#government — Public Fediverse posts

Pro-Palestine Action group to rally at Labour conference led by Burnham | Israel-Palestine conflict News

London, United Kingdom – Protesters are planning to gather in support of Palestine Action at the first Labour…
#NewsBeep #News #Headlines #Europe #Government #Israel-Palestineconflict #UnitedKingdom #World
https://www.newsbeep.com/747842/

https://www.europesays.com/people/244466/ Burnham should call snap election on reversing Brexit, Lammy’s foreign policy guru says #AndyBurnham #BoardOfPeace #Business #BusinessFinanceAndIndustry #BusinessPerson #ColorImage #founder #Government #horizontal #meeting #NewYorkCity #Photography #Politics #PoliticsAndGovernment #PrimeMinisterOfTheUnitedKingdom #talking #TonyBlair #USA

Powerful storm brings flooding, power outages to northeastern United States | Weather News

Authorities warn residents that extreme conditions could pose risks and urge people to stay away from flooded areas.…
#NewsBeep #News #Headlines #Climate #government #TopNews #TopStories #UnitedStates #US&Canada #Weather
https://www.newsbeep.com/ca/911077/

Powerful storm brings flooding, power outages to northeastern United States | Weather News

Authorities warn residents that extreme conditions could pose risks and urge people to stay away from flooded areas.…
#NewsBeep #News #Headlines #CA #Canada #Climate #Government #UnitedStates #US&Canada #Weather
https://www.newsbeep.com/747710/

(paywall) Well, I guess someone is getting ready for a government bailout. We can't support people experiencing poverty or provide proper medical care in this country, but somehow there is money to help billionaires destroy the planet.
#AI #government #bailout
https://www.nytimes.com/2026/09/26/opinion/ai-bubble-banking-crisis-too-big-to-fail.html

(paywall) Well, I guess someone is getting ready for a government bailout. We can't support people experiencing poverty or provide proper medical care in this country, but somehow there is money to help billionaires destroy the planet.
#AI #government #bailout
https://www.nytimes.com/2026/09/26/opinion/ai-bubble-banking-crisis-too-big-to-fail.html

(paywall) Well, I guess someone is getting ready for a government bailout. We can't support people experiencing poverty or provide proper medical care in this country, but somehow there is money to help billionaires destroy the planet.
#AI #government #bailout
https://www.nytimes.com/2026/09/26/opinion/ai-bubble-banking-crisis-too-big-to-fail.html

(paywall) Well, I guess someone is getting ready for a government bailout. We can't support people experiencing poverty or provide proper medical care in this country, but somehow there is money to help billionaires destroy the planet.
#AI #government #bailout
https://www.nytimes.com/2026/09/26/opinion/ai-bubble-banking-crisis-too-big-to-fail.html

Could the Government Have Built an AI-Like Telephone Intelligence System in 1989?

https://youtu.be/jt3DQ5vT9s8

Slug: telephone-intelligence-ai-1989
Meta description: Could 1989 technology have screened phone calls automatically? A source-driven look at keyword spotting, vectors, expert systems and their hard limits.

Could an intelligence agency in 1989 have built a primitive ancestor of today’s AI-driven communications analysis systems? A 66-page technical reconstruction, using only publicly documented technology, suggests that nearly every required component already existed: keyword spotting and DSP arrays, vector retrieval, latent semantic analysis, graph analytics, expert systems and automated reporting.

What if, in 1989, a government intelligence organization had wanted to build a machine that could listen to telephone traffic? A machine that could detect selected words, compare communications mathematically, analyze who was calling whom, flag unusual activity, and produce reports for human analysts automatically?

Not with modern artificial intelligence. Not with Whisper, transformer models or cloud computing. And certainly not with large language models.

Just with the technology that was publicly documented at the time.

That is the question explored in our research paper, A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989. The study does not claim that such a classified system existed. It asks a narrower and more defensible question:

What could have been built?

The distinction matters. The paper is a technical plausibility reconstruction. It draws on peer-reviewed research, technical reports, patents, vendor documentation, government records and the limited declassified material in the public record.

The result is more interesting than a simple yes or no. By the end of the 1980s, virtually every major component needed for a primitive machine-assisted communications-intelligence system already existed. The surprising part is not any one technology. It is what happens when they are connected together.

The Wrong Question: “Could a 1989 Computer Transcribe Every Phone Call?”

If we frame the problem that way, the answer is simple: no.

Large-vocabulary recognition of ordinary conversational telephone speech was nowhere near good enough. The first published attempt at it came in 1993, and it got roughly 78 percent of the words wrong.

So the science-fiction version, a gigantic 1989 machine perfectly transcribing every call in America, does not survive technical scrutiny. But that is not how an efficient intelligence system would have been designed. The better question is this:

Could thousands of specialized systems work together as a funnel, examining enormous volumes of communications cheaply and forwarding only a tiny fraction for deeper processing?

That reframes the problem. The reconstruction finds that a cascaded architecture could cut computational demand by roughly 90 to 400 times, depending on the scenario. It also finds that keyword screening of tens of thousands of simultaneous calls was arithmetically possible with large numbers of late-1980s digital signal processors (DSPs), even though full conversational transcription was not.

The system would not need to understand everything. It would need to know what to ignore.

Stage One: The Telephone Network Was Already Becoming Digital

The first important fact is easy to overlook. By 1989, much of the telephone system was already becoming a digital data network.

The Bell System put the T1 carrier into commercial service in 1962. A single T1 stream carried 24 voice channels; higher-capacity DS3 links carried 672. Standard digital telephone audio ran at 64 kilobits per second per channel.

That meant that wherever a hypothetical collector had access to a digital trunk line, the voice did not need to be recorded from an analog wire and digitized afterward. It was already data. A digital carrier could be split back into individual voice channels for machine processing.

And the network provided something even easier to analyze than voice: metadata. Stored-program telephone switches already generated machine-readable call records. These held the calling number, called number, date, start time and duration.

From an engineering perspective, that is the obvious first filter. Before asking a machine to recognize speech, ask cheaper questions. Who is calling, and who are they calling? How often, and at what time? Is the number already known? Is it connected to other known numbers?

The reconstruction concludes that metadata would have been the cheapest and most practical first-stage selector in a 1989 screening architecture. That changes the scale of the voice-processing problem dramatically. You do not listen deeply to everything. You select first, and then you listen.

Stage Two: Machines Had Been Learning to Recognize Speech Since the 1950s

Machine speech recognition was not new in the 1980s. Bell Labs demonstrated AUDREY in 1952, which could recognize spoken digits from a trained speaker. IBM followed with the Shoebox in the early 1960s. By the 1970s, speech recognition had become a major research field involving ARPA, Carnegie Mellon, IBM, Bell Labs, BBN and others.

One of the most important systems was Carnegie Mellon’s HARPY. It handled a vocabulary of 1,011 words using a network of roughly 15,000 states. It was also very expensive to run: about 28 million instructions for every second of speech. On its PDP-10 hardware, it ran about 80 times slower than real time.

That sounds unusable. But HARPY showed something critical: the problem was no longer theoretical. Machines could already turn speech into constrained symbolic interpretations. The remaining problem was cost, and cost can sometimes be attacked with architecture.

Stage Three: Don’t Transcribe the Call — Search for Words

This is where the reconstruction becomes much more practical. A machine searching for ten or twenty specific words does not need to solve the entire speech recognition problem. It needs to solve a smaller one: keyword spotting.

Researchers were already working on this in the 1970s. The public record includes:

• elastic-template detection of words in running speech (1973)
• a Rockwell International patent describing keyword detection in continuous speech (1975)
• keyword spotting with linear predictive coding and dynamic time warping (1977)
• keyword-recognition research by a defense contractor, ITT (1985)
• keyword spotting with hidden Markov models at AT&T Bell Labs and MIT Lincoln Laboratory (around 1990)

The underlying idea is simple. Treat most speech as irrelevant filler and search only for specific acoustic patterns. If a confidence score crosses a threshold, forward the call; if not, discard it. This is far cheaper than full transcription.

Specialized DSP chips became common in the 1980s. Once they did, large numbers of relatively inexpensive processors could handle the front-end signal work in parallel. That gives us the foundation of the funnel.

The Computational Funnel

The hypothetical system reconstructed in the paper has six tiers, which fall into five broad stages.

Collection and metadata filtering. Digital carriers are split into individual channels and call records are captured. Known numbers, routes or circuits are selected. Voice-activity detection separates speech from silence, and tone detectors reject modem, fax and other non-speech audio.

Regional speech processing. DSP arrays extract acoustic features, keyword spotters search for selected terms, and only calls that cross set thresholds go on to more expensive recognition systems.

Central analysis. Partial transcripts are normalized and their terms weighted. Call records enter databases. Graph statistics reveal new contacts and unusual calling structures, and statistical systems flag changes over time.

High-end numerical processing. Vector supercomputers periodically run heavy mathematics such as singular value decomposition (SVD). Massively parallel machines run similarity searches across large collections.

Rules and reporting. An expert system assigns priority, and a template system generates a report. A human analyst receives the audio, transcript fragments, scores, graph relationships and alerts.

Each class of hardware gets the work it suits: DSPs for signal processing, workstations for recognition, database systems for records, supercomputers for SVD and statistics, and rule engines for final triage.

That division of labor is the key. The system does not depend on one impossibly advanced computer. It depends on many ordinary specialized ones.

Then Something Unexpected Happens: Speech Becomes a Vector

Today we talk casually about embeddings and vector databases, but representing documents mathematically is much older than modern AI. Gerard Salton’s information-retrieval work helped establish the vector-space model decades ago. A document could be represented as a vector based on the words it contained, using term frequency, inverse document frequency and cosine similarity. Computers could then compare documents mathematically rather than simply matching exact strings.

Now imagine feeding a noisy, imperfect telephone transcript into that system. The transcript does not need to be perfect. If enough important terms survive recognition, it can still become a weighted vector and be compared with categories or earlier communications.

The published evidence backs this up. In 1993, a system identified the topic of conversational phone calls 88 percent of the time among ten topics, even though it got most of the words wrong. Topics survive recognition errors far better than transcripts do, because the evidence adds up across many words.

Then, in 1988, another key piece appeared publicly: latent semantic indexing (LSI). LSI used SVD to project the relationships between terms and documents into a lower-dimensional mathematical space. It was not a neural embedding, but it solved a related problem. Documents with statistically related patterns of language could move closer together mathematically, even if they did not share the same words.

That date is one reason the paper’s target period matters. By 1988, every major stage of the hypothetical pipeline had appeared in the public technical literature. The reconstruction identifies 1988–1989 as the earliest period when the complete architecture, including LSI, could plausibly have been prototyped. A version without LSI may have been technically possible several years earlier.

