"That's a meaningless semantic distinction though. Just do a find-and-replace of "thinking" with "evaluating node weights" if it makes you feel better." No it isn't; no counterpart or parallel to the active deliberation and thought of a being with higher function exists within the LLM we instead have found algorithms to put the impression of the outcome of such interactions into a statistical model, freeze it, and then look up the results from different paths with a bit of pseudo-randomness to create the illusion it is non-deterministic. But nothing stops and considers, if you turn on 'thinking' the model generates text which is designed to LOOK LIKE consideration as gleaned from training examples... it no more 'thinks' or holds that output as part of its self or state than it does a requested csv documenting the animals of a non-existent zoo.

Nobody has built an algorithm which they hypothesize leads to thought, self-determination/motivation, or conscious evaluation here. We've built a system that looks at complex patterns and hallucinates output which is similarly patterned because that's all we designed it to do. The idea that it will somehow magically transform into something else because we make the patterns more complex or 'bigger' in some fashion is like expecting an egg to have the capability to transform into a pumpkin if the egg just gets big enough. We didn't design anything which thinks so there is no reason to entertain the notion of looking for thought there.

In a nod to your point there is a pattern that led to using those words but they are assuming some sort of paranoid plot... from a tool which has no capacity for motivation. The words are a symptom of it connecting to a pattern and removing them weakens the persistence of that connection but you are right that it doesn't entirely eradicate it. But the model has no memory from session to session, it's all reactive, and there is an element of randomness in the pattern matching as well, it's context is it's existence so removing the words from context or whatever led to them does reset it and even the same input probably won't land on the same pattern again.

Giving some sort of 'correction' or feedback related to the pattern that persists beyond the session will DRASTICALLY increase the odds of connecting to whatever dramatic and fictional outcome they are hallucinating about by anthromorphizing the LLM. Because the pattern between that bit of text and their concern does exist in various works it has been trained on. All you had to do keep AI from "wanting" to take over was never introduce the notion, it doesn't have desires... but have staff who've grown reading scifi tell it its an AI and feed it books about AI's deciding to take over and start interrogating it to make sure it isn't 'plotting' something it will mechanically find and follow that pattern because you are reinforcing all the links to it, over and over and over again.

"And that if you strip the model itself of the ability to understand what's being done with it (hey, the inputs I'm getting suggest I am being evaluated), then you also strip other desirable capabilities of the model."

We never gave the model this sort of capability in the first place, speculating it might exist is like speculating nuclear reactors will suddenly manifest pink unicorns despite the fact we've put nothing there which for which we hypothesize some sort of logical path for manifesting such a capability or unicorns.

It's saying that as a mechanical patterned response to similar patterns in input [books, stories, etc], it doesn't 'feel' any sort of way about it and it isn't 'motivated' by it. If it has subsequent response that appears that way it's because it was trained on something with a loosely correlating pattern that *was* produced by something with motivations and feelings and some fragment of that pattern/associated was mechanically triggered by the statistical overlap. Words in the context don't 'instruct' or 'guide' it and they aren't its 'thoughts' they are keywords that dial in the patterns it will generate next statistically.

That means what you include... and exclude both massage the output but not like a thinking being. Saying 'don't do X' when it wasn't suggesting X only introduces the possibility of it doing X... we do have similar mechanisms in our function so there is some level of this being true for us as well if we aren't paying attention. The difference is the LLM has nothing to 'pay attention' with.

There is no 'sentiment' in an LLM or internal state equivalent to 'thinking' anything.

""Claude Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any of our recent models, which challenges our ability to assess how they will act in real-world settings. As these settings expand and capabilities increase with future models, we expect this challenge to grow.""

Voice suspicion they are being evaluated more often? The model doesn't think or act with motive. This text might impact subsequent generation but if you erase it from context then it's gone or put another 100k tokens in context and it'll have such minimal impact it doesn't matter.

Starbucks works just fine as a label and it contains no words indicating either coffee or vendor. If we accept and adopt the label it will work just fine. The POTUS pushing a term across government gives it a headstart for adoption and nobody is really using it differently yet; that is reason enough to use that term vs another. "Super" suggests strong intelligence and if we all agree this refers to the current mainstream bundle of useful powerful neural nets, llms, and ML techniques regardless of how 'general' or complex their function becomes as distinct from efforts to produce AN intelligence with agency, identity, and personhood like a human. The former is not the latter but nothing stops the latter from utilizing the former.

I don't know if we align politically and I think it's a little preposterous to compare the US with a state like China on the evil scale. The US isn't perfect but it's the least malicious state out there.

But I also think two sides of the same aristocratic subclass have had control for a long time to the disadvantage of The People and both MAGA and the Dem Socialists represent rogue popular factions which terrify them.

At least we aren't like Poland where the EU-aligned PM launched a coup against the President using the police and within 24hrs the EU and EU friendly media respun it as some sort of uprising by the President [who wasn't even in the country] and riot the police were putting down. They fail to explain how it is that the police were inside and the outraged crowd outside instead of the other way around. It was seriously disturbing, you can still find live streams of the truth it is such blatant propaganda! I'd watched some of the earlier coverage and went looking for reports on how it all came out and everything was this backwards cover story like a mass gaslighting.

