LLMs are training on what people have WRITTEN about a topic, not by observing people practicing whatever the topic is.
So when there's a gap between theory and practice, AI will completely align with theory.
So when there's a gap between theory and practice, AI will completely align with theory.
Sep 14, 2026 · 12:05 AM UTC
84
101
985
23,924
@JohnDCook What terrifies me is that LLMs have been trained on religious text and many of these texts hold that a just and loving God would condemn some humans to be tortured for all eternity.
1
0
2
176
@JohnDCook Ha -- and that's how we win!
"In theory, theory and practice are the same. In practice, they are not" A.Einstein, most likely.
"In theory, theory and practice are the same. In practice, they are not" A.Einstein, most likely.
2
0
16
517
@JohnDCook That was the first step, now they are training on user/agent interactions with LLMs, which is essentially *the* practice. Of course still works best with coding or math than anything else (immediate effect).
1
0
8
391
@JohnDCook Not just what they've written, but what they've written publicly.
For example, there is reams and reams of good writing about flight safety at Boeing and Airbus. But AI is trained on plane-spotter reddits.
For example, there is reams and reams of good writing about flight safety at Boeing and Airbus. But AI is trained on plane-spotter reddits.
1
0
5
140
@JohnDCook I would respectfully disagree. Autonomous cars, robots, cameras and simulation. We have labs and factory floors and R&D plus real world interaction beginning to scale. There is a whole world happening we don’t see day-to-day but it exists.
5
0
4
375
@JohnDCook LLM's provide better search. The whole notion of AGI is about selling not truth.
0
0
3
73
@JohnDCook This is only partially true, likely a lot of model training data is now coming from things like YouTube videos and Podcasts. Additionally, I'm sure the major labs are actively working to get more and more real-world audio/video recordings they can use for training.
1
0
2
106
@JohnDCook Yes, I've discussed this problem with LLMs. True masters (tracking, wildlife behavior, working dogs, intuitive archery, hunting/trapping) aren't the type to write books... and the people who do know very little. LLM + Robotics + Sensors will fix this eventually.
0
0
1
88
@JohnDCook there is a weird, slightly cannibal loop on top of that, where the post LLM internet is largely people using LLMs, which then becomes reference/training for the LLM models
1
0
2
87
@JohnDCook And what's been deemed to be 'good enough' to publish, or what the winners wanted to be the official and last word on a subject. Published works are just the tip of the iceberg of human knowledge.
0
0
2
228
@JohnDCook They're artificial wordcels.
0
0
1
7
@JohnDCook That's why humanoid robots are obvious next breakthrough in AI.
0
0
1
12
@JohnDCook Yes, which is why they will do well in junk science fields, especially medicine.
0
0
1
9
@JohnDCookis is one of the reasons egocentric data collection (via glasses, etc…) is so valuable. I’d even argue Meta’s EMG wearable provides them with valuable training data by capturing small muscle movements on the dominate hand (works even when the hand is outside the camera view).
0
0
1
194
@JohnDCook we gotta turn this fact into some kind of personal-agent discussion platform! This is how we could scale the knowledge that is needed to update the current software for the coming swarm
optimally even monetised, but I don't really see yet how this could work
optimally even monetised, but I don't really see yet how this could work
1
0
1
361
@JohnDCook No they're looking at your transcripts to find out where it didn't work.
They will get better
They will get better
1
0
1
162
@JohnDCook Yeah, writing my emails and creating my documents assuming AI will read a reuse the data.
0
0
0
28
@JohnDCook The second L stands for Language.
0
0
0
1
@JohnDCook Video and audio feeds from smart devices all around us beg to differ?
0
0
0
31
@JohnDCook lots of people have written about their experiences. Lots of videos show their experiences.
0
0
0
51
@JohnDCook Is this based on observing practical responses or what you've read?
0
0
0
35
@JohnDCook In theory, theory and practice are the same; but in practice they're not.
0
0
0
4
@JohnDCook incredible to constantly read AI takes that haven't been even vaguely factually correct for like 3-4 years.
0
0
0
35
@JohnDCook Thus it is also influenced by who writes, who is allowed to write, and which writing sources are used as training data.
Or as the old expression goes, garbage in, garbage out.
Or as the old expression goes, garbage in, garbage out.
0
0
0
26