Hacker Newsnew | past | comments | ask | show | jobs | submit | ep103's commentslogin

> For example, in a situation where you can strongly benefit from the Napoleon technique, and all the potential negative outcomes are minor and unlikely to occur, you will almost always want to implement this technique. Conversely, in a situation where there is even a moderate likelihood that this technique will lead to serious negative outcomes, you will likely want to avoid using it, even if it has some potential positive outcomes.

I swear, AI is decreasing everyone's reading and writing abilities.

Well written language conveys maximum information (or emotional impact, or etc) with minimum verbosity. AI is incentivized to do the exact opposite, and results in slop like the above.

The quoted paragraph above takes 71 words to say "You should do this technique if the positive potential outcomes outweigh the negative ones," which is such a banal thought as to have been a waste of the reader's time, the writer's time, and the electricity it took to run an AI to generate those sentences.


The text was first linked on HN during September 2020. ChatGPT became public access in November 2022.

The paragraph you criticized was part of the original text: https://web.archive.org/web/20200909104647/https://effectivi...

So: Yes, it could have been more concise. Nope, we humans can write much too long text for the sake of writing text, which some of us can do better than others (e.g., better than me), and we can do that with no artificial assistance or substitute - we do it just fine using our own (in)ability ;-)


It is everywhere, but SF really brings it to another level. Its wild.

You and I see the tiktok slop. But as that functionality improves, its going to make its way into the toolchain of every digital image and video editing software in existence, the same way that its finding its way into programming IDEs. And that type of feature build is worth $. It might be a matter of time until we get to the point where we start seeing major Hollywood movies (for example) doing things that were unthinkable the same way that CGI revolutionized cinema in the 80s. Even if it doesn't, from my layman perception, it seems that Hollywood has spent the last ~20 years differentiating itself from the rest of global cinema largely based on a moat built on IP ownership and capital intensive production value (largely around name brand actors and expensive CGI). AI already threatens to remove one of those pillars, which I have to think in turn makes it very valuable.


Call me a conspiracy theorist, but I think you're right, and I think the reason for it is because Google has historically had an extremely effective astroturf marketing team for Chrome


I have always been in favor of changing the definition if incorporation to ensure that over time ownership transfers slowly but increasingly to the employees of the corporate entity. How that would work, though, would require detailed thought by experts more knowledgeable than i :)


You should look up something called the "Rehn–Meidner model:"

https://en.wikipedia.org/wiki/Rehn%E2%80%93Meidner_model

Sounds similar to what you're asking for.


Its so nice to see this echo'd somewhere. This has been what I've been calling them for a while, but it doesn't seem to be the dominant view. Which is a shame, because it is a seriously accurate one.


This is my wife starting up a 20 minute conversation the moment the first actor shows up on the screen xD

Don't worry, I love her anyway. But yes, we're restarting the movie because no, I don't have any idea what happened either, you were talking. ahahaha


The benefit of cloud has always been that it allows the company to trade capex for opex. From an engineering perspective, it trades scalability for complexity, but this is a secondary effect compared to the former tradeoff.


"trade capex for opex"

This has nothing to do with cloud. Businesses have forever turned IT expenses from capex to opex. We called this "operating leases".


I’ve heard this a lot, but… doesn’t Hetzner do the same?


Hetzner is also a cloud. You avoid buying hardware, you rent it instead. You can rent either VMs or dedicated servers, but in both cases you own nothing.


How are you guys spinning up vms, specifically windows vms, so quickly? I used to use virtual box back in the day, but that was a pain and required a manual windows OS install.

I'm a few years out of the loop, and would love a quick point in the right direction : )


A lot of the world has moved on from virtualbox to primarily qemu+kvm and to some extent xen. Usually with some higher-level tool on top. Some of these are packages you can run on your existing OS and some are distributions with hypervisor for people who use VMs as part of their primary workflows. If you just want quick-and-easy one-off Windows VM and move on, check out quickemu.

Libvirt and virt-manager https://wiki.archlinux.org/title/Libvirt

Quickemu https://github.com/quickemu-project/quickemu

Proxmox VE https://www.proxmox.com/en/proxmox-ve

QubesOS https://qubes-os.org

Whonix https://whonix.org

XCP-ng https://xcp-ng.org/

You can also get some level of isolation by containers (lxc, docker, podman).


You take the time to set one up, then you clone it and use the clones for these things.


Windows does have a builtin sandbox that you can enable. (it also enables copy-paste to it)


Not sure about windows but I solved it for myself with basic provisioning script (could be an ansible playbook also) that installs everything on a fresh linux vm in a few minutes. For macos, there is tart vm that works well with arm64 (very little overhead compared to alternatives). Could be a rented cloud vm in a nearby location with low latency. Being a neovim user also helped not to having to worry about file sync when editing.


For coding I normally run Linux VMs. But Windows should be doable as well. If you do a fresh install every time then sure it takes a lot of time, but if you keep the install in VirtualBox then it's almost as fast as you rebooting a computer.


Also, you can spin up an ec2/azure/google vm pretty easy too. I do this frequently and it only costs a few bucks. Often more convenient to have it in the data center anyway.