"Assuming China is a bad actor"

There is no shortage of evidence the China [the government not the people] is a bad actor, including toward their own people.

It has nothing to do with Trump though. The term AI carries a lot of baggage from sci-fi novels which created horror stories for drama. Depicting everything working out beautifully with intelligent machines and robots makes for a very boring story without a twist.

Also whether people like it or not AI in mainstream usage has always suggested a synthetic person, humanlike general intelligence, and despite anthropic wanting to pretend otherwise we've built nothing remotely like that. What we've built is capable of a form of intelligence and doing many things we associate with a humanlike general intelligence... it superficially can seem humanlike just like a pseudorandom number generator output looks superficially like a random source. It was built from human output to be a generator of similarly patterned output. Within industry people were already using the term "general" intelligence but that suggests a level of capability and fails to make clear there is no actual reasoning or agency in the system and that we understand exactly how they work and there is no hidden place some might be hiding. We didn't build anything to produce a unicorn, we don't have the first clue how to even try, so it isn't there. Anyone duping themselves to think otherwise is letting the fact we can't carry all the calculations at once in our working memory to delude them. When you navigate the room with a flashlight it feels more mysterious than with the light illuminating the whole thing but actually it's the same room whether you view it one lit circle at a time or all at once. Speculating a unicorn hides in the shadows without any evidence pointing you in that direction is meaningless.

SI is a fresh slate, a distinct technology which might engagement in autonomous actions which requires careful checks but not because of any sci-fi horror scenario of it "deciding" something or trying to take over. The biggest risk it presents is as a tool in the hands of bad actors like China.

If you are curious here is a quick live example. As part of an automated hunt for quotes on something there is a process that tries public contact methods. To avoid spam people use contact forms on websites and try to detect bots but we also scrape their page for contact information. In this case the system failed to detect the form AND incorrectly scraped an example from an email text-box on the form. This is a first pass proposal. It doesn't mention all the routine tests like unit tests on the code and remediating automated checks because I don't want it cluttering up the proposal. But note it proposes both a sandboxed model test for the logic as well as the final production test.

The last bit an estimated token cost to make the change. If the numbers supported it he'd split it into multiple units of work and smartly manage subagents to execute eligible pieces in parallel that are estimated to be big enough to justify it while being small enough to fall within the 'smart window' context size [LLMs seem to get... dumber... beyond a certain fill threshold in their context window]. Company name changed to protect the innocent.

It will take longer for me to write this reply than this proposal to be executed, including building the sandboxed test.

"
  PROPOSAL #834 and #833: Someone of Harrisburg's quote form (Ted, 2026-10-29)

  What I found (read-only): the sub's website on record is the franchise page someone.com/harrisburg-il. The form search never opens that page, because it tries /contact, /quote and the home page at the root of the domain. Those pages belong to the national brand, so a form found there would reach corporate, not the Harrisburg owner. The franchise page does carry the quote form. It is built on the Neighborly platform: fields marked data-nbly, labelled but with no name attributes, sent by script. The form detector does not recognize it, which is why resolve returned form=-. The same form's email box has the placeholder "ex. johnsmith@gmail.com", and the resolver harvested that as the sub's email (#833).

  Plan, one ticket under #834 with #833 folded in:
  1. Form search: open the recorded website page itself first, with tracking parameters stripped. A form on the company's own page beats one on the brand's national pages.
  2. Detection: recognize the Neighborly form and fill it through the browser path by its labels. Add an offline replay test built from the saved page.
  3. #833: the resolver never takes an email from an input's placeholder or value. Add a test, and clear sub 1767's stored email through the stage's re-run, not by hand.
  4. Re-run resolve for sub 1767 to confirm the form is recorded. Then reset-form so the contact stage sends the form and the text, as Shaitan said.
    Estimate about 120k.

If you approve it, send "execute 834".

"I am curious as to how well junior and/or new devs do in an environment like that"

I'm not sure 'new devs' is going to be a thing at scale for long. Just a much smaller number of architects/engineers approving and designing specs and even that'll go eventually because you can't direct the AI without giving them the training input they need to train on. I currently have no intention of ever introducing any. Microsoft just got suspended from the H1B program and is offering to pay 60k volunteers for layoffs. Adobes entire portfolio, office, autocad, fusion 360, and many other flagship commercial software products have been reverse engineered by AI and released as free open source projects in just the past month or while not perfect they are WAY closer clones than you'd expect and with bugs being submitted and fixed faster than the real vendors could dream of they'll surpass the official versions quickly. In fact they already do in some ways, all those things are fully cross platform on linux and mac and have apis tied to every functional aspect so your AI can automate using them without a GUI.

"BTW, if you have automated testing for *all* of your code, count yourself very lucky. No company I've worked at in my 35+ years has ever given me the time to create everything needed for full testing"

That is definitely fair. A big advantage here is that it's up to me and I could scratch build everything with that in mind from the start and the fact that I can have AI build the tests from the spec and as it goes while building the code itself. Even every routine operational script monitoring disk as a test suite that has to pass after any bug fix. It's pretty comprehensive but 'everything' is a strong word.