No LLM Required

One of the most striking conclusions is that the system would not need anything resembling modern generative AI. It could be almost entirely deterministic.

Expert systems were mature enough by the 1980s to encode large sets of explicit rules. A communications-intelligence system could use rules like this:

IF a selected number contacts a new number,
AND the call contains multiple keyword hits,
AND the contact pattern deviates from baseline,
THEN increase priority.

Add network statistics, time-series anomaly detection and category similarity, then produce a structured report. No generative prose, no hallucinations, no hidden reasoning. Every decision could be logged, every score could be traced to its source, and every forwarded call would have a mechanical explanation.

The machine’s purpose would not be to replace the intelligence analyst. It would be to reduce volume. Millions of events enter, thousands survive, hundreds receive deeper processing, and a smaller number reach human beings.

That is a very different idea of “artificial intelligence.” It is closer to an industrial sorting machine.

The Biggest Limitation Wasn’t Necessarily Compute

This is where the reconstruction becomes especially interesting. You might expect raw computing power to be the fatal constraint. It was certainly a major one. But the paper concludes that an even worse problem appears as the system scales: false alarms.

Suppose a word spotter has what seems like a low error rate. Multiply that error across huge numbers of calls, then again across dozens of monitored keywords, and the false alerts begin to overwhelm the system. With 50 keywords and just one false alarm per keyword per hour, about 92 percent of ordinary three-minute calls would be flagged. Keyword spotters of the period did no better than that. With watch lists of that size, the funnel stops narrowing.

This is the base-rate problem. If the overwhelming majority of calls are irrelevant, even an impressive detector produces mostly false positives. That is why targeting by metadata matters so much. The architecture works far better when it starts from a small set of numbers, routes, individuals or circuits that are already of interest.

The system becomes plausible as a prioritization tool. It becomes much less plausible as an all-hearing national listener.

The paper is blunt about this. The funnel can cut computational demand roughly 90 to 400 times, but it cannot remove the false-alarm and base-rate problem. Effective selectivity requires targeting before any speech analysis.

Storage Was Another Brutal Constraint

Even if recognition had been perfect, storing everything would have been enormously expensive. At standard digital telephone quality, one continuously active channel produces roughly:

• 28.8 MB per hour
• 691 MB per day
• about 0.25 TB per year

Scale that to 1,000 channels and you get 691 GB per day. At 10,000 channels, it is 6.91 TB per day. Those numbers were punishing for 1980s storage systems.

The answer is again the funnel. Do not archive everything. Keep short buffers, extract features, store metadata and keyword hits, retain selected conversations, and discard most of the raw audio. The paper concludes that keeping raw audio long term beyond a few thousand channels would have been impractical with period storage, while selective retention was feasible.

This is not just an implementation detail. It tells us what the architecture would have had to look like. A system of that period would have been designed around forgetting.

So Could It Have Existed?

This needs a careful answer. According to the research:

• A research testbed covering tens to low hundreds of channels is technically plausible by 1988–1989.
• A targeted system covering hundreds to low thousands of metadata-selected channels is also plausible as a specialized operational capability.
• A system that indiscriminately screens tens of thousands to a million simultaneous channels with useful selectivity is another matter. That is not demonstrably plausible at that scale in 1989.

The limits were not only hardware. They included recognition accuracy, false alarms, base rates, storage, the long-haul links needed to move audio, and access to the traffic itself. The paper places genuinely effective untargeted large-scale deployment no earlier than the late 1990s, and more defensibly in the 2000s.

The final classification is surprisingly nuanced:

• Small research prototype: plausible.
• Narrowly targeted operational system: plausible.
• National full-transcription system: computationally impractical and inaccurate.
• Untargeted national automated listener: not demonstrably plausible in 1989.

And most importantly, there is no credible public documentary evidence that the complete integrated system described here existed in 1988–1989. A 1999 European Parliament technical assessment went further. It reported that effective automatic word-spotting of telephone calls still was not available, “despite 30 years of research.”

The Part That Should Make Us Think

The most interesting lesson here is not really about surveillance. It is about technological history.

We often imagine major systems appearing when one breakthrough invention arrives. Real systems rarely evolve that way. Technologies develop independently: speech recognition in one laboratory, information retrieval in another, digital telephony somewhere else, and expert systems in another field entirely. Add graph theory, statistical process control, signal processing, supercomputing and database systems.

Each looks ordinary in isolation. Then someone connects them, and the system can suddenly do something none of the individual parts seemed capable of.

That is the genuinely open question in this reconstruction. By 1988, the public literature contained the pieces. The unresolved historical question is not whether those pieces existed; they did. It is whether anyone outside the public literature assembled them earlier than the public record shows.

The available evidence does not answer that question, and responsible research should not pretend it does. But technically, the threshold is much earlier than most people would guess.

The Bottom Line

The paper’s central conclusion comes down to this. A recognizable ancestor of a modern machine-assisted communications-intelligence pipeline could plausibly have been assembled from publicly documented technology by 1988–1989.

It would not have been a modern AI. It would not have understood language the way current models do, and it would not have flawlessly transcribed a nation. Instead, it would have been a distributed computational funnel:

metadata → acoustic filtering → keyword spotting → partial recognition → vectorization → similarity analysis → graph analysis → anomaly detection → expert-system triage → deterministic reporting → human analyst

It is a primitive architecture by today’s standards, but an unmistakably familiar one. And perhaps that is the most important observation of all. Sometimes the ancestor of a modern AI system does not look like artificial intelligence. It looks like a telephone switch, a rack of DSP boards, a database, a vector supercomputer and a rule engine, quietly passing information from one stage to the next.

Research basis: A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989: A Source-Driven Technical Plausibility Reconstruction Using Only Publicly Documented Technology.

-Sean 9-26-2026

Could the Government Have Built an AI-Like Telephone Intelligence System in 1989?

https://youtu.be/jt3DQ5vT9s8

Slug: telephone-intelligence-ai-1989
Meta description: Could 1989 technology have screened phone calls automatically? A source-driven look at keyword spotting, vectors, expert systems and their hard limits.

Could an intelligence agency in 1989 have built a primitive ancestor of today’s AI-driven communications analysis systems? A 66-page technical reconstruction, using only publicly documented technology, suggests that nearly every required component already existed: keyword spotting and DSP arrays, vector retrieval, latent semantic analysis, graph analytics, expert systems and automated reporting.

What if, in 1989, a government intelligence organization had wanted to build a machine that could listen to telephone traffic? A machine that could detect selected words, compare communications mathematically, analyze who was calling whom, flag unusual activity, and produce reports for human analysts automatically?

Not with modern artificial intelligence. Not with Whisper, transformer models or cloud computing. And certainly not with large language models.

Just with the technology that was publicly documented at the time.

That is the question explored in our research paper, A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989. The study does not claim that such a classified system existed. It asks a narrower and more defensible question:

What could have been built?

The distinction matters. The paper is a technical plausibility reconstruction. It draws on peer-reviewed research, technical reports, patents, vendor documentation, government records and the limited declassified material in the public record.

The result is more interesting than a simple yes or no. By the end of the 1980s, virtually every major component needed for a primitive machine-assisted communications-intelligence system already existed. The surprising part is not any one technology. It is what happens when they are connected together.

The Wrong Question: “Could a 1989 Computer Transcribe Every Phone Call?”

If we frame the problem that way, the answer is simple: no.

Large-vocabulary recognition of ordinary conversational telephone speech was nowhere near good enough. The first published attempt at it came in 1993, and it got roughly 78 percent of the words wrong.

So the science-fiction version, a gigantic 1989 machine perfectly transcribing every call in America, does not survive technical scrutiny. But that is not how an efficient intelligence system would have been designed. The better question is this:

Could thousands of specialized systems work together as a funnel, examining enormous volumes of communications cheaply and forwarding only a tiny fraction for deeper processing?

That reframes the problem. The reconstruction finds that a cascaded architecture could cut computational demand by roughly 90 to 400 times, depending on the scenario. It also finds that keyword screening of tens of thousands of simultaneous calls was arithmetically possible with large numbers of late-1980s digital signal processors (DSPs), even though full conversational transcription was not.

The system would not need to understand everything. It would need to know what to ignore.

Stage One: The Telephone Network Was Already Becoming Digital

The first important fact is easy to overlook. By 1989, much of the telephone system was already becoming a digital data network.

The Bell System put the T1 carrier into commercial service in 1962. A single T1 stream carried 24 voice channels; higher-capacity DS3 links carried 672. Standard digital telephone audio ran at 64 kilobits per second per channel.

That meant that wherever a hypothetical collector had access to a digital trunk line, the voice did not need to be recorded from an analog wire and digitized afterward. It was already data. A digital carrier could be split back into individual voice channels for machine processing.

And the network provided something even easier to analyze than voice: metadata. Stored-program telephone switches already generated machine-readable call records. These held the calling number, called number, date, start time and duration.

From an engineering perspective, that is the obvious first filter. Before asking a machine to recognize speech, ask cheaper questions. Who is calling, and who are they calling? How often, and at what time? Is the number already known? Is it connected to other known numbers?

The reconstruction concludes that metadata would have been the cheapest and most practical first-stage selector in a 1989 screening architecture. That changes the scale of the voice-processing problem dramatically. You do not listen deeply to everything. You select first, and then you listen.

Stage Two: Machines Had Been Learning to Recognize Speech Since the 1950s

Machine speech recognition was not new in the 1980s. Bell Labs demonstrated AUDREY in 1952, which could recognize spoken digits from a trained speaker. IBM followed with the Shoebox in the early 1960s. By the 1970s, speech recognition had become a major research field involving ARPA, Carnegie Mellon, IBM, Bell Labs, BBN and others.

One of the most important systems was Carnegie Mellon’s HARPY. It handled a vocabulary of 1,011 words using a network of roughly 15,000 states. It was also very expensive to run: about 28 million instructions for every second of speech. On its PDP-10 hardware, it ran about 80 times slower than real time.

That sounds unusable. But HARPY showed something critical: the problem was no longer theoretical. Machines could already turn speech into constrained symbolic interpretations. The remaining problem was cost, and cost can sometimes be attacked with architecture.

Stage Three: Don’t Transcribe the Call — Search for Words

This is where the reconstruction becomes much more practical. A machine searching for ten or twenty specific words does not need to solve the entire speech recognition problem. It needs to solve a smaller one: keyword spotting.

Researchers were already working on this in the 1970s. The public record includes:

• elastic-template detection of words in running speech (1973)
• a Rockwell International patent describing keyword detection in continuous speech (1975)
• keyword spotting with linear predictive coding and dynamic time warping (1977)
• keyword-recognition research by a defense contractor, ITT (1985)
• keyword spotting with hidden Markov models at AT&T Bell Labs and MIT Lincoln Laboratory (around 1990)

The underlying idea is simple. Treat most speech as irrelevant filler and search only for specific acoustic patterns. If a confidence score crosses a threshold, forward the call; if not, discard it. This is far cheaper than full transcription.

