yeck

By that do you mean that you believe AI is safe by default? Or that AI safety is not possible?

I've seen some likening Sam to Caesar, being betrayed and such. Certainly there are communities that match what the parent comment describes. HN has generally been more balanced than I would have though, though. The fan boys I've personally encountered have been more in the lands of YT and Reddit.

Not OP, but the choice of phrasing in a number of reports implied that the board was uncertain of their decision. Normal sensationalization if the same limited information that was publicly available at the time.

Anyone find any sources with new information? Now that we are well past the 5PM (PST) deadline I'm curious if anything has emerged.

That's 4x the amount of buy-in than is required to fire most people.

Aspects of this article were surprisingly good. This part was not one of them, though. What's already been released publicly is feels "rooted in the realm of science fiction".

You are chaining some big assumptions together to make this conclusion. We can suppose AGI is around the corner, but these other assumptions are massive leaps that would need strong arguments to back them: - AGI thinks similar to humans - AGI knowing we are afraid of it will make it consider humans a threat

Unpacking that second point are the implications that: - AGI considering humans a threat is conditional on our fearing it - AGI seeing humans as a threat is the only reason it would harm humans

I feel like I can rule out these last 3 points just by pointing out that there are humans that see other humans as a threat even though there is not a display of fear. Someone could be threatening because of greed, envy, ignorance, carelessness, drugs, etc.

Also humans harm other humans all this time in situations where there was not a perceived threat. How many people have been killed by cigarettes? Car accidents? Malpractice?

And this is going off the assumption that AGI thinks like a human, which I'm incredibly skeptical of.

If Microsoft ends up essentially at the helm of OpenAI after this all this then I'll know we are living in a tragic comedy.

Yeah, I suspect there will be a strong over reaction regardless of whether Altman is reinstated or not.

To be fair it is a very subjective term, god-like. You could make a claim for many different technical advancements to represent god-like capabilities. I'd claim that many examples exist to day, but many of them are not readily available to most people for inherent or regulatory reasons.

Now, I feel even just "OK" agential AI would represent god-like abilities. Being able to spawn digital homunculi that do your bidding for relatively cheap and with limited knowledge and skill required on the part of the conjuror.

Again, this is very subjective. You might feel that god-like means an entity that can build Dyson Spheres and bend reality to it's will. That is certainly god-like, but just a much higher threshold than what I'd use.

Hah, yes. I chuckle to myself every time.

I suspect some do and some don't. Hard to know what the ratio is.

I feel that people have a right to life and liberty, but liberty does not mean access to god-like powers.

There are many people that would do great things with god-like powers, but more than enough that would be terrible.

I wouldn't call GPT-4 sub-human intelligence. While it's intelligence is less robust aggregate human intelligence, I don't think there is any one person alive who can compete with the breadth of GPT-4 knowledge.

I also think that the potential of what currently is possible with existing models has not been fully realized. Good prompting strategies and reflection may already be able to produce a system that is effectively AGI. Might already exist in several labs.

> models to be (easily) swayed is a different problem

No, this is the alignment problem at a high level. You want a model to do X but sometimes it does Y.

Mechanistic interpretability, one area of study in AI alignment, is concerned with being able to reason about how a network "makes decisions" that lead it to an output.

If you wanted an LLM that doesn't succumb to certain prompt injections, it could be very helpful to be able to identity key points in the network that took the AI out of bounds.

Edit: I should add, I'm not referring to AI safety, I'm referring to AI alignment.

Was the article changed for this? Used to be this one from the verge: https://www.theverge.com/2023/11/18/23967199/breaking-openai... but was since changed to https://www.forbes.com/sites/alexkonrad/2023/11/18/openai-in...

Isn't that a massive quality improvement though? How many applications for LLMs are not feasible right now because of the ability for models to be swayed of course by a gentle breeze? If AI is a ship with a sail, data is the wind and alignment is the equivalent of a rudder.

Until I can trust that when I send an AI agent off to do something that I will be successful without me babysitting and watching over it constantly AI won't truly be transformative (since the human bottleneck will remain).

This is one of the core promises of alignment. Without it how can there be trust? While there are probably short term slow downs with an alignment focus, ultimately it is necessary to avoid throwing darts in the dark.

> the runner is closed source

My team runs our GHA runners on our own infrastructure using Actions Runner Controller (ARC).

> The fact that you have to push a commit to the repo to test a change, and then wait for a runner delays everything indefinitely, and makes iterating really painful.

Sounds like a mistake. Steps in jobs can often be run locally, and you could make choices in your jobs to make that more or less true. While there are certainly some things that you can only realistically test by triggering a CI job, many things can be tested locally.

