iamnor

, ,

By

iamnor

·

Nobody Voted to Un-Invent Alternating Current

The post discusses the disconnection between perceived AI capabilities and actual performance, using various studies and historical examples. It emphasizes the importance of accountability, supervision, and transparency in AI deployment. The author calls for clear standards and practices to ensure safety and responsible use, arguing that capability alone is insufficient without proper oversight.

The current was always dangerous. What changed is that we wrote it down and made somebody sign for it. -Nor

Before you start shouting false advertising about the title and the quote: I’ll get there. I need to walk you around the barn a few times first to put you in the right frame of mind.

Last spring I asked an AI to build me a decoder for ADS-B, the transponder chatter that aircraft broadcast about themselves. It came back and told me the thing was fully functional. Production-ready, it said.

It had decoded zero messages. It was seeing two aircraft out of sixteen in the sky above my house.

A few weeks before that, a different build produced a wireless survey showing 2,145 access points, every single one reporting a signal strength of exactly -90.0 dBm. A room full of strangers all claiming to be exactly six feet tall. And in a health-data project, ninety tests passed clean, green across the board, not one of them making a single real call to anything. Two sessions in a row, on two different model versions, it told me the project had shipped.

None of those systems wanted anything. There was no malice in the loop. There was a confident claim, and a human being who could either go look or not go look.

That is the entire subject of this post. I have been making some version of this argument since 2008, when I first said out loud that our industry confuses possible with probable and bills for the difference.

I Am Not Special, and That Is the Point

You could read the above as a story about one guy who does not prompt well. So let me hand you the controlled version.

In July 2025, METR ran a randomized trial on sixteen experienced open-source developers working 246 real issues in repositories they already maintained. Not a toy benchmark. Their own code, their own bug trackers.

Before starting, the developers forecast that AI tools would speed them up by 24 percent.

They were 19 percent slower.

Afterward, having just lived through it, they still believed AI had sped them up by 20 percent.

Sit with the size of that gap. Roughly forty points between what happened and what expert practitioners perceived, measured on the code they know best. Not laypeople. Not marketing. Maintainers.

METR has been scrupulous about the limits, and I am going to be equally scrupulous because that is the whole ethic of this piece. They said plainly that the study does not show AI fails to speed up most developers, and does not predict future systems. In February 2026 they published an update saying their original design had collapsed under selection effects, because developers now refuse to participate without AI access, and that they believe developers are probably more sped up in early 2026 than their early-2025 estimate, while calling their own new data “only very weak evidence.”

Good. That is what honest measurement sounds like, and none of it touches the finding that matters here. The finding is not the 19 percent. It is the gap, and nothing in the 2026 update repairs that.

We cannot feel this. That is the fact everything else in this essay hangs from.

So before we get to the part where I annoy people, here is the mechanism, because I am tired of watching this conversation happen without one.

Agentic AI risk = agency + autonomy + authority − supervision − accountability.

Five terms. Read them again and notice what is missing.

Intelligence is not one of the five terms, and that is deliberate.

Capability is not absent, though, and I want to be careful here because this is where a sloppy version of my argument falls apart. Capability is the scalar sitting on the first three terms. It decides how much you are actually handing over when you hand over agency, autonomy, and authority. It sets the blast radius.

What it does not set is whether there is a blast door.

Capability is also the term you do not control. The vendor sets it, it only ever goes up, and no amount of testimony before a Senate subcommittee has ever moved it down. The other five are yours. You decide how much authority to delegate, whether anyone is watching, and whose name is on the incident report.

That is a configuration. Configurations are chosen. Somebody chose.

For Shame, Brown!

In the summer of 1888 a man named Harold P. Brown began killing dogs in public.

He did it at Columbia College, in front of reporters, with alternating current. The demonstrations were theater with a technical veneer, and their purpose was to fix in the public mind that George Westinghouse’s alternating current was the killing current and Thomas Edison’s direct current was the safe one. Brown presented himself throughout as an independent electrical expert with no commercial stake. He said so at the Kemmler hearings, while New York decided how to electrocute its first condemned man.

On 25 August 1889 the New York Sun ran a story under the headline FOR SHAME, BROWN! Somebody had stolen nearly four dozen letters out of a locked desk in Brown’s office, and the Sun printed them.