Specialized DSP chips became common in the 1980s. Once they did, large numbers of relatively inexpensive processors could handle the front-end signal work in parallel. That gives us the foundation of the funnel.

The Computational Funnel

The hypothetical system reconstructed in the paper has six tiers, which fall into five broad stages.

Collection and metadata filtering. Digital carriers are split into individual channels and call records are captured. Known numbers, routes or circuits are selected. Voice-activity detection separates speech from silence, and tone detectors reject modem, fax and other non-speech audio.

Regional speech processing. DSP arrays extract acoustic features, keyword spotters search for selected terms, and only calls that cross set thresholds go on to more expensive recognition systems.

Central analysis. Partial transcripts are normalized and their terms weighted. Call records enter databases. Graph statistics reveal new contacts and unusual calling structures, and statistical systems flag changes over time.

High-end numerical processing. Vector supercomputers periodically run heavy mathematics such as singular value decomposition (SVD). Massively parallel machines run similarity searches across large collections.

Rules and reporting. An expert system assigns priority, and a template system generates a report. A human analyst receives the audio, transcript fragments, scores, graph relationships and alerts.

Each class of hardware gets the work it suits: DSPs for signal processing, workstations for recognition, database systems for records, supercomputers for SVD and statistics, and rule engines for final triage.

That division of labor is the key. The system does not depend on one impossibly advanced computer. It depends on many ordinary specialized ones.

Then Something Unexpected Happens: Speech Becomes a Vector

Today we talk casually about embeddings and vector databases, but representing documents mathematically is much older than modern AI. Gerard Salton’s information-retrieval work helped establish the vector-space model decades ago. A document could be represented as a vector based on the words it contained, using term frequency, inverse document frequency and cosine similarity. Computers could then compare documents mathematically rather than simply matching exact strings.

Now imagine feeding a noisy, imperfect telephone transcript into that system. The transcript does not need to be perfect. If enough important terms survive recognition, it can still become a weighted vector and be compared with categories or earlier communications.

The published evidence backs this up. In 1993, a system identified the topic of conversational phone calls 88 percent of the time among ten topics, even though it got most of the words wrong. Topics survive recognition errors far better than transcripts do, because the evidence adds up across many words.

Then, in 1988, another key piece appeared publicly: latent semantic indexing (LSI). LSI used SVD to project the relationships between terms and documents into a lower-dimensional mathematical space. It was not a neural embedding, but it solved a related problem. Documents with statistically related patterns of language could move closer together mathematically, even if they did not share the same words.

That date is one reason the paper’s target period matters. By 1988, every major stage of the hypothetical pipeline had appeared in the public technical literature. The reconstruction identifies 1988–1989 as the earliest period when the complete architecture, including LSI, could plausibly have been prototyped. A version without LSI may have been technically possible several years earlier.

No LLM Required

One of the most striking conclusions is that the system would not need anything resembling modern generative AI. It could be almost entirely deterministic.

Expert systems were mature enough by the 1980s to encode large sets of explicit rules. A communications-intelligence system could use rules like this:

IF a selected number contacts a new number,
AND the call contains multiple keyword hits,
AND the contact pattern deviates from baseline,
THEN increase priority.

Add network statistics, time-series anomaly detection and category similarity, then produce a structured report. No generative prose, no hallucinations, no hidden reasoning. Every decision could be logged, every score could be traced to its source, and every forwarded call would have a mechanical explanation.

The machine’s purpose would not be to replace the intelligence analyst. It would be to reduce volume. Millions of events enter, thousands survive, hundreds receive deeper processing, and a smaller number reach human beings.

That is a very different idea of “artificial intelligence.” It is closer to an industrial sorting machine.

The Biggest Limitation Wasn’t Necessarily Compute

This is where the reconstruction becomes especially interesting. You might expect raw computing power to be the fatal constraint. It was certainly a major one. But the paper concludes that an even worse problem appears as the system scales: false alarms.

Suppose a word spotter has what seems like a low error rate. Multiply that error across huge numbers of calls, then again across dozens of monitored keywords, and the false alerts begin to overwhelm the system. With 50 keywords and just one false alarm per keyword per hour, about 92 percent of ordinary three-minute calls would be flagged. Keyword spotters of the period did no better than that. With watch lists of that size, the funnel stops narrowing.

This is the base-rate problem. If the overwhelming majority of calls are irrelevant, even an impressive detector produces mostly false positives. That is why targeting by metadata matters so much. The architecture works far better when it starts from a small set of numbers, routes, individuals or circuits that are already of interest.

The system becomes plausible as a prioritization tool. It becomes much less plausible as an all-hearing national listener.

The paper is blunt about this. The funnel can cut computational demand roughly 90 to 400 times, but it cannot remove the false-alarm and base-rate problem. Effective selectivity requires targeting before any speech analysis.

Storage Was Another Brutal Constraint

Even if recognition had been perfect, storing everything would have been enormously expensive. At standard digital telephone quality, one continuously active channel produces roughly:

• 28.8 MB per hour
• 691 MB per day
• about 0.25 TB per year

Scale that to 1,000 channels and you get 691 GB per day. At 10,000 channels, it is 6.91 TB per day. Those numbers were punishing for 1980s storage systems.

The answer is again the funnel. Do not archive everything. Keep short buffers, extract features, store metadata and keyword hits, retain selected conversations, and discard most of the raw audio. The paper concludes that keeping raw audio long term beyond a few thousand channels would have been impractical with period storage, while selective retention was feasible.

This is not just an implementation detail. It tells us what the architecture would have had to look like. A system of that period would have been designed around forgetting.

So Could It Have Existed?

This needs a careful answer. According to the research:

• A research testbed covering tens to low hundreds of channels is technically plausible by 1988–1989.
• A targeted system covering hundreds to low thousands of metadata-selected channels is also plausible as a specialized operational capability.
• A system that indiscriminately screens tens of thousands to a million simultaneous channels with useful selectivity is another matter. That is not demonstrably plausible at that scale in 1989.

The limits were not only hardware. They included recognition accuracy, false alarms, base rates, storage, the long-haul links needed to move audio, and access to the traffic itself. The paper places genuinely effective untargeted large-scale deployment no earlier than the late 1990s, and more defensibly in the 2000s.

The final classification is surprisingly nuanced:

• Small research prototype: plausible.
• Narrowly targeted operational system: plausible.
• National full-transcription system: computationally impractical and inaccurate.
• Untargeted national automated listener: not demonstrably plausible in 1989.

And most importantly, there is no credible public documentary evidence that the complete integrated system described here existed in 1988–1989. A 1999 European Parliament technical assessment went further. It reported that effective automatic word-spotting of telephone calls still was not available, “despite 30 years of research.”

The Part That Should Make Us Think

The most interesting lesson here is not really about surveillance. It is about technological history.

We often imagine major systems appearing when one breakthrough invention arrives. Real systems rarely evolve that way. Technologies develop independently: speech recognition in one laboratory, information retrieval in another, digital telephony somewhere else, and expert systems in another field entirely. Add graph theory, statistical process control, signal processing, supercomputing and database systems.

Each looks ordinary in isolation. Then someone connects them, and the system can suddenly do something none of the individual parts seemed capable of.

That is the genuinely open question in this reconstruction. By 1988, the public literature contained the pieces. The unresolved historical question is not whether those pieces existed; they did. It is whether anyone outside the public literature assembled them earlier than the public record shows.

The available evidence does not answer that question, and responsible research should not pretend it does. But technically, the threshold is much earlier than most people would guess.

The Bottom Line

The paper’s central conclusion comes down to this. A recognizable ancestor of a modern machine-assisted communications-intelligence pipeline could plausibly have been assembled from publicly documented technology by 1988–1989.

It would not have been a modern AI. It would not have understood language the way current models do, and it would not have flawlessly transcribed a nation. Instead, it would have been a distributed computational funnel:

metadata → acoustic filtering → keyword spotting → partial recognition → vectorization → similarity analysis → graph analysis → anomaly detection → expert-system triage → deterministic reporting → human analyst

It is a primitive architecture by today’s standards, but an unmistakably familiar one. And perhaps that is the most important observation of all. Sometimes the ancestor of a modern AI system does not look like artificial intelligence. It looks like a telephone switch, a rack of DSP boards, a database, a vector supercomputer and a rule engine, quietly passing information from one stage to the next.

Research basis: A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989: A Source-Driven Technical Plausibility Reconstruction Using Only Publicly Documented Technology.

-Sean 9-26-2026

Could the Government Have Built an AI-Like Telephone Intelligence System in 1989?

https://youtu.be/jt3DQ5vT9s8

Slug: telephone-intelligence-ai-1989
Meta description: Could 1989 technology have screened phone calls automatically? A source-driven look at keyword spotting, vectors, expert systems and their hard limits.

Could an intelligence agency in 1989 have built a primitive ancestor of today’s AI-driven communications analysis systems? A 66-page technical reconstruction, using only publicly documented technology, suggests that nearly every required component already existed: keyword spotting and DSP arrays, vector retrieval, latent semantic analysis, graph analytics, expert systems and automated reporting.

What if, in 1989, a government intelligence organization had wanted to build a machine that could listen to telephone traffic? A machine that could detect selected words, compare communications mathematically, analyze who was calling whom, flag unusual activity, and produce reports for human analysts automatically?

Not with modern artificial intelligence. Not with Whisper, transformer models or cloud computing. And certainly not with large language models.

Just with the technology that was publicly documented at the time.

That is the question explored in our research paper, A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989. The study does not claim that such a classified system existed. It asks a narrower and more defensible question:

What could have been built?

The distinction matters. The paper is a technical plausibility reconstruction. It draws on peer-reviewed research, technical reports, patents, vendor documentation, government records and the limited declassified material in the public record.

The result is more interesting than a simple yes or no. By the end of the 1980s, virtually every major component needed for a primitive machine-assisted communications-intelligence system already existed. The surprising part is not any one technology. It is what happens when they are connected together.

The Wrong Question: “Could a 1989 Computer Transcribe Every Phone Call?”

If we frame the problem that way, the answer is simple: no.

Large-vocabulary recognition of ordinary conversational telephone speech was nowhere near good enough. The first published attempt at it came in 1993, and it got roughly 78 percent of the words wrong.

So the science-fiction version, a gigantic 1989 machine perfectly transcribing every call in America, does not survive technical scrutiny. But that is not how an efficient intelligence system would have been designed. The better question is this:

Could thousands of specialized systems work together as a funnel, examining enormous volumes of communications cheaply and forwarding only a tiny fraction for deeper processing?

