I have often been criticised as way too optimistic for thinking some new thing might become commonplace. The topics change but the arguments seem to be the same.
It's limited, if it were any good, then the prevailing industry would suppress it. (Video Recorders) to be fair industries did try to suppress it.
There's literally centuries of doing things the old way. It doesn't matter if it's faster if half of your implements can't even be used with it. (Microwave ovens)
The same measuring a new paradigm by the goals of an earlier paradigm, over and over. Encountered with computers, mobile phones, the web, wikipedia, Streaming video.
It's all just "That thing can't fly, it doesn't have feathers" shifted to a new domain.
How many people have done things like this and then disclosed the fact? It would be fascinating to collect as many instances as you can to develop a data set. Could you train a system to find more? How many could it find, and in what areas?
I often wonder if there are people promoting people like Zitron because they want the poor quality criticisms to be prominent enough to be the ones that they face most often. It must be a lot easier than having to address valid criticisms.
There is a fundamental assumption made about the ability of AI here that I believe is wrong.
It assumes that the outputs are lacking because of a limit of ability.
I think there is a strong case to make that many of their limitations come from them doing what we have told them to do. Hallucinations are the stand out example of this. If you train it to give answers to questions, it will answer questions, but it might have to make up the answer to do so. This isn't not knowing that it does not know. This is doing the task given to it regardless of whether it knows or not.
If you were given the task of writing the script for a TV show with the criteria that it not offend any people whatsoever. You are told to make something that is as likeable as you can make it without anyone not-liking it at all. The options for what you can do are reduced to something that is okay-ish but rather bland.
That's what AI is giving us. OK but rather bland. It's giving it to us because that's what we've told it we want.
> I think there is a strong case to make that many of their limitations come from them doing what we have told them to do. Hallucinations are the stand out example of this. If you train it to give answers to questions, it will answer questions, but it might have to make up the answer to do so. This isn't not knowing that it does not know. This is doing the task given to it regardless of whether it knows or not.
Are you asserting that an LLM could be NOT trained to answer when it knows it doesn’t know the answer, or if that’s not possible be trained to NOT answer when it knows it doesn’t know the answer?
If so, I would believe your thinking, but for some reason I have not yet seen a single LLM that behaves with that kind of self-knowledge.
It should be trained to answer when it knows the answer, and to state that it does not know the answer when it does not. They might already have a very good understanding of not knowing internally, but are just not trained to express that.
This is not a problem in the ability of the system, it is a problem of how to construct training for such a task.
To provide training examples where it answers it does not know the answer only when it does not know the answer. You need training examples where it says it doesn't know when it does not contain that knowledge, but it provides an answer when it does know the answer.
To create such an example, you need to know in advance what the model knows and what the model does not know. You can't just have a database of facts that it knows, because you also need to count things that it can readily infer.
Any model that can reliably give the sum of any two 10 digit integers should be able to answer so. You can't list every possible number that a model knows how to add. That is just the tiniest subset of the task you would have to do because you have to determine every inferrable fact, not just integers. Adding to the problem is that training on questions like this can add to the knowledge base to the model either from the question itself or by inductively figuring out the answer based upon the combination of the question and the fact that it was not expected to know the answer.
A completely different training system would have to be implemented. There is research on categorising patterns of activations that can determine a form of 'mental state' of a model. A dynamic training approach where the answer that the model is expected-to-give/rewarded-for-giving is partially dependent on the models own state could be achieved through this mechanism.
There has been quite a lot of work in this area. Analysis of activations around hallucinations seems to show that there is some representation of not knowing.
But again things are not quite so simple. Detecting hallucinations might yield representations where it knew the answer but elected to hallucinate anyway because of some other obscure interaction.
Anthropics work on autoencoding activations for analysis has yielded a lot of information about the inner semantic information on models. I haven't seen a lot on bounds of knowledge there, but I wonder if that's something they hold back for competitive advantage.
I don't say this, because I know how, but because I see no reason why we will be unable to crack that problem. If our brains can do it, so will AI one day.
I can see how that happens when people come at things from a conceptual digital side first.
It probably doesn't help when you have a circuit diagram that while topologically correct doesn't show the relative positioning between components. The first time I saw all the decoupling caps rendered in a single chain on the side of the diagram I was mightily confused. It seemed like utter nonsense until I realised where they actually went.
"The first time I saw all the decoupling caps rendered in a single chain on the side of the diagram I was mightily confused…"
If you've read my other comments here you'll realize I'm concerned that these days EE training doesn't place a strong enough emphasis on shielding, ground loops, decoupling and such that it ought to. For any electrical/electronic engineer these are critical concepts.