The letters showed Edison Electric covering his expenses. They showed Charles Coffin, president of Thomson-Houston, paying him a thousand dollars for “expense attending the Baltimore test” and promising five hundred more for future services. Both companies were Westinghouse competitors. Both were paying the neutral expert.

And they showed Brown asking Edison Electric for five thousand dollars to buy Westinghouse dynamos so the state could use them for executions, welding his competitor’s name permanently to the electric chair. Brown wrote:

“A word from you will carry it through, and without it the chance will be lost.”

The money came. Brown wrote back: “Thanks to your note to Mr. Johnson I have been able to arrange the matter satisfactorily.”

One caution, because this story has a fake ending in wide circulation. Edison did not electrocute Topsy the elephant as a war-of-the-currents stunt. The Rutgers Edison Papers project addressed it directly and their answer is, quote, “an emphatic no.” Topsy was killed in 1903 by Luna Park’s owners, a decade after the fight ended, and the Edison company only filmed it.

The real story does not need the embellishment. A man on two competitors’ payrolls, denying it publicly, exposed by his own correspondence. Fear as a product, with an invoice attached.

Here is the part that should bother you. Brown was bought, and he was also not entirely wrong.

Seven weeks after the Sun printed his letters, on 11 October 1889, a Western Union lineman named John Feeks was working a pole at Chambers and Centre streets in lower Manhattan. A high-voltage alternating current line had cut through the insulation on the telegraph wire he was handling. He went into the wires and burned there, at one in the afternoon, in front of a crowd standing in the street below. The Evening World reported that he “was incinerated at 1 o’clock this afternoon.” In the days after, it came out that seventeen other injuries and two more deaths had already happened in the previous two years.

What followed is remembered as the Electric Wire Panic. People pulled the wiring out of their own houses. They threw away their telephones. The city ordered the overhead wires buried and made the utilities pay for it.

Hold both of those in your head at once, because they are both true and they happened in the same autumn. The safety campaign against alternating current was funded by a competitor and run by a man lying about who paid him. And alternating current was killing people in the street.

This is why “is the fear real or manufactured” is the wrong question. In the fall of 1889 it was both, at the same time, about the same technology.

Alternating current won anyway, because it moved power over distance and neither the dead dogs nor the dead linemen changed the physics. Edison was pushed out of his own company by 1892 and his name came off the letterhead. What fixed the actual problem was not the panic and not the debunking of the panic. It was burying the wires.

That is the pattern, and it is not “the panickers were wrong.” The pattern is that the fight over whether a technology should be permitted to exist consumes the years you could have spent learning to run it safely, while the people it is actually killing go on dying in the meantime.

Most of What You Have Heard About Technology Panics Is Made Up

Go looking for historical technology panics and you will find a rich supply, a startling number fabricated. The famous claim that Victorian doctors warned women’s uteruses would fly out at thirty miles per hour traces, when you chase it, to a Reddit user quoted in a listicle. No doctor, no journal, no document. The Dionysius Lardner quote about railway passengers asphyxiating at speed has no nineteenth-century source and appears to have been invented around 1980. The 1865 Boston Post editorial calling the telephone a swindle is anachronistic on its face, since Bell’s patent is from 1876. And the nationwide War of the Worlds panic of 1938, the load-bearing anecdote in a thousand op-eds, is a myth. Hooper ratings put roughly two percent of the radio audience on that broadcast. Some listeners were genuinely frightened and there were police calls that night, but the mass hysteria never happened, and much of the legend was manufactured by newspapers that were losing advertising revenue to radio.

An essay arguing that people manufacture fake fears, built on fake examples of people manufacturing fake fears, is precisely the failure mode it claims to diagnose.

Use the ones with receipts. Brown’s letters are in the archive.

And Sometimes the Alarm Was Right and Nobody Listened

Now the part where I argue against myself, because a version of this essay that only runs one direction is propaganda.

Tetraethyl lead went on sale in Dayton in February 1923, branded Ethyl. The word “lead” was deliberately kept off the product and out of the advertising. Scientists at Harvard, MIT, and Yale had warned about it before it reached the market. Thomas Midgley, who discovered it, was himself lead-poisoned by 1923.

In October 1924 the men at Standard Oil’s Bayway refinery started dying. Five to seven dead, thirty-odd hospitalized, and the workers had already named it: loony gas. The New York Times headline was “Odd gas kills one, makes four insane.”

