gsuuon

I always thought furniture music was such a pragmatic description of his work. Every few years I make a half-hearted attempt to learn Gymnopedie 1 on guitar but can never seem to follow through.

I will now write the corrected code.You are right. I have been making a mess of this. I will get it right this time.

It keeps putting import statements in the middle of files or duplicating edits at the wrong places, but it's hard to get frustrated since it seems pretty self-aware. I didn't say it was making a mess - it's just being hard on itself.

A short session ended up sending over 3 million tokens though - wonder how the economics of this work out for google?

Pretty sure OP was talking about OpenAI expiring their credits (just had mine expire).

Btw, unsurprisingly, the time for expiry appears to be in UTC in case anyone else is in the situation of trying to spend down their credits before they disappear.

This makes sense, effective post-mortems don't focus on assigning blame. They try to identify and fix the problem. The issue with AI is that it's such a black-box right now, this process is likely not feasible. If an AI makes a decision that turns out to be wrong, there's no reliable way to identify and fix the problem. You can prompt engineer or finetune or re-train the model, but in the end you can only hope that issue has been fixed. I think the black-box nature of AI makes it different from other systems we use.

I think this could be one of the more legitimate uses of blockchain - distributed communications, contacts, and a refundable pay-per-call system to make spam calling uneconomical. Communication in general does desperately need an overhaul, phones are effectively useless as phones nowadays.

Is it just me or does it seem like the Catholic church might have a better grasp on technology than the US government?

  46. While responsibility for the ethical use of AI systems starts with those who develop, produce, manage, and oversee such systems, it is also shared by those who use them. As Pope Francis noted, the machine “makes a technical choice among several possibilities based either on well-defined criteria or on statistical inferences. Human beings, however, not only choose, but in their hearts are capable of deciding.”[92] Those who use AI to accomplish a task and follow its results create a context in which they are ultimately responsible for the power they have delegated. Therefore, insofar as AI can assist humans in making decisions, the algorithms that govern it should be trustworthy, secure, robust enough to handle inconsistencies, and transparent in their operation to mitigate biases and unintended side effects.[93] Regulatory frameworks should ensure that all legal entities remain accountable for the use of AI and all its consequences, with appropriate safeguards for transparency, privacy, and accountability.[94] Moreover, those using AI should be careful not to become overly dependent on it for their decision-making, a trend that increases contemporary society’s already high reliance on technology.

I also really liked:

  62. In light of the above, it is clear why misrepresenting AI as a person should always be avoided; doing so for fraudulent purposes is a grave ethical violation that could erode social trust. Similarly, using AI to deceive in other contexts—such as in education or in human relationships, including the sphere of sexuality—is also to be considered immoral and requires careful oversight to prevent harm, maintain transparency, and ensure the dignity of all people.[124]

Really interesting read overall, thanks for sharing.

I'm confused, is he saying that the other voice on the call is google assistant voice ai? Or the assistant just routed the call through the google number?

I'm really hopeful for e-ink or low-fidelity devices to help ween us off media addiction. Hopefully Nothing pursues something in that space since it aligns with their mission. Would love to switch most of my work screens to e-ink and only have 'normal' screens for explicit recreation time.

I was impressed this is open source, then impressed it was done by one person, then impressed it only took 6 months, and eventually somehow impressed again that it was a _high schooler_. My mind is blown. Kudos for managing to do something insane like this. Very inspirational.

I tried this via the chat website and it got it right, though strongly doubted itself. Maybe the specific wording of the prompt matters a lot here?

https://gist.github.com/gsuuon/c8746333820696a35a52f2f9ee6a7...

I'm a little suspicious of the Isaac Newton example. The values of the better answer are very close, I wonder if the ordering holds up against small rewordings of the prompt?

Another approach if you're working with a local model is to ask for a summary of one word and then work with the resulting logits (wish I could find the article/paper that introduced this). You could compare similarity by just seeing how many shared words are in the top 500 of two queries, for example.

I wonder if "build it and they will come" is just flat out wrong, or only correct for certain products? Is there anything one can "just build" now and expect some market adoption?

Fantastic that progress is being made on this. Hopefully it's enough to stem the tide, though consumer behavior wrt calls has probably already fundamentally shifted. It'll take a long _long_ time before folks who have stopped picking up are comfortable answering a random call again.

The ability to edit parts of the model after the fact using prompts is pretty incredible

This is generally a pro for me. I can skip to the bottom of a file and see what the main purpose of it is, knowing that it's the only place that could reference everything else in the file.

This was cool (spacelancers didn't work in Arc but did in Chrome, forest demo 403's) - audio and fullscreen fail because they aren't initiated by a user gesture. Impressive how quick it is to get into a game with this level of fidelity, normally you'd need several minutes of downloading.

~9% of young people leave Bhutan

It's worse than that:

9% of the country's population, most of them young people

Young people want adventure, but all their homeland is offering is contentment. They need to account for the desire for opportunity in their GNH metric.

I really like this idea and have (briefly) tried something similar with a 'personal' discord server. I think this sort of thing would be great to rebuild local community over the internet instead of mostly faceless or parasocial interaction.

I default to never answering now. Every time I feel like it might be an actual person on the other end and risk answering, it has turned out to be a scammer or spam call. Not sure why governments don't do more against them - scam and spam callers have destroyed phone communication.

I (too recently) realized "shoulda, coulda, woulda"'s are generally useless, and to focus on only "need, want"s. If it doesn't fit into these two categories then it's not important enough to think about.

I'm really confused by this policy - isn't security and enforcement the value prop of the app store in the first place? Why are they offloading the work of accountability to user mob-rule by basically saying "here's where they live, go get them"? Who does this actually benefit? No sane disgruntled user is ever physically showing up at a developers home.

#5 is the only one I do, but because I never thought about it as a fear mechanic - I'll definitely avoid doing this now. That said, what's a good alternative? Sometimes you don't have time to bargain, is picking them up kicking and screaming actually better?

This is nice - I for one appreciate the results not being ranked by points, since some interesting stories don't get any traction.

I made a neovim plugin awhile back with that explicit purpose: being a toolbox to assemble your own AI stuff with. I struggled a lot to make AI useful and was hoping if the tools were there to make it easy to play around with, folks would figure out how to utilize LLMs effectively and share their results. Til now though, I'm not sure if anyone's even using the plugin beyond the starter prompts (which were only meant as examples). Maybe the API sucks.. idk.

The docs say that cache_control is just ignored[1] if less than 1024 tokens, maybe it's a bug if it's erroring instead?

[1] https://docs.anthropic.com/en/docs/build-with-claude/prompt-...

".. and let that be a lesson to the rest of you"

There should be some sort of side-project tinder for starter engineers and finisher engineers.

I really like her content, especially the earthship videos. The 'home engineering' that goes into these buildings is awesome. Though I'll probably never get to build one, would happily play a game based on this (like some sort of physics based survival game? Is this a thing?)

This is (well -- ad-libs) what I based the name of my fill-in-the-blank-with-llm ts library on https://github.com/gsuuon/ad-llama/

Sometimes complex topics are really like rubric's cubes - some changes here break things over there and then you need to make a bunch of turns to fix things elsewhere. Thinking through writing is necessary for these, because they look much simpler until they aren't and all the gory details start tripping over each other. The unfortunate part is that it feels very difficult to 're-enter' the topic as if reading for the first time, so the writing can easily become difficult to understand for a fresh-reader since it was edited by someone who's read it dozens of times in various incarnations and orders.

1) that sounds like a Montessori school? 2) I feel like Walter White is one of the more memorable character names (w/ the alliteration, no?)