|
Post by slh1234 on Feb 17, 2023 18:41:36 GMT
Earlier in my lifetime, the elder people of the day were telling us how TV wasn't going to end well for humanity, and were telling us what was going to happen to us because of TV. It doesn't seem to have come to pass.
TV is a communication tool. AI is a tool. A gun is a tool. They can be incredibly useful, or they can be used for evil, and none of them know one from the other. It's up to people how it will be used.
|
|
|
Post by missouriboy on Feb 17, 2023 21:38:32 GMT
Yes, I look forward to using the new Bing Chat. I am on the waiting list. The warped offspring of a Woke programmer? Or an in-bot mutation?
Have I heard mentally-ill people being described in a similar way?
Microsoft's chief technology officer, Kevin Scott, told the New York Times that "the further you try to tease it down a hallucinatory path, the further and further it gets away from grounded reality."
|
|
|
Post by slh1234 on Feb 17, 2023 22:13:48 GMT
Yes, I look forward to using the new Bing Chat. I am on the waiting list. The warped offspring of a Woke programmer? Or an in-bot mutation?
Have I heard mentally-ill people being described in a similar way?
Microsoft's chief technology officer, Kevin Scott, told the New York Times that "the further you try to tease it down a hallucinatory path, the further and further it gets away from grounded reality."
Are you looking for reasons to dislike it because you think it is this? If you think AI is developed by some programmer, there is the first place you're going off track. What is your definition of an "in-bot mutation?" The two articles are not objective articles. The first one is referencing the conversation previously with the same New York Times reporter. The second one is a mischaracterization of a quote. IOW, a straw man. Neither of them show the slightest understanding of what they are dealing with. It seems they're just looking for a reason to be afraid or to feel doomed or persecuted. I offered to explain the processes of machine learning and explain in more depth what AI is. Instead of this, people on here are going out to find yet another advocacy article from someone else who doesn't know what AI or ML is and bringing it back. Is it to try to draw a connection of woke-ism and the development of a new AI based chatbot? The AI/ML discussion doesn't have to be 500 level to get a better understanding of the processes behind it. So serious question, do you want to know a little bit about AI? or is it really the issue and not the answer that you seek?
|
|
|
Post by walnut on Feb 18, 2023 2:40:28 GMT
I already do understand it pretty well. I have some programming background too. I think that the "technology reporter" probably did manage to tease it into a strange place. But that does not bother me. I am very impressed actually. I know that I should buy some more Microsoft while it is down on all the blah blah blah. Maybe first of next week.
|
|
|
Post by ratty on Feb 18, 2023 4:15:59 GMT
Earlier in my lifetime, the elder people of the day were telling us how TV wasn't going to end well for humanity, and were telling us what was going to happen to us because of TV. It doesn't seem to have come to pass. TV is a communication tool. AI is a tool. A gun is a tool. They can be incredibly useful, or they can be used for evil, and none of them know one from the other. It's up to people how it will be used. I recall the evil Rock 'n' Roll too.
|
|
|
Post by ratty on Feb 18, 2023 22:51:14 GMT
|
|
|
Post by missouriboy on Feb 18, 2023 23:24:08 GMT
"All that any computer model can do is solidify and provide false support for the understandings, misunderstandings, and limitations of the modelers and the input data."
|
|
|
Post by youngjasper on Feb 19, 2023 1:05:07 GMT
"All that any computer model can do is solidify and provide false support for the understandings, misunderstandings, and limitations of the modelers and the input data." I did a lot of computer modeling in my former career. There is no doubt someone that is familiar with a model, and familiar with the inputs, and has a dashboard of outputs can make the model give the output they want. In cases where there is not enough sensitivity, a little reprogramming will take care of it. It amazes me anyone puts faith in a model, especially with one that fails over the test of time and/or attempts to predict a chaotic system like the climate, or a model that attempts to answer questions that clearly result in a biased response. It is nonsense.
|
|
|
Post by walnut on Feb 19, 2023 1:14:21 GMT
Chat GPT drank the whole pitcher of kool-aid. But I think that there is still hope for it.
|
|
|
Post by slh1234 on Feb 19, 2023 4:01:21 GMT
That is not what a "computer model" does at all.
There are two main categories of machine learning models: Supervised and unsupervised. The difference is whether there is an outcome that can be known and used with the training data. They are used for different things.
Unsupervised learning includes such things as clustering of data. This gives insights into natural groupings that are not otherwise obvious. Unsupervised learning may be used for just such purposes, or it may also be used in data exploration for supervised learning.