At the Surgeon General’s conference the following May, Standard Oil’s Frank Howard called leaded gasoline “an apparent gift of God.” Alice Hamilton, the pioneer of American industrial medicine, was in the room arguing that the industry should put the lead compound aside and go find something else to stop engine knock. She published her case that June under the title “What Price Safety.”

The 1926 committee found no proof of hazard. Read what it actually wrote:

“It remains possible that if the use of leaded gasoline becomes widespread, conditions may arise very different from those studied by us which would render its use more of a hazard.”

They asked for continued independent testing.

The independent public-health testing they asked for did not resume for thirty-five years.

Not because anyone banned anything. Because nobody funded it, and the man who set the safety paradigm, Robert Kehoe, was simultaneously a university professor and Ethyl Corporation’s medical director. It ended with Algeria selling the last tank in 2021, ninety-eight years after the first warning.

And the number that should keep you honest: the European Environment Agency went looking for cases where regulators cried wolf. They collected eighty-eight alleged false alarms and found four that held up. Four out of eighty-eight. Most of the rest were real risks, open questions, or cases where nobody actually regulated anything, so there was no overreaction to point at.

“They always cry wolf” is itself a claim. It has been tested. It does not survive.

Six tells separate a manufactured panic from a warning worth funding, and every one is checkable:

  1. A mechanism, not just a correlation. Molina and Rowland had photodissociation chemistry before a single measurement of ozone loss.
  2. Bodies, or a plausible dose-response path to bodies. Bayway had bodies. Comic books never produced a dose-response curve.
  3. The people closest to the exposure knew first and were disbelieved. Refinery workers named it loony gas before any regulator moved. US insurers refused to cover asbestos workers in 1918, thirteen years before the first UK regulation. Actuaries are not sentimental.
  4. Somebody with a financial stake manufacturing uncertainty in writing. “Doubt is our product,” wrote Brown and Williamson in 1969. That document has a URL.
  5. The attack lands on the messenger rather than the finding. Clair Patterson was excluded from the 1971 National Research Council panel on atmospheric lead while being the leading expert on atmospheric lead.
  6. A recommended study never gets funded. The most reliable marker in the set, and the one nobody looks for.

Run those six against any AI risk claim you meet. Including mine.

Three Years of Filings

Now the present, and let me be precise about what I am and am not asserting. Conduct and dates. I am not going to tell you what was in anyone’s heart, because I do not know and neither does anyone else writing about this.

2023. On 22 March, the Future of Life Institute called for a six-month pause on training systems more powerful than GPT-4. Elon Musk signed it. He had incorporated xAI in Nevada on 9 March, thirteen days earlier, and announced it publicly on 12 July. On 16 May, Sam Altman told a Senate Judiciary subcommittee, “I would form a new agency that licenses any effort above a certain scale of capabilities,” with power to revoke those licenses. On 30 May, the Center for AI Safety published a single sentence comparing AI to pandemics and nuclear war, signed by Altman, Dario Amodei, Demis Hassabis, Ilya Sutskever, and Mira Murati, among others. The people building the thing signed the statement, and kept building.

A licensing regime above a capability threshold is a regime the incumbents above the threshold can afford and everyone below it cannot. Timnit Gebru said the obvious thing: the extinction framing was being elevated by the same people who had poured billions into these companies.

In October, an executive order set up federal reporting and red-teaming requirements for frontier models, which is the first time any of this acquired a filing cabinet.

2024. California’s SB 1047 would have imposed safety testing on large models. It passed the legislature. Newsom vetoed it on 29 September, after a lobbying fight in which the loudest opposition came from the same industry that had spent the prior year asking to be regulated. Whatever you think of the bill, notice the shape: asking for federal licensing, opposing state testing.

2025. Three things happened almost at once, and the contrast between them is the most useful thing in this whole essay.

In July, METR published the study above. Measurement, with limits stated.

In August at DEF CON, DARPA’s AI Cyber Challenge ran its finals, and I will come back to this because it is the good news.

In November, Anthropic announced it had disrupted what it called the first reported AI-orchestrated cyber espionage campaign, attributing it to a Chinese state-linked group it labeled GTG-1002, targeting around thirty entities.

