[ih] History of AI and Internet

[Jack Haverty]

Thanks, Arun.

I agree that network operators have always used available tools, 
whatever they might be called.  One could even argue that the ARPANET, 
50+ years ago, used AI -- the routing mechanisms and algorithms could 
alter traffic paths much faster than a human with a patch panel even 
then, and did so without human intervention required at all.

In the early days of The Internet, circa late 1970s, connectivity 
between Europe and the US was provided by SATNET, which used a channel 
on the Intelsat IV-A satellite for transatlantic paths.  At the time, 
the bulk of other users' traffic using Intelsat IV-A was video feeds and 
telephone calls.  SATNET was likely one of the first to send data over 
the satellite, with computers at each end (called SIMPs for Satellite IMPs).

The SIMPs performed constant checks on the error behavior of the 
satellite channel, just as the IMPs in the ARPANET measured circuit 
behavior.  Such checks were performed constantly and consistently, not 
just when a (human) technician was investigating a problem.  The BBN NOC 
monitored SATNET as well as ARPANET, so the NOC operators would be 
informed with a loud bell from the TTY in the NOC when a circuit was 
degrading either in ARPANET or SATNET.

ARPANET circuit degradations were usually referred to ATT Long Lines, 
where most circuits were provisioned.  SATNET circuit degradations were 
reported to the Intelsat IV-A NOC.

At first, the Intelsat NOC simply dismissed such reports.  It was 
difficult for them to believe that some unknown company (BBN) in 
Massachusetts could know anything about the performance of their 
satellite channel between ground stations in West Virginia and Goonhilly 
Downs in Europe.   Users of video or telephony didn't notice a few 
glitches in their screens or audio, until it got bad enough.  But the 
SIMPs did notice right away.  It didn't take long for the Intelsat NOC 
to start paying attention to such reports, which often preceded some 
kind of major failure in their satellite service.  The Internet became 
an "early warning" service for video and telephony services through the 
satellite.

I would characterize such early mechanisms as "expert systems" variety 
of AI.  They all did what humans would do but, unlike humans, they could 
do the work much faster and work 24 hours a day.

Along with the memory of the ARPA project to use expert systems 
techniques in a Morse Code communications environment, these experiences 
motivated the pursuit of the "Automated Network Management" project, to 
apply such expert systems techniques to the operation and management of 
networks.   I believed it was possible even in 1982.   Still do.

Fast forward to modern times....

I've been thinking about network problems for quite a while now, but 
from an end-users' perspective.   End-users have lots of "network 
problems", and as a local tech nerd (surrounded in my neighborhood by 
non-techies) I get such questions.  I've been experimenting with various 
AIs to learn what they can do, so I just asked one of them a typical 
end-user Networking question and asked it how to fix it.

The answer I got was unsurprising.  My summary of the AI -- "There is no 
one to call and report your problem.  It could be a number of things.  
You're on your own to figure it out."  If anyone's curious, here's the 
conversation I just had: 
https://claude.ai/share/468f2576-bd1b-4dbb-9a48-941a94051b2b

Surely The Internet and AI can do better...?   I still think some Expert 
System approach might be appropriate, even if it is an ancient idea.

---------------------

One possible cause of the 50 year timeline for AI to be more broadly 
used in Networking is what I term "Technology Silos".

During the 1970s/1980s, there was a lot of research on both AI and 
Networking.   At MIT both were quite active, but in separate groups.  
The MIT AI Lab was the locus of AI research.  Licklider's group and a 
few others (such as Multics) did lots of Network research projects.  The 
groups were friendly and collaborated on many things, such as our 
underlying operating system (ITS).

But there was little crosstalk in core research issues of either 
Networking or AI.  AI was interesting to Networkers (like me), but not 
our Mission.  Networking was interesting to the AI groups, but not their 
Mission either.   It wasn't until the Morse Project appeared in our 
Networking group that I really had a reason to understand what "Expert 
Systems" was all about.   We successfully used "Expert Systems" to 
create a computer system.  The government report concluded 'Using 
available AI techniques, a successful automatic "Morse Code reader" was 
developed by the MIT group and picked up quickly by NSA.'   I've always 
thought that the goal of Research should include getting the results 
into actual use "in the field."   The introduction of the "Technology 
Silos" of Morse Code (a form of Networking) and AI was the enabler.    
Having a distinct goal of "Understanding Morse Code" focussed our 
research on the actual problem the customer wanted to be solved.

I encountered another "Technology Silo" in 1990, when I joined Oracle in 
Silicon Valley.  The group I joined developed software for use on all 
types of networks.  But it also had responsibility for operating the 
corporate "intranet", which had cisco routers deployed in more than 100 
countries.  As you might expect, there were lots of "network problems" 
and our operators struggled to diagnose and fix them.  The tools 
available from the Networking industry weren't much help.

One afternoon, two of us sat down at a console to help diagnose some 
problem.  I knew a lot about Networking from my years at BBN and 
Internet experience.  My colleague, like everyone else at my new 
employer, knew a lot about databases.

We couldn't modify any of the routers of course, so changing any of the 
existing code or protocols was impractical.  But I mentioned SNMP as a 
mechanism we had used in The Internet to gather data about what was 
happening in the network.  IIRC we were sitting at a Sun Workstation 
(1990 vintage), and it had tools such as "snmpget".  The 1990 routers 
also "supported SNMP", whatever that meant at the time.  We even found a 
few host computers (such as the Sun) which had some implementations of 
SNMP, critical for figuring out what TCPs were doing.