A docker container isn’t as bulletproof as a VM but it would certainly block this kind of attack. They’re super fast and easy to spin up.


It would not block many other attacks.


Can you give some examples? I think of my containers as decently good security boundaries, so I'd like to know what I'm missing.


Containers share resources at the OS level, VMs don't. That's the crucial difference.


Containers share the whole kernel (and more) so there's a massive attack surface.


If you're on a Mac, you probably want OrbStack nowadays. It's fabulous!


Yesterday I used ChatGPT to transform a csv file. Move around a couple of columns, add a few new ones. Very large file.

It got them all right. Except when I really looked through the data, for 3 of the excel cells, it clearly just made up new numbers. I found the first one by accident, the remaining two took longer than it would have taken to modify the file from scratch myself.

Watching my coworkers blindly trust output like this is concerning.


After we fix the all the simple specious reasoning of stuff like Alexander-the-great and agree to out-source certain problems to appropriate tools, the high-dimensional analogs of stuff like Datasaurus[0] and Simpson's paradox[1] etc are still going to be a thing. But we'll be so disconnected from the representation of the problems that we're trying to solve that we won't even be aware of the possibility of any danger, much less able to actually spot it.

My take-away re: chain-of-thought specifically is this. If the answer to "LLMs can't reason" is "use more LLMs", and then the answer to problems with that is to run the same process in parallel N times and vote/retry/etc, it just feels like a scam aimed at burning through more tokens.

Hopefully chain-of-code[2] is better in that it's at least trying to force LLMs into emulating a more deterministic abstract machine instead of rolling dice. Trying to eliminate things like code, formal representations, and explicit world-models in favor of implicit representations and inscrutable oracles might be good business but it's bad engineering

[0] https://en.wikipedia.org/wiki/Datasaurus_dozen [1] https://towardsdatascience.com/how-metrics-and-llms-can-tric... [2] https://icml.cc/media/icml-2024/Slides/32784.pdf


> it just feels like a scam aimed at burning through more tokens.

IT IS A SCAM TO BURN MORE TOKENS. You will know when it is no longer a scam when you either:

1) pay a flat price with NO USAGE LIMITS

or

2) pay per token with the ability to mark a response as bullshit & get a refund for those wasted tokens.

Until then: the incentives are the same as a casino's which means IT IS A SCAM.


Ding ding ding! We have a winner!


>it just feels like a scam aimed at burning through more tokens.

I have a growing tin foil hat theory that the business model of LLM's is the same as 1-900-psychic numbers of old.

For just 25¢ 1-900-psychic will solve all your problems in just 5 minutes! Still need help?! No problem! We'll work with you until you get your answers for only 10¢ a minute until your happy!

eerily similar


To me it’s a problem of if a piece of information is not well represented in the training data the llm will always tend towards bad token predictions for related to said information. I think the next big thing in LLM’s could be figuring out how to tell if a token was just a “fill in” or “guess” vs a well predicted token. That way you can have some sort of governor that can kill a response if it is getting too guessy, or atleast provide some other indication that the provided tokens are likely hallucinated.

Maybe there is some way to do it based on the geometry of how the neural net activated for a token, or some other more statistics based approach, idk I’m not an expert.


A related topic you might want to look into here is called nucleus sampling. Similar to temperature but also different.. it's been surprising to me that people don't talk about it more often, and that lots of systems won't expose the knobs for it.


It sometimes happens with simple things. I once pasted the announcement for an event in Claude to check for spelling and grammar.

It had a small suggestion for the last sentence and repeated the whole corrected version for me to copy and paste.

Only last sentence slightly modified - or so I thought because it had moved the date of the event in the first sentence by one day.

Luckily I caught it before posting, but it was a close call.


Yup, I always take editing suggestions and implement them manually, then re-feed the edited version back in for new suggestions if needed. Never let it edit your stuff directly —— the risk of stealth random errors sneaking in is too great.

Just because every competent human we know would edit ONLY the specified parts, or move only the specified columns with a cut/paste operation (or similar deterministically reliable operation), does not mean an LLM will do the same, in fact, it seems to prefer to regenerate everything on the fly. NO, just NO.


Tool use seems like a much better solution in theory. I wonder how it works out IRL?


> Yesterday I used ChatGPT to transform a csv file. Move around a couple of columns, add a few new ones. Very large file.

I'm struggling with trying to understand how using an LLM to do this seemed like a good idea in the first place.


When you have a shiny new hammer, everything around you takes on a nail-like aspect.


the safe way to do this is to have it write code to transform data, then run the code

I expect future models will be able to identify when a computational tool will work, and use it directly


I don't mean to be rude, but this sounds like user error. I don't understand why anyone would use an LLM for this - or at least, why you would let the LLM perform the transformation.

If I was trying to do something like this I would ask the LLM to write a Python script, validate the output by running it against the first handful of rows (like, `head -n 10 thing.csv | python transform-csv.py`).

There are times when statistical / stochastic output is useful. There are other times when you want deterministic output. A transformation on a CSV is the latter.


Because it markets and presents itself as deterministic and honest. That's the whole issue. AI is unethically marketed and presented to the public.


iPod marketing presented then as a device that made you cool. I just used mine to listen to music though


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: