Little more hands-on with a certain #LLM for some time now.
Once I learned what the "stop sequence" is actually good for, my instinctive ascription of at least a little bit of personality to the thing disappeared immediately.
I find the "oh, I know them, they're good people" argument about A.I. researchers to be such a tired and garbage statement. It doesn't matter how good the scientists who worked on the atomic bomb were, they still ushered in the age of mutually assured destruction.
I'd like to reiterate that my concern about #AI is not that it's going to become conscious, or that it's going to displace a huge portion of the workforce (it will) but before it does any of that it's going to be leveraged by bad actors to flood society with truthy sounding bullshit the likes of which we've never seen and it will lead to killing in the streets.
@mmitchell_ai maybe we could help kick start the collaboration of leading scientists on ways of controlling #LLM technology by improving the discourse on this very platform in such away that it brings together the many sources of academic expertise that are already here?
People overestimate what "AI" of today is and what it can actually do, because memes are spreading that describe any machine-assisted process in a highly glossed-over form, ignoring the required human effort to make it work.
The naive impression is that you just gave some generative engine a prompt and the result came out fully formed, when the actual process is that the people behind the project used multiple purpose-built engines, for each of the engines they iterated on prompts that would output something semi-coherent, and then they used human efforts to tie the result together.
This is currently spreading as "this AI-generated pizza commercial" with no further explanation, but Tom's Hardware interviewed the actual people who made it work:
Menschen empfinden als harmonisch, was sie gewohnt sind. Sie mögen, was sie kennen, am liebsten leicht geremixt, aber wiedererkennbar.
Auftritt "#KünstlicheIntelligenz": Bildgeneratoren rekombinieren wohlbekannte Bildinhalte und Stile. #LLM-Textgeneratoren ermitteln aus riesigen Textkorpora, welche Wörter am gewöhnlichsten aufeinander folgen. Sie sind beliebt und angenehm, weil sie harmonische Ergebnisse erzeugen - außer dem Betrachter fällt eine Disharmonie auf, insbesondere eine zur Realität.
Are LLMs generic writers - or rather generic books? Would anyone say of a book, it had theory of mind or feelings?
What if the #LLM is a compressed storage of utterances of human actors? The utterances are what we are used to see as signs of the actors’ knowledge or theory of mind or feeling.
That compression is not lossless. The decompression, triggered by prompt, does not restore the original faithfully. We misunderstand the errors of the book as creativity of a writer.
#LLM are brute-forcing their way through absurd amounts of data to generate an autocomplete output for any given input that approximates outputs a human might give instead.
They lack a few distinct properties of human cognition, including language, that more brute force alone cannot compensate for. Because they can only ever internalize and compute intratextual context.
Incidentally, humans need much less input(!) to learn language. Probably because they can contextualize across domains.
This is interesting. Yousaf Shah demonstrating how to use ChatGPT to build JSON for a hypothetical fast food ordering system. At the London Xojo conference.
Just ran into this entertaining and accessible explainer by #LiveOverflow about why large language models like #ChatGPT sometimes 'misbehave' and present output to the user that they're not supposed to see.
Long story short: both the system's 'filters' and the user input are presented as one big prompt to the model, which means you can influence the filters.
We're finally in the age where we can talk to machines.
I regularly discover ways to improve my ChatGPT prompts. Consequently, I went from thinking, "LLMs are interesting but mostly useless" to "How did I work without LLMs" quite quickly. This newsletter is my attempt to share my findings.
@womble@marcelsalathe@fj
Maybe it can do a lot better already: computers, like most tools, do far better than humans in the tasks they were design to perform, e.g. number crunching, data storage, communications etc. It's just that this #LLM technology, as good as it may be in language and pattern recognition, it is not capable of functioning the way they advertise it. The problem is not the technology itself, the problem is the #snakeoil salesmen that advertise it as being human like.
The #AI Act is a flagship legislative proposal to regulate #ArtificialIntelligence based on its potential to cause harm. The #European Parliament is now inching toward formalising its position on the file, after #EU lawmakers reached a political agreement on Thursday (27 April). #machinelearning#llm#LLMs
I'm impressed and quite enthusiastic about GrammarlyGO. Even after a few uses (10 of 500 monthly prompts used so far) I can see this being an LLM that I use to assist with non-fiction writing.
"When it comes to data storage, GrammarlyGO does not retain customer data for longer than necessary for processing. All sensitive data is anonymized and de-identified before being passed onto Azure OpenAI. In addition, we do not allow Azure OpenAI to retain or train on customer data.”
• Generative AI models learn from mass data scraped from web
• Indigenous groups fear losing control over their data
• Some move to protect their information from commercial use
"When U.S. tech firm OpenAI rolled out Whisper, a speech recognition tool offering audio transcription and translation into English for dozens of languages including Māori, it rang alarm bells for many Indigenous New Zealanders.
"Whisper, launched in September by the company behind the ChatGPT chatbot, was trained on 680,000 hours of audio from the web, including 1,381 hours of the Māori language."