HN user
My father worked in the HPCL refinery in Chembur. I got to go visit on Republic day when I was a kid, but they stopped doing visits. He worked in the distillation tower at first, but then moved into diesel desulphurization. I wish it wasn't but its a dangerous job, and he narrowly escaped several accidents, including a horrible naphta fire that took many lives.
As a young grad student, I remember going to a talk by Bennett where he explained how a Quantum Computer allows manipulation in a 2^N dimensional hilbert space, while the outputs measurements give you only N bits of information. The trick is to somehow encode the result in the final N bits.
I felt this was a much better layman explanation of what a quantum computer does than simply saying a quantum computer runs all possible paths in parallel.
I feel like there’s a lesson to be learnt by reading Lord of the Rings and seeing what happens to Saruman and Denethor.
I don't think we get rid of the CPU. But the relationship will be inverted. Instead of the CPU calling the GPU, it might be that the GPU becomes the central controller and builds programs and calls the CPU to execute tasks.
What if we create a situation in a lab that can be labelled as a collapse of the wave function by interaction with a macroscopic object. Except the macroscopic object is under our control and we can reverse the collapse.
A quantum computer is such a macroscopic state.
There’s papers that “derive” Born’s rule from the many worlds interpretations, e.g. https://arxiv.org/abs/1405.7907
I don’t claim to understand them though. I have tried.
What you said here makes sense. Forgive me, but I have trouble even articulating what it is that I don’t understand correctly.
Maybe what I meant was this: if I perform a quantum experiment where the spin measurement of an electron could be spin up or spin down, the future me would end up in one of two branches: I measure spin up, or I measure spin down. There wouldn’t be any possible world where I measure a superposition of spin up and spin down, because such a a state is going to decohere rapidly. This makes sense. What I’m unable to grasp is that even though the wave function of the universe contains both branches, “I” somehow experience only one of the two branches.
The answer to that I guess if that the two branches are nearly orthogonal they will merrily evolve independent of each other. But somehow “I” only experience only one of them.
Sorry for the rambling. I’m not able to articulate what I don’t understand.
The part that I have trouble wrapping around with many worlds interpretation is how I as an observer end up in one of the many bifurcations. Any links you can share that will help me with understanding that is welcome!
The Stanford Encyclopedia of Philosophy (https://plato.stanford.edu/entries/qm-manyworlds/) goes into this in some depth, and it seems like the right way to think about it is say that "I" in one branch is a different entity than the "I" in a different branch. I have somehow not been able to grok it yet.
And I agree about the naming. I really dislike the name "many worlds interpretation", which seems to imply that we have to postulate the existence of these additional worlds, whereas in fact they are branches of the wavefunction exactly predicted by standard quantum mechanics.
I agree. My suspicion is that token efficiency is what will drive more efficient tool calls, and tool building. And we want that. Agents should rely less on raw intelligence (ability to hold everyting in context), and more on building tools to get the job done.
These seem ideal for robotics applications, where there is a low-latency narrow use case path that these chips can serve, maybe locally.
I had the same question.
Maybe the train is software that's built by SWEs (w/ or w/o AI help). Specifically built for going from A to B very fast. But not flexible, and takes a lot of effort to build and maintain.
I didn’t think of that, but you are right. At some point I thought I understood templates r-value references work but now I’ve forgotten.
Maybe a std::take to pair with it?
Maybe a compiler error that a const object cannot be “moved”?
That would force the programmer to remove the std::move, making it clear that its a copy.
How strict was your elimination of sugar? Did you find a gradual trend of your lipid profile over the course of a year, or was it more sudden?
Pre LLM agents, a trick that I used was to type in
auto var = FunctionCall(...);
Then, in the IDE, hover over auto to show what the actual type is, and then replace auto with that type. Useful when the type is complicated, or is in some nested namespace.
"I personally don’t touch LLMs with a stick. I don’t let them near my brain. Many of my friends share that sentiment."