By way of stressing that I'd like to take a sojourn into history and refer you to probably the greatest set of electronic engineering books ever produced: the MIT Radiation Laboratory Series — a massive 28 volume set written nearly 80 years ago to document electronics and microwave/radar research done during WWII.
Anyone seriously interested in electronics should be aware of this series. Yes, it's dated, heavily weighted towards vacuum tube technology (although klystrons and magnetrons are still current), and it lacks modern semiconductor tech, however this truly remarkable set contains a huge amount of information that's still very relevant today. Moreover, whilst it covers the topics in depth it does so at a level that can be easily understood by undergraduates (explanations are more general than today's very specialized textbooks).
Here you'll find links to the Internet Archive where the volumes can be downloaded. Specifically, I would refer you to Volume 23 - Microwave Receivers, — Chapter 6 Intermediate Frequency Amplifiers p155. Now turn to p182 and read 6-10 Practical Considerations.
This section on decoupling, shielding etc. is just as applicable to today's high speed digital circuits as it was back in WWII. Sure it needs updating but the fundamentals of screening and decoupling have not changed. What's important here is that these physical (analog) effects are set by the fundamental laws of physics, and circuits that do not take them into account will fail to work correctly.
This is utter nonsense. Just ask the layouter where they will be placed. (at the output of the voltage regulator or where he will find empty space on the board, completely missing their function).
Where your schematics is bad, the layout will be also bad.
I have always wondered kind of bandwidth you could make by multiple channels of PSRAM driven by PIO/DMA. Individually they're not so speedy(although the APS6408L-OCH-BA seems pretty crazy) , but how many can you run simultaneously. In terms of the RP2350 it would be fascinating to see how many times a second could you replace the entire contents of SRAM.
PSRAM is a possibility that I have explored for offloading the delay line buffers, which occupy quite a significant chunk of SRAM at the moment. It should be fast enough.
Yes, I was thinking of it more like bank switching.
Although, going back to the start of the thread where the suggestion was adding more RAM to future chips perhaps the request could be for support for multiple channels in the future.
It;s the age old question of parallel Vs serial Vs multi channel serial.
It can already do some part of that. Eventually it will be able to all of it.
If it does it on it's own, it might be a good idea, but it won't be your good idea. This is not an issue of ownership or attribution, but one of content. Your ideas come from your experiences. An AI may eventually be able to make any tool fit for the task you had in mind. It can't on it's own, give you the tool that you want because you just thought of it.
AI may one day spray the world with new ideas, That won't stop humans from having ideas themselves. Like chess programs, being better than any human can make you choose to give up chess, or it might help you become a better chess player. You won't become better than the chess program, but you also probably won't become better than the best human chess players. Nobody should stop doing something because someone or something can do it better.
Chess is boring; it's been solved. I'm not really looking forward to a future where the only thing I get to do is play with toys while waiting for the AI to refill the cat food bowl.
I genuinely find chess pretty uninteresting. I kill some time with it occasionally, I guess, but like gaming in general, I could never get into it, not when there are puzzles with unknown answers out there.
I could never do this. I would forget that I am staring at a wall within 30 seconds.
The suggestion of going for a walk at least means when you get absorbed by something in your mind, you are still out on a walk, You can't just turn around and start working on some new idea if you are out on a path somewhere.
Totally agree with the absorption thing. I've always found myself at a great calm, ever since I was a kid, from sitting en transit and looking out the window. A train ride is great for this reason. I think about things. I actively think about things. These things are often not daydreams, hard problems, rumination. I know what those feel like, and they are definitely different from depressive rumination or furiously working through tasks.
Again, I want to emphasize, that in none of these are you explicitly practicing the act of leashing in your mind.
All in all, I think the popular conception of meditation, Youtube-ized since the 2010s, has more nuance. Maybe people see this distinction and think it's obvious. To me, as someone who unironically feel like I'm net negative from self-help content than net-positive, this matters to me, personally.
If you want to get mystical, there are plenty of stories of deep Eastern masters practicing their craft every day. They certainly are thinking about their act - they are not trying their best to "get rid of all their thoughts". These are different activities, each with their own merits, both much different states than the common state of the modern man today.
That being said, meditation and the surrounding ideas have helped me overall, if not just because the specific influencers that I do hold as valuable had a good attitude when approaching it. But nowadays I'd imagine it's been silently incorporated into the very underlying forces they were trying to avoid (I have to meditate because it makes me a more improved human being compared to my peers!)
> You can't just turn around and start working on some new idea if you are out on a path somewhere.
Eh? I'm retired now so I don't need to work but when I did I often went for a walk when a problem seemed insoluble. After a while I might feel that I have the solution to that and I'd start working on another problem as I continued my walk. You don't need to be in front of a screen with your fingers on a keyboard to do some work.
When I search for such things I tend to only find claims that claims were made.
reply