> writing pipelines in YAML is just painful... I invested 3 days to write a small DSL that compiles to GHA

Yes, that can definitely be the case. But the thing about YAML is that it can easily be made a target for DSLs, as you demonstrated. I'd rather GHA use something common like YAML than force another DSL on to the world. It is better for users to decide if/what abstractions to use in front of YAML.

I've seen this sentiment mentioned many times. That at school "abuse is more likely to be noticed". But how much more likely? How likely is noticed abuse to be resolved in a meaningful way?

As an anecdote I have close family member that teaches younger elementary school grades. They have told me about signs of possible abuse they've seen but no available recourse.

I'm also aware that school is a also an environment where abuse occurs. How can we say that we are not exposing public school children to more abuse per capita than homeschooled? There is a school in my neighbourhood that often has the police called in. Lockdowns over students bringing in knives and attacking each other.

It would be good to know these things if making an argument that pivots on abuse.

One last anecdote. All the experiences I can remember having that I'd consider abuse occurred in either: school, work or public transit. Some were bad interactions with strangers in passing, other were more chronic (generally at school).

While it is true terrible things can happen to kids at home, you need stats to compare this with the alternative before a stance should be taken.

As users' codebases grown over time, the average usage of users will probably continue to grow. By a static price increase alone won't be a great model. They need to make their models and infrastructure more efficient first.

Or create a pricing model where the monthly sub covers a certain amount of tokens and then beyond that is pay-as-you-go. Sucks for users, but a scalable model.

Controlling for academic experience probably raises the average accuracy of labelling, but by how much? Clearly having a degree will not make you omniscient in your major, let alone other subjects.

K8S maintains the internal state of the system in JSON which is easy to convert to YAML which is seen as more readable in comparison. I think their choice to support JSON and YAML as an interface for configuring K8S is because of this.

I'd also guess that the expectation was that abstractions would form around YAML to make it more powerful. Helm uses Go templating language which supports logic and loops (but is very unpleasant to write complex configs with) and the operator pattern is also popular for more advanced cases. Ultimately both end up exposing some JSON/YAML interface for configuring, though. It is up to you to decide how you create that JSON/YAML.

Code also requires a runtime or needs to be compiled to something static, like YAML or JSON, in order to be consumed. That latter option is essentially how things are today given that templating tools the generate configurations are plentiful.

I think that this requirement makes broad adoption a lot more difficult for code-based configuration.

Bun is getting a lot of hype due to all the press, but almost everyone is just echoing Bun's docs and marketing material and lack direct experience using it. I haven't done a deep dive into Bun yet, but I have tried porting a legacy project to it. Hit a lot of issues after a few hours of hacking around. It wouldn't correctly install node modules because of opt-in dependency lifecycles (and maybe other issues?) and then I couldn't use it to execute the existing webpack build without encountering opaque runtime errors.

The promise of Bun seems pretty good, but I think their marketing material is further along than their core projects atm.

The gist of this is that ff those who are concerned with AGI catastrophe are wrong but try to prevent the perceived looming disaster then they might do bad things. Bad because they caused harm to people by were unnecessary.

I don't think anyone would disagree with that, but the author side-steps the most important dynamic here, which is whether there really are extreme risks from AGI. The closest we see to an argument for why we should share the author's implied assumption that AGI won't be a problem is just flippant remarks by framing those who argue that there are great risk are "believers".

The funny thing is that a lot of the ppl on social media that I've seen taking this stance against "longtermism" tend to also hold their own long term agendas about utopia. This "conflict" feels really contrived to me.

> health benefits are dubious

Are they? Certainly if you stand still for hours I'd be skeptical for that, but while you are standing at your desk you are able to do all kinds of movements that you are not while sitting. And you are also not able to slouch (which I personally do like crazy after sitting for only a few minutes).

Anecdotally I've been using a standing desk for a couple weeks. For most of this year I've been having difficult sleep due to my back several days a week, but in the last week I haven't. Just generally my back has been fine, which I haven't been able to say for a while.

Upgrading to a WI-FI mesh that worked well for my home has had the largest impact for me. I've always liked to be flexible with where I'm working in the house but many spots had very degraded WI-FI signals. Working in the backyard on a nice day, going upstairs and doing some laundry while in a more passive meeting or temporarily changing my "office" are all fine from a WI-FI perspective now.

I've also picked up a standing desk recently. While it is too soon to say, I think I'm already seeing a positive impact on my back. Getting that ergonomic desk setup is probably the more universally helpful thing.

Yeah, it is harder than other things, but if we can train a model to explain collections of pixels in human language then we might be able to do similar with collections of activations.

I don't know if that is the direction, but just an example that comes to mind easily.

If someone figures out how to do this, I think their models will be far more capable and reliable.