Any software engineer who shares this sentiment is doing their career a disservice. LLMs have their pitfalls, and I have been skeptical of their capabilities, but nevertheless I have tried them out earnestly. The progress of AI coding assistants over the past year has been remarkable, and now they are a routine part of my workflow. It does take some getting used to, and effectively using an AI coding assistant is a skill in and of itself that is worth mastering.
Right, but it needs to be formalized.
The reason code can serve as the source of truth is that it’s precise enough to describe intent, since programming languages are well-specified. Compilers have freedom in how they translate code into assembly and two different compilers ( or even different optimization flags) will produce distinct binaries. Yet all of them preserve the same intent and observable behaviour that the programmer cares about. Runtime performance or instruction order may vary, but the semantics remain consistent.
For spec driven development to truly work, perhaps what’s needed is a higher level spec language that can express user intent precisely, at the level of abstraction where the human understanding lives, while ensuring that the lower level implementation is generated correctly.
A programmer could then use LLMs to translate plain English into this “spec language,” which would then become the real source of truth.
Found an article which talks about the phase margin, referencing 80 degrees (pi). https://en.wikipedia.org/wiki/Phase_margin
A wild guess as to what is happening. I haven’t actually tested this hypothesis so I could be completely wrong.
In feedback systems, the gain is a function of frequency, and typically decreases when going from low frequency to high frequency. This is often accompanied by a phase delay.
So if the overall gain of the system is high enough, there will be some high frequency where the gain is 1, and the phase is 180 degrees. This would result in positive feedback, amplifying noise at that frequency.
Maybe that’s what’s happening in the latest AirPods? If Apple is aggressive cranking up the gain of the noise cancellation system, there’s some high frequency where the noise gets amplified rather than suppressed.
The solution would be to either reduce the gain (which reduces the noise cancellation), or to add some differential gain in the system which pushes out the unity gain frequency to higher frequencies.
The scenario the author describes is bound to happen more and more frequently, and IMO the way to address it is by evolving the culture and best practices for code reviews.
A simple solution would be to mandate that while posting coversations with AI in PR comments is fine, all actions and suggested changes should be human generated.
They human generated actions can’t be a lazy: “Please look at AI suggestion and incorporate as appropriate. ”, or “what do you think about this AI suggestion”.
Acceptable comments could be: - I agree with the AI for xyz reasons, please fix. - I thought about AIs suggestions, and here’s the pros and cons. Based on that I feel we should make xyz changes for abc reasons.
If these best practices are documented, and the reviewer does not follow them, the PR author can simply link to the best practices and kindly ask the reviewer to re-review.
I do most of my banking and investments on my phone. I don’t think I’m in the minority here.
I've never understood the argument against lidars (except cost, but even that you can argue can come down).
If a sensor provides additional data, why not use it? Sure, humans can drive withot lidars, but why limit the AI to using human-like sensors?
Why even call it a crutch? IMO It's an advantage over human sensors.
Layoffs contribute to the average worker taking a paycut. And we are seeing layoffs even in a market that is soaring. Why do you think that workers wouldn't be affected during a downturn?
My guess is that more than the raw capabilities of a model, users would be drawn more to the model's personality. A "better" model would then be one that can closely adopt the nuances that a user likes. This is a largely uninformed guess, let's see if it holds up well with time.
I subscribe to this view. LLMs can be tools for thinking, not substitutes to thinking. In fact, I feel that with time, the way we think may fundamentally change, just like how reading and writing changed thinking. An LLM can be an ideal foil to test out ideas, and the process of thinking could be an iterative process with the LLM as an active participant.
My concern is more about being questioned or unnecessarily held at the border.
That would be super cool if it works! I’ve also wondered the same thing about activation functions. Why not let the algorithm learn the activation function?
Right. This makes sense. But why Fourier space in particular. Why not, for example, a wavelet transform.