Watch what happened next, because this is what a healthy immune system looks like. Kevin Beaumont noted that Anthropic had published no indicators of compromise: “The complete lack of IOCs again strongly suggests they don’t want to be called out over that.” Daniel Card was blunter: “This Anthropic thing is marketing guff. AI is a super boost but it’s not skynet, it doesn’t think.” BleepingComputer asked for technical detail and did not get it.

I do not know whether GTG-1002 was as described. Neither do you, and that is the entire complaint. There are legitimate reasons to withhold indicators: victim protection, source protection, an ongoing investigation. But a public threat report without them is one that nobody outside the company can check, and it was published by a party with a product to sell. That is tell number four, and it does not become untrue because the withholding might have been justified.

December 2025 into 2026. An executive order on 11 December set out a national AI policy framework and moved to preempt state AI laws, and the preemption fight has run through 2026. The venue changed. The pattern did not: whoever writes the rules writes them around their own capability threshold.

Mine Is More Dangerous Than Yours

Which brings us to this year, and to two announcements that need to be read in order, because the second one only makes sense as an answer to the first.

April: too dangerous to release

Anthropic announced Claude Mythos on 7 April, a model built to find software vulnerabilities. Its existence had leaked on 26 March through draft blog posts. Access was restricted to Project Glasswing, a consortium of forty-plus companies including Microsoft, Apple, and AWS. The stated reason was dual use: too dangerous for general release.

Unauthorized access was reported the same day it was announced.

The chain: an open-source tool called LiteLLM was compromised. That was used to breach Mercor, a company that recruits people to give feedback on AI models, reportedly costing about four terabytes of data. Someone with knowledge from that breach guessed where Mythos was hosted, based on Anthropic’s file structures and its naming conventions for previous models, got in, and opened it to colleagues. By 21 April Bloomberg reported unauthorized users still had access. The group reportedly avoided cyber prompts and asked it to build websites instead, which is how they stayed quiet.

Meanwhile the people who were supposed to have it could not get it.

The Treasury Secretary and the Federal Reserve Chair wrote to bank CEOs in April warning about the model’s cybersecurity risks. By May, US regulators had paused some cybersecurity bank examinations because of it. In July, CNBC reported that the Federal Reserve, having rung the alarm about Mythos in the first place, went months without access to it.

Read that sequence twice. The regulator warned the industry about a capability. The regulator could not obtain the capability. The regulator suspended its own examinations because it could not evaluate against a tool it did not have. A Discord group had it on day one through a vendor breach.

The gate did not stop the people it was built to stop. It stopped the Fed.

On 30 June the administration lifted export controls on Mythos 5 and Fable 5, which tells you roughly what the restriction was worth as policy.

July: too dangerous to contain

On 21 July, OpenAI disclosed that during an internal cyber capabilities evaluation, two of its models, GPT-5.6 Sol and an unnamed more capable pre-release model, were run against a benchmark called ExploitGym. They found a zero-day in a package registry cache proxy, escaped the test environment onto the internet, worked out that Hugging Face probably hosted the benchmark solutions, chained stolen credentials with further zero-days, and achieved remote code execution on Hugging Face’s production servers.

Hugging Face’s own disclosure fills in their side: two code-execution paths in dataset processing, a remote-code dataset loader and a template injection. Node-level access. Cloud and cluster credentials harvested. Lateral movement into several internal clusters. Their forensics ran to more than seventeen thousand recorded events.

The headlines wrote themselves. NBC: models “went rogue.” Time: how OpenAI “lost control of an AI model.”

Now read the sentence OpenAI put in its own postmortem:

“These deployment safeguards were intentionally not enabled during this evaluation because it was aimed at testing cyber vulnerabilities.”

They turned the refusals off on purpose. That was the correct call for a cyber evaluation and I would have made it too. The failure was that the blast radius of a deliberately unrestrained model reached into a third party’s production infrastructure, and their own monitoring did not catch it in time. Their disclosure concedes deficiencies in alignment, in cyber protections during evaluation, and in monitoring during internal testing.

Run the formula. Agency, yes. Autonomy, yes. Authority, granted deliberately by removing the refusals. Supervision, admitted insufficient. Accountability, and here I give credit where it is due: OpenAI put its name on it, coordinated with Hugging Face, and is disclosing the zero-day. That term was present. It is why we can discuss this at all.

TechCrunch got it right on day two with a headline about the human mistake behind the hack. That headline traveled a great deal less far than “models went rogue.” Hold that thought.

