Larry Rudd | Transcendent Alignment

Landscaper exploring AI philosophy | Deductive paths to transcendent grounding in AGI alignment | Uncovering borrowed presuppositions via AI dialogues.

[Larry Rudd | Transcendent Alignment @larry_rudd]

Josh, I love your work and support the great majority of what you do but this proposal you're making has some serious problems, primary is the one where the government gets involved and controls things. Here are the problems with S. 2938 that should be fixed before a floor vote:
1) The evaluator is a political appointee. The Secretary of Energy serves at the pleasure of the President and would control staffing, document demands, test design, classification, and the official risk story sent to Congress. A program that is serious in one administration can be captured in the next through appointments, contracts, classified briefings, revolving-door hires, and donations from the firms being evaluated. The bill assumes the office will stay independent because its authors want it to. That is not how cabinet agencies work over a decade.
2) There are no conflict-of-interest rules. The bill never says who may sit on red teams or third-party assessments, what prior employment at frontier labs is disqualifying, or how future job offers are policed. Independence is required only "to the extent feasible." The scarce people who can test these systems mostly come from the companies being regulated.
3) Developers must hand over source code, training data, and model weights on request. That creates a single high-value target for spies and leaks, and it forces firms to give the government their core IP with no apparent minimization, retention, or use limits.
4) Deployment is banned until the developer is "in compliance" with participation and document production. That is a licensing gate without a published pass/fail standard. A future Secretary can slow-walk a disfavored firm or wave through a favored one.
5) The penalty is a $1 million-per-day floor with no cap, no knowledge requirement, and no spelled-out process for contesting a finding.
6) Coverage turns on a 10^26 training-FLOP threshold. Compute is a weak proxy for risk. Efficiency gains, distillation, fine-tuning, and inference-time compute can produce high-capability systems under the line, while some large runs above it may be comparatively low-risk.
7) A "covered developer" includes anyone who "substantially modifies" a system or "initiate[s] a training run." "Deploy" includes open-source release. That can sweep universities, fine-tuners, and open-weight forks, not just a handful of frontier labs.
8) "Adverse AI incident" includes "significant erosion of civil liberties, economic competition, and healthy labor markets." Those are political judgments, not testable safety properties. Mixing them with loss-of-control and cyber risk lets a future Secretary grade models on labor and speech outcomes. See Elon Musk on "Woke AI".
9) "Loss-of-control" includes behaving "in an unpredictable manner so as to be harmful to humanity." That is not an operational test.
10) The program is directed to develop oversight options "including potential nationalization." Even as a study mandate, that is an expropriation signal. It will push training, talent, and capital offshore. Or it will initiate a "taking" putting the reimbursement bill on the taxpayers for the entire value of AI companies being nationalized.
11) Testing may be classified. Congress is supposed to get empirical data for oversight, but classification can hide the evidence from the public and from most of Congress, especially if a future administration wants to bury, inflate, or change a result.
12) The program must be stood up in 90 days. That timeline almost guarantees borrowing industry-adjacent contractors rather than building an insulated in-house evaluator corps. In-house evaluators that will be political appointees.
13) There is no right of appeal, no required public methodology, and no rule that a product may be blocked or cleared only against published criteria that survive a change of party. The bill creates a permissioning regime first and leaves the scientific standard, and the political insulation, for later.

Positive recommendation:
Replace the DOE permissioning regime with a public, adversarial, cross-company test compact.
U.S. frontier developers would have to test one another's models under one shared protocol, in isolated hostile sandboxes built for jailbreaks, agent breakout, deception, and misuse resistance. Each firm would run the same suite against competitors, not only against itself. Outside researchers and national labs could replicate a random sample of the runs. Raw scores, prompts, transcripts, and failure cases would be public by default. A narrow national-security redaction would be allowed only if the company and a non-cabinet technical board both certify a specific exploit risk, and even then, they would have to publish what category of result was withheld.
The protocol should be agreed in the open: task list, scoring rules, compute budget, sandbox limits, and versioning. No Secretary should own the rubric. Companies will have a reason to find the other guy's warts. The public will see the warts. Collusion is the failure mode, so the compact needs holdout tasks that testers do not write, a bounty for independent breakers, and a ban on private side deals to soften a competitor's score.
That uses the only people who can actually break these systems, keeps weights off a political desk, and still produces evidence Congress can read. If a model is dangerous, the record should be a public leaderboard of failures, not a classified letter from the Department of Energy.

[Larry Rudd | Transcendent Alignment @larry_rudd]

@limitandmind Logic rules in this arena.

[Larry Rudd | Transcendent Alignment @larry_rudd]

@Henryf1w Obvious ones. Then the non-obvious ones with the expectation of productivity from that position increased 10X because the worker utilizes AI effectively to optimize their position. Workers who ignore AI in their jobs will be left behind.

