HN user
It feels like having the memories explicitly stored by the agent itself as natural language is quite limiting.
Would it be possible to add one or several memory components to the LLM model itself while training it ? Example: https://en.wikipedia.org/wiki/Differentiable_neural_computer
I guess doing so would make the training less parallelizable ?
For some reason, it is unable to solve this query. All code produced gives wrong results, and it didn't correct it when given clues:
Write a python program that efficiently computes the probability that a random permutation of r ones and w zeros contains k consecutive ones.
Wow thank you for your work and providing it for free. Is there a way to submit correction PRs ? maybe a github ?
What are the killswitches acting on ? Are they acting only on the dedicated "enable" or "poweroff" inputs of various chips ? Are they cutting the power supply of the chips ? Are they cutting both the power and logic busses, thus isolating the target chips (at least electrically) ?
I remember finding OpenPCR too expensive for a student project, and I ended up designing and building a cruder/cheaper alternative for that project: http://2013.igem.org/Team:Paris_Saclay/PS-PCR/detailed_descr...
The Bretton Woods Agreement pegged the dollar to gold ($35 for an ounce of gold). I understand that this can work if the central bank emits $35 for every ounce of gold they store.
But how does that account for destroyed/damaged/lost dollar bills ? The loss is very hard quantify and monitor, but needs to be compensated either through re-printing of new dollars, or through the destruction of stored gold to maintain the desired exchange rate.
Could someone with better knowledge than me explain how this works ?
Here is also a Nature paper about using synchronization phenomena in coupled solid state spintronic nano-oscillators to achieve vowel recognition:
https://www.nature.com/articles/s41586-018-0632-y
Arxiv PDF: https://arxiv.org/ftp/arxiv/papers/1711/1711.02704.pdf
This work focuses on steady-state computing, but it could also be interesting to use transient physical behaviors to process time-varying signals
Maybe by modelling dynamical systems as "neural nets" as in: https://arxiv.org/abs/1806.07366 and https://arxiv.org/abs/1808.08412
Or by using complicated physical systems we don't even understand to build Echo State Networks: http://www.scholarpedia.org/article/Echo_state_network
Here is also a lesser known open source project that demonstrates the use of proof-of-stake for energy efficiency, combined with a fixed-width directed acyclic grah with transaction sharding for parallelization. Furthermore, "final" and "stale" blocks are dynamically forgotten to increase storage efficiency.
It achieves 10,000 tx/s on chain and transaction times below 1 minute.
Differentiable does not mean easy to optimize. One could imagine implementing sha-256 using differentiable operators, and yet the system as a whole would not be optimizable at all. It would be interesting to have compilers that optimize the "optimizability" of differentiable programs tho...
Also, here are two interesting examples of differentiation through physical systems for classification:
https://arxiv.org/pdf/1808.08412.pdf
https://innovate.ee.ucla.edu/wp-content/uploads/2018/07/2018...
I used a naive forward-differentiate-through-euler-integration scheme to optimize networks of coupled oscillators to perform classification ( https://arxiv.org/pdf/1808.08412.pdf ), but your approach seems to be miles away in terms of performance. Congratulations to the authors, this is amazing work, I see great potential applications in physics-based computing with nanotechnology, and I may actually use it myself in this context !
Couldn't it be that the measurement device (or any other interacting object) "falls" into all of the possible states of the measured system, therefore we see the measurement device itself as a quantum object when we don't interact with it, but every possible "outcome" of the measurement device sees the system as classical. At a higher scale, a human being itself is an object interacting with the measurement device, and every possible "outcome" of the human being sees the device as a classical system, showing a classical system. At least this explanation does not involve "quantum brains" nor enthropocentrism.