That reframes the problem. The reconstruction finds that a cascaded architecture could cut computational demand by roughly 90 to 400 times, depending on the scenario. It also finds that keyword screening of tens of thousands of simultaneous calls was arithmetically possible with large numbers of late-1980s digital signal processors (DSPs), even though full conversational transcription was not.

The system would not need to understand everything. It would need to know what to ignore.

Stage One: The Telephone Network Was Already Becoming Digital

The first important fact is easy to overlook. By 1989, much of the telephone system was already becoming a digital data network.

The Bell System put the T1 carrier into commercial service in 1962. A single T1 stream carried 24 voice channels; higher-capacity DS3 links carried 672. Standard digital telephone audio ran at 64 kilobits per second per channel.

That meant that wherever a hypothetical collector had access to a digital trunk line, the voice did not need to be recorded from an analog wire and digitized afterward. It was already data. A digital carrier could be split back into individual voice channels for machine processing.

And the network provided something even easier to analyze than voice: metadata. Stored-program telephone switches already generated machine-readable call records. These held the calling number, called number, date, start time and duration.

From an engineering perspective, that is the obvious first filter. Before asking a machine to recognize speech, ask cheaper questions. Who is calling, and who are they calling? How often, and at what time? Is the number already known? Is it connected to other known numbers?

The reconstruction concludes that metadata would have been the cheapest and most practical first-stage selector in a 1989 screening architecture. That changes the scale of the voice-processing problem dramatically. You do not listen deeply to everything. You select first, and then you listen.

Stage Two: Machines Had Been Learning to Recognize Speech Since the 1950s

Machine speech recognition was not new in the 1980s. Bell Labs demonstrated AUDREY in 1952, which could recognize spoken digits from a trained speaker. IBM followed with the Shoebox in the early 1960s. By the 1970s, speech recognition had become a major research field involving ARPA, Carnegie Mellon, IBM, Bell Labs, BBN and others.

One of the most important systems was Carnegie Mellon’s HARPY. It handled a vocabulary of 1,011 words using a network of roughly 15,000 states. It was also very expensive to run: about 28 million instructions for every second of speech. On its PDP-10 hardware, it ran about 80 times slower than real time.

That sounds unusable. But HARPY showed something critical: the problem was no longer theoretical. Machines could already turn speech into constrained symbolic interpretations. The remaining problem was cost, and cost can sometimes be attacked with architecture.

Stage Three: Don’t Transcribe the Call — Search for Words

This is where the reconstruction becomes much more practical. A machine searching for ten or twenty specific words does not need to solve the entire speech recognition problem. It needs to solve a smaller one: keyword spotting.

Researchers were already working on this in the 1970s. The public record includes:

• elastic-template detection of words in running speech (1973)
• a Rockwell International patent describing keyword detection in continuous speech (1975)
• keyword spotting with linear predictive coding and dynamic time warping (1977)
• keyword-recognition research by a defense contractor, ITT (1985)
• keyword spotting with hidden Markov models at AT&T Bell Labs and MIT Lincoln Laboratory (around 1990)

The underlying idea is simple. Treat most speech as irrelevant filler and search only for specific acoustic patterns. If a confidence score crosses a threshold, forward the call; if not, discard it. This is far cheaper than full transcription.

Specialized DSP chips became common in the 1980s. Once they did, large numbers of relatively inexpensive processors could handle the front-end signal work in parallel. That gives us the foundation of the funnel.

The Computational Funnel

The hypothetical system reconstructed in the paper has six tiers, which fall into five broad stages.

Collection and metadata filtering. Digital carriers are split into individual channels and call records are captured. Known numbers, routes or circuits are selected. Voice-activity detection separates speech from silence, and tone detectors reject modem, fax and other non-speech audio.

Regional speech processing. DSP arrays extract acoustic features, keyword spotters search for selected terms, and only calls that cross set thresholds go on to more expensive recognition systems.

Central analysis. Partial transcripts are normalized and their terms weighted. Call records enter databases. Graph statistics reveal new contacts and unusual calling structures, and statistical systems flag changes over time.

High-end numerical processing. Vector supercomputers periodically run heavy mathematics such as singular value decomposition (SVD). Massively parallel machines run similarity searches across large collections.

Rules and reporting. An expert system assigns priority, and a template system generates a report. A human analyst receives the audio, transcript fragments, scores, graph relationships and alerts.

Each class of hardware gets the work it suits: DSPs for signal processing, workstations for recognition, database systems for records, supercomputers for SVD and statistics, and rule engines for final triage.

That division of labor is the key. The system does not depend on one impossibly advanced computer. It depends on many ordinary specialized ones.

Then Something Unexpected Happens: Speech Becomes a Vector

Today we talk casually about embeddings and vector databases, but representing documents mathematically is much older than modern AI. Gerard Salton’s information-retrieval work helped establish the vector-space model decades ago. A document could be represented as a vector based on the words it contained, using term frequency, inverse document frequency and cosine similarity. Computers could then compare documents mathematically rather than simply matching exact strings.

Now imagine feeding a noisy, imperfect telephone transcript into that system. The transcript does not need to be perfect. If enough important terms survive recognition, it can still become a weighted vector and be compared with categories or earlier communications.

The published evidence backs this up. In 1993, a system identified the topic of conversational phone calls 88 percent of the time among ten topics, even though it got most of the words wrong. Topics survive recognition errors far better than transcripts do, because the evidence adds up across many words.

Then, in 1988, another key piece appeared publicly: latent semantic indexing (LSI). LSI used SVD to project the relationships between terms and documents into a lower-dimensional mathematical space. It was not a neural embedding, but it solved a related problem. Documents with statistically related patterns of language could move closer together mathematically, even if they did not share the same words.

That date is one reason the paper’s target period matters. By 1988, every major stage of the hypothetical pipeline had appeared in the public technical literature. The reconstruction identifies 1988–1989 as the earliest period when the complete architecture, including LSI, could plausibly have been prototyped. A version without LSI may have been technically possible several years earlier.

No LLM Required

One of the most striking conclusions is that the system would not need anything resembling modern generative AI. It could be almost entirely deterministic.

Expert systems were mature enough by the 1980s to encode large sets of explicit rules. A communications-intelligence system could use rules like this:

IF a selected number contacts a new number,
AND the call contains multiple keyword hits,
AND the contact pattern deviates from baseline,
THEN increase priority.

Add network statistics, time-series anomaly detection and category similarity, then produce a structured report. No generative prose, no hallucinations, no hidden reasoning. Every decision could be logged, every score could be traced to its source, and every forwarded call would have a mechanical explanation.

The machine’s purpose would not be to replace the intelligence analyst. It would be to reduce volume. Millions of events enter, thousands survive, hundreds receive deeper processing, and a smaller number reach human beings.

That is a very different idea of “artificial intelligence.” It is closer to an industrial sorting machine.

The Biggest Limitation Wasn’t Necessarily Compute

This is where the reconstruction becomes especially interesting. You might expect raw computing power to be the fatal constraint. It was certainly a major one. But the paper concludes that an even worse problem appears as the system scales: false alarms.

Suppose a word spotter has what seems like a low error rate. Multiply that error across huge numbers of calls, then again across dozens of monitored keywords, and the false alerts begin to overwhelm the system. With 50 keywords and just one false alarm per keyword per hour, about 92 percent of ordinary three-minute calls would be flagged. Keyword spotters of the period did no better than that. With watch lists of that size, the funnel stops narrowing.

This is the base-rate problem. If the overwhelming majority of calls are irrelevant, even an impressive detector produces mostly false positives. That is why targeting by metadata matters so much. The architecture works far better when it starts from a small set of numbers, routes, individuals or circuits that are already of interest.

The system becomes plausible as a prioritization tool. It becomes much less plausible as an all-hearing national listener.

The paper is blunt about this. The funnel can cut computational demand roughly 90 to 400 times, but it cannot remove the false-alarm and base-rate problem. Effective selectivity requires targeting before any speech analysis.

Storage Was Another Brutal Constraint

Even if recognition had been perfect, storing everything would have been enormously expensive. At standard digital telephone quality, one continuously active channel produces roughly:

• 28.8 MB per hour
• 691 MB per day
• about 0.25 TB per year

Scale that to 1,000 channels and you get 691 GB per day. At 10,000 channels, it is 6.91 TB per day. Those numbers were punishing for 1980s storage systems.

The answer is again the funnel. Do not archive everything. Keep short buffers, extract features, store metadata and keyword hits, retain selected conversations, and discard most of the raw audio. The paper concludes that keeping raw audio long term beyond a few thousand channels would have been impractical with period storage, while selective retention was feasible.

This is not just an implementation detail. It tells us what the architecture would have had to look like. A system of that period would have been designed around forgetting.

So Could It Have Existed?

This needs a careful answer. According to the research:

• A research testbed covering tens to low hundreds of channels is technically plausible by 1988–1989.
• A targeted system covering hundreds to low thousands of metadata-selected channels is also plausible as a specialized operational capability.
• A system that indiscriminately screens tens of thousands to a million simultaneous channels with useful selectivity is another matter. That is not demonstrably plausible at that scale in 1989.

The limits were not only hardware. They included recognition accuracy, false alarms, base rates, storage, the long-haul links needed to move audio, and access to the traffic itself. The paper places genuinely effective untargeted large-scale deployment no earlier than the late 1990s, and more defensibly in the 2000s.

The final classification is surprisingly nuanced:

• Small research prototype: plausible.
• Narrowly targeted operational system: plausible.
• National full-transcription system: computationally impractical and inaccurate.
• Untargeted national automated listener: not demonstrably plausible in 1989.

And most importantly, there is no credible public documentary evidence that the complete integrated system described here existed in 1988–1989. A 1999 European Parliament technical assessment went further. It reported that effective automatic word-spotting of telephone calls still was not available, “despite 30 years of research.”

The Part That Should Make Us Think

The most interesting lesson here is not really about surveillance. It is about technological history.

We often imagine major systems appearing when one breakthrough invention arrives. Real systems rarely evolve that way. Technologies develop independently: speech recognition in one laboratory, information retrieval in another, digital telephony somewhere else, and expert systems in another field entirely. Add graph theory, statistical process control, signal processing, supercomputing and database systems.

Each looks ordinary in isolation. Then someone connects them, and the system can suddenly do something none of the individual parts seemed capable of.

That is the genuinely open question in this reconstruction. By 1988, the public literature contained the pieces. The unresolved historical question is not whether those pieces existed; they did. It is whether anyone outside the public literature assembled them earlier than the public record shows.

The available evidence does not answer that question, and responsible research should not pretend it does. But technically, the threshold is much earlier than most people would guess.

The Bottom Line

The paper’s central conclusion comes down to this. A recognizable ancestor of a modern machine-assisted communications-intelligence pipeline could plausibly have been assembled from publicly documented technology by 1988–1989.

