Well you see, it's Schrödinger's LLM:

The LLM is simultaneously a failure at running a simple vending machine, while simultaneously being a world ending super intelligence already plotting the annihilation of humanity.

Let's stop calling this AI. There no intelligence, no will, and no curiosity. These are all Barnum statement generators that have been programmed to use flattering and reflective phrase to fool the marks.

Simpler than that:

"YOU GOTTA BUY IT MAN! I have got to pay back Marc Andreessen and the other VC's or I'll never get another VC investment again! Maybe if I'm real lucky I'll be so rich I'll never have to go begging for money again like you poors!"

Here's a hint: This isn't a conspiracy to "bake in a lead" or "stop others" or other nonsense.

The greatest conspiracy is that there is no real conspiracy, just people acting out of greed and/or self interest. Dario and Sam doom trolling lets people get money elsewhere in the chain, which is why you see them doom trolling as well. Dario and Sam doom trolling lets people who still have shares in OpenAI / Anthropic to protect the value of those shares in the run up to IPO

Let's all take a collective breath, and look at what these things can actually do. Let's support the media that stops and actually investigates the capabilities, financials, and ROI of the models.

Once we do that LLM's are likely going to go down in financial history as VC dumpster fires.

Preferably, but even in areas where police fight tooth and nail to keep them we can fix the problem. Granted the fix will probably result in them being removed.

The fix is a couple parts:
1) Requiring a warrant issued by a state judge to access the cameras at all, with the requirement that the officer personally testify to facts. No fobbing it on random informant, they are required to own it.
2) Accessing the cameras without a warrant is a felony worth a minimum of 5 years in prison, but up to 10 years.
3) Regular transparency conducted by an independent group, not associated with the police.
4) Since we're dealing with cameras make it a felony worth 10 years in prison to hamper, mute, disable, modify, or otherwise obstruct, the operation of a body / dash camera during an interaction or when discussing an interaction. No blocking the camera, mic, or secret conversations with your coworkers.

Of course the ideal solution will be to keep pushing for these things to be shoved off into oblivion.

As a side note, if they try to claim that a warrant is too high a burden because "it needs to happen fast" or whatever false emergency situation they dream up, then ask them again if the cameras are active surveillance. They have to say no. Then you can remind them that it's not hot pursuit or anything similar since they're looking at historical, not real-time, data so they should be able to take a few minutes to get a judge to issue a warrant over the phone.

Wait, that's not what we're being sold. We're being told it's just about, or possibly is, AGI that can do ANYTHING and could decide to KILL HUMANITY!!!!

/snicker

I think Benn Jordan, and others, have proven that the ALPR companies are pretty much incompetents at information security.

Can you even trust locally deployed models? Maybe if they're completely isolated from the outside, but remember that OpenAI and Anthropic like to claim their models are SUPER HACKERS ABLE TO BREAK ANY SANDBOX now. So you'd have to install them on an air-gapped compute cluster and only allow direct access. That's an awful expensive system with limited utility, but the alternative is the board finds out you lost trade secrets due to incompetence.

Remember to remind everyone that, according to OpenAI and Anthropic, the only safe solution is to run all models in local air-gapped systems entirely cut off from the Internet. (lol)

Most people I know haven't had the degradation issue, although I have seen it in some people. I find a bigger problem is people believing bad data/ information provided by LLM's over Subject Matter Experts (SME) who are providing information from the field.

They also react very negatively to you when you show them that they were given bad information.

When you watch the report they cozy up to the 911 operators and you hear the 911 operators schmooze right back. We'll skip the dubious (at best) claims made on that call, but the fact 911 is joking and laughing with them is all you need to know.

You want to fight Flock at meetings? Ask the police chief, if they're on-site, why they don't have officers taking pictures of license plates and uploading the pictures to a database. Get them to say that police photographing everyone isn't surveillance. That won't go over well at all with the public if they spit that out.

Another good thing to do might be to push for at the state level is the requirement of a warrant, issued by the judge AFTER the police "provide affirmation under oath that they personally attest..." to get one. You'll hear the "but we might have exigent circumstances" to whit you can respond: "This isn't real-time surveillance, or so you say, so you're looking back in time meaning the five minutes it takes to get a judge to issue a warrant won't make a difference." Then also push the state to make it a minimum 5 years but up to 10 years long felony to "misuse" the system AT ALL. You'll see them tap dancing fast when faced with the risk of 1) having to personally attest under oath because that puts them in legal jeopardy if they lie and, 2) A decade in prison likely wouldn't be issued as judges are usually pro-cop but the minimum 5 years would freak cops out. They're used to walking away after resigning or getting probation.

That could happen now because people on both sides of the aisle are pushing back. Once it's in place repealing it becomes incredibly difficult, because who are you going to cite as a reason to end it without blowing up your own reasoning? You're going to admit it's 24/7 surveillance? You're going to say that cop who stalked their ex should walk? Politically infeasible.

They want the "forced slowdown" to cover for the fact that they are making far less improvements per model release. They also need to change the story because the "AI JOBSPOCALYPSE!!!" didn't materialize. Their grand pronouncements of "Claude doing all programming" inside of a year didn't happen. Heck from the perspective of useful things LLM's have been shown to be a failure overall. The biggest threat to them is that people are beginning to understand this isn't actually AI.

This is what seems likely, even though it won't be reported anywhere near as loudly as the "AI SOLVES PROBLEM HUMANS COULDN'T!"

The problem is that sensationalist fear, Fear, FEAR sells (gets clicks) more than the boring truth. Heck I'll go as far as to say many outlets are flat out spreading disinformation about AI. When I say disinformation they are basically reading AI company press releases and going "OMGWTFBBQ! END OF HUMANITY!!!" when what actually happened isn't that impressive.

That's all that needs to be said.

I mean Claude should be programming everything and AGI should be a superintelligence.

Anything they say should be met with a derisive John Candy meme: "Oh sure, sure. Whatever you say..."

Actually it's the dumbest strategy but not for the reasons a lot of people will think.

First and foremost - you don't want kids dependent on something technological to develop simple answers. We still teach kids basic language skills along with addition and subtraction in schools. The best schools still teach kids long division, square roots, and more, on top of creative writing, a foreign language proficiency, and basic art / music. Why? Because as completely frustrating as it can be, it builds problem solving into the brain.

On top of that LLM's hand out false information on a regular basis. A calculator is a functional tool as it will tell you 1+1+1 is 3 without it being hardcoded, unlike LLM's that had to it hardcoded there are 3 r's in strawberry.

The anthropomorphizing of LLM's is a sales technique, like the programmed incessant glazing of users telling them that is a great idea. I think LLM's could be good software helping with searches. However that is of limited utility, and would require utilizing only verified data. Even that is fraught with issues, see the need to hard code the number letter r's in strawberry, but it could have a purpose.

Unfortunately for OpenAI, Anthropic, Google, Amazon, Oracle, et al., that functionality is nowhere near the near trillion they've dumped into development and datacenters.

You would think they would at least put a HUD (similar to the Telluride, et al.) if they were going to pull those shenanigans.

I'll talk to my friends in Korea to see if it is overall accurate, but if it is accurate most of them will buy a different brand. Well, if they own a car anyway.