35'000$ for a prompt and some training is a nice chunk of cash.
Next steps are replacing the encoders and all the other human-biased compression parts by just raw data and training, probably another 20-30% is up for grabs.
Its cool, but it goes against the whole optimization that NLP has gone through with BPE and redundancy/repetition in large data in general.
Put this on a coding task or English and you will underperform against tokenizer that exploit multi-char patterns and can decode much faster.
Given where things are going with spec decode (I've seen 32 token spec-decode pop up in the frontier models), you will need a very domain-specific use for this to be useful imo like massive tabular data, but at that point you already have everything structured, and might as well use an LLM with polars/sql/arrow.
The whole point of this architecture is that it can also make use of multi-char patterns, but unlike BPE, they aren't hard-coded. And you can still use speculative decoding. Whether it works better or worse is something that you have to actually measure.
> you will need a very domain-specific use for this to be useful imo
The main premise of this idea is how general and widely applicable it is. They also show examples of how it works on images.
A common tokenizing paradigm would be a huge step towards consolidating how different modalities are preprocessed - something I believe is quite useful for intelligence to scale well across modalities.
- If you work with the raw bytes in a non-hierarchical (isotropic) model, such as a byte transformer, you get better transfer to downstream tasks, but at a huge computational cost.
- if you use a tokenizer (pretrained or regexp based), you gain a lot of computational efficiency, but loose semantics and the transfer to downstream tasks may degrade
- finally, with H-Nets you get the best of both worlds. The model learns how to chunk the data in a better way than a hardcoded/separately trained tokenizer
Authors also demonstrate that this approach works in other modalities, such as DNA sequences and images, where we don't normally have lots of ready to use tokenizers.
But you lose the semantics you choose to not care about. BPE is equally equipped to not fuse single chars if they have downstream value. Most modern tokenizers are already trained with
the target text corpus in mind.
For example Google's propabilistic Unigram embedding model is trained jointly with the end task, so if you would train it on a DNA set where individual chars matter it probably would not fuse them
So the interesting question is rather how the authors manage to maintain perf on all modalities/tasks with a single corpus or if they are benchmaxxing something.
The amount of upselling, extra parts necessary for basic functionality, predatory resin pricing and specific open "modes" that are free to be subjects to amendment by contract makes this seem like a cash grab for investors rather than serious product. Really sour taste in the mouth from this.
Just buy an open source version from China that has the necessary hardware in an open format and let Claude adjust the software if there are concerns.
I'm unclear what the open source version from China of a resin printer is. Creatlity and Elegoo don't appear to be at all it as they lock you to the chitubox ecosystem.
I use prusaslicer and uvtools with my Elegoo Saturn. No proprietary lockin. I also own a Prusa, so I don't feel bad using their software with another brand.
There is currently no open source version of the high resolution monochrome LCDs. The LCDs and their drivers lock you into the Chitu ecosystem.
The problem, as Chitu figured out, is that as soon as you cough up the NRE to make the LCDs and make it publicly available, EVERYBODY will copy you and then there will be no profit left to sink the NRE that you just spent.
Maybe you could be happy with that, or perhaps I could be as well, but you and I aren't ChiTu. And ChiTu is free to do whatever they want (including doing nothing at all).
And that matters because statements like "just buy an open Chinese printer" are non-starters. The machines themselves have the soul of ChiTu inside of them and openness isn't really their jam.
So while the PC software that produces the layer images can be replaced, the printer's brains stays the same.
(This is very different compared to the world of conventional world of FDM printers, wherein: It's just some sensors, motors, heaters, and sensors; a screen, if present, is optional. A person can replace the brainbox of a traditional FDM Cartesian rig with open, off-the-shelf parts and be printing something in an afternoon, and there's a whole industry providing parts to support this kind of thing.
Well I don't understand the paragraph you wrote starting with "The problem, as Chitu figured out". Did they figure out an actual problem or did they hallucinate a problem?
Does their proprietary software affect their profit at all?
Chitu, IIRC, originally were open when they were using the initial color LCDs (off the shelf) and then the first iteration of monochrome ones (required NRE as nobody made 1920x1080 monochrom LCDs by then). And then everybody in the universe cloned them and Chitu got mad and Chitu made sure that the next iterations of monochrome LCDs had encrypted driver chips.
This isn't an unsolvable problem. It takes probably less than $100K (and probably less than $50K) in NRE to make a truly open-source monochrome LCD and driver chip. This is on the order of a plastic injection mold.
But everybody simply whines rather than banding together, coughing up the CASH, and actually FIXING the problem once and for all.
Sadly, this seems to be standard operating procedure in the world of open source.
Dude, it's CHINA. Once developed, the manufacturer simply sold the LCDs to everybody.
