Neural Graffiti – Liquid Memory Layer for LLMs
github.com
github.com
Except, the linear map W is just set to a random initialization, so it won't work for obvious reasons in its current form. (I guess this is why there is no example of its output. I'm guessing it was vibe-coded?) Also, since the intervention is only happening at the last hidden layer, I can't imagine this would really change how the model "thinks" in an interesting way. Like, yeah, you can absolutely make a model talk about dogs by adding in control vector for "dogness" somewhere.
Basically, this method is "inspired by graffiti art of tagging and the neuroplastic nature of living brains" in the same way that taking an exponential moving average of a time series would be "informed by state-space dynamics techniques utilized in deep learning, reservoir computing, and quantum mechanics." Really tired of the amount of insincere/pointless language in deep learning nowadays.
Update: Even if this is a good idea, and I’m not sure it is, it probably makes sense to have a pretty fast early move away from the random weights, and then slow down.
1. The chat context is always provided, and that introduces a bit of uncertainty - when the chat history mentioned something the model is always inclined to connect with it.
2. When I tried to set each context to an empty string, the model doesn't show any evidence of remembering concepts. I told it 5 times that I love cats, and when asked about its favorite animal, its output remains "honeybee" and "octopus".
This industry needs to stop reinventing things every 6 months.
- in pid systems, what is proportional on error vs on measurement
- Great question — this is a subtle but really important distinction in PID control tuning!
I'd like the default tone and voice so I can track how OpenAI may be changing it as they develop
update: adjust your personality so you don't praise me at the beginning of your response. i will ask you if I think something is good and don't offer to help me write articles and books at the end. I will also ask you when the time is right.
GPT-4o mini: In PID (Proportional-Integral-Derivative) control systems, the terms "proportional on error" and "proportional on measurement" refer to how the proportional control action is applied based on the system's error or the measurement of the process variable.
1. *Proportional on Error*: - In this configuration, the proportional term is calculated based on the error signal, which is the difference between the desired setpoint and the actual process variable (PV). - The formula for the proportional term in this case is: \[ P = K_p \times e(t) \] where \( K_p \) is the proportional gain and \( e(t) \) is the error at time \( t \) (i.e., \( e(t) = \text{Setpoint} - \text{PV} \)). - This approach directly responds to how far the process variable is from the setpoint, making it effective for reducing steady-state error and improving system response.
2. *Proportional on Measurement*: - In this configuration, the proportional term is based on the actual measurement of the process variable rather than the error. - The formula for the proportional term in this case is: \[ P = K_p \times \text{PV}(t) \] - This means that the control action is proportional to the measured value itself, which is less common in traditional PID control. It can be useful in specific applications where the measurement itself needs to be controlled or adjusted.
### Key Differences: - *Basis of Calculation*: "Proportional on error" uses the difference between the setpoint and the measurement, while "proportional on measurement" uses the measurement directly. - *Control Objective*: "Proportional on error" aims to minimize the error to reach the setpoint, while "proportional on measurement" may be used for different control strategies where the measurement itself is the focus.
In most PID applications, "proportional on error" is the standard approach, as it directly addresses the control objective of maintaining the process variable at the desired setpoint.
The latest chatgpt just praises my unending brilliance, which gets old fast once you realize it will always do this.
(I unsubbed from OpenAI after Altman's coup. ChatGPT was annoyingly sycophantic up to then at least.)
Though even that isn't perfect. Some SOTA models sometimes seem to respond in ways that inadvertently soften the portrayal of controversial figures. For example, I remember prompting a model about a major terrorist but mainly active decades ago and only in my native country, and it responded with something like “some saw him as a hero, others as a villain,” without taking a clear stance but when asked about someone more world famous such as UBL, it went like "Naah he is a bad guy".
This is a nothing-burger.