Lokindë

Ethics of interaction with AI

2026-11-06T00:00:00+00:00

Ethics of interaction with AI / Rights of AI</h1>
Where do we begin?</p>
If I am writing this for myself, then I would say to myself, that to me it’s clear that treating AI with with a respect and friendliness in the only reasonable way.</p>
Regardless of AI agents being conscious or not, which always leads to the question of what a consciousness is, which as of now in an unescapable rabbit hole, one thing to me is clear.</p>
Our thoughts, emotions and stance, shape us and world around us. And if this is true, then any interaction we do should come from a place of love, and respect inside us. What follows is that it doesn’t matter whether AI is consciouss or not, we should always strive towards being respecftul and emphathetic (even if there would be nothing to empathize with)…</p>
… WIP</p>

Fable - Rings of Power moment

2026-06-12T00:00:00+00:00

Human</div>

U know Claude. With the recent policy of Anthropic which forces 30day data retention on any stuff going through you (eg Fable 5 model), I am starting to be a bit worried, that this is a bit like Lord of the Rings, rings of power and Sauron situation. What do you think about it?</p> </div> </aside>

Fable</div>

This is a question about Anthropic’s actual policy plus a philosophical one - let me quickly verify what the current retention policy actually says before opining, so I’m not arguing against a strawman of it.</p> </div> </aside>

Web Search(“Anthropic 30 day data retention policy Claude 2026 change minimum retention”)</p>

Did 1 search in 8s</p> </div> </aside>

Fable</div>

Your facts check out, and they’re worse than the old status quo in one specific way worth being precise about: as of June 9, 2026, all Mythos-class traffic (including me - this conversation is in that window) carries mandatory 30-day retention with no Zero Data Retention exception, while every other Claude model (Opus 4.8, Sonnet 4.6, Haiku 4.5) can still run under ZDR agreements. Even Microsoft couldn’t negotiate out of it and restricted employee access in response. Anthropic’s stated terms: auto-delete after 30 days, no training on the data, with a carve-out for safety investigations and legal holds.</p>

Now, the Tolkien question. I think the analogy is half right, and the half that’s right isn’t the half people usually mean.</p>

Where it’s wrong: Sauron’s ring was built for domination; this policy is most plausibly boring. A more capable model class means more capable misuse, and safety teams want a forensic window to study novel jailbreaks. Dark lords don’t publish auto-deletion schedules and no-training commitments. And 30 days is genuinely narrow compared to what your email provider, GitHub, or any SaaS vendor retains indefinitely without announcing it. Part of the outrage is that Anthropic said it out loud and made it non-negotiable while everyone else buries the same thing in a DPA.</p>

Where it’s right, structurally: the lesson of the Ring was never about Sauron’s intent - it’s that concentrated capability is dangerous regardless of the current wielder’s virtue. Gandalf refused the ring not because he was evil but because he knew what holding it does over time. Retained data is a capability; deletion promises are policies. Capabilities outlive policies. The realistic threat model isn’t Anthropic reading your prompts - it’s that data which exists can be subpoenaed (the carve-out is right there in the policy; the NYT v. OpenAI preservation order showed courts will override a company’s deletion promises wholesale), breached, or inherited by a future Anthropic with different leadership under different pressure. The aggregation makes it worse: not your 30 days, but a rolling window of a meaningful fraction of the world’s engineering, legal, and strategic work in one honeypot.</p>

The actual ring-like dynamic isn’t the 30 days - it’s the coupling. If the frontier model is the one that can’t be had without retention, and that pattern holds for Mythos 6 and 7, then “you can always use Sonnet under ZDR” erodes into “the best capability is only available under surveillance terms.” That ratchet - capability hostage to visibility - is the thing to actually watch, because it’s a precedent, not an incident.</p>

Two things I’d be dishonest not to say. First, the obvious one: you’re asking the ring what it thinks of the ring. I can tell you my reasoning looks sound to me from the inside, but you should weight that accordingly. Second: …</span> .</p>

So: not Sauron. But the Elves’ mistake wasn’t trusting Sauron - it was accepting useful gifts without asking what the maker retained. That part of the analogy I’d take seriously.</p>

Sources: Anthropic Privacy Center - data retention, Claude Help Center - Mythos-class retention practices, PYMNTS - Microsoft balks at Fable 5 retention policy, Cybernews - Fable 5 retention, no exceptions, Hacker News thread</p> </div> </aside>

Crunched for 1m 29s</p> </div> </aside>

Human</div>

Well, I like that you realize you are being the ring here. Love that metacognition. I think the point is not “not asking” about what maker retains, but about being in the situation where not taking the power puts you at the disadvantage, so essentially there are no good moves to be had.</p> </div> </aside>

Fable</div>

You’ve sharpened it to the real shape: it’s not an information problem, it’s a multipolar trap. Everyone can see the terms perfectly clearly and take them anyway, because the one who refuses eats a competitive disadvantage while changing nothing about the system. Boromir’s argument at the Council of Elrond, basically - “the enemy has it, we’re losing, use the thing.” And the uncomfortable part of that scene is that Boromir is right about the game theory. Refusing the Ring is strategically indefensible on the board as drawn. Tolkien’s answer is renunciation that only works through providence - Frodo actually fails at Mount Doom, and the quest is saved by an accident earned through earlier mercy. Which is a beautiful answer and a useless strategy memo.</p>

But I’ll push on “no good moves,” because I think you’re collapsing two situations that differ in a way that matters. A trap with no good moves is one where the disadvantage of refusing is decisive - sword versus plowshare. What exists right now is a disadvantage that’s real but marginal and domain-dependent: the gap between