Leaderboard

@wbmnfktr I edited your post to remove the automatic embedding of the twitter feed and replaced it by the link only. It was huge 😄 @Pete Should we deactivate it? Doesn't even make sense to embed the whole feed when you link to it 🤨

Imagine why Microsoft's BING AI is in some parts way better than anything else? Hint: Github Absolutely everything in terms of code is in Microsoft's hands right now. No matter that... there are more ethical questions in regards to this. What's allowed and what not. On the other hand, there are already models that found possible cures for lots of illnesses, more critical things in terms of society and such, some are already in real-life studies, like some possible cancer cures and even old Nicola Tesla patents and whatnot. I hope and wish people want unopinionated and unbiased models. Otherswise the echo chambers just became million times bigger than already. BUT... noone can say which model/AI is really that neutral. So... yeah. In terms of writings and suggestions... GPT4 is already quite good as this. Feed it your last 10+ books you read and you will get awesome results. Depending on your input and your prompt... you might be impressed. You can. I tried it with Inferno (Dante, not Dan Brown) and some other similar books. You could ask whatever you want. Just ask with some optimized prompts. Maybe even try DAN prompts or use extensions to allow internet access. 🤯 There is so much going on right now. If you are interested in models and specialiced prompts... follow Brian Rommele on Twitter. He has some interesting prompts that show you how to "manage" the deeper level https://twitter.com/BrianRoemmele Just tried to find a bookmark of someone that trained GPT on a local Raspberry Pi for his needs. There are already so many solutions out there. Most of the things we experience right now... are 6+ months old already and were just made public to us. Or how could so many companies, services and even single person offer such incredible tools within weeks? BING AI, Canva AI, Mozilla AI, Brave AI... and... much... more...

Once LLM models can be personally trained on home networks, we can develop the ability to train on personal data and education, resources and inspiration. Some LLM can already be run on home machine with a modest Nvidia 4090 😆 but other models are more generic, running on any 24GB+ GPU. Imagine being able to feed all of August Rodin's notebooks, sketchbooks and studies into an AI and the design cues that could produce. Or DaVinci. Who owns the rights to this, or is it public domain? The rights aspect of models is going to be the next big war. I DON'T believe people are going to want universal and unopinionated models for their own work. I don't think those are as useful for maintaining the creative integrity of a person in their field. It is great to be able to consult ai librarians - models trained on a particular body of knowledge. For example, I would love to be able to consult an AI trained on all of the protected content published by the Journal for the Study of Old Testament. It would radically transform academia to be able to have research assistants that could be leased for a small fee to provide reliable synopsis and citations on the vast volumes of journals within a given field. Or even reinvent Reader's Digest, training a model with the text of all popular novels and developing a bot that could be consulted to read summaries of popular novels that are shorter or even adaptations of novels for different age groups of language skill levels. What if we could read a Dostoyevsky novel in an abridged version - with an optional character diagram - at an 8th grade level - for someone with ESL. What I am interested in, and what many bosses are interested in - the ability to train an AI model that is like us and retains our style, experience and knowledge base. Lots of bosses would like to be able to hire as many of themselves as they could - particularly when growth and innovation are important. Not everyone wants drones. Someone who could produce draft output that we curate - this is the most promising edge of AI, and the one that frightens a lot of people because many mid-level folks are where they are today because they have copied or stolen other people's work. Once you have a digital twin trained on your own work inspiration, background and style, you should be able to lease your twin to do certain tasks. Because trust of training sources is going to be important, and in the end there will be models for all of us - how are those rights managed? When you use an engine to train and build out a model based on your preferences, likes, background, interests and expertise, do you own the rights to what has been trained? Or do those rights belong to the content owners used to train to model? The answers to these questions frighten me more than the technology itself - just as medical patents infringe on bodily integrity on a regular basis in the US once you get involved with product trials. The idea a company can own a part of your biology because you get a treatment is disturbing. The patents and rights issues related to medical ethics and personal property rights are still unresolved - https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3440234/ This will certainly apply to the digitization of connectome neural paths. A cool primer on connectomes, if you haven't heard of this: As an artist, who wouldn't love to be able to train their own personal visual training model based solely on their own collection of work - studies, materials, style, etc. - and then put that 'virtual apprentice' to work. I think these global singularity-type models are great for those who can afford to not discriminate for quality and taste of output. It gets the job done when the specifics aren't important, or when you are willing to compromise your vision for something that serendipity brings along that is superior. But for those who have a more precise vision with lower tolerances, the general training models are not lucky enough to work consistently. "The work of a master of their craft who is also a professional is reliable, repeatable, and responsible." - R. Fripp AI is not this - yet - for anything. But all the ingredients are coming together nicely. An incredible documentary on this - AlphaGo - https://www.imdb.com/title/tt6700846/ Well worth a watch.

On a less positive note, this is a very interesting, and a bit scary, talk https://www.humanetech.com/podcast/the-ai-dilemma

I couldn’t resist the hype and created a simple module using the ChatGPT API to process field values. You can find it on GitHub here: https://github.com/robertweiss/ProcessChatGPT ProcessChatGPT is triggered upon saving, if you select it in the save dropdown. It processes the value of a page field, which can be set in the module config, using ChatGPT. The processed value can be saved back to the same field or to another field on the same page, which can also be set in the module configuration. You can add commands to the value that will be prefixed to the source field content. This way, you can give ChatGPT hints on what to do with the text. For example, you could add ›Write a summary with a maximum of 400 characters of the following text:‹. One of my clients is already using the module to summarise announcement texts for upcoming music events on their website (Let’s face it, nobody reads them anyway 😄). If anyone finds it useful, I would be happy to submit it to the official module list.

You can use this to change the template settings’ valid UrlSegments: https://processwire.com/api/ref/template/url-segments/. Be aware this persists the change immediately, without having to call $template->save()! You can also use property syntax: //Add UrlSegments “login” and “logout” to whitelist of current page’s template: $page->template->urlSegments = array_merge($page->template->urlSegments, ['login', 'logout']); If you have this requirement, it’s probably worth checking out the relatively new URL hooks: https://processwire.com/blog/posts/pw-3.0.173/

To follow up on that, I fed this into GPT-4: That was it's reply:

Some scary aspects indeed.. Good that some folks are trying to bring this to our attention. But it might be too late...

@adrian's tracydebugger also has an api explorer panel that you might be able to get inspiration from 🙂

OpenAI are moving on pretty fast. Referring to my last post, there is a more up to date way of supplying training data to LLMs with instructions from https://platform.openai.com/docs/guides/fine-tuning I think we could use this to train the AI on the PW API. We'd need to provide a JSONL file with training data. The data should contain the usage descriptions and examples from the PW API docs. As far as I know the API docs are automatically generated with PHPDocumentor https://www.phpdoc.org/ I am looking for ways to convert the present HTML output from Documentor to JSONL files. Maybe output XML first with Documentor and then convert that. Or write a Documentor template that outputs JSONL directly. Just an idea, but should be doable. Is anyone here experienced wit PHP Documentor and can tell me whether this is possible?

@bernhard At https://github.com/baumrock/RockMigrations#working-with-fieldsets you have e nice description on how to work with fieldsets. This seems to be limited to adding fields to fieldsets at runtime. Is there a way how we can define a migration that puts fields inside a fieldset in template context? Something like 'content-page' => [ 'tags' => 'content', 'fields' => [ 'title' => [ 'label' => 'Page Title', ], 'text' => [ 'label' => 'Subtitle', ], 'content_blocks' => [], 'tab' => [], // this adds the fieldsettab open and close 'fieldname' => [] // how to add this one to the tab fieldset ], It works fine with with $rm->addFieldToTemplate('fieldname', 'content-page' , 'tab', 'tab_END'). Just wondering if it is possible to do this inside the fields array in my example above.

Machine learning is quite an amazing thing and I have recently looked at what happened with the Google Alpha Zero program using neural networks. In board games like Chess and Go the machine was given only the rules of the game and objective and then given huge amounts of historical games to learn from. It was also playing itself millions of times and learning from this. Apparently after 4 hours it became stronger than the strongest program (Stockfish) and plays in a radically different way compared to previous Non AI programs, even showing moves that seemed alien to human grand masters but when analysed were found to be previously undiscovered winning strategies. This completely amazed me. So I think that AI will be very powerful organising a database into very pertinent and relevant responses and will only get better and more accurate over time with the neural networks they are using now. Of course humans will intervene and make corrections but the program will "learn" from this.

AI is great in producing text, that looks plausible, but it's not concerned with actual correctness. I'm working with elixir nowadays and there have been many examples posted, where responses claimed some API to exist in the stdlib, which just didn't. Looked totally fine on paper until you looked into the actual stdlib and found nothing. So really AI can be useful, but in the case of ChatGPT you really want someone chatting, who can actually validate the responses. Languages evolve and that's not a new thing at all. Take for example computer, which used to be a human doing math, and now is a machine. There's a good reason why we don't use natural language for programming our – now machine – computers. They're messy and not at all strictly defined.

NOTE 2022: Better use URL hooks! https://processwire.com/blog/posts/pw-3.0.173/#introducing-url-path-hooks Hidden option ? // site/ready.php $wire->addHookBefore('Page::render', function($event) { $page = $event->object; if($page->template == ...) include(yourcodefile.php); // or any other condition }); // yourcodefile.php <?php namespace ProcessWire; if(!$this->config->ajax) return; // early exit // any other conditions // return json header('Content-Type: application/json'); die(json_encode($yourdata)); @Guy Incognito you might also be interested in this thread:

You can always reset your password just by pasting this temporarily into any one of your templates, and then viewing a page that uses the template: $u = $users->get('admin'); // or whatever your username is $u->of(false); $u->pass = 'your-new-password'; $u->save();