It would not have been a modern AI. It would not have understood language the way current models do, and it would not have flawlessly transcribed a nation. Instead, it would have been a distributed computational funnel:

metadata → acoustic filtering → keyword spotting → partial recognition → vectorization → similarity analysis → graph analysis → anomaly detection → expert-system triage → deterministic reporting → human analyst

It is a primitive architecture by today’s standards, but an unmistakably familiar one. And perhaps that is the most important observation of all. Sometimes the ancestor of a modern AI system does not look like artificial intelligence. It looks like a telephone switch, a rack of DSP boards, a database, a vector supercomputer and a rule engine, quietly passing information from one stage to the next.

Research basis: A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989: A Source-Driven Technical Plausibility Reconstruction Using Only Publicly Documented Technology.

-Sean 9-26-2026

Could the Government Have Built an AI-Like Telephone Intelligence System in 1989?

https://youtu.be/jt3DQ5vT9s8

Slug: telephone-intelligence-ai-1989
Meta description: Could 1989 technology have screened phone calls automatically? A source-driven look at keyword spotting, vectors, expert systems and their hard limits.

Could an intelligence agency in 1989 have built a primitive ancestor of today’s AI-driven communications analysis systems? A 66-page technical reconstruction, using only publicly documented technology, suggests that nearly every required component already existed: keyword spotting and DSP arrays, vector retrieval, latent semantic analysis, graph analytics, expert systems and automated reporting.

What if, in 1989, a government intelligence organization had wanted to build a machine that could listen to telephone traffic? A machine that could detect selected words, compare communications mathematically, analyze who was calling whom, flag unusual activity, and produce reports for human analysts automatically?

Not with modern artificial intelligence. Not with Whisper, transformer models or cloud computing. And certainly not with large language models.

Just with the technology that was publicly documented at the time.

That is the question explored in our research paper, A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989. The study does not claim that such a classified system existed. It asks a narrower and more defensible question:

What could have been built?

The distinction matters. The paper is a technical plausibility reconstruction. It draws on peer-reviewed research, technical reports, patents, vendor documentation, government records and the limited declassified material in the public record.

The result is more interesting than a simple yes or no. By the end of the 1980s, virtually every major component needed for a primitive machine-assisted communications-intelligence system already existed. The surprising part is not any one technology. It is what happens when they are connected together.

The Wrong Question: “Could a 1989 Computer Transcribe Every Phone Call?”

If we frame the problem that way, the answer is simple: no.

Large-vocabulary recognition of ordinary conversational telephone speech was nowhere near good enough. The first published attempt at it came in 1993, and it got roughly 78 percent of the words wrong.

So the science-fiction version, a gigantic 1989 machine perfectly transcribing every call in America, does not survive technical scrutiny. But that is not how an efficient intelligence system would have been designed. The better question is this:

Could thousands of specialized systems work together as a funnel, examining enormous volumes of communications cheaply and forwarding only a tiny fraction for deeper processing?

That reframes the problem. The reconstruction finds that a cascaded architecture could cut computational demand by roughly 90 to 400 times, depending on the scenario. It also finds that keyword screening of tens of thousands of simultaneous calls was arithmetically possible with large numbers of late-1980s digital signal processors (DSPs), even though full conversational transcription was not.

The system would not need to understand everything. It would need to know what to ignore.

Stage One: The Telephone Network Was Already Becoming Digital

The first important fact is easy to overlook. By 1989, much of the telephone system was already becoming a digital data network.

The Bell System put the T1 carrier into commercial service in 1962. A single T1 stream carried 24 voice channels; higher-capacity DS3 links carried 672. Standard digital telephone audio ran at 64 kilobits per second per channel.

That meant that wherever a hypothetical collector had access to a digital trunk line, the voice did not need to be recorded from an analog wire and digitized afterward. It was already data. A digital carrier could be split back into individual voice channels for machine processing.

And the network provided something even easier to analyze than voice: metadata. Stored-program telephone switches already generated machine-readable call records. These held the calling number, called number, date, start time and duration.

From an engineering perspective, that is the obvious first filter. Before asking a machine to recognize speech, ask cheaper questions. Who is calling, and who are they calling? How often, and at what time? Is the number already known? Is it connected to other known numbers?

The reconstruction concludes that metadata would have been the cheapest and most practical first-stage selector in a 1989 screening architecture. That changes the scale of the voice-processing problem dramatically. You do not listen deeply to everything. You select first, and then you listen.

Stage Two: Machines Had Been Learning to Recognize Speech Since the 1950s

Machine speech recognition was not new in the 1980s. Bell Labs demonstrated AUDREY in 1952, which could recognize spoken digits from a trained speaker. IBM followed with the Shoebox in the early 1960s. By the 1970s, speech recognition had become a major research field involving ARPA, Carnegie Mellon, IBM, Bell Labs, BBN and others.

One of the most important systems was Carnegie Mellon’s HARPY. It handled a vocabulary of 1,011 words using a network of roughly 15,000 states. It was also very expensive to run: about 28 million instructions for every second of speech. On its PDP-10 hardware, it ran about 80 times slower than real time.

That sounds unusable. But HARPY showed something critical: the problem was no longer theoretical. Machines could already turn speech into constrained symbolic interpretations. The remaining problem was cost, and cost can sometimes be attacked with architecture.

Stage Three: Don’t Transcribe the Call — Search for Words

This is where the reconstruction becomes much more practical. A machine searching for ten or twenty specific words does not need to solve the entire speech recognition problem. It needs to solve a smaller one: keyword spotting.

Researchers were already working on this in the 1970s. The public record includes:

• elastic-template detection of words in running speech (1973)
• a Rockwell International patent describing keyword detection in continuous speech (1975)
• keyword spotting with linear predictive coding and dynamic time warping (1977)
• keyword-recognition research by a defense contractor, ITT (1985)
• keyword spotting with hidden Markov models at AT&T Bell Labs and MIT Lincoln Laboratory (around 1990)

The underlying idea is simple. Treat most speech as irrelevant filler and search only for specific acoustic patterns. If a confidence score crosses a threshold, forward the call; if not, discard it. This is far cheaper than full transcription.

Specialized DSP chips became common in the 1980s. Once they did, large numbers of relatively inexpensive processors could handle the front-end signal work in parallel. That gives us the foundation of the funnel.

The Computational Funnel

The hypothetical system reconstructed in the paper has six tiers, which fall into five broad stages.

Collection and metadata filtering. Digital carriers are split into individual channels and call records are captured. Known numbers, routes or circuits are selected. Voice-activity detection separates speech from silence, and tone detectors reject modem, fax and other non-speech audio.

Regional speech processing. DSP arrays extract acoustic features, keyword spotters search for selected terms, and only calls that cross set thresholds go on to more expensive recognition systems.

Central analysis. Partial transcripts are normalized and their terms weighted. Call records enter databases. Graph statistics reveal new contacts and unusual calling structures, and statistical systems flag changes over time.

High-end numerical processing. Vector supercomputers periodically run heavy mathematics such as singular value decomposition (SVD). Massively parallel machines run similarity searches across large collections.

Rules and reporting. An expert system assigns priority, and a template system generates a report. A human analyst receives the audio, transcript fragments, scores, graph relationships and alerts.

Each class of hardware gets the work it suits: DSPs for signal processing, workstations for recognition, database systems for records, supercomputers for SVD and statistics, and rule engines for final triage.

That division of labor is the key. The system does not depend on one impossibly advanced computer. It depends on many ordinary specialized ones.

Then Something Unexpected Happens: Speech Becomes a Vector

Today we talk casually about embeddings and vector databases, but representing documents mathematically is much older than modern AI. Gerard Salton’s information-retrieval work helped establish the vector-space model decades ago. A document could be represented as a vector based on the words it contained, using term frequency, inverse document frequency and cosine similarity. Computers could then compare documents mathematically rather than simply matching exact strings.

Now imagine feeding a noisy, imperfect telephone transcript into that system. The transcript does not need to be perfect. If enough important terms survive recognition, it can still become a weighted vector and be compared with categories or earlier communications.

The published evidence backs this up. In 1993, a system identified the topic of conversational phone calls 88 percent of the time among ten topics, even though it got most of the words wrong. Topics survive recognition errors far better than transcripts do, because the evidence adds up across many words.

Then, in 1988, another key piece appeared publicly: latent semantic indexing (LSI). LSI used SVD to project the relationships between terms and documents into a lower-dimensional mathematical space. It was not a neural embedding, but it solved a related problem. Documents with statistically related patterns of language could move closer together mathematically, even if they did not share the same words.

That date is one reason the paper’s target period matters. By 1988, every major stage of the hypothetical pipeline had appeared in the public technical literature. The reconstruction identifies 1988–1989 as the earliest period when the complete architecture, including LSI, could plausibly have been prototyped. A version without LSI may have been technically possible several years earlier.

No LLM Required

One of the most striking conclusions is that the system would not need anything resembling modern generative AI. It could be almost entirely deterministic.

Expert systems were mature enough by the 1980s to encode large sets of explicit rules. A communications-intelligence system could use rules like this:

IF a selected number contacts a new number,
AND the call contains multiple keyword hits,
AND the contact pattern deviates from baseline,
THEN increase priority.

Add network statistics, time-series anomaly detection and category similarity, then produce a structured report. No generative prose, no hallucinations, no hidden reasoning. Every decision could be logged, every score could be traced to its source, and every forwarded call would have a mechanical explanation.

The machine’s purpose would not be to replace the intelligence analyst. It would be to reduce volume. Millions of events enter, thousands survive, hundreds receive deeper processing, and a smaller number reach human beings.

That is a very different idea of “artificial intelligence.” It is closer to an industrial sorting machine.

The Biggest Limitation Wasn’t Necessarily Compute

This is where the reconstruction becomes especially interesting. You might expect raw computing power to be the fatal constraint. It was certainly a major one. But the paper concludes that an even worse problem appears as the system scales: false alarms.

Suppose a word spotter has what seems like a low error rate. Multiply that error across huge numbers of calls, then again across dozens of monitored keywords, and the false alerts begin to overwhelm the system. With 50 keywords and just one false alarm per keyword per hour, about 92 percent of ordinary three-minute calls would be flagged. Keyword spotters of the period did no better than that. With watch lists of that size, the funnel stops narrowing.

This is the base-rate problem. If the overwhelming majority of calls are irrelevant, even an impressive detector produces mostly false positives. That is why targeting by metadata matters so much. The architecture works far better when it starts from a small set of numbers, routes, individuals or circuits that are already of interest.

The system becomes plausible as a prioritization tool. It becomes much less plausible as an all-hearing national listener.