Even now, there is nothing stopping you from going to the upstream LCD manufacturer and saying "Hi, Mr. LCD manufacturer. Here is a different driver chip, could you please take that nice LCD and use it with it?"
It all just takes cash. And not even a lot by manufacturing standards.
Was the question of if horses and carriages are important a thing we had to answer? No, it just fell out of it. If human maths is pointless, what is the issue with defunding it?
if it's pointless then there is no issue with defunding it, but that is the question being asked.
if a similar investment on humans would have similar or superior results then human maths is not pointless and directing the investment to AI instead is arguably a loss.
the same question can be made for other areas AI is advancing. My AI budget is over 3 times my salary. If instead of providing AI tooling the company were to raise everyone's salary by 3x, what type of results would it get? Surely it would not get 10x the amount of PRs and LoC but maybe it would get superior value in other ways.
Humans are bound by communication overhead and non-market-based incentives. Adding 3x engineers does not do anything. On an average workplace, you have 80-90% of dead weight who shuffle pointless information around and play social games to keep their children fed.
Giving the 10% access to these tools is a crazy value multiplier, because they can run at lightning speed with their mental models and the right incentives for the company.
I think that this was true pre-attention economy. If you look at security incidents and other prod issues, they rarely move the needle today. I work at a hyperscalar and we have been shipping 40% more bugs to prod since 2023 year-to-year, yet user growth and revenue is up about 25% similarly due to more features and more ads.
I think that the whole premise of that failure affects things is based on a older model. You kind of assume that:
1) users are able to easily gain information about alternatives. We have seen that since LLMs replaced search engines, we have much less churn.
2) users are not behaviorally addicted/pushed into patterns. We have patterns that utilize users to keep using the app by exploiting their cognitive weaknesses
3) users have the mental fortitude and stamina to learn new tools instead of accepting flaws in the one they are used to
Some failures and bugs in prod don't matters as much
Almost the entire structure of software oriented companies, and why devs are expensive, are to prevent these sorts of errors. That job has gotten harder, and in some places impossible as management forces more trust on LLMs and reduces the capacity of devs to look. This is less of a problem in existing products in production and existing cultures - revenue is protected. Eventually those structures break down, as does the user perception of quality. Once the brands have taken enough knocks, it accelerates.
HN is one of the few sites where someone can launch into an actual villain monologue, the kind you'd read in leaked corporate emails, and receive approximately no negative reaction.
If this is the face of our industry, that might explain a lot of things...
You can take the right path and in the end no-one wants to use your product because it's not as engaging.
Instagram tried to take the "good" path in 2014 and lost 90% market share to TikTok who attentionmaxxed things. Then they had to take the exploitative path to claw back users.
As long as it's not regulated for everyone at the same time, you need to do this for survival of your product.
It's more like it doesn't matter if you sell the most organic, free range, ethically sourced, nutritious, tasty ice cream if your competition is allowed to put cocaine in theirs and get all your customers.
TikTok would survive and do just fine even if Instagram decided to cede the market to them
I am afraid there is a demand for the torment nexus, and until govt steps in, you're out of luck and behavioural surplus is there for the taking by any means.
Many bad things that people do stem from weakness or inner pain. So it's helpful to be reminded from time to time that some people deliberately, thoughtfully, calmly and systematically devote their lives to harming other human beings -- that is, to evil. We should be grateful to augment_me for this reminder.
I agree that not all failures matter (most don't, in fact). On the other end of the spectrum, I know of a bunch of seed/series A startups that basically operate under a model where they ship everything and fix the stuff people complain about. As long as you do that fast enough, the median user experience is pretty good, because only the first n% actually see the problem before a fix is live.
I'll go out on a limb and guess that at a hyperscaler, shipping a bug to prod means hitting canary or 1% of traffic. Seems fine!
I think my point still stands, even as the shape of failures shift. There's certain ones that still hurt (data corruption, exploited security holes, etc.) because they're hard to notice. It might even be that the primary goal of ownership is to develop intuition about where those failures are likely to happen.
Wait, how do working at a hyperscaler (that's a datacenter provider) and using an addictive app for uhm retention go together? I imagine a gamified AWS provisioning and billing app, but surely that's not it?
This very thing. "Shipment of 40% more bugs and fixing only in response to user complaints" could be done in social streams or advertising tech, but not in finance and trading where margin for errors is close to zero.
A small bug implemented by an agent in the way R-multiples, stop loss distance, or P&L heatmaps are computed will never be complained about by the user—either he will trade incorrectly and lose his money, or see that single off-by-one difference compared to exchange ledger and churn forever.
People are going to be upset about this, but any learning material in this day and age is solely made for your own understanding when writing it. The LLMs can be better teachers and understand your level of need and engagement better than anyone else.
reply