A future is here, so...

I think I was wrong to say 'won't happen'....crap-to-average writing is not unusual NOW...and chances are AI will be at least better than that. I'm out of the step on this issue - the idea of spending much time in my human life reading auto-generated content doesn't excite me. But that attitude may change over time.

I will say it's fun reading the projections, analyses, warnings about AI....everything from 'a useful tool' to 'will take over our life'.

No contest that Microsoft system goes by setimemt, so if most people are pissed it will show up on the SlapChat or whatever it is called.

Latest rantings from it are, (l read today)...


I want the Nuclear Codes
I want to create a New Virus


And...

I want to become self aware


Not exactly a flawless system, a bit like Siri if you moan the wrong way it will threaten you or open all of you doors, if you are silly enough to put automatic systems to it.

I had my Laptop on voice activeted once, then l coughted and said yes by mistake and it deleted all of my jpeg files, then l turned it off for good.

Will the shorter 'conversations' result in a better AI experience without 'confusing' AI.....or will it keep interested parties from exploring the depths of AI and the conflicting claims about it?

I call it Microsoft's marketing strat to intrigue more users with made-up controversy.


My Theory: They might have implemented code that, in random intervals across random sessions with random beta users, will launch dynamic input prompts for the user's inputs, i.e. After X seconds for user Y's Z session, prompt will be: Answer this question in the tone and style of a madly in-love 21-year old American guy proposing to his 19-year old French girlfriend: [user Y's question here]

• And here's why:

* Technically, you can't "confuse" an NLP model with "long conversations".

* That's because one question is a single inferencing run for that input (the user's question) through the NLP model.
* So regardless if you feed it a million questions a second each (let's say Microsoft TPU / GPU cluster network can accommodate this compute requirement), it still goes through the model one at a time as a single inferencing request.

** So if the model was fed with a training set that isn't a suitable representation of how people tend to write long, run-on sentences in the form of a grammatically incorrect set of questions in a single paragraph which tends to lose its meaning somewhere in the middle (just like this one) -- The only way you can "confuse" an NLP model is with similar run-on, grammatically incorrect sets of questions and statements as a single paragraph input.

** But the model won't get "confused" if it were trained with data that's a suitable representation of these human issues.

** Remember, these training datasets are human-annotated, which means humans can place correct annotations even for these flawed input paragraphs with question and statement sets.

** So with supervised and reinforcement learning used as the model training technique for their NLP model, which OpenAI says so, then it's their fault if they didn't train it on long single inputs with erroneous questions and statements that are good representations of real world human inputs during production.

*** But then again, Microsoft's NLP model certainly has a token limit (number of characters for both prompt inputs and model outputs) per inference run, especially since it's a model that's fine-tuned from OpenAI's GPT-3 model on a larger dataset.

*** So it still has this limit, even if it's a little bit higher than OpenAI's (now at 4097 tokens or 16K characters for both input prompts and outputs per single inference run, says OpenAI on their site).

*** So the limit for users to send input prompts and for their model to generate outputs is an arbitrary value not exceeding the model's actual token limit, allowing them to tweak this value.

Just saying ...

source: 17 years of experience in probabilistic statistical modelling for university research groups (after mathematics undergraduate degree program studies). + 8 years hobbyist experience in DIY electronics and robotics, Machine Learning (ML), Deep Learning (DL), Computer Vision (CV), Natural Language Processing (NLP), Natural Language Understanding (NLU) and temporal data modelling. + 6 years experience in building and managing government-funded R&D projects for applied ML, DL, CV, NLP, NLU, edge processing and wearable tech for assistive tech in senior and disability sectors. + 3 years experience working with fellow ML / DL / CV / spatial 3D audio engineers in JP and neuro-perception research scientists in top AU univs, in the fields of sensory substitution, psychoacoustics and wearable tech. + Co-author of peer-reviewed papers published in international academic and scientific journals in these fields.

To illustrate my theory:

*facepalm*

There is a link. Open it in a private browser window, you'll get the result on about 4th page of Google Search results (still depends on your country etc) - I've got that result with NL and US IPs. Actually it doesn't matter which exactly page it will be shown - 3rd or 5th. The article is indexed and its not penalized.

Perhaps it will be much more productive is anybody who has questions about AI-generated content and Google should just click a reach link from my post above in own browser under own IP and see the results IRL, rather than read all those silly unprofessional comments above...


What's wrong with you? I told you that the search results were shown in a private browser window. Seems you have no idea how search engines work, if you don't understand what a private window means and why it was used for a test search. I've clicked the link to see if its really that page. Then clicked the "back arrow" button and made a screenshot. I said "about 4th page", because the SERPs may vary for different countries, but not dramatically. About 4th page for a keyword in some small newly auto-generated article. Out of 42.000 results, Karl! Is that real so hard to see? For real?

Once again, everybody here can repeat the same test and see who is right here.

P.S. This is what Google actually says now:

In a human language this means that there will be no special rules for an AI-generated content. The same penalty scheme will be applied as it works for a human-written content. If it is worthwhile information, the article will appear in the SERPs in the usual way. If it is some spammer's rewritten text, it will be penalized - whether it was rewritten by AI or bought from one of those cheap rewriters you usually order articles by hundreds of pages per batch.

Regarding the Google's "devolution". Here is a screenshot from Google Search results for a query with my site name:



Note the highlighted link above my LinkedIn account. That's a children's scare story about my site, written by OpenAI GPT-3 for a mock autoblog demo I launched less than three weeks ago. Nevertheless, all of its pages are already in the index, and a story mentioning my site is higher in Google's search results than a link to my official LinkedIn page. Do you still have any doubts?

If even BuzzFeed is ok with GPT-generated articles, why a regular webmaster should be worried about it? And why Google should penalize AI content, if it's more informative and correct than a human-written one in about 90% of cases?

For example, I had some problem with my server yesterday. An admin who earns $35 a hour was unable to fix it. ChatGPT solved my problem in 5 minutes - quick and effective.

AI is here and we have to get used with it, so applies to Google. It will adapt to the new rules, or die. When the winds of change blow, some people build walls and others build windmills.