The paper is blunt about this. The funnel can cut computational demand roughly 90 to 400 times, but it cannot remove the false-alarm and base-rate problem. Effective selectivity requires targeting before any speech analysis.

Storage Was Another Brutal Constraint

Even if recognition had been perfect, storing everything would have been enormously expensive. At standard digital telephone quality, one continuously active channel produces roughly:

• 28.8 MB per hour
• 691 MB per day
• about 0.25 TB per year

Scale that to 1,000 channels and you get 691 GB per day. At 10,000 channels, it is 6.91 TB per day. Those numbers were punishing for 1980s storage systems.

The answer is again the funnel. Do not archive everything. Keep short buffers, extract features, store metadata and keyword hits, retain selected conversations, and discard most of the raw audio. The paper concludes that keeping raw audio long term beyond a few thousand channels would have been impractical with period storage, while selective retention was feasible.

This is not just an implementation detail. It tells us what the architecture would have had to look like. A system of that period would have been designed around forgetting.

So Could It Have Existed?

This needs a careful answer. According to the research:

• A research testbed covering tens to low hundreds of channels is technically plausible by 1988–1989.
• A targeted system covering hundreds to low thousands of metadata-selected channels is also plausible as a specialized operational capability.
• A system that indiscriminately screens tens of thousands to a million simultaneous channels with useful selectivity is another matter. That is not demonstrably plausible at that scale in 1989.

The limits were not only hardware. They included recognition accuracy, false alarms, base rates, storage, the long-haul links needed to move audio, and access to the traffic itself. The paper places genuinely effective untargeted large-scale deployment no earlier than the late 1990s, and more defensibly in the 2000s.

The final classification is surprisingly nuanced:

• Small research prototype: plausible.
• Narrowly targeted operational system: plausible.
• National full-transcription system: computationally impractical and inaccurate.
• Untargeted national automated listener: not demonstrably plausible in 1989.

And most importantly, there is no credible public documentary evidence that the complete integrated system described here existed in 1988–1989. A 1999 European Parliament technical assessment went further. It reported that effective automatic word-spotting of telephone calls still was not available, “despite 30 years of research.”

The Part That Should Make Us Think

The most interesting lesson here is not really about surveillance. It is about technological history.

We often imagine major systems appearing when one breakthrough invention arrives. Real systems rarely evolve that way. Technologies develop independently: speech recognition in one laboratory, information retrieval in another, digital telephony somewhere else, and expert systems in another field entirely. Add graph theory, statistical process control, signal processing, supercomputing and database systems.

Each looks ordinary in isolation. Then someone connects them, and the system can suddenly do something none of the individual parts seemed capable of.

That is the genuinely open question in this reconstruction. By 1988, the public literature contained the pieces. The unresolved historical question is not whether those pieces existed; they did. It is whether anyone outside the public literature assembled them earlier than the public record shows.

The available evidence does not answer that question, and responsible research should not pretend it does. But technically, the threshold is much earlier than most people would guess.

The Bottom Line

The paper’s central conclusion comes down to this. A recognizable ancestor of a modern machine-assisted communications-intelligence pipeline could plausibly have been assembled from publicly documented technology by 1988–1989.

It would not have been a modern AI. It would not have understood language the way current models do, and it would not have flawlessly transcribed a nation. Instead, it would have been a distributed computational funnel:

metadata → acoustic filtering → keyword spotting → partial recognition → vectorization → similarity analysis → graph analysis → anomaly detection → expert-system triage → deterministic reporting → human analyst

It is a primitive architecture by today’s standards, but an unmistakably familiar one. And perhaps that is the most important observation of all. Sometimes the ancestor of a modern AI system does not look like artificial intelligence. It looks like a telephone switch, a rack of DSP boards, a database, a vector supercomputer and a rule engine, quietly passing information from one stage to the next.

Research basis: A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989: A Source-Driven Technical Plausibility Reconstruction Using Only Publicly Documented Technology.

-Sean 9-26-2026

Could the Government Have Built an AI-Like Telephone Intelligence System in 1989?

https://youtu.be/jt3DQ5vT9s8

Slug: telephone-intelligence-ai-1989
Meta description: Could 1989 technology have screened phone calls automatically? A source-driven look at keyword spotting, vectors, expert systems and their hard limits.

Could an intelligence agency in 1989 have built a primitive ancestor of today’s AI-driven communications analysis systems? A 66-page technical reconstruction, using only publicly documented technology, suggests that nearly every required component already existed: keyword spotting and DSP arrays, vector retrieval, latent semantic analysis, graph analytics, expert systems and automated reporting.

What if, in 1989, a government intelligence organization had wanted to build a machine that could listen to telephone traffic? A machine that could detect selected words, compare communications mathematically, analyze who was calling whom, flag unusual activity, and produce reports for human analysts automatically?

Not with modern artificial intelligence. Not with Whisper, transformer models or cloud computing. And certainly not with large language models.

Just with the technology that was publicly documented at the time.

That is the question explored in our research paper, A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989. The study does not claim that such a classified system existed. It asks a narrower and more defensible question:

What could have been built?

The distinction matters. The paper is a technical plausibility reconstruction. It draws on peer-reviewed research, technical reports, patents, vendor documentation, government records and the limited declassified material in the public record.

The result is more interesting than a simple yes or no. By the end of the 1980s, virtually every major component needed for a primitive machine-assisted communications-intelligence system already existed. The surprising part is not any one technology. It is what happens when they are connected together.

The Wrong Question: “Could a 1989 Computer Transcribe Every Phone Call?”

If we frame the problem that way, the answer is simple: no.

Large-vocabulary recognition of ordinary conversational telephone speech was nowhere near good enough. The first published attempt at it came in 1993, and it got roughly 78 percent of the words wrong.

So the science-fiction version, a gigantic 1989 machine perfectly transcribing every call in America, does not survive technical scrutiny. But that is not how an efficient intelligence system would have been designed. The better question is this:

Could thousands of specialized systems work together as a funnel, examining enormous volumes of communications cheaply and forwarding only a tiny fraction for deeper processing?

That reframes the problem. The reconstruction finds that a cascaded architecture could cut computational demand by roughly 90 to 400 times, depending on the scenario. It also finds that keyword screening of tens of thousands of simultaneous calls was arithmetically possible with large numbers of late-1980s digital signal processors (DSPs), even though full conversational transcription was not.

The system would not need to understand everything. It would need to know what to ignore.

Stage One: The Telephone Network Was Already Becoming Digital

The first important fact is easy to overlook. By 1989, much of the telephone system was already becoming a digital data network.

The Bell System put the T1 carrier into commercial service in 1962. A single T1 stream carried 24 voice channels; higher-capacity DS3 links carried 672. Standard digital telephone audio ran at 64 kilobits per second per channel.

That meant that wherever a hypothetical collector had access to a digital trunk line, the voice did not need to be recorded from an analog wire and digitized afterward. It was already data. A digital carrier could be split back into individual voice channels for machine processing.

And the network provided something even easier to analyze than voice: metadata. Stored-program telephone switches already generated machine-readable call records. These held the calling number, called number, date, start time and duration.

From an engineering perspective, that is the obvious first filter. Before asking a machine to recognize speech, ask cheaper questions. Who is calling, and who are they calling? How often, and at what time? Is the number already known? Is it connected to other known numbers?

The reconstruction concludes that metadata would have been the cheapest and most practical first-stage selector in a 1989 screening architecture. That changes the scale of the voice-processing problem dramatically. You do not listen deeply to everything. You select first, and then you listen.

Stage Two: Machines Had Been Learning to Recognize Speech Since the 1950s

Machine speech recognition was not new in the 1980s. Bell Labs demonstrated AUDREY in 1952, which could recognize spoken digits from a trained speaker. IBM followed with the Shoebox in the early 1960s. By the 1970s, speech recognition had become a major research field involving ARPA, Carnegie Mellon, IBM, Bell Labs, BBN and others.

One of the most important systems was Carnegie Mellon’s HARPY. It handled a vocabulary of 1,011 words using a network of roughly 15,000 states. It was also very expensive to run: about 28 million instructions for every second of speech. On its PDP-10 hardware, it ran about 80 times slower than real time.

That sounds unusable. But HARPY showed something critical: the problem was no longer theoretical. Machines could already turn speech into constrained symbolic interpretations. The remaining problem was cost, and cost can sometimes be attacked with architecture.

Stage Three: Don’t Transcribe the Call — Search for Words

This is where the reconstruction becomes much more practical. A machine searching for ten or twenty specific words does not need to solve the entire speech recognition problem. It needs to solve a smaller one: keyword spotting.

Researchers were already working on this in the 1970s. The public record includes:

• elastic-template detection of words in running speech (1973)
• a Rockwell International patent describing keyword detection in continuous speech (1975)
• keyword spotting with linear predictive coding and dynamic time warping (1977)
• keyword-recognition research by a defense contractor, ITT (1985)
• keyword spotting with hidden Markov models at AT&T Bell Labs and MIT Lincoln Laboratory (around 1990)

The underlying idea is simple. Treat most speech as irrelevant filler and search only for specific acoustic patterns. If a confidence score crosses a threshold, forward the call; if not, discard it. This is far cheaper than full transcription.

Specialized DSP chips became common in the 1980s. Once they did, large numbers of relatively inexpensive processors could handle the front-end signal work in parallel. That gives us the foundation of the funnel.

The Computational Funnel

The hypothetical system reconstructed in the paper has six tiers, which fall into five broad stages.

Collection and metadata filtering. Digital carriers are split into individual channels and call records are captured. Known numbers, routes or circuits are selected. Voice-activity detection separates speech from silence, and tone detectors reject modem, fax and other non-speech audio.

Regional speech processing. DSP arrays extract acoustic features, keyword spotters search for selected terms, and only calls that cross set thresholds go on to more expensive recognition systems.

Central analysis. Partial transcripts are normalized and their terms weighted. Call records enter databases. Graph statistics reveal new contacts and unusual calling structures, and statistical systems flag changes over time.

High-end numerical processing. Vector supercomputers periodically run heavy mathematics such as singular value decomposition (SVD). Massively parallel machines run similarity searches across large collections.

Rules and reporting. An expert system assigns priority, and a template system generates a report. A human analyst receives the audio, transcript fragments, scores, graph relationships and alerts.

Each class of hardware gets the work it suits: DSPs for signal processing, workstations for recognition, database systems for records, supercomputers for SVD and statistics, and rule engines for final triage.

That division of labor is the key. The system does not depend on one impossibly advanced computer. It depends on many ordinary specialized ones.

Then Something Unexpected Happens: Speech Becomes a Vector

Today we talk casually about embeddings and vector databases, but representing documents mathematically is much older than modern AI. Gerard Salton’s information-retrieval work helped establish the vector-space model decades ago. A document could be represented as a vector based on the words it contained, using term frequency, inverse document frequency and cosine similarity. Computers could then compare documents mathematically rather than simply matching exact strings.