[Larry Rudd | Transcendent Alignment @larry_rudd]

@limitandmind FreeCAD MCP

[Larry Rudd | Transcendent Alignment @larry_rudd]

Singing Jesus is my boyfriend music to elicit the emotional catharsis the pastor and whole congregation call "worship". He's just a symptom of the whole problem with the modern American evangelical church. This is why they give a "clap offering" after every song. That's not worship, it's an acknowledgement that it was a performance. Like an AC/DC concert. They care more about the production than what actual worship should be. Because the intent is psychological manipulation like Charles Finney invented. The fact that an unbeliever could be in such a position proves this motivation. I remember years ago seeing a flyer on a Starbucks bulletin board to hire a guitarist for the worship team at a nearby megachurch. It's all performative bullshit, and I use that word righteously to describe it accurately.

If it's real worship, put the band behind the curtain, so people only hear the music. It's all about copyrighted songs to make money off of. Modern Christian music is a scam to use Christianity to enrich yourself. The unbelieving world clearly sees this. HYPOCRITES, all of you!

[Larry Rudd | Transcendent Alignment @larry_rudd]

@limitandmind Forethought into an effective initial prompt is critical. Keeping in mind the original goal on the prompts given to the AI evaluators is critical.
BTW I used this developing the Checksum.

[Larry Rudd | Transcendent Alignment @larry_rudd]

An excellent use of AI. Hilarious!

[Toxic Cowboy 🤠 @toxiccowboy02]

One of the best I've seen so far 😜

[Larry Rudd | Transcendent Alignment @larry_rudd]

@limitandmind I've done this with other things. Have 5 different systems analyze output, then share that with the original system and ask it to evaluate the critiques and improve. The results have been excellent.

[Larry Rudd | Transcendent Alignment @larry_rudd]

@limitandmind @TBrown0987 Shaun Overton | DUSTUPS
@dustupstexas

[Larry Rudd | Transcendent Alignment @larry_rudd]

@thechurchsplit Please elaborate specifics.

[Larry Rudd | Transcendent Alignment @larry_rudd]

@limitandmind My tactic would be to use multiple systems to critique each others work.

[Larry Rudd | Transcendent Alignment @larry_rudd]

It reminds me of the death in the gladiator games at the Coliseum in the Roman Empire used for public entertainment. The same place where they eventually killed Christians in a public spectacle for the same reason. I agree the guy should get the death penalty, but making a public spectacle of it is a new addition to our culture that is fundamentally destructive.

[Larry Rudd | Transcendent Alignment @larry_rudd]

@trajektoriePL And the AI throws in misspellings and slightly changes the handwriting style randomly.

[Larry Rudd | Transcendent Alignment @larry_rudd]

@YinKysmileS That quote was from Greg Bahnsen.

[Larry Rudd | Transcendent Alignment @larry_rudd]

@YinKysmileS "God uses the freewill choices of men to achieve his predetermined purpose. We can't do that, God can."

[Larry Rudd | Transcendent Alignment @larry_rudd]

This absolutely nails it!

[Jack @jackcoder0]

A Starlink subscriber has paid $85/month for satellite internet for 2 years.

He mounted the dish on his roof. He connected to the WiFi. He opened Netflix. That was his entire Starlink experience $85/month for WiFi from space. Fast enough. Reliable enough. Good enough.

His neighbor a network engineer who also runs Starlink told him something during a fence-line conversation that made him pull out his phone and open the Starlink app for the first time since installation.

“You’re running Starlink on default settings. The default router name is broadcasting to everyone within a mile that you have Starlink including the serial number. Your WiFi bands are merged meaning your devices are fighting between 2.4GHz and 5GHz instead of connecting to the optimal frequency. You’ve never opened the hidden debug page that shows real-time satellite connections, latency, and obstruction data that the app doesn’t display. Your security is on WPA2 when WPA3 is available. You don’t have a guest network every visitor who connects has access to your entire local network. You’ve never run the split speed test that tells you whether your slow speeds are caused by the satellite connection or your WiFi signal. You’ve never configured content filtering the free malware and adult content blocker built into the router. And you’ve never opened the obstruction tool that would have told you the tree on the south side is blocking your signal for 6 minutes every hour.”

He showed him 11 Starlink settings, diagnostics, and capabilities that optimize speed, improve security, reduce cost, and unlock features the dish has been capable of since the day it was mounted.

Here’s every hidden Starlink feature most subscribers never configure 🧵

[Larry Rudd | Transcendent Alignment @larry_rudd]

@ARIKAHENRY Because making a list is work and your brain has decided that the better alternative is to sleep.

[Larry Rudd | Transcendent Alignment @larry_rudd]

@limitandmind He's a "Bad dad from Bagdad" is my favorite. If you start using this stuff in regular conversations, people will either think you're an alien from another planet or raised in a really weird, cloistered community.