For supervised training, you have data that you know the outcomes for, and the data is labeled. When a data scientist starts with a data set, they begin with things like looking for high cardinality values, or values with colinearity, etc. because this can cause a model to be impossible to train. Additionally, there is feature engineering such as normalization (putting it on a scale of 0 - 1) or deriving such things as relative length instead of absolute values, or putting things on the same scale (dollars and euros don't mean the same thing, for example). You look for correlation in the data and the outcomes, and based on this, null and alternate hypotheses are established, and an acceptance criteria is established based on metrics such as area under curve, or accuracy. These things vary depending on whether the question is classification, regression, or other. Everybody knows that a model will not be 100% correct, but acceptance criteria has to be established before training and testing occurs. (As an example, the models I developed for malaria only had about 1% false negative, and about 4% false positive. The cost of a false negative is far greater than a false positive, so this is built into the acceptance criteria, and changed the method so that I needed to use a raw image, but also create a model that tested on HSV enhanced versions of the image, and also using a gaussian blurred version of the same image. Because of relative risk, if any of the 3 models predicted the cell to be parasitized, the final result would be "parasitized." This minimized the number of false negatives, and the higher number of false positives was deemed acceptable. Of course, more than one cell is tested for any one suspected case, so the number of cases missed was much less than this. FWIW, this is better than humans can do without the assistance of AI.)
Once the data is reduced to values with a correlation, feature engineering is complete, etc., You split the data into training and test data sets, and this is commonly done with 70-30 or 80-20 splits depending on such things as size of data set. The training process is an automated process of testing relative weights for each feature (data element that is relative to the final prediction) across the testing data set. In cases of continuous values, it is not possible to test every possibility, but the testing process uses different methods (depending on the type of algorithm or neural net), and in some cases multiple runs (epochs) are made through the training data with each set in order to change the order of evaluations (neural nets) or try different structures (decision trees) or combinations of structures (random forests) to return one model that performs best with that data set and that set of hyperparameters. Sometimes, multiple algorithms, and almost always, multiple sets of hyperparameters will be tested to arrive at the best performing model according to the chosen metrics.
From here, the test data is run through the trained model to validate it. There is a concept of overfit, and a concept of underfit. It if performs well against the training data, but not against the test data, then this is "overfit," and it indicates that the model fit too closely to the training data and doesn't generalize. This will not be a useful model, and it needs to go back to prior steps. If it performs poorly against the training, then this is underfitting, and it will never progress to test data because it doesn't even perform well with the data it is using to build the weightings. Depending on acceptance criteria based on different costs for different kinds of errors, multiple models may be developed like what I described above. For complex tasks, different models are needed for evaluations in different parts of the tasks.
But as for "standing the test of time," no model does this, and if you think about the world, it makes sense. You can predict gas mileage in a car based on engine displacement, weight, and other such factors. However; if you are using a model trained on cars built in the 1970s, it will not perform well with cars built in the 2010s. Engineering has changed a great deal in that time, and this leads to something called "data drift." Data scientists use automated processes (usually unsupervised training) to detect data drift. This means that models need to be re-trained periodically because virtually any data set, even diabetes data sets measuring such things as BMI and tricep thickness will drift over time and this means the model(s) must be retrained periodically. Whatever the case, once the model is trained, you're not going to make the model give you any output you want based on a spreadsheet of inputs and outputs (that works for rules-based applications, but that's not what ML is).
Every ML/AI process has a feedback loop and retraining, and this includes language models (which are VERY complex). That's it in a nutshell. Of course, there is a great deal more complexity in each step. There is a Continuous Integration/Continuous Deployment (CI/CD) process set up for the feedback loop and retraining for every model type, and it is different from application development, and that is why this is typically set up to be delivered and called via a web endpoint so that new deployment of the application is not necessary when you need to deploy a new version of the model - they have different release cycles. (The translation in MS Word, or the speech to text built into Power Point or MS Teams uses this. These features will only work for you when you are connected to the internet because the processing through the model to perform these things is online in Office 365, and not on your desktop).
There is a cute adage that "all models are wrong. Some models are useful." From the description I gave, I think you can understand the meaning of that. If it fits data even in a seemingly chaotic system adequately (as defined by acceptance criteria prior to experimentation), then it is deemed useful. I can give you examples where Chat GPT answered incorrectly, but Bing chat answered correctly - it's been trained. I've also noticed differences in behavior in Bing chat over the last two weeks - likely CI/CD. It's expected. It's still a great tool, and there is no sentience there. In fact, developers often do cute things like alternating among similar scoring candidates in language models in order to give it the appearance of sentience, but all it is is probabilistic analysis and sometimes a few rules based evaluations like randomization to give it this appearance just to make it seem more natural to the end users.
