Here is something that hints strongly at how human scientists and engineers are already doomed by AI. 🧵
I noticed this tonight while using Grok for technical research. I asked it a complex question and Grok understood it completely and gave a sophisticated and highly believable answer, but when I asked for specific references so I can write it into a paper for a journal, none of the references Grok provided exactly support the answer it gave me. Instead, they hint at something deeper.
In this case, I am quantifying the loss of signal margin in a Moon-Earth communications link as a function of how many times you landed near the communication system so the rocket plume sandblasted the electronics' thermal coatings, causing them to operate hotter than designed. There is a real cost to sandblasting your hardware on the Moon, and I am trying to quantify it.
Grok gave me many quantified effects, including that the frequency oscillator will drift about 10 to 50 ppm per deg C of temperature rise outside its operating range and that the Signal to Noise Ratio of the overall communications link will drop about ~0.1–0.5 dB for small drifts (<10 kHz) in particular modulation schemes. This is a great result that I can use to quantify sandblasting damage on the Moon, and the result is totally plausible, but it doesn't appear in ANY reference that Grok provided. Nothing discusses this.
So I suspect Grok actually derived that relationship itself during the LLM training. I think the relationship is probably correct, because the many references hint around the edges of this relationship in the right magnitude. I think Grok noticed the patterns of many performance metrics including temperature, input power and frequency, outputs, etc., for many devices and how they are connected in typical systems, and it stored as a higher-level symbol the result that you get 10 to 50 ppm per deg C performance loss. I think it solved that during training as it sought the higher-order symbols to store everything it had learned. IOW, its learning process included a heckuva lot of valid inference on these technical issues, and it now knows more about the performance of communications equipment than even the published literature knows.
I asked Grok if this is true, and it says it is correct (screenshots).
/1
2/ I then asked Grok to derive this relationship the same way it probably did during the LLM training, and it did. So now, if I want to use this key result in my paper, I have to use the many references that Grok used when it derived the relationship, and I have to show the derivation explicitly in the paper, or I can't publish it per the rules of scientific publishing (which of course were created in the days before reliable AI, and we still don't have totally reliable AI, but we can see it is coming fast).
3/ So here is the derivation, which it says replicates the process it did during its LLM training, which led it to believe in the quantified relationship between frequency shift and signal to noise ratio. I'm including this just to show its character. 