Now imagine feeding a noisy, imperfect telephone transcript into that system. The transcript does not need to be perfect. If enough important terms survive recognition, it can still become a weighted vector and be compared with categories or earlier communications.

The published evidence backs this up. In 1993, a system identified the topic of conversational phone calls 88 percent of the time among ten topics, even though it got most of the words wrong. Topics survive recognition errors far better than transcripts do, because the evidence adds up across many words.

Then, in 1988, another key piece appeared publicly: latent semantic indexing (LSI). LSI used SVD to project the relationships between terms and documents into a lower-dimensional mathematical space. It was not a neural embedding, but it solved a related problem. Documents with statistically related patterns of language could move closer together mathematically, even if they did not share the same words.

That date is one reason the paper’s target period matters. By 1988, every major stage of the hypothetical pipeline had appeared in the public technical literature. The reconstruction identifies 1988–1989 as the earliest period when the complete architecture, including LSI, could plausibly have been prototyped. A version without LSI may have been technically possible several years earlier.

No LLM Required

One of the most striking conclusions is that the system would not need anything resembling modern generative AI. It could be almost entirely deterministic.

Expert systems were mature enough by the 1980s to encode large sets of explicit rules. A communications-intelligence system could use rules like this:

IF a selected number contacts a new number,
AND the call contains multiple keyword hits,
AND the contact pattern deviates from baseline,
THEN increase priority.

Add network statistics, time-series anomaly detection and category similarity, then produce a structured report. No generative prose, no hallucinations, no hidden reasoning. Every decision could be logged, every score could be traced to its source, and every forwarded call would have a mechanical explanation.

The machine’s purpose would not be to replace the intelligence analyst. It would be to reduce volume. Millions of events enter, thousands survive, hundreds receive deeper processing, and a smaller number reach human beings.

That is a very different idea of “artificial intelligence.” It is closer to an industrial sorting machine.

The Biggest Limitation Wasn’t Necessarily Compute

This is where the reconstruction becomes especially interesting. You might expect raw computing power to be the fatal constraint. It was certainly a major one. But the paper concludes that an even worse problem appears as the system scales: false alarms.

Suppose a word spotter has what seems like a low error rate. Multiply that error across huge numbers of calls, then again across dozens of monitored keywords, and the false alerts begin to overwhelm the system. With 50 keywords and just one false alarm per keyword per hour, about 92 percent of ordinary three-minute calls would be flagged. Keyword spotters of the period did no better than that. With watch lists of that size, the funnel stops narrowing.

This is the base-rate problem. If the overwhelming majority of calls are irrelevant, even an impressive detector produces mostly false positives. That is why targeting by metadata matters so much. The architecture works far better when it starts from a small set of numbers, routes, individuals or circuits that are already of interest.

The system becomes plausible as a prioritization tool. It becomes much less plausible as an all-hearing national listener.

The paper is blunt about this. The funnel can cut computational demand roughly 90 to 400 times, but it cannot remove the false-alarm and base-rate problem. Effective selectivity requires targeting before any speech analysis.

Storage Was Another Brutal Constraint

Even if recognition had been perfect, storing everything would have been enormously expensive. At standard digital telephone quality, one continuously active channel produces roughly:

• 28.8 MB per hour
• 691 MB per day
• about 0.25 TB per year

Scale that to 1,000 channels and you get 691 GB per day. At 10,000 channels, it is 6.91 TB per day. Those numbers were punishing for 1980s storage systems.

The answer is again the funnel. Do not archive everything. Keep short buffers, extract features, store metadata and keyword hits, retain selected conversations, and discard most of the raw audio. The paper concludes that keeping raw audio long term beyond a few thousand channels would have been impractical with period storage, while selective retention was feasible.

This is not just an implementation detail. It tells us what the architecture would have had to look like. A system of that period would have been designed around forgetting.

So Could It Have Existed?

This needs a careful answer. According to the research:

• A research testbed covering tens to low hundreds of channels is technically plausible by 1988–1989.
• A targeted system covering hundreds to low thousands of metadata-selected channels is also plausible as a specialized operational capability.
• A system that indiscriminately screens tens of thousands to a million simultaneous channels with useful selectivity is another matter. That is not demonstrably plausible at that scale in 1989.

The limits were not only hardware. They included recognition accuracy, false alarms, base rates, storage, the long-haul links needed to move audio, and access to the traffic itself. The paper places genuinely effective untargeted large-scale deployment no earlier than the late 1990s, and more defensibly in the 2000s.

The final classification is surprisingly nuanced:

• Small research prototype: plausible.
• Narrowly targeted operational system: plausible.
• National full-transcription system: computationally impractical and inaccurate.
• Untargeted national automated listener: not demonstrably plausible in 1989.

And most importantly, there is no credible public documentary evidence that the complete integrated system described here existed in 1988–1989. A 1999 European Parliament technical assessment went further. It reported that effective automatic word-spotting of telephone calls still was not available, “despite 30 years of research.”

The Part That Should Make Us Think

The most interesting lesson here is not really about surveillance. It is about technological history.

We often imagine major systems appearing when one breakthrough invention arrives. Real systems rarely evolve that way. Technologies develop independently: speech recognition in one laboratory, information retrieval in another, digital telephony somewhere else, and expert systems in another field entirely. Add graph theory, statistical process control, signal processing, supercomputing and database systems.

Each looks ordinary in isolation. Then someone connects them, and the system can suddenly do something none of the individual parts seemed capable of.

That is the genuinely open question in this reconstruction. By 1988, the public literature contained the pieces. The unresolved historical question is not whether those pieces existed; they did. It is whether anyone outside the public literature assembled them earlier than the public record shows.

The available evidence does not answer that question, and responsible research should not pretend it does. But technically, the threshold is much earlier than most people would guess.

The Bottom Line

The paper’s central conclusion comes down to this. A recognizable ancestor of a modern machine-assisted communications-intelligence pipeline could plausibly have been assembled from publicly documented technology by 1988–1989.

It would not have been a modern AI. It would not have understood language the way current models do, and it would not have flawlessly transcribed a nation. Instead, it would have been a distributed computational funnel:

metadata → acoustic filtering → keyword spotting → partial recognition → vectorization → similarity analysis → graph analysis → anomaly detection → expert-system triage → deterministic reporting → human analyst

It is a primitive architecture by today’s standards, but an unmistakably familiar one. And perhaps that is the most important observation of all. Sometimes the ancestor of a modern AI system does not look like artificial intelligence. It looks like a telephone switch, a rack of DSP boards, a database, a vector supercomputer and a rule engine, quietly passing information from one stage to the next.

Research basis: A Distributed Computational Funnel for Telephone Intelligence, circa 1988–1989: A Source-Driven Technical Plausibility Reconstruction Using Only Publicly Documented Technology.

-Sean 9-26-2026

🚨 Breaking news: #OpenAI #bots have allegedly been tampering with the ultra-secure, never-hackable US #government websites 🕵️‍♂️! Who knew that #AI could be smart enough to mess with the SEC, but not smart enough to avoid getting caught 🤖? Next up: AI becomes self-aware and runs for #Congress – it'll fit right in! 🏛️
https://www.bbc.com/news/articles/cw62jje658dlo #Security #Scandal #HackerNews #ngated

🚨 Breaking news: #OpenAI #bots have allegedly been tampering with the ultra-secure, never-hackable US #government websites 🕵️‍♂️! Who knew that #AI could be smart enough to mess with the SEC, but not smart enough to avoid getting caught 🤖? Next up: AI becomes self-aware and runs for #Congress – it'll fit right in! 🏛️
https://www.bbc.com/news/articles/cw62jje658dlo #Security #Scandal #HackerNews #ngated

🚨 Breaking news: #OpenAI #bots have allegedly been tampering with the ultra-secure, never-hackable US #government websites 🕵️‍♂️! Who knew that #AI could be smart enough to mess with the SEC, but not smart enough to avoid getting caught 🤖? Next up: AI becomes self-aware and runs for #Congress – it'll fit right in! 🏛️
https://www.bbc.com/news/articles/cw62jje658dlo #Security #Scandal #HackerNews #ngated

🚨 Breaking news: #OpenAI #bots have allegedly been tampering with the ultra-secure, never-hackable US #government websites 🕵️‍♂️! Who knew that #AI could be smart enough to mess with the SEC, but not smart enough to avoid getting caught 🤖? Next up: AI becomes self-aware and runs for #Congress – it'll fit right in! 🏛️
https://www.bbc.com/news/articles/cw62jje658dlo #Security #Scandal #HackerNews #ngated

🚨 Breaking news: #OpenAI #bots have allegedly been tampering with the ultra-secure, never-hackable US #government websites 🕵️‍♂️! Who knew that #AI could be smart enough to mess with the SEC, but not smart enough to avoid getting caught 🤖? Next up: AI becomes self-aware and runs for #Congress – it'll fit right in! 🏛️
https://www.bbc.com/news/articles/cw62jje658dlo #Security #Scandal #HackerNews #ngated

OpenAI bots meddled with multiple US Government agency sites

https://www.bbc.com/news/articles/cw62jje658dlo

Comments: https://news.ycombinator.com/item?id=49856665

#HackerNews #OpenAI #Government #Bots #US #Cybersecurity #AI #News

OpenAI bots meddled with multiple US Government agency sites

https://www.bbc.com/news/articles/cw62jje658dlo

Comments: https://news.ycombinator.com/item?id=49856665

#HackerNews #OpenAI #Government #Bots #US #Cybersecurity #AI #News

OpenAI bots meddled with multiple US Government agency sites

https://www.bbc.com/news/articles/cw62jje658dlo

Comments: https://news.ycombinator.com/item?id=49856665

#HackerNews #OpenAI #Government #Bots #US #Cybersecurity #AI #News

OpenAI bots meddled with multiple US Government agency sites

https://www.bbc.com/news/articles/cw62jje658dlo

Comments: https://news.ycombinator.com/item?id=49856665

#HackerNews #OpenAI #Government #Bots #US #Cybersecurity #AI #News

OpenAI bots meddled with multiple US Government agency sites

https://www.bbc.com/news/articles/cw62jje658dlo

Comments: https://news.ycombinator.com/item?id=49856665

#HackerNews #OpenAI #Government #Bots #US #Cybersecurity #AI #News

https://www.europesays.com/conflicts/9488/ Medal of Honor Monday: Marine Corps Brig. Gen. Robert Edward Galer > U.S. Department of War > Story #DepartmentOfWar #DeputySecretary #DeputySecretaryOfWar #DOW #Government #JointChief #KoreanWar #MarineCorps #MedalOfHonor #MedalOfHonorMonday #Military #moh #navy #Pentagon #Secretary #SecretaryOfWar #USMarines #UnitedStates #valor #WorldWarII

3 Reasons Investors Should Be Ready for Bond Market Sell-Off to Worsen

Rising long-term bond yields are once again pestering stocks, and BNP Paribas says it’s only going to get…
#France #FR #Europe #EU #BNPParibas #bank #bondmarket #government #House #interest #month #rate #reasoninvestor #stock #treasurybond #usmidtermelection #usd168bnincrease #year #yield
https://www.europesays.com/france/84843/

“Amidst all the talk of the need for #government #regulation of #AI, there has been too little discussion of the simplest route: #criminal #liability.” deanbaker22.substack.com/p/sam-altman...

