426 karma · joined January 15, 2021
- "Our coin hit $100M daily volume, get on this rocketship before it's too late!"
- "Our exchange does $1B annually, so you know we're trustworthy!"
- "Hey investors, look at the massive demand for our GPUs (driven by the company we invested $100B)!"
> Regardless, again: if the AI is so smart, and it somehow needs something akin to MCP as input (which seems silly), then we can use the AI to take, as input, the human readable documentation -- which is what we claim these AIs can read and understand -- and just have it output something akin to MCP.
This example is like telling someone who just wants to check their email to build an IMAP client. It's an unnecessary and expensive distraction from whatever goal they are actually trying to accomplish.
As others have said, models are now being trained on MCP interactions. It's analogous to having shared UI/UX patterns across different webapps. The result is we humans don't have to think as hard to understand how to use a new tool because of the familiar visual and interaction patterns. As the design book title says, 'don't make me think.'
Our DNA specifies the model structure and hyperparameters for our brains. Then it is compiled and instantiated into a physical medium, our bodies, and our connectome is trained.
If you want to make a comparison about the quantity of information contained in different components of an artificial and a biological system, then it only makes sense if you compare apples to apples. DNA:Code :: Connectome:Weights
Then we compile and run that source code and our individual lived experience is the training data for the instantiation of that architecture, e.g. our brain.
It's two different but interrelated training/optimization processes.
I take this to mean during weight updates, e.g. training.
> "runtime train of thought"
I take runtime here to mean inference, not during RL. What does runtime mean to you?
Previous approaches [0] successfully used inference time chain of thought to improve model responses. That has nothing to do with RL though.
The grandparent is wrong about the paper. They are doing chain of thought responses during training and doing RL on that to update the weights, not just during inference/runtime.
One is talking about an improvement made by making control flow changes during inference (no weights updates).
The other is talking about using reinforcement learning to do weight updates during training to promote a particular type response.
OpenAI had previously used reinforcement learning with human feedback (RLHF), which essentially relies on manual human scoring as its reward function, which is inherently slow and limited.
o1 and this paper talk about using techniques to create a useful reward function to use in RL that doesn't rely on human feedback.
It also sounds like you have a PM on your team, but they actually have the title of SWE or Eng. Mgr. They probably spend > 50% of their time on the above listed responsibilities rather than engineering. Hopefully they don't get docked in their performance reviews for essentially performing the duties of a PM rather than those of an engineer.
Literally the first line in the comment that started this thread.
People are being dismissive of your comments because to say that proteins are niche in the context of pharma is like saying advertising is niche in the context of Meta and Google.
Which as we know puts you down an irreversible ideological path of... something?
We could also talk about how the word 'executes' implies some kind of agency which computers lack. It's like saying a rock just executes the laws of physics when it rolls down a hill...
Well in the case of Python, the interpreter is indeed interpreting.
One followup: does it take into account grammatical errors, incorrect conjugations and that kind of thing?
This downplays the importance of the set of systems for which his proof holds and makes it sound like it applies to some obscure branch of mathematics.
It applies to a huge set of important systems, not least of which is any system that is sufficiently expressive as to uniquely identify the natural numbers.
Can you give any examples of those types of problems you've encountered?
They only need to be better than a few dozen or hundreds of others at the required skills. The average 7 footer in the NBA is a way better shooter than the average player in the general population and probably better than any average college player. But shooting skill is inversely proportional to height in the NBA for the exact reason the GP stated.
To be something like 6'3'' or shorter in the NBA means you need to be an insane outlier in other skills, better than literally millions of others at your height and hundreds of thousands with your general athleticism.
The upshot is that once you've found an architecture that is already biased towards solving a specific problem, then the training of the weights is faster and results in better performance.
From the abstract, "...In this work, we question to what extent neural network architectures alone, without learning any weight parameters, can encode solutions for a given task.... We demonstrate that our method can find minimal neural network architectures that can perform several reinforcement learning tasks without weight training. On a supervised learning domain, we find network architectures that achieve much higher than chance accuracy on MNIST using random weights."
He was notorious for doing just this. He would make what have now been engine analyzed as sub-optimal moves but which lead to such complex and dynamic positions that his opponent would eventually slip and he would exploit their error in dramatic fashion. This is one reason why he is considered one of the most creative attacking players ever.
After reading the description of the SOP at this shop, the idea that the OP would be able to introduce an additional layer of process requiring multiple stakeholders and management seemed like a bridge too far in my mind :).
> other than in end-game
This is frequently where such scenarios occur. Many end games are difficult for humans to play with absolute precision, even seemingly simple ones like knight and bishop vs. king. But when there are fewer pieces on the board is exactly when computers are able to perform incredibly deep calculations.
A good example is the notorious 30 move forced mate that Caruana "missed" in game 6 of his world championship match with Magnus, which occurred with only 3 pieces and 3 pawns left on the board.
It's a similar property to "code quality" in that regard. There are metrics like 'test coverage' or 'dependency graph complexity' but these can also be gamed intentionally or unintentionally much like LoC.
There are lots of descriptive metrics you can come up with that can be useful in exploring how you are performing or how your performance is changing with regard to quality or velocity. But treating such a metric as an actual proxy for the underlying property is not useful and sometimes counterproductive as you point out.
I think this is where the notion of software craftsmanship appears. We are still at the point where it requires an expert who is engaged in the process to essentially sense how they or their team is performing on these properties. They can use metrics to inform that sense or to help in determining what actions should be taken.
But this is frustrating for larger organizations where frequently the people who want to measure a certain property are not actually participating in the process because they are separated by multiple layers of management. You essentially have to trust that every management layer below you is able to accurately assess a qualitative property, which is difficult to standardize and scale.