Grumpy Monoceros
@grumpymonoceros.bsky.social
Madisonian, software engineer, and general nerd. I don't know what I'm doing here
That said I do agree that Amazon - and most of the other big tech companies - are massively overpromising what they're able to deliver right now and it's going to burn them before long because people just won't trust them.
December 23, 2025 at 9:56 PM
That said I do agree that Amazon - and most of the other big tech companies - are massively overpromising what they're able to deliver right now and it's going to burn them before long because people just won't trust them.
It may be true that "LLMs" can do it, but only with the right tweaking and context. An LLM is really just a software component, and developers have a lot of options that affect whether it will be "good" at something or not.
December 23, 2025 at 9:56 PM
It may be true that "LLMs" can do it, but only with the right tweaking and context. An LLM is really just a software component, and developers have a lot of options that affect whether it will be "good" at something or not.
Amazon also didn't do themselves any favors by aggressively badgering everyone to upgrade before they worked out all the issues.
December 23, 2025 at 4:21 PM
Amazon also didn't do themselves any favors by aggressively badgering everyone to upgrade before they worked out all the issues.
The bad news is that it's hard, and these are basically brand new systems rewritten from scratch. So people who upgrade early are seeing a lot of rough edges. Presumably Amazon/Google/etc will do the work to make them better over time but that's up to them
December 23, 2025 at 4:17 PM
The bad news is that it's hard, and these are basically brand new systems rewritten from scratch. So people who upgrade early are seeing a lot of rough edges. Presumably Amazon/Google/etc will do the work to make them better over time but that's up to them
The good news is that there's a ton of room to tune how they work - there's a lot that Amazon/Google/etc can do to make the experience better and more consistent; potentially so consistent you don't notice or care.
December 23, 2025 at 4:17 PM
The good news is that there's a ton of room to tune how they work - there's a lot that Amazon/Google/etc can do to make the experience better and more consistent; potentially so consistent you don't notice or care.
That randomness is how they are so good at understanding the fullness of human language, and you don't have to memorize specific keywords to use them like you did with old Alexa and similar systems. But they're not as deterministic as older systems.
December 23, 2025 at 4:17 PM
That randomness is how they are so good at understanding the fullness of human language, and you don't have to memorize specific keywords to use them like you did with old Alexa and similar systems. But they're not as deterministic as older systems.
A LLM works at a very fundamental level by just predicting the next word based on everything that comes before it (including instructions and the last word it said). There's a significant amount of randomness; that's why if you ask ChatGPT the same thing in 3 tabs you get 3 different responses
December 23, 2025 at 4:17 PM
A LLM works at a very fundamental level by just predicting the next word based on everything that comes before it (including instructions and the last word it said). There's a significant amount of randomness; that's why if you ask ChatGPT the same thing in 3 tabs you get 3 different responses
I wouldn't expect it to "lose" its core instructions after awhile (device makers get to decide what goes in the context after all). But figuring out what should be included to get a good result can be very tricky, especially when the user can ask for literally anything.
December 23, 2025 at 3:37 PM
I wouldn't expect it to "lose" its core instructions after awhile (device makers get to decide what goes in the context after all). But figuring out what should be included to get a good result can be very tricky, especially when the user can ask for literally anything.
Different models have different context limits but there is always some kind of limit.
December 23, 2025 at 3:37 PM
Different models have different context limits but there is always some kind of limit.
is it though? People were big mad about inflation during the Biden years even when average wages were increasing even more. The problem is that people attribute wage increases to their own hard work but price increases to a failure in The System.
December 17, 2025 at 10:03 PM
is it though? People were big mad about inflation during the Biden years even when average wages were increasing even more. The problem is that people attribute wage increases to their own hard work but price increases to a failure in The System.
