[Elecraft] The future of AI

[David Gilbert]

You know that eventually (probably already) the developers of AI are 
going to be using AI to enhance ... maybe even direct ... the AI coding 
of the future, right?

Apparently you don't.

Dave   AB7E



On 9/29/2025 2:47 PM, Karl W Hubbard wrote:
> Stalin: What matters is who counts the votes.
>
> AI: What matters is who writes the programs.
>
>
>
>
> Sent from Yahoo Mail for iPhone 
> <https://mail.onelink.me/107872968?pid=nativeplacement&c=Global_Acquisition_YMktg_315_Internal_EmailSignature&af_sub1=Acquisition&af_sub2=Global_YMktg&af_sub3=&af_sub4=100000604&af_sub5=EmailSignature__Static_&af_ios_store_cpp=9d3a686e-218d-4849-8298-b480188dc8ac&af_android_url=https%3A%2F%2Fplay.google.com%2Fstore%2Fapps%2Fdetails%3Fid%3Dcom.yahoo.mobile.client.android.mail%26listing%3Demail_signature>
>
> On Monday, September 29, 2025, 2:24 PM, David Gilbert via Elecraft 
> <elecraft at mailman.qth.net> wrote:
>
>
>     As others have said, AI today merely replies based upon the data
>     it was
>     trained on, aided by a very fast and very broad real time online
>     search
>     capability.  That search capability is one reason I use ChatGPT
>     ... its
>     replies are faster and broader than Google, for me at least.
>
>     Sometimes when I ask ChaGPT a question it gives me an inaccurate
>     response and I need to correct it to get the proper result.  ChatGPT
>     remembers that (for me) and (usually) doesn't make the same mistake
>     again if I ask it a related question, but  AI currently doesn't
>     aggregate what it learns from multiple users. In other words, the
>     feedback it gets from its interactions with one batch of users is not
>     used to "teach" it how to more accurately respond to a similar
>     question
>     from other users.  At least part of that limitation is per the
>     decision
>     of the developers, such as for user privacy concerns.
>
>     However ... that is not necessarily how it will be in the future.
>     Humans
>     learn from others, not just their own perceptions and observations of
>     the world around them ... i.e., the raw "data".  AI could do the same
>     ... and then who knows what we will get.  So I asked ChatGPT this
>     question:
>
>     "How long do you think it will be before AI is able to aggregate
>     what it
>     learns from a broad spectrum of interactions with multiple users?"  I
>     got the response below.  It's a cut&paste so hopefully it makes it
>     through the reflector.
>
>     Dave   AB7E
>
>     p.s.  The comment about "drift" is a real issue I hadn't considered.
>     Pretty interesting.
>
>
>           *Technically:* we’re already there (in a limited way)
>
>     OpenAI, Google, Anthropic, etc., already do something like this
>     today —
>     but not by letting the model itself remember. Instead:
>
>       *
>
>         They *collect and anonymize* user interactions.
>
>       *
>
>         Those logs feed into *future training runs or fine-tunes* of
>     the model.
>
>       *
>
>         The new model you get months later has improved capabilities based
>         on /aggregate/ patterns — but not on any specific identifiable
>         conversation.
>
>     So in effect, “learning from a broad spectrum of interactions”
>     already
>     happens; it’s just *offline* and at periodic update intervals, not
>     continuous and in real time.
>
>     ------------------------------------------------------------------------
>
>
>           *What isn’t here yet:* true “online lifelong learning”
>
>     That would mean the model you’re chatting with today is directly
>     shaped
>     by everyone’s conversations from yesterday. That’s technically harder
>     (you risk drift, bias, privacy leaks), and it’s a regulatory
>     minefield.
>
>     To do it safely you’d need:
>
>       *
>
>         Very robust anonymization and filtering.
>
>       *
>
>         Real-time evaluation for harmful or biased shifts.
>
>       *
>
>         Clear user consent and opt-out mechanisms.
>
>     ------------------------------------------------------------------------
>
>
>           *Timeline guess (if done responsibly):*
>
>       *
>
>         *Within 2–4 years*: Large commercial AIs may begin limited, opt-in
>         “online fine-tuning” where they adapt weekly or monthly using
>         anonymized data.
>
>       *
>
>         *5–8 years*: We’ll probably see truly “continuously learning”
>     public
>         models, but with heavy privacy layers and regulatory oversight.
>
>       *
>
>         *Beyond that*: Personal AIs that combine global learning with your
>         private, encrypted data for hyper-personalized experiences,
>     without
>         leaking your details to others.
>
>     ------------------------------------------------------------------------
>
>     So the short version:
>
>       *
>
>         The /aggregate learning/ you’re imagining already happens
>     today, but
>         in batches and offline.
>
>       *
>
>         Real-time, ongoing learning from everyone at once is coming — but
>         probably in a *measured, opt-in* way over the next decade, not
>     as a
>         sudden flip.
>
>
>
>     ______________________________________________________________
>     Elecraft mailing list
>     Home: http://mailman.qth.net/mailman/listinfo/elecraft
>     Help: http://mailman.qth.net/mmfaq.htm
>     Post: mailto:Elecraft at mailman.qth.net
>
>     This list hosted by: http://www.qsl.net
>     Please help support this email list: http://www.qsl.net/donate.html
>     Message delivered to karlwhubbard at yahoo.com
>