lpearson.bsky.social
70 posts
16 followers
53 following
Active Commenter
comment in response to
post
Hotel Keurigs DO work for making hot water. Works great with a travel aero press and good grounds.
Side note: coffee grounds + TSA can flag a security check. Where they open the bags and then everyone develops a sudden strong urge for coffee.
comment in response to
post
I think that’s the exception that proves the rule.
Experts can use gen AI to speed up workflows because they can easily assess output quality and course correct as needed. It’s like hiring staff you will supervise. It’s when we treatAI *as* the expert that it collapses.
comment in response to
post
OR the polls don’t ask questions well in order to actually understand what people really think about complex issues and people almost always substitute easy questions for hard/complex ones.
comment in response to
post
See: Venmo default settings inexplicably being public
comment in response to
post
Yes!
It’s so frustrating hearing people talk about LLMs like they are pseudo-human, when they are just advanced mathematical models trained to put their output into natural language.
No true or false, only mathematical similarity.
comment in response to
post
Probably some real questions about how we treat celebrities as abstract concepts vs people (unless the research got his permission for this.) But it wasn’t people offering unsolicited opinions.
comment in response to
post
Context: it was an online researcher posting pictures of Ollie Murs (apparently an English singer) before and after a 12 week gyn program and asking men and women for their preferences between the two.
comment in response to
post
The problem is you still need to check details.
My first AI gaslighting: asked it to take Arizona Trail web info and make a table. It looked right, but the actual data were wrong. 2nd pass gave me random plant names. 3rd pass it said it never gave me that data. Context is better now, but still.
comment in response to
post
LLMs can actually be very helpful for surfacing information - the pattern recognition capabilities are immense and could be very helpful. (Obviously still with caution).
The problems seem to be with generative AI / outsourcing analysis.
comment in response to
post
FB and IG can still be good and useful tools, but the key seems to be the type / character of the groups and feeds. The useful ones seem to be firmly grounded in some real thing and have strict “no politics” rules.
comment in response to
post
Minor point of nuance: I’m not sure “like” is the right framing for this.
The political divide is roughly urban v rural. And at this point EVs work great in urban-ish areas but have significant cons in rural ones.
comment in response to
post
New Mexico is a wonderful state and the “Land of Enchantment” moniker is well earned.
And it’s very much *not* *Texas* as anyone who has made that drive can attest.
comment in response to
post
And can we discuss how not-in-Texas it is? The landscape is a huge character and it’s just so obviously NOT west Texas.
comment in response to
post
Heat injury is no joke. The older I get the more cautious I get about that line between Type 2 and Type 3 fun.
A hard workout / You feel like dying / brain working well? Good.
Too hot / very dehydrated / other people worry you could *actually* die? Bad.
comment in response to
post
I am really hoping this works better for economics than it has for climate science.
comment in response to
post
Hopefully they will roll those out as they are fairly straightforward to build. It’s just a RAG (retrieval augmented generation) model trained on internal documentation.
comment in response to
post
Yes - I know, but some prompt engineering techniques can improve the quality of outputs and reduce hallucinations. Linking an article of tips that seems decent (I just did a quick search and found something recent since my knowledge is ~year old.)
www.zdnet.com/article/how-...
comment in response to
post
Sometimes explicitly asking for sourcing in your prompt works. Asking LLMs to break things down / show their work is another tactic for decreasing hallucinations, though based on the responses it looks like they’ve started just building that in.
comment in response to
post
A small amount. Possibly enough to understand what it’s attempting to do and appreciate how little I know.
I also have a growing collection of generative AI gaslighting examples 😂
comment in response to
post
I like *using* all the mkdocs-material based sites, so that’s what we are trying to use.
But no software alive can make humans actually document stuff.
comment in response to
post
My expectation:
AI comes out of NLP (natural language processing) work and will be good for (usually simple) tasks where the computer needs to understand the ask and provide the response in human vs machine language. And there will still be lots of nuanced items that need actual humans.
comment in response to
post
Feature Request: add a “per capita” version, or at minimum add county population in the pop out box.