I don't even think it's just pirated music, most of the Internet + smartphones have always operated on a freemium-with-ads business model (e.g. all social media). We've convinced ourselves that digital content is supposed to be free or very, very cheap
December 10, 2025 at 8:38 PM
I don't even think it's just pirated music, most of the Internet + smartphones have always operated on a freemium-with-ads business model (e.g. all social media). We've convinced ourselves that digital content is supposed to be free or very, very cheap
Or worse, trying to solve real problems and doing an extremely bad job of it (Google AI overviews)
December 9, 2025 at 11:37 PM
Or worse, trying to solve real problems and doing an extremely bad job of it (Google AI overviews)
Instead of made-up problems like "I can't write a fan letter to a favorite athlete" or "there's not enough short-form video for me to spend my day scrolling through"
December 9, 2025 at 11:37 PM
Instead of made-up problems like "I can't write a fan letter to a favorite athlete" or "there's not enough short-form video for me to spend my day scrolling through"
It's not an accurate, nuanced, or helpful conclusion, but I think if you're trying to market to the public we (tech people) need to do a much better job of demonstrating how it might actually solve real problems people actually have
December 9, 2025 at 11:37 PM
It's not an accurate, nuanced, or helpful conclusion, but I think if you're trying to market to the public we (tech people) need to do a much better job of demonstrating how it might actually solve real problems people actually have
If your primary exposure to LLMs is watching your social media feeds be overrun with low-effort slop, watching scams suddenly be a lot more convincing, and seeing OpenAI release what is basically AI TikTok, it's understandable how one might come to a highly pessimistic conclusion
December 9, 2025 at 11:37 PM
If your primary exposure to LLMs is watching your social media feeds be overrun with low-effort slop, watching scams suddenly be a lot more convincing, and seeing OpenAI release what is basically AI TikTok, it's understandable how one might come to a highly pessimistic conclusion
That sucks. I've never understood becoming emotionally invested in a game you don't like, as opposed to just...recognizing it's maybe not for you and just not playing it.
I hate Skyrim but I have precisely 0 desire to spend time crapping on Skyrim online to people who don't
I hate Skyrim but I have precisely 0 desire to spend time crapping on Skyrim online to people who don't
December 7, 2025 at 5:55 PM
That sucks. I've never understood becoming emotionally invested in a game you don't like, as opposed to just...recognizing it's maybe not for you and just not playing it.
I hate Skyrim but I have precisely 0 desire to spend time crapping on Skyrim online to people who don't
I hate Skyrim but I have precisely 0 desire to spend time crapping on Skyrim online to people who don't
What's the reasoning? I've become pretty disillusioned to steam reviews as a measure. You just starting with things that are uncontroversially good?
December 7, 2025 at 3:17 PM
What's the reasoning? I've become pretty disillusioned to steam reviews as a measure. You just starting with things that are uncontroversially good?
For example, a chess AI beat Gary Kasparov at chess in 1997.
Deep Blue versus Garry Kasparov - Wikipedia
en.wikipedia.org
December 4, 2025 at 7:09 PM
For example, a chess AI beat Gary Kasparov at chess in 1997.
Artificial Intelligence as a term has been around for decades if not more and has a very established (if vague) definition that has never required actual human cognition.
December 4, 2025 at 7:04 PM
Artificial Intelligence as a term has been around for decades if not more and has a very established (if vague) definition that has never required actual human cognition.
Helldivers has this problem. The community is filled with people who lose their minds when they struggle instead of just lowering the difficulty. It's infuriating
November 28, 2025 at 4:28 PM
Helldivers has this problem. The community is filled with people who lose their minds when they struggle instead of just lowering the difficulty. It's infuriating
Every time I reread this my soul dies a little more
November 24, 2025 at 5:43 PM
Every time I reread this my soul dies a little more
Are you just describing RAG or something else? It kinda does exist, it's just buried under a landslide of AGI hype and brainrot
November 24, 2025 at 2:07 PM
Are you just describing RAG or something else? It kinda does exist, it's just buried under a landslide of AGI hype and brainrot
All this to say - today, the most powerful models and infrastructure running most workloads are controlled by billionaires, but their stranglehold isn't as strong as one might think.
November 21, 2025 at 1:21 AM
All this to say - today, the most powerful models and infrastructure running most workloads are controlled by billionaires, but their stranglehold isn't as strong as one might think.
To an extent, but most of the research (& also many of the models) are actually freely available. The major innovation that enabled this stuff was a research paper from 2018 anyone can go read.
November 21, 2025 at 1:21 AM
To an extent, but most of the research (& also many of the models) are actually freely available. The major innovation that enabled this stuff was a research paper from 2018 anyone can go read.