Sam Altman is Driving Drunk, L...

“Amidst all the talk of the need for #government #regulation of #AI, there has been too little discussion of the simplest route: #criminal #liability.” deanbaker22.substack.com/p/sam-altman...

Sam Altman is Driving Drunk, L...

“Amidst all the talk of the need for #government #regulation of #AI, there has been too little discussion of the simplest route: #criminal #liability.” deanbaker22.substack.com/p/sam-altman...

Sam Altman is Driving Drunk, L...

“Amidst all the talk of the need for #government #regulation of #AI, there has been too little discussion of the simplest route: #criminal #liability.” deanbaker22.substack.com/p/sam-altman...

Sam Altman is Driving Drunk, L...

“Amidst all the talk of the need for #government #regulation of #AI, there has been too little discussion of the simplest route: #criminal #liability.” deanbaker22.substack.com/p/sam-altman...

Sam Altman is Driving Drunk, L...

Switzerland is piloting open source workplace software for 3,000 federal employees, replacing Microsoft 365 in parallel. 🇨🇭
The CHF 9 million program uses openDesk and targets completion by the end of 2027. 🐧

🔗 https://itsfoss.com/news/switzerland-replace-microssoft-pilot/

#TechNews #Switzerland #OpenSource #DigitalSovereignty #openDesk #Microsoft #Microsoft365 #LibreOffice #Linux #FOSS #Privacy #Security #Technology #Cloud #Software #Europe #EU #Government #Freedom #IT #PublicSector #Office #Tech

Switzerland is piloting open source workplace software for 3,000 federal employees, replacing Microsoft 365 in parallel. 🇨🇭
The CHF 9 million program uses openDesk and targets completion by the end of 2027. 🐧

🔗 https://itsfoss.com/news/switzerland-replace-microssoft-pilot/

#TechNews #Switzerland #OpenSource #DigitalSovereignty #openDesk #Microsoft #Microsoft365 #LibreOffice #Linux #FOSS #Privacy #Security #Technology #Cloud #Software #Europe #EU #Government #Freedom #IT #PublicSector #Office #Tech

Switzerland is piloting open source workplace software for 3,000 federal employees, replacing Microsoft 365 in parallel. 🇨🇭
The CHF 9 million program uses openDesk and targets completion by the end of 2027. 🐧

🔗 https://itsfoss.com/news/switzerland-replace-microssoft-pilot/

#TechNews #Switzerland #OpenSource #DigitalSovereignty #openDesk #Microsoft #Microsoft365 #LibreOffice #Linux #FOSS #Privacy #Security #Technology #Cloud #Software #Europe #EU #Government #Freedom #IT #PublicSector #Office #Tech

Switzerland is piloting open source workplace software for 3,000 federal employees, replacing Microsoft 365 in parallel. 🇨🇭
The CHF 9 million program uses openDesk and targets completion by the end of 2027. 🐧

🔗 https://itsfoss.com/news/switzerland-replace-microssoft-pilot/

#TechNews #Switzerland #OpenSource #DigitalSovereignty #openDesk #Microsoft #Microsoft365 #LibreOffice #Linux #FOSS #Privacy #Security #Technology #Cloud #Software #Europe #EU #Government #Freedom #IT #PublicSector #Office #Tech

Switzerland is piloting open source workplace software for 3,000 federal employees, replacing Microsoft 365 in parallel. 🇨🇭
The CHF 9 million program uses openDesk and targets completion by the end of 2027. 🐧

🔗 https://itsfoss.com/news/switzerland-replace-microssoft-pilot/

#TechNews #Switzerland #OpenSource #DigitalSovereignty #openDesk #Microsoft #Microsoft365 #LibreOffice #Linux #FOSS #Privacy #Security #Technology #Cloud #Software #Europe #EU #Government #Freedom #IT #PublicSector #Office #Tech

https://www.europesays.com/africa/444558/ PAM wins big in Morocco parliament elections but falls short of majority | Elections News #élections #Europe #Government #MiddleEast #Migration #Morocco #news #politics #Spain #WesternSahara #women

$10M grant announced to address Arkansas food insecurity https://www.diningandcooking.com/us/31976/ #Arkansas #ArkansasCommunityFoundation #ArkansasDining #ArkansasFood #ArkansasFullTableFund #ArkansasGroceryStores #ArkansasRestaurants #Children+Families #economy #FoodInsecurity #Government&Politics #JulieGehrki #SarahFord #SarahHuckabeeSanders #WalmartFoundation

$10M grant announced to address Arkansas food insecurity https://www.diningandcooking.com/us/31976/ #Arkansas #ArkansasCommunityFoundation #ArkansasDining #ArkansasFood #ArkansasFullTableFund #ArkansasGroceryStores #ArkansasRestaurants #Children+Families #economy #FoodInsecurity #Government&Politics #JulieGehrki #SarahFord #SarahHuckabeeSanders #WalmartFoundation

$10M grant announced to address Arkansas food insecurity https://www.diningandcooking.com/us/31976/ #Arkansas #ArkansasCommunityFoundation #ArkansasDining #ArkansasFood #ArkansasFullTableFund #ArkansasGroceryStores #ArkansasRestaurants #Children+Families #economy #FoodInsecurity #Government&Politics #JulieGehrki #SarahFord #SarahHuckabeeSanders #WalmartFoundation

[#TRADESHOW] #Security #China 2026 – China #International #Exhibition on #Public #Safety and Security will be held from November 3 to 6, 2026, at China International Exhibition Center (Shunyi Hall), #Beijing. As one of China’s most #influential #events dedicated to public safety, #intelligent security #technologies, and #digital #governance, the #expo brings together #government authorities, #industry #leaders, and #global #technology #providers. https://cnbusinessforum.com/event/security-china-2026-china-international-exhibition-on-public-safety-and-security/

[#TRADESHOW] #Security #China 2026 – China #International #Exhibition on #Public #Safety and Security will be held from November 3 to 6, 2026, at China International Exhibition Center (Shunyi Hall), #Beijing. As one of China’s most #influential #events dedicated to public safety, #intelligent security #technologies, and #digital #governance, the #expo brings together #government authorities, #industry #leaders, and #global #technology #providers. https://cnbusinessforum.com/event/security-china-2026-china-international-exhibition-on-public-safety-and-security/

[#TRADESHOW] #Security #China 2026 – China #International #Exhibition on #Public #Safety and Security will be held from November 3 to 6, 2026, at China International Exhibition Center (Shunyi Hall), #Beijing. As one of China’s most #influential #events dedicated to public safety, #intelligent security #technologies, and #digital #governance, the #expo brings together #government authorities, #industry #leaders, and #global #technology #providers. https://cnbusinessforum.com/event/security-china-2026-china-international-exhibition-on-public-safety-and-security/

https://www.europesays.com/conflicts/8621/ Air Force Reservist Leads Joint Team During NATO Pentathlon in Romania > U.S. Department of War > Story #AirForce #Army #DepartmentOfWar #DeputySecretary #DeputySecretaryOfWar #DOW #Government #JointChief #Military #MilitaryCompetition #NATO #Pentagon #ReserveComponents #Secretary #SecretaryOfWar #sports #USMilitaryNews #UnitedStates

https://www.europesays.com/africa/444353/ Are Ethiopians marching back to war in Tigray? #Ethiopia #fighting #Government #TheGovernment #tigray #TigrayPeople'sLiberationFront #TigrayanForces #TPLF #WesternTigray

https://www.europesays.com/conflicts/8565/ Are Ethiopians marching back to war in Tigray? #Ethiopia #Fighting #Government #News #TheGovernment #Tigray #TigrayPeople'sLiberationFront #TigrayanForces #TPLF #War #WarNews #WesternTigray

The Most Exclusive Restaurant in the World > U.S. Department of War > Story https://www.diningandcooking.com/us/31669/ #DepartmentOfWar #DeputySecretary #DeputySecretaryOfWar #DOW #government #JointChief #military #Navy #Pentagon #Secretary #SecretaryOfWar #SouthDakota #SouthDakotaDining #SouthDakotaFood #SouthDakotaGroceryStores #SouthDakotaRestaurants #submarine #UnitedStates

The Most Exclusive Restaurant in the World > U.S. Department of War > Story https://www.diningandcooking.com/us/31669/ #DepartmentOfWar #DeputySecretary #DeputySecretaryOfWar #DOW #government #JointChief #military #Navy #Pentagon #Secretary #SecretaryOfWar #SouthDakota #SouthDakotaDining #SouthDakotaFood #SouthDakotaGroceryStores #SouthDakotaRestaurants #submarine #UnitedStates

The Most Exclusive Restaurant in the World > U.S. Department of War > Story https://www.diningandcooking.com/us/31669/ #DepartmentOfWar #DeputySecretary #DeputySecretaryOfWar #DOW #government #JointChief #military #Navy #Pentagon #Secretary #SecretaryOfWar #SouthDakota #SouthDakotaDining #SouthDakotaFood #SouthDakotaGroceryStores #SouthDakotaRestaurants #submarine #UnitedStates

The Most Exclusive Restaurant in the World > U.S. Department of War > Story https://www.diningandcooking.com/us/31669/ #DepartmentOfWar #DeputySecretary #DeputySecretaryOfWar #DOW #government #JointChief #military #Navy #Pentagon #Secretary #SecretaryOfWar #SouthDakota #SouthDakotaDining #SouthDakotaFood #SouthDakotaGroceryStores #SouthDakotaRestaurants #submarine #UnitedStates

https://www.europesays.com/conflicts/8501/ Army Lasers Blast Drones, Proving Directed-Energy Weapons Are Future > U.S. Department of War > Defense Department News #ArmedWithScience #Army #BorderSupport #DepartmentOfWar #DeputySecretary #DeputySecretaryOfWar #DOW #DOWBorderSecurity #DOWDroneDominance #DOWModernization #DroneDominance #drones #GoldenDome #Government #innovation #JointChief #laser #Military #MilitaryWeapons #NorthernCommand #Pentagon #Secretary #SecretaryOfWar #Technology #UnitedStates