A lot of rural counties will be hit disproportionately hard, but that doesn’t show without any per capita normalization
comment in response to
post
And would love to see that map including population. Because a lot of those “white” counties likely have so few people they were already at 0.
comment in response to
post
To be clear: not calling you one. But the “it’s not *really* a gun” arguments are too cute and unconvincing here for regular-person litmus tests.
comment in response to
post
I 100% get the law needs to nail down particulars and edge cases, and that’s a hard and messy process. But when Hello Fresh sends a box of ingredients and a recipe, everyone knows they are sending dinner. And denying that’s what they intended is just acting a fool.
comment in response to
post
I was being a tad facetious. But considering most adults I know struggle with Ikea’s directions, and the Costco shed in pieces on my garage floor is almost certainly more difficult to put together, I don’t think it negates my point at all. The items are still bookcases and sheds. And guns.
comment in response to
post
I mean - sure. But “a kit with instructions my 5 year old could follow to put together” is WAY past the line for all the non-lawyer folks. That seriously feels like a “yeah right, nice try, go to your room” level of argument to make. (Obviously IANAL so just going off common sense.)
comment in response to
post
So the ghost gun manufacturer argument is that the vast majority of IKEA products aren’t furniture because they are sold flat-packed and you have to assemble it yourself? That’s really the argument they went with?
comment in response to
post
A “best use case” for AI are probably RAG (Retrieval Augmented Generation) models. To create: take LLM, train on internal documentation, provide search tool with sourcing enabled. Enables improved natural language queries of source materials to find inputs. Human still takes that to create output.
comment in response to
post
Nuance: generative AI (the kind you can chat with) is a flavor of AI. The Large Language Models (LLMs) behind it can be used for other purposes, and some they are better suited for.
I think a bit like dogs - problematic when set free, immensely useful when well trained and under human direction
comment in response to
post
You also have to know enough to spot when they provide non-truths.
LLMs don’t lie like humans do - they absorb bias from training but don’t have an inherent sense of moral truth to violate. But the process can easily produces false outputs, and it takes knowledge to spot those.
comment in response to
post
Yup. Walking up on Bison is always a little intimidating, but somewhere on the plains it feels right. On Catalina it’s just *weird*.
comment in response to
post
Thank you! I am hoping to get back into the Sierra this year with a basic trip plan that amounts to “wandering around in wonder”. I think that might be the perfect accompaniment.
comment in response to
post
I think they hide just around the corner in the fog
comment in response to
post
That is beautiful, and I will now go find more of his writing. Thank you for sharing.
comment in response to
post
I think of it often these days too.
Also my realization a few years ago that the Lord’s Prayer is as much a demand that we act with grace as a supplication to receive it.
“forgive us our trespasses **as we forgive** those who trespass against us”
Do unto me as I do to them.
comment in response to
post
I think disappearing people is bad. Full stop.
All the rest just raises the level of bad.
comment in response to
post
Things would get much better very quickly if we stopped asking AI to give us answers and instead treated it like a tool to source relevant information for our questions.
comment in response to
post
AI is very helpful at *surfacing* information. When used to generate responses it falls down because there’s no truth validation / judgement. It relies on the majority of info being correct. And should always be treated with at minimum a “trust but verify” stance.
comment in response to
post
d’oh. yeah. that’s a not fun lesson in variable naming for sure!
comment in response to
post
Yup. You learn quickly to make that find all a little more precise with a “reader =“
comment in response to
post
It’s like Strava mapping of bases, but for human vs physical resources.
comment in response to
post
I am sure that the cost/benefit analysis for WFH varies by job. But it’s mind boggling to me that the capability for remote work hasn’t been more strongly championed as a way to revitalize rural communities.
comment in response to
post
You can also take an LLM as a base model and create RAGs (retrieval augmented generation) by training them on your own corpus of documentation. Can work great for stuff like querying internal documentation - can envision a boosted LexisNexis. Just can’t trust them to do all the analysis.
comment in response to
post
So actually? You’re not far off for an actual improvement. Asking the LLM to cite sources is one way to reduce hallucinations (bonus: gives you what you need to fact-check.) Also breaking the prompt down into stepwise asks helps - a “show me your work” - really does improve outputs.
comment in response to
post
He’s not wrong if you define ‘education’ narrowly.
However the “ask the right questions” is doing a huge amount of lifting there. Because that’s the whole crux of the ball game and what people really struggle with.
The “curiosity and willingness to learn” also powerlifting here.
comment in response to
post
Also why it’s good to have people old enough to remember recent history.