So I assert back that assertions like "All that any computer model can do is solidify and provide false support for the understandings, misunderstandings, and limitations of the modelers and the input data" is either someone that does not really understand ML modeling, or someone looking for an excuse because the model did not return the answer they were hoping to get. Actually, that's why the experimentation process is in place. In the case of the responses given by ChatGPT, they were what you would expect with the data it is working with - it is a language model and not a model of a climate system. It is giving answers based on text used to train it, and you know, that text didn't support what someone wanted to hear, so now, it appears they will attack ChatGPT.
I can tell you a question that ChatGPT confidently answers wrong, and I wonder where it got that answer because it's not a matter of interpretation - it tells me that Cosmos DB can join documents in different containers, but that is not possible. This is one that Bing Chat has right, though, thus reflecting additional training in the model - something that Bing Chat will continue to undergo going forward just because all commercial software, ML, and AI products do if they are to be viable.
|
|
|
Post by slh1234 on Feb 19, 2023 4:07:08 GMT
Shortening something from the previous answer:
If you're looking for someone to agree with you on what you believe on the climate, you need a buddy to sit on the back porch with you, or go grab a beer with you. That's not what a language model will do, and that's why you're frustrated with the answer that ChatGPT gave. ChatGPT is functioning as a language model, and it is doing it pretty well. When carried forward to Bing Chat, it is doing even better as a language model to find answers available in text on the web, but not to run your climate experiments independently. It's not going to be your buddy, and it's not going to drink a beer with you while you discuss your climate beliefs, either.
|
|
|
Post by ratty on Feb 19, 2023 4:30:57 GMT
|
|
|
Post by slh1234 on Feb 19, 2023 4:47:14 GMT
On questions of models being useful, or just reinforcing someone's misunderstanding:
When Mrs. SLH scheduled her knee replacement, the hospital sent her a 25 page PDF document to prepare for the surgery, help her understand what to expect, explain the nerve block she would have in her leg when she came home, and how she should proceed with recovery. The document was in English. Mrs. SLH's first language in Korean. She speaks English, but is not as good as she is in Korean. The document was daunting, difficult, and she was afraid that even if she could read through all of it, she might not understand it well enough. My first language is English and Korean is my second language. When I looked at it, I was not sure I could translate it, and even if I could, it would likely take me days to do it. It was more than I could take on.
You can open pdf files in MS word, and I know that MS Word with Office 365 has translation built in to it, so I opened that pdf in word, saved it as a word document, then used the translate feature to translate the entire 25 page document into Korean. I was actually a bit tentative because the language used was very specific to a medical context, but I gave the document to Mrs. SLH and asked her if it was understandable.
She disappeared upstairs for about an hour, then came back downstairs with one word: "Perfect." I reviewed some things with her, and she did, indeed, understand all of the critical points she needed to. (I don't think "perfect" meant every word was perfect and every sentence sounded professional. I think that meant she could understand it perfectly.)
So was this translation model useful? or was it just providing false support for my misunderstandings and limitations of the modelers and/or input data?
On the same wise, I have legal matters I sometimes need to take care of in Mexico. Spanish is my third language and to be honest, I'm not that good at it. I can go shopping and take care of 80% of what I need to do downtown quite comfortably, but on legal matters, there is no way I can understand it. I use Microsoft Edge specifically because I know the translate feature in it uses Microsoft's translation models. I use this to translate web pages to ensure I understand what I need to do, and to ensure I am providing the correct information.
So is this translation model useful? or is it just providing false support for my misunderstandings and limitations of the modelers and/or input data?
Likewise, expecting a language model designed to carry on a conversation with you and give information obtained on the internet to confirm your beliefs for you is a fundamental misunderstanding of what a model is.
I gave 2 specific examples of AI models that are really useful. I have others in areas of environment narrators for visually impaired, screen readers for visually impaired, voice to text in Power Point or Teams (and other apps) which are good for hearing impaired including me, and I can assure you they are quite useful. For that matter, voice enabled chatbots like Alexa, Siri, Cortana, or the Google Assistant (whom I call "Miss Google" when using Android Auto) are quite useful, and since they are models similar to ChatGPT, they will likely just return to you the data they are trained to use as a source which will be found on the internet, so even though they're useful, they're not going to tickle your ears when you want someone to tell you what you want to hear. Maybe someone can build a speech enabled "buddy model" for the times when you need the model to tell you what you want to hear on a subject.
Note that I'm skeptical of manmade global warming - I think the jury is still out. But I don't expect a natural language model to confirm my belief for me.
|
|
|
Post by ratty on Feb 19, 2023 9:40:49 GMT
|
|
|
Post by walnut on Feb 19, 2023 14:40:13 GMT
Completely changed the topography of that place.
|
|