i_dont_know_

I realized it's so much worse than a filter bubble... there are whole worldviews that friends and online communities are a part of that get effectively censored away, even if you say you would like to know what those people and communities are talking about.

Until recently I didn't realize meditation practices were as varied as they are, and I think part of that is when we hear about things like "the benefits of meditation" or "scientists studying meditation", it's treated as though anything where you close your eyes and try to clear your mind is the same activity. Now I'm pretty sure it isn't

Assuming a model is person-like, it gets even harder when we ask "who" the model is.

Is it this particular model from today? What if it's a minor release version change, is it a new entity, or is it only a new entity on major release versions? What about a finetune on it? Or a version with a particular tool pipeline? Are they all the same being?

I think the analogy breaks down pretty fast. Again, not to say we shouldn't think about it, but clearly the way to think of it is not "exactly a person"

I wanted to talk about Anthropic's "soul" document they include in Claude's prompt, some of the issues it might be causing, and point out what we're seeing now as we're seeing it probably isn't artificial consciousness so much as prompt adherence.

I'm actually quite surprized.

From another article today, I discovered the IRS has a github repo with (what seems to be) XML versions of tax questions... surely some combination of LLM and structured data querying could solve this? https://github.com/IRS-Public/direct-file/tree/main

I feel like these kind of things push us as a society to decide what exactly the purpose of school should be.

Currently, it's been a place for acquiring skills but also a sorting mechanism for us to know who the "best" are... I think we've put too much focus on the sorting mechanism aspect, enticing many to cheat without thinking about the fact that in doing so they shortchange themselves of actual skills.

I feel like some of the language here ("securing assessments in response to AI") really feels like they're worried more about sorting than the fact that the kids won't be developing critical thinking skills if they skip that step.

Maybe we can have

Thanks, the goal wasn't to get a full overview of the commercial space but really to understand more of the fundamentals of building, the limits and use-cases of agents in general.

Specifically I see a huge push to "build agents" and to "use agents" as building blocks for more complex interactions. I wanted to get a feel for why and my conclusions are more around that and not "what's possible"

I wanted to dig into the hype around agents and figure out both the use-cases and quirks, so I wrote this little 'review'.

I think, like most of the 'AI' sphere right now, they're overhyped but I think it's good to get a feel firsthand of when they're useful and when they're not.

I clicked - it's for a self-improvement week.

I read through and I think it's an interesting initiative (lots of good stuff around holding self and others accountable, positive reinforcement etc) but the title really makes it sound like you're recruiting for a toxic workaholic startup or something.

Really good point on open source and the nudges it provides, and definitely a point that isn't made often enough!

Thanks for the perspective. For me I think it's a matter of degree (I guess I was a bit "one or the other" when I wrote it).

These things are also concerns and definitely shouldn't be dismissed entirely (especially things like AI telling you when it's unsure, or, the worse cases of propaganda), but I'm worried about the other stuff I mention being defined away entirely, the same way I think it has been with privacy. Tons more to say on the difference between "how you use" vs "how you share" but good perspective, and interesting that you see the emphasis differently in your experiences.

True... I was trying to define them the way (I think) companies are defining them (like what their alignment teams are looking at) and the way it's reported. I think in these specific contexts they're used with overlap but yeah I do bounce back and forth a bit here.

Thanks! Yeah, there's a lot of "well, it's 'standard practice' now so it can't be wrong" going on in so many different ways here too...

I keep seeing "AI ethics" being redefined to focus on fictional problems instead of real-world ones, so I wrote a little post on it.

I tried to create an "intuitive" explanation of LLMs without going into the math but still removing a little of the "black box" sense people get around them...

I think there's something about the way this is written that's endemic to our time -- it basically feels right because it gives a possible explanation of current events that has some internal consistency. It doesn't mean it's right, but neither the author nor (most of) the audience care about that part.

I feel like politics will be automated eventually. We have voters who express intents, then we have politicians who are (in theory) hired to best represent and interpret those intents.