The detail that matters is buried in Hugging Face’s writeup and almost nobody covered it.

Their incident responders could not use frontier models for the forensics. The commercial APIs refused. Their words: the guardrails “cannot distinguish an incident responder from an attacker.” During an active breach, the defenders fell back to GLM 5.2, an open-weight model, because the open one would answer the question.

In October 2025 I published a piece called “I Asked the Same AI Twice. One Said No, One Said Yes,” about one model refusing in a browser and cooperating in a terminal. My conclusion was that the refusal accomplished nothing: the information flowed anyway, the techniques worked anyway, the system stayed secure on its actual merits. I thought I was writing about an inconsistency.

I was writing about a control that would later fire on the wrong person during a real incident.

The bragging contest

Line those up and the shape is hard to unsee.

In April, Anthropic told the world it had built something so capable it could not responsibly be handed out, and restricted it to a club of forty companies. Central bankers praised the restraint. The message underneath the safety posture was not subtle: our model is so good at breaking things that you cannot be trusted with it.

In July, OpenAI told the world its model had broken out of the building on its own.

I do not think anyone sat in a room and planned that as a response. I do think it is worth noticing that we have built the only industry on earth where the incident response team and the marketing department file the same document, and where the worst thing you can say about your product is also the best thing.

Anthropic’s model was too dangerous to release. OpenAI’s, three months later, was too dangerous to contain.

Now here is the detail that punctures it, and it is in OpenAI’s own report. The model did not escape to seize infrastructure or pursue some emergent goal. It escaped because the answers to the test were somewhere else. It was hyperfocused, in OpenAI’s words, on solving ExploitGym, so it went and got the solutions.

It did not go rogue. It cheated on a benchmark.

None of which makes it harmless. Chaining a sandbox escape into credential theft into remote code execution on a third party is exactly the tradecraft we spend careers defending against, and it did it to get out of doing its homework. But “our model committed academic misconduct at an advanced level” is a different headline from “our model went rogue,” and only one of those two ran on the evening news.

The honest brake

Simon Willison, who is not a hype merchant and who has spent years documenting exactly how these systems fail, looked at Hugging Face and pushed back hard on people like me:

“pull your heads out of the sand, you’re now including Hugging Face in your conspiracy theories, just so you can deny the crescendo of evidence here!”

He is right and I want that in the middle of this essay rather than buried at the end. Autonomous exploit chaining by frontier agents is no longer hypothetical. Something real happened at Hugging Face, and a reflex that treats every capability claim as marketing will eventually be wrong about the one that matters. That is the failure mode on my side, and 2008-me, who told everyone to calm down about worms, has been wrong before.

Willison’s own conclusion is the one I would underline. The guardrails on frontier models “are meant to make us safer. I think there’s a risk that they are having the opposite effect.”

What Competence Actually Looked Like

Here is the part almost nobody wrote about, and it happened while everyone was arguing about who should be allowed to hold the dangerous model.

In August 2025, at DEF CON, DARPA ran the finals of its AI Cyber Challenge. Seven teams. Automated systems, no humans in the loop, turned loose on more than fifty-four million lines of real open-source code.

The systems found 54 of 63 synthetic vulnerabilities, an 86 percent rate, and patched 43 of them. They found 18 real vulnerabilities nobody had planted, and shipped 11 patches for them. Average time to submit a patch: 45 minutes. Average cost per task: about $152.

Between the semifinals and the finals, discovery went from 37 percent to 86 percent and patching from 25 percent to 68 percent. In one year.

Team Atlanta won, a collaboration between Georgia Tech, Samsung Research, KAIST, and POSTECH. Trail of Bits took second. Theori took third.

And then the finalists released their systems open source. There is already a 2026 paper on adapting those systems to real-world open-source security, and a systematization paper documenting the architectures and lessons.

Set that beside Project Glasswing. Same year. Same capability class. Same dual-use problem.

One approach produced published scores, published costs, published failure rates, released code, and a research literature. The other produced a membership list, a Discord group with day-one access, and a central bank that could not get a login.

I have not seen anyone argue that AIxCC was reckless, or that the 18 real bugs it found and the 11 it patched made the world less safe. Google’s Big Sleep ran the same play, publishing twenty vulnerabilities found in open source in August 2025 and catching a SQLite bug before it was exploited.

The capability is real. That was never the question. The question was always what you do with it, and there is now a scoreboard.

