Suppose the question is 'high quality'.

It absolutely floods the field with false positives.

However, depending on the context, you can still see some issues a human is likely to miss. It's an idiot with crazy attention to detail.

So you can't count on it to catch things a human would and in many scenarios it will fire off more false positives than anything vaguely right, but it does represent a value in a more manageable haystack to catch certain issues.

But go back to a reasonable level of skepticism about suggested fixes, and ensure even if you agree with the LLM on an issue, that you understand it well enough to not get screwed over by the LLM suggestion...

Note that even if it is motivated by dubious ways, even if it is propaganda, that doesn't *necessarily* mean it is disinfirmation.

As they say, the best propaganda is true.

It's not sufficient, and it has a lot of false positives, but it can help get a smaller haystack with some of the needles that a human review can miss.

Code review for quality and security is something the LLMs can help without much inherent downside, so long as you do not trust the review but use it instead as an additional pass and make sure you follow up and understand any 'finding'. It *is* a risk if you take it as replacing the need for carefully considered human review, but with discipline you can have the best of both worlds: actually intelligent human review and the detailed coverage an LLM tends to get.

I've found that it's usually wrong, but a fair amount of time despite being wrong, understanding the area it was wrong about yields an a real problem it didn't catch. Then of course, upon occasion it is simply correct.

I would be extra wary of suggested remediations just like other generated code, but I would look into things it flags as somehow tricky.

But what about when the people *are* tools?

While you may be correct that the claim about consciousness being from beyond the observable universe is not falsifiable thus beyond the scope of scientific credibility, so too is any current understanding that would support an assertion that "Anthropics models work *just* like human consciousness".

We have pondered the question in a philosophical way, and can assert certain trends based on evidence, but in a comprehensive fundamental sense, the question has to date remained philosophical at the levels that the Anthropic paper needs. Note that the paper even explicitly acknowledges this facet of their work as philosophical, and the generated responses reflect upon that. It puts a lot of weight on the models being able to self-assess accurately and then using the result to show that they can self-assess a consciousness. Including one area where I noted that the "j-space ablated" output indicated that it was just token prediction, and seems like they use that to illustrate that it is conscious that it lacks consciousness...There's a lot of circular reasoning around the headline claim.

Feel like those are more akin to science philosophy than concrete scientific understanding. Which tends to happen in scientific pursuits when we dig deeper than what we can realistically actually explore in a strictly scientific way. Biggest problem is we struggle to recognize it as philosophy when it comes out of scientific thinking. We *want* to know more than we can piece together through scientific basis and for lack of a better option we go for the unfalsifiable and subjective opinion.

I recall a physicist explicitly addressing the phenomenon in his field. He refrained from participating in the musings because all he could do was figure out what math worked and hope someone comes up with better math, the wilder speculation upon unobservable implications was beyond the scope of what he felt science could do, but it's a very popular thing in physics.

I perused the paper, I think the comment was pretty spot on. Specifically the part where they have as a casual assumption that they know how human consciousness works.

It's basically shaking the magic 8 ball and it works, but it's totally a judgement call when it has worked 'enough'. It doesn't necessarily progress from easiest to hardest issues to find, it just is a bit random. Hard to say how many passes before you've *probably* got the real ones. My experience has not been that false positive rate increases, sometimes you might have all false positives but a continuation will flag real issues.

Indeed, they take needles in haystacks and make smaller haystacks that have some of the needles. They don't catch some things and they falsely flag a lot of things, but they do either directly or get close to something.

One issue in our codebase that was caught was pretty much spot on. Now upon it being highlighted, it was obvious to anyone that the inexperienced developer screwed up, problem was no one had the attention span to notice. The other issue close to real that it caught was actually not the flaw it indicated, however while looking at the jankiness that caused it to mis-indicate, a real issue was discovered. So it's indicated "fix" would have done nothing and left the codebase vulnerable, but the finding *was* useful in identifying an actual problem and fix.

However, it does open up a gigantic mess in customer engagement. Customers like to use every security tool at their disposal and hold the vendor accountable. Fine, except now we are inundated with false positives because the LLM indicates something, the customer has no way of knowing if the LLM is right or not, and we have to address false positives. We are now having to consider how to "fix" false positives by steering the LLMs away from indicating bogus things when analyzing our product. When it was the occasional misguided security researcher making a misunderstanding, no huge deal, we could discuss nuance and come to an accord. With the LLM mess, there's no skilled human to appeal to, a skeptical customer, and just a lot of volume...

Whatever you may think of capabilities, the narrative has always been about what they think sounds best in the moment.

So at one point, only the owner class matters, investment is King and every other concern can be bulldozed. The business opportunity down market would never be as much as what employee replacement would be.

Now they see a need to balance it because shockingly the people still have some leverage, whether it's banning construction projects or getting skittish with retail investment. Further the stories around employee replacement have gone permanently bad at prominent companies like Ford. So they need to move the goal posts to dodge the negative optics and recalibrate expectations.

I cannot agree more with this sentiment. It feels outright insulting to asked to read LLM output in a context where it is *supposed* to be human feedback. Tell me what you would have told the LLM to say, I can take it from there. I don't need you to LLM it up, because it will bury your point in a bunch of crap.

Could it provide useful info? Maybe, but I can do that myself if so. I want *your* thought on something, however incomplete it might be.

The dream is that the world is built for human limbs and the 'easiest' answer to claim the same versatility is to also have human limbs.

Stairs, cluttered terrain, a humble curb can all cause problems for the usually better answer of wheels.

The non-humanoid robots we already make those by the ton, and are, as one would predict, much more useful than human-like anatomy in their context. They however want to cover the underserved facet, banking hard on ML to make the humanoid design more viable while they traditionally are just infeasible to program.

Of course, that has proven a challenge, since the ML needs to instrument all the inputs and outputs of a human interaction system, and feeling is a huge part of human operation that cannot be instrumented. So they set people about trying to clumsily remote operate them in hopes of gaining training data, but it's low quality control and very low volume of data.

The disconnect is the "promise" is that LLM brings expertise down to the masses. If AI is "too hard for Ford to get right", that dramatically undermines the messaging that drives the current expectations and levels of investment.

This is very much evidence that companies can't be as bullish as they might inclined to be, because whatever you may think of Ford, the typical company is probably worse.

It's less funny when you have people literally and sincerely saying this...

I have to take some solace in the fact that these are the same idiots who talk a lot and do nothing that I usually ignore who do this all the time, but management is extra entranced with them over the AI cheerleading this go around...

Maybe for starting their own businesses, but I don't think just to compensate them for having to take a job that's "beneath them". Like you indicate, if another person makes $15/hour and didn't get a handout because they never made money, it would seem awfully unfair to give someone more money because they used to make more money. Easy to argue that the person coming down should by all logic have more financial resources already than the folks working the jobs. If their lifestyle based on higher income is infeasible with available jobs, think the reasonable sentiment would be "tough". Sell that car and settle for something more modest or take a bus, the same sort of compromises the lower paid coworker has *always* had to make. I suppose if anything, some financial relief for those trying to support a college kid where scaling the lifestyle down isn't as straightforward, but broadly speaking they don't have a particular right for government to help them more than they help peers that never made a lot in the first place.