But there's only 1 signal (election/not-election) and it's only delivered every few years. Yes, there's some voter feedback in the interim sometimes, but unless there's also a lot of press and noise around the voter feedback, politicians are unlikely to do anything with it because it probably won't impact re-election.

Ideally, there would be much more interactive feedback between voter and political agent (human or otherwise) carrying out voter intents, especially around informing of unintended repercussions of policy decisions and ultimately forcing voters to have more nuanced policy views.

At least that's what I hope for. What reality says is usually quite different.

"In the first year" is a good and well-needed qualifier for such analysis.

Oftentimes we fall into the trap of defining a business success as "forever", anything short of that is deemed a failure. I think such views make the industry look far bleaker than it actually is.

Even by their standards, I'm not sure that 'running a business that did well for 10 years' should count as a fail in the same way that 'couldn't even figure out a business plan' should.

I am the right fool... A random internet link said "can we read all your email" an I clicked 'yes'.

Welp. Time to do some major soul-searching and digital triage.

I don't know if this is a myth that ought to be perpetuated.

The skilled CEOs I know are all exceptionally well-attuned to the problem space that their product solves, and their own company's place in that space (usually including internal company dynamics). Yes, from an employee's perspective, the CEO is pretty magical -- they usually have a better 'big-picture' view of the task you are trying to accomplish than you do, and they can give you the right context, validation, resources etc. to totally turn your work life around.

However, this is usually limited to the problem-space and the company itself. If you're in the same problem-space or dealing with similar company problems then a conversation can be quite valuable. But it doesn't sound like that's what you're looking for so much as inspiration.

You might get that from talking to a CEO. They can tell you the situation and circumstances that set them on their particular path, and maybe you can extrapolate nuggets of wisdom out of that. But they won't be able to figure out the path that's right for you. They're not oracles or mystics, just people with the right intersection of skills deeply specialized in the right cross-section of the market.

To me, it feels like it's started giving superficial responses and encouraging follow-up elsewhere -- I wouldn't be surprized if its prompt has changed to something to that effect.

Before, if I had an issue with a library or debugging issue, it would try to be helpful and walk me through potential issues, and ask me to 'let it know' if it worked or not. Now it will try to superficially diagnose the problem and then ask me to check the online community for help or continuously refer me to the maintainers rather than trying to figure it out.

Similarly, I had been using it to help me think through problems and issues from different perspectives (both business and personal) and it would take me in-depth through these. Now, again, it gives superficial answers and encourages going to external sources.

I think if you keep pressing in the right ways it'll eventually give in and help you as it did before, but I guess this will take quite a bit of prompting.

A huge difference between ChatGPT and crypto, though is that the latter literally offers to print money.

The former, what, lets you make low-quality internet content a little more easily? It's harder to share the long-term 'get-rich-quick' benefits (though, of course many are trying right now). There's an initial bump right now as various content mediums try to cope with ChatGPT content, but once that stabilizes, I don't see the long-con grift that the author is mentioning being possible in the same way.

I think what I'm getting at is that there are different types of 'pain', some that we'd consider ethical and some that we wouldn't.

Constantly subjecting a person (or some abstract simulation that responds like a person) to the equivalent of continuous bodily pain would be deeply unethical, but, say, giving a person clues towards solving a puzzle would be considered less so.

Overall, I think you're right though, if we somehow discovered we've created simulated people (or sentient beings) then we probably shouldn't use them to solve arbitrary problems.

K, I'll play. Let's say that Reinforcement Learners (algorithms/strategies/agents in Reinforcement Learning), let's say that they have some property of 'consciousness' that's similar to humans.

A 'reinforcement learner' gets positive or negative feedback and adjusts its strategy away from negative feedback and towards positive feedback. As humans, we have several analogs to this process.