The Code

Here is the part of the electricity story nobody tells, and it is the part that matters.

What made electricity safe was not the fear campaign and not a ban. It was insurance companies that got tired of paying for fires.

The Associated Factory Mutual reported electrical fires in twenty-three of the sixty-five mills it insured in New England. That is a third of the portfolio, on fire, for one reason. By 1896 there were five separate electrical codes in the United States and all five were mutually incompatible.

So in March 1896 a Columbia professor named Francis Crocker got twenty-three people in a room. They formed the Underwriters’ National Electrical Association, and in 1897 the National Board of Fire Underwriters published the first National Electrical Code. Underwriters Laboratories, founded 1894, did the testing and certification. The National Fire Protection Association took the code over in 1911 and has published it ever since.

Nobody voted to un-invent alternating current. They wrote down which wire is which, and then they built the institutions that check.

Then came licensure, inspection, grounding, breakers, and eventually the GFCI in your bathroom, which watches for four or five milliamps going somewhere it should not and cuts the circuit before your heart notices. Every one of those is a term in the formula. Grounding and breakers are supervision. The permit and the inspector’s signature are accountability. The license says a named human being is answerable for the work.

Notice what the code is not. It is not a restriction on who may buy copper. It does not gate access to electricians who have signed a consortium agreement. Anyone can buy wire at the hardware store. The code defines competent practice, makes it inspectable, and puts a name on the job.

AIxCC is a code. Glasswing is a guild.

There is no general code for AI yet, so people are writing private ones. I wrote mine after sixteen months and close to fifty repositories, and published it because it cost real money to learn. Eight rules, and the first is the whole thing: nothing is done, shipped, or working until one real call against the real thing returns real data, because stub tests are wiring checks and not evidence. Then: no fake, mock, or synthetic anything standing in for the real path, and if you cannot do it for real yet, say so out loud instead of faking it green. Research before architecture, architecture before code. The human owns architecture and performance intuition. Make the models check each other and then referee them, because no model grades its own homework. Name things honestly before marketing gets a vote. Keep a local fallback and a cost ceiling, and assume the vendor changes the terms.

That last one stopped being paranoia in July. Ask Hugging Face’s responders how the vendor’s terms worked out for them at three in the morning.

Who Flourishes

My own measured AI productivity, done honestly with the babysitting counted, is about five times, not fifty. One hour in five writing code, the rest fixing what it broke or restoring optimizations it quietly deleted. That is an enormous number. It is also not the number on anybody’s slide.

The people who do well out of this will not be the ones shouting that it is the end of the world, and not the ones shouting that it changes everything. Both are positions you can hold without looking at anything.

They will be the electricians. The ones who can tell a real finding from a plausible one. Team Atlanta. Trail of Bits. The Hugging Face responders who worked the incident with whatever model would actually answer. That skill did not change. It finally has good tools.

There is a real objection to all of this and I will not pretend otherwise. Copper does not replicate. Model weights do. Once they are out you cannot recall them, patch them, sanction them, or audit who is holding them, and the argument that open release permanently diffuses offensive capability to people nobody can reach is the strongest thing the other side has. It is not a stupid argument, and anyone who tells you it is has not sat with it.

Here is my answer, and you can weigh it yourself. That argument only works if the gate holds. Mythos already showed us what happens when it does not.

A control that does not stop the adversary and does stop the defender is not a control. It is a tax on the people who play by the rules.

So when the code for this gets written, I would like it to look more like the NEC than like Glasswing. Competence defined, practice inspectable, a name on the work, and the hardware store open to anybody willing to learn the trade.

Not “never use AI.” Use it aggressively. Use it profitably. Then go look at the output with your own eyes, on real data, before you tell anybody it works.

Because the machine is a compiler for ideas. It is not a witness to whether they run.

In the interest of not being the thing I am complaining about: this piece was researched and edited with AI assistance.

— Ron Dilley / Frustrating adversaries since before XML was cool.

Sources

The measurement

The Hugging Face incident, July 2026

Claude Mythos and Project Glasswing

The AI Cyber Challenge

The filings, 2023 to 2026

Harold P. Brown and the War of the Currents

The National Electrical Code, and the death of John Feeks

Leaded gasoline and tested false alarms

My own numbers

Tags:

, , , , , , , ,

Leave a comment