yeck

If you could know that this is the case with interpretability tools than we would be able to train new models with purposeful decisions to reduce or remove hallucinations. Narrow the range of the tests and experiments you need to do to solve the problem. Otherwise we are mostly speculating about why stuff doesn't work and play a game of darts in the dark.

I have a hard time understanding why mechanistic interpretability has so few eyes on it. It's like trying to build a complex software system without logging or monitoring. Any other improvements you want to make on the system are going to just be trail and error with luck. The hallucination problem is one where interpretability of a model might be able to identify the failure modes that we need to address. Really any AI problem could likely be aided by a scalable approach to interpretability that is just as mundane feeling as classical software observability.

I've noticed the same recurring debates centred around the existence and severity of AI risks. I'd like to see more discussions that start with the assumption that risks are significant so that we could have a debate on what might be done to mitigate risks.

> I would assume that we would also be in that deal, until a time is met that they no longer need us

Isn't that what Eliezer is suggesting? The AI will be the moon to our Earth until it has determined it is capable and ready to transition into a sun?

> I think it makes zero sense to fight or compete

I'm not sure Eliezer is thinking that there will be a fight or competition. Mostly I think he believe we won't really be much of a consideration to a super intelligence at all. For instance, if a super intelligence in our solar system decided to harness all the energy radiating from the Sun, Earth would freeze over in a matter of days (hours?). That scenario probably doesn't have any more of a conflict than what you'd expect to see between a family of deer living in a forest and a construction company clearing that forest for a development project.

Is "fireguard" a typo or a distinct thing?

I think these days getting kids involved in non-tech based activity groups could help a child have a more balanced relationship with technology.

There are also options like homeschooling, but that is obviously a massive commitment and different life journey for a family to take.

Then ofc there are rules and enforcement mechanisms. Rules can really be whatever you like. Enforcement can happen at your router, DNS or maybe firewall. If you like to configure those sorts of things you could block access to certain services or websites based on time of day, or target specific devices in you network. When they are not home you won't be able to do this so easily. There is also risks of coming off too adversarial and damaging your relationship with your child.

edit: typo

If you see where the packets are going you might be able to block at the DNS level.

Maybe `< HackerOne`

And why not just go with _hackers_?

I'd say they are a winner right now. Selling some of the best shovels and sieves in the gold rush hype.

Edit: if AMD plays their cards right, I'd expect that they can get a lot more in on the action, too.

What if the output is part of an `eval` not just an `echo`? People want to be able to do this, because there is massive potential, but they can't so long as there are reliable ways to steer outputs toward undesired directions. A lot of money is behind figuring this out.

If you spend timing considering how to use LLMs as part of another product or system you realize quickly that there are a lot of interesting things that you just can safely do because you can't control the LLMs inputs or outputs. I have personally examined use cases for products that would be used by children and I cannot use LLMs as it currently stands because I have no way to ensure the content generated is age appropriate.

They are proving that the present techniques being used to control their models can be reliably bypassed. Regardless of what you want your model to be able to do, there may be things that you don't want it to do. For instance, if you are making a product for children you'd probably want it to avoid certain content.

If you are training you own model, it would be nice to know what, if any, techniques you could employ to balance the effectiveness of it with generality.

Sure, these particular examples are concerned with morality, but the problem is more general and limits the value of language models because they can be hacked for other purposes. A good example that's been going around is having an agential model that manages you emails. Someone sends you an email using prompt injection to compel the agent to delete all your emails. Or forward all you emails to another address.

If there isn't a way to secure the behaviour of AI models against reliable exploits then the utility of the models is dramatically limited.

Imo, the strength of containers is portability, not reproducible builds. Even portability isn't perfect, since images are sensitive to CPU architectures and containers can still actually rely on host system configurations (mounting part of fs, privileged mode).

I hadn't been following it all that closely, so the repeat sightings of black squares in a transparent sphere was news to me.

The more interesting one is seeing the signs that there is growing momentum in the government to gain more clearance and potentially make more classified information public. So perhaps this is the amuse-bouche?

Hearing just concluded. If you listened to it, what are your thoughts?

Convincing, y/n? How come?

The duplication is not about being becoming smarter but increasing the amount of concurrent tasks that can be performed.

Perceivable latency isn't the only consideration. Depending on your business, compute for your proprietary workloads might be one of your biggest expenses. You could see an order of magnitude improvement to resource consumption depending on the type of workload and the language it is written in.

Ultimately these these choices are all about trade-offs. Maybe python is fine for them, maybe they've built themselves into a corner. Time will tell.

What sort of part time roles do you have? What are your requirements for part time, like hours per week and availability?

I think it is wise to use Nix for a bit before trying to manage all your user configs with it. Using `nix-shell` for project dependencies is a good starting point. You can't really make any costly mistakes here since the scope of any given project is likely much less than all your system and user configurations.

Nix shell plus something like direnv using flakes or nix expressions (non-flake approach) would give you a chance to learn the language and ecosystem. Once you feel comfortable with this, home-manager should be much more approachable.

I just came across this yesterday. While I'm pretty good at keeping my code <= 80 columns, I also write a lot of markdown where lines often wrap. Remapping `j` and `k` to `gj` and `gk` is a really nice quality of like improvement :D

(If doing such a keybind, you'd want to check to make sure count is 0 or things will get wonky for you)

I think you are forgetting that the "modern society" you are familiar with is not ubiquitous and in our present moment there are many people that lead completely different lives than you. If you are not ignorant of this, then you are supposing that no one in a different circumstance may prefer it and fight for it.

That is post-agricultural revolution humanity. Sure it is ancient by the standards of an individual, but it is only a relatively recent and small part of the more than 100 thousand years of human pre-history.

Which claim does the burden of proof land on? That an artificial super intelligence can easily be controlled or that it cannot? And what is your rational for deciding?

I didn't end my Rust learning with that. I did a bit of a deep dive into articles about linked lists in Rust out of curiosity before pivoting to more fruitful projects. My favourite "project" being last year's Advent of Code.

Generally in agreement with the article. This bit was confusing to me, though.

> Moving with the arrow keys, erasing text by holding backspace for an eternity feels slow, but it is predictable and reliable. Vim or Emacs-style editing, where you fly through the buffer with keybindings, feels faster, but a single wrong keypress puts your editor in some unpredictable state it can be hard to recover from.

I've been using neovim for about a year now. I don't think I've ever entered a state from a mis-click that is hard to recover from yet. The worst case scenarios tend to result in something taking comparable time to if a mouse was used.

I find it so troubling that the most common HN response to this isn't to engage with the ideas or logic behind the concerns, but simply to speculate on the unknowable intentions of those that signed the letter.

We can base our arguments on the unprovable, some specific person's secret intentions, or we engage the their ideas. One is lazy and meaningless, the other actually takes effort.

Then there is this lazy and false equivalency between corporations being interested in market capture and AI risks being exaggerated.

Hah, yes. I thought I'd get started with Rust going through building some basic data structures. Linked lists were near the top of my list so I didn't get very far with that first approach XD