I think he is being over dramatic. In the space of about four years, LLMs progressed from mediocre high school student to Ph.D. graduate in every field. That's impressive, but there is no evidence yet they can outperform or outsmart humans. Their biggest advantage for tasks such as proving theorems or long coding sessions is that they don't get tired.
I have yet to see this in my field. Maybe like a PhD student who bullshits their way through. LLMs still can't make correct decisions, only as useful as the person who uses them. To me, LLMs are only useful for making some mundane tasks faster.
You arguably need a lot more intelligence to assemble furniture.
Sometimes it gets wrong things that I had spelled out already.
It may be the Doomsday machine, but it is a very silly one. If it kills humans it will do so by mistake.
"You are completely right! Humans cannot breathe sulfur dioxide! My mistake, and I take complete responsibility"
Can you give a simple example?
I would have agreed 2 years ago, but it's extremely rare I see a frontier model making a silly mistake these days.
To be fair it is no where near what we had just one year ago and the rate of change only seems to be increasing.
Also, I don’t have 30 million dollars to spare spawning tens of thousand sub agents like what they did with Navier-Stokes so I’m clearly not testing the full capabilities of these models.
It was to build an extremely simple job using an internal framework to walk through a table and log the ids os some records that have a certain scenario.
There's a ton of jobs exactly like this in the codebase, and the framework code is in the codebase as well.
It was so silly I even thought of writing it myself, probably took me longer to steer claude to do it for me.
Anyway, it refused to use a method from the framework to retrieve the parameter as a list, it wanted to retrieve it as a string and parse the commas. I had spelled out in the initial prompt what method it should use.
I really don't like Claude much. Frontier my ass.
Earnest question: by what mechanism that exists today would the achieve that in a way humans on top top of the situation could not curtail?
All of this runs on top of compute in meatspace that humans can disconnect.
Here's my answer, as a non-superintelligent human: "see to it that the humans on top of the situation have a compelling financial interest in the systems not disconnecting".
In nuclear engineering, where safety is taken seriously, it's not enough to end the conversation at "the humans in charge can always simply shut down the reactor during a meltdown" or "a meltdown has never happened before, so we don't have to design safety systems before one does".
If you turn off the power cooling a data center, the servers within rapidly stop doing any computing.
Positive feedback loops are dangerous. Negative ones self-regulate.
For example, if the head of cyber security at your company suggested there's no need to worry about hacker infiltration or worms because one can always unplug one's computer as the primary defense mechanism, you might find that a little lacking. Will you be able to unplug the computer before the damage is done? Will it spread to other systems before you detect it? How will you unplug the computer if the attack is from an external facility? What if an attack happens but the boss says the computers have to keep running because an important customer is monitoring uptime? What if the attack goes unnoticed because it looks like a benign service?
Now imagine the head of cyber security answers by saying "actually you don't even need to unplug them, you can just wait for the computers to overheat, thus solving all concerns."
We have your daughter / dog / Amazon delivery. If you ever want to see her / him / it again, plug this USB drive into the control panel at your station / let off the parking brake roll your car into this substation / change the meatpacking thermometers to read 8C lower than calibrated / ground your vessel on this sandbank / send an envelope of white powder to these addresses / set fire to the following hospitals / …
So far around one in 350,000 PhD math grads solve a millenium prize problem (Perelman).
I would say exponential.
I mean, unless you see clear reasons for them to stop getting better _right now_, this is not very comforting.