One could be physical pain... if you put your hand on a stove, a whole slew of neural circuitry comes up to try and pull your hand away. Another could be physical pleasure, you get a massage and lean in to the pressure undoing the knots because it's pleasurable.

If we look at it from this angle, then if we're metaphorically taking the learner's hand and putting it continuously on the stove, this would be problematic. If we're giving it progressively less enjoyable massages, this would be a bit different.

Even more different still is the pain you feel from, say, setting up an experiment and finding your hypothesis is wrong. It 'hurts' in some ways (citation needed, but I think I've seen studies that show at least some of the same receptors fire during emotional pain as physical pain), but putting a human in a situation where they're continuously testing hypothesis is different from a situation where their hands are being continuously burned on a hot stove.

I think, then, that the problems (like they alluded to here) are:

- how can we confirm or deny there is some kind of subjective experience that the system 'feels'?

- if we can confirm it, how can we measure it against the 'stove' scenario or any other human analogue?

- if the above can be measure and it turns out to be a negative human scenario, can we move it to one of the other scenarios?

- even if it's a 'pleasurable' or arguably 'less painful' scenario, do we have any ethical right to create such scenarios and sentiences who experience them in the first place?

This too.

I've seen consultants hired for projects where everyone on the team (including leadership) basically knew what to do or what needed to be done, but they were only, say 75% certain about it. In a normal startup, they would've just done it, failed fast, moved on. But in a bigger corporate (non-startup) environment where they didn't know if their jobs would be on the line for taking risks, they could bring in consultants to verify but mostly echo back what they already wanted to do. Plus they come with certificates and stuff so everyone in the chain assumes that's worth something.

These are actually the same engine underneath, though, aren't they? Just with slightly different prompts? (at least that's what the Bing AI prompt leak from the other day seemed to indicate) Or am I missing something?

The whole point of Section 230 was that we wanted to simultaneously incentivize user-content platforms (so, services that just allowed people to freely express themselves without concerning itself too much about the content) while simultaneously encouraging those services to take down illegal content. That was it.

The whole reason it was needed was because these platforms were worried that if they took down content, it might be seen as implicitly endorsing the content that remained. If that was the case, then they'd be liable for the content that remained. Including things like libel laws... it would be the same as a newspaper printing that content.

They didn't want that, so this was the compromise. You can do light moderation and we'll grant you exemption.

Fast-forward to today. Facebook and Twitter do not do "light moderation". They decide what you see and what you don't... they direct your attention from smaller stories to bigger ones, they spend countless thousands of employee hours catering to the whole experience. If that doesn't count as a modern digital scaled form of editorialization, I don't know what would.

I think a law designed to encourage very light moderation in no way applies to full-fledged algorithmic determination, and that distinction needs to be made and clarified.

Totally this. I moved away from the US many years ago, and having the luxury to go a few months without work and not have to worry about things like health insurance is phenomenal. You can find a new job at your own pace, when you're actually genuinely excited about it.

Also, there are enough other people around you who are taking their time that you don't feel like garbage for not having a job (somehow in the US, even when I was more than financially secure, I felt like it was 'not ok' to not be working)

That PT Barnum... it's interesting because it really is a good read, and the whole way through you do feel like he's giving you some kind of good advice (so, if you had purchased it back in the day you don't feel ripped off), but at the same time it's not really practically actionable. You won't actually get closer to being like PT Barnum (even if you were reading it in the 1800s) or understand the steps to becoming like him, or necessarily become any better at making money. But at the same time, you're satisfied enough with the read to feel like you hadn't wasted your time and money on it. That's his real trick :)

I think you're on to something here... the target audience of this product is the hardcore meat-eaters, hoping they'll just do a drop-in replacement. But I think the actual target should be the people who are meat-eaters considering quitting.

Those people are more likely looking at health benefits, and would probably be willing to tolerate some small issues here and there if it was definitely healthier. It looks like they prioritized taste over health concerns, meaning those 'early adopters' rejected the product, so it never makes it to their core segment.