as the "AI trade" unravels and Wiley Coyote discovers that he is, in fact, over the cliff, i want to recall how the narrative has shifted over the last few years.
-
as the "AI trade" unravels and Wiley Coyote discovers that he is, in fact, over the cliff, i want to recall how the narrative has shifted over the last few years.
-
as the "AI trade" unravels and Wiley Coyote discovers that he is, in fact, over the cliff, i want to recall how the narrative has shifted over the last few years.
remember when "prompt engineer" was gonna be a thing? that's because everyone thought these models were capable of anything, you just had to prompt them right.
-
remember when "prompt engineer" was gonna be a thing? that's because everyone thought these models were capable of anything, you just had to prompt them right.
then it turned out they weren't capable of anything, no matter how hard you prompted them, so we were told they would inevitably get better as they got bigger.
-
then it turned out they weren't capable of anything, no matter how hard you prompted them, so we were told they would inevitably get better as they got bigger.
then as they got bigger, it turned out they had problems with just making shit up-- problems that were fundamental and based on the design of the technology, not the size of the models. but we were told that techniques like retrieval-augmented generation (RAG) would fix this.
-
then as they got bigger, it turned out they had problems with just making shit up-- problems that were fundamental and based on the design of the technology, not the size of the models. but we were told that techniques like retrieval-augmented generation (RAG) would fix this.
well, it turned out RAG was way more finicky than expected, but then we were told that tool-calling, MCP servers, and "reasoning" models were going to **finally** make LLMs genuinely useful.
-
well, it turned out RAG was way more finicky than expected, but then we were told that tool-calling, MCP servers, and "reasoning" models were going to **finally** make LLMs genuinely useful.
then "reasoning" turned out to be bullshit, plus the bigger models were insanely expensive to train and basically not much better than the previous ones, so we arrived to the final boss of hype: AGENTS.
-
then "reasoning" turned out to be bullshit, plus the bigger models were insanely expensive to train and basically not much better than the previous ones, so we arrived to the final boss of hype: AGENTS.
and it was agents everything! agents calling agents calling agents. agents supervising agents, agents **voting**, little agents checking big agents calling tools sending queries to vector databases in an endless loop, and if it didn't work? don't worry, we will add **more iterations**
-
and it was agents everything! agents calling agents calling agents. agents supervising agents, agents **voting**, little agents checking big agents calling tools sending queries to vector databases in an endless loop, and if it didn't work? don't worry, we will add **more iterations**
and that's the state of the technology now as the financial markets have finally decided "enough."
-
and that's the state of the technology now as the financial markets have finally decided "enough."
i am pretty sure they'll bounce back from this one and maybe have another run at reinflating the bubble this summer, but there's honestly nothing else on the horizon. everyone's just building giant, insanely-complicated webs of agentic workflows around LLMs and pretending it might eventually work. no one talks about AGI anymore, no one really thinks any of this is intelligent, even the top researchers are saying LLMs is a dead end.
-
i am pretty sure they'll bounce back from this one and maybe have another run at reinflating the bubble this summer, but there's honestly nothing else on the horizon. everyone's just building giant, insanely-complicated webs of agentic workflows around LLMs and pretending it might eventually work. no one talks about AGI anymore, no one really thinks any of this is intelligent, even the top researchers are saying LLMs is a dead end.
maybe they will convince investors to shoot some servers into space and **that** will be the last, stupid, crazy gambit that everyone falls for. but the final reckoning is near.
-
and it was agents everything! agents calling agents calling agents. agents supervising agents, agents **voting**, little agents checking big agents calling tools sending queries to vector databases in an endless loop, and if it didn't work? don't worry, we will add **more iterations**
@peter when I first heard about “agents,” I was like “oh they’re doing ensemble methods again.”
In older methods, ensembles get you nice performance gains—a few percent, usually. Idk how much they improve LLMs because no one measured it.
-
@peter when I first heard about “agents,” I was like “oh they’re doing ensemble methods again.”
In older methods, ensembles get you nice performance gains—a few percent, usually. Idk how much they improve LLMs because no one measured it.
@thedansimonson it definitely makes it **feel** like it's working better lol
-
maybe they will convince investors to shoot some servers into space and **that** will be the last, stupid, crazy gambit that everyone falls for. but the final reckoning is near.
@peter great writeup, thank you, you voiced the feelings I had but couldn't put into words exactly
-
i am pretty sure they'll bounce back from this one and maybe have another run at reinflating the bubble this summer, but there's honestly nothing else on the horizon. everyone's just building giant, insanely-complicated webs of agentic workflows around LLMs and pretending it might eventually work. no one talks about AGI anymore, no one really thinks any of this is intelligent, even the top researchers are saying LLMs is a dead end.
@peter So it is like SARS-CoV-2 variants.
-
@peter when I first heard about “agents,” I was like “oh they’re doing ensemble methods again.”
In older methods, ensembles get you nice performance gains—a few percent, usually. Idk how much they improve LLMs because no one measured it.
@thedansimonson @peter there’s a bit of that, but another way to see it (and a reason it kinda works better), is that it’s just introducing a large deterministic layer (the tools the agents use), and (in the few cases where agents outperform) a sort of cross-validation method (whatever measurable task the agents are assigned to). In any case, we’re back to good ol ML with a fuzzy layer and lots of hype…
-
@peter So it is like SARS-CoV-2 variants.
-
maybe they will convince investors to shoot some servers into space and **that** will be the last, stupid, crazy gambit that everyone falls for. but the final reckoning is near.
@peter
From 1634 to 2026... The story may be different but the plot remains the same #TulipMania
https://en.wikipedia.org/wiki/Tulip_mania -
remember when "prompt engineer" was gonna be a thing? that's because everyone thought these models were capable of anything, you just had to prompt them right.
@peter no wanting to derail your thread (it was great, by the way) but your "prompt engineer" made me remember what I saw in a newsletter this morning: "Why Your LLMs Need CONTEXT ENGINEERING" (all caps mine for emphasis)
I was and still am laughing out loud
-
@peter no wanting to derail your thread (it was great, by the way) but your "prompt engineer" made me remember what I saw in a newsletter this morning: "Why Your LLMs Need CONTEXT ENGINEERING" (all caps mine for emphasis)
I was and still am laughing out loud
@iclavijos they are frantic to come up with a New Thing lol
-
@iclavijos they are frantic to come up with a New Thing lol
@peter hey, maybe there's a business opportunity in there. We must come up with some made up titles (Hallucination Detector Engineer, for instance), trademark them and voilà!