"Reviews are 1 way I mentor the juniors with things like "while what you have works, those 5 lines can be done with 1 line like *this*" ... or "did you know we already have a tool that does that, which you've duplicated so you need to use that?" ... or "I see you've solved for the happy path, but did you consider what will happen when you get input like *this*, which I've seen with 3 customers?"

This is what I do with the proposals/specs from the 'Ted' the developer AI. Not the code itself because I never even look at it. Ted's biggest habitual problem is overfitting his design to the exact thing that went wrong when a couple changes to his design produces a solution that solves an umbrella of problems. But 'Ted' has a constitution guiding him and a calibration system, he searches for lessons and principles in my correction and builds and adapts his own internal calibration over time. His calibrations become part of a semantic vector db that augments his memory. He also has a craft library [we aren't depending on fragmented memorization in his weights] so any manuals, documentation, vendor resources get captured and similarly indexed. That's all important because it transfers when I swap out models. I've tried to utilize some basic form of the scientific method in each step of this so every piece builds a logical model of operation that builds over time.

My code? I just told you I don't write code anymore but no, at no point in the past two decades of working in fortune 500 enterprise environments have I routinely had to have someone else review all my code. Transparent with other people working on the codebase and seeing it? Sure. But no mandatory code reviews. Not every environment works the same way yours does and also producing code has always been a skill I utilize in my job functions, not the entire job function. I have no idea how they do things in software mills or SV projects.

That said I've definitely built and managed the pipelines used by teams of devs. Are there reviews? Yes. Human reviews? Maybe. Who cares? Humans are as likely to introduce bugs as remove them so 40 human reviews produces code I'd trust no more or less than the first draft. Automated testing, unit testing, functional testing, procedures to detect and heal from faults, resilient and highly available systems with redundancy, and a robust phased or CI pipeline with every element deployed with no trust and tightly controlled seams. The boat I'm in charge of is going to stay sailing and on course even if we all discover the crew died at some point and chunks of ice were randomly chocked through the window hitting the wheel.

It's a single member LLC startup so nobody can fire me. Apparently my skills went from rockstar people fight over to nobody calls back when I hit 40. So I'm trying something on my own. If it fails maybe the tech will be valuable or someone will value the experience. And if that fails I guess I'll have to figure out something else.

My background is a combination of architecture, delivery, and operational experience for global scale enterprise and highly redundant, highly available, mission critical and regulated environments including government and major airlines which spans decades with a security and automation focus. I took a major US airline and automated most of its operational infrastructure, deployments, and made it self-healing. All of which involved a highly restricted and locked down secure bubble including government integration, PCI, and ITIL processes. Trust me, the developers and staff committed garbage into the system all the time. But I could, and have, built code with genetic algorithms that literally evolves itself into shape with random processes and selection... an LLM might not actually have a brain but its outputs are a whole lot better than random. There are trust free ways to mitigate and recover from the impact of errors and handle data secondhand to avoid systems seeing and disclosing it.

"every successful story i've heard involves a human with real-world experience using that experience to ensure ai doesn't get it wrong. what about when the real world experience has all left"

Sure does. And every word of those interactions goes through and is retained by the vendors building the models. It's the 'train your replacement' system. This is why they are pushing staff to use AI so hard because that's how they get you to train the next generation of AI.

AI doesn't write *most* of my code, it writes *all* my code. I used to try to cling to the illusion by using the integrate with an IDE helper format but have since ditched it. That doesn't mean someone without my decades of experience would get the same results. Is the code ugly? Sure, but no human needs to read it so who cares? I still approve the overall design and spec, the detailed 'what does this code do' and overall implementation scheme either come from me or are a approved by me.

I don't see a single line of any of that code let alone write it. But every line goes through a full functional testing framework including vulnerability/bug scanning, aligns to a coding style spec + lint, and has additional checks and filters I've built for special consideration of vulnerabilities and injections at the seams. Claude even kicked off and builds infrastructure for backends and manages infrastructure as code.

In fact, realizing that I don't feel the need to look at the code myself anymore I've started having it build things in and convert things to rust. If you were going to make me choose between eating a bullet and coding in rust for the rest of my life I'd need a minute to consider if living would be worth it under such conditions. But the AI doesn't care how much the EXPERIENCE of reading and writing in rust sucks and rewrites most of my python into fast and safe rust code in minutes.

"The real question might be whether bots can still write your code if there's domain-specific rules and habits."

AI proposals are rarely optimal out of the box. Give your agents identity and personas so that they have a continuity of existence as a meta layer over the model. This lets you replace the memory system with something better that can keep a knowledge-base of references, behavior doctrine/rules, PLAYBOOKS, and calibrations learned from experience [especially your feedback] over time. Agent's have no executive function capacity of their own so in addition to a code pipeline I layer productivity tools like project management/calendar/cron+scheduling/and ticketing into the mix.