In a few hours, literally, we had written some small shell scripts that 
continuously probed routers and hosts -- anything that would respond to 
SNMP interactions.  All the data was entered into a database.  IIRC 
every tidbit of SNMP data from the field was simply entered as a row in 
a table.

Once the data was in a database table, it looked just like all other 
sorts of data that companies used to manage their business operations - 
inventory, shipping, customer records, etc.

The ARPANET, and The Internet, were created as ARPA projects, which 
funded much of the early Networking implementations.   Most Networkers 
may not know that database technology also had its basis in 
government-funded projects.   For example, Oracle's genesis (long before 
I got there) involved funding from the Intelligence parts of the US 
government (as I suspect the Morse Code project was).

So by 1990 there had been more than a decade of databases being evolved 
and used to manage all sorts of business data.  Networking data gathered 
through SNMP was just another form of business data.  Database people 
have long experience in managing all sorts of business data.

I looked into the NOC at Oracle a week or so later, and was wowed at 
what I saw.  Database people had clearly been there, and all of Oracle's 
technology was of course available for them to use.   I saw a screen 
with a scrolling display showing network performance data, constantly 
being updated as more data arrived from SNMP.  It looked like the 
various medical screens you see in hospitals monitoring lifesigns or in 
Investment Banks watching stock prices and trends.    You could also 
easily bring up a screen showing current performance and also display 
similar graphs of past performance to see how things were different when 
there were no problems.

None of this was what I would call AI, even 35 years ago.  But it was 
useful.  The key enabling action was to link together the two 
longstanding "Technology Silos", of Networking and Databases, which had 
both existed for a long time but apparently not been coupled together.

AI has now advanced to include the ability to interact with natural 
human languages.  There's no longer a need to learn some programming 
language such as LISP or whatever.  That makes AI much more accessible 
to people outside of the AI community.  Computing costs have also come 
way down.

So perhaps "Technology Silos" of AI and Networking have been a 
longstanding obstacle which is now disappearing.

/Jack Haverty




On 6/25/26 09:19, Arun Welch wrote:
>
>> On Jun 24, 2026, at 7:36 PM, Brian E Carpenter via Internet-history <internet-history at elists.isoc.org> wrote:
>>
>> On 25-Jun-26 10:22, Sivan via Internet-history wrote:
>>> Dear Jack Haverty,
>>> Your question "For example, when there are problems in today's Internet,
>>> are AI techniques and tools used to diagnose and repair them?   What's the
>>> History of such things?" is immensely interesting. Concerns about A.I.
>>> momentarily set aside, are there initiatives underway to positively use
>>> A.I. tools to "diagnose and repair" problems in the Internet? For example,
>>> using A.I. to scan for malware, bots, phishing and other forms of technical
>>> and non-technical Abuse? Or using A.I. to scan and detect barriers to
>>> network protocols such as vpns? Or even using A.I. to scan and detect
>>> non-human content and other forms of Abuse?
>> I can't imagine that the answer to any of those questions is "No".
>> I wouldn't have said that a year ago, but progress is very fast.
> Large operators have always used every technology available, and this has included AI (for whatever the current interpretation of that term is) for a long time. You simply cannot run at scale using only human support. Reducing noise and swivel-chair in the NOC, rapid issue resolution, stability, etc become harder as the network gets larger.
>
>
>> Of course, operators will see this as a competitive advantage and may
>> choose not to publicise such AI deployments. But there is a lot of
>> work in progress on agent-to-agent communication.
>
> In our case, we see talking about it as a competitive advantage, and our VP in charge of all this regularly speaks about what we’re doing at conferences, on LinkedIn, etc. As one of the people in our company who spends a lot of time talking to customers about what we’re doing, I can tell you that they absolutely love to hear about this.
>
> To provide some concrete examples on how we’re using AI today:
>
> a) Incident correlation: If the circuit is down, you don’t need to troubleshoot BGP. You’ll get alerts for both, and choosing which to safely ignore is important.
> b) Predictive analytics: We’re gathering metrics on lots of things, and can predict that a laser diode is going to fail because we’ve seen the light levels drop with this pattern before. We can schedule an outage and replace the hardware in a controlled manner rather than an unplanned outage.
> c) Chatbots for our service and NOC operators: “What’s the status of site/circuit/incident X?”. This uses agents to gather information across multiple systems to present a single cohesive answer, reducing swivel chair. In addition to our internal monitoring systems we’re also querying weather, 811 (call before you dig) registries, power systems, etc. to get a comprehensive view. For example, if there’s a regional power outage then you probably know why a site is down. Some of this is also customer-facing, so that a customer can send an email asking “what’s the status of my outage at 100 Main Street” resulting in a response that lays out all the information known about the issue.
> d) Analyzing traffic patterns to shut down botnets.
> e) Writing automation workflows using AI tools. Spec-driven development has some interesting implications here.
> f) Using AI coding tools to develop other sorts of things. This is particularly true for doing dashboards and analysis.
>
> The list goes on, this is not everything we’re doing, AI in many forms is used for all sorts of things.
>
> I’m not speaking in any official capacity here, obviously.
>
> …arun

-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature.asc
Type: application/pgp-signature
Size: 665 bytes
Desc: OpenPGP digital signature
URL: <http://elists.isoc.org/pipermail/internet-history/attachments/20260625/e3d0b4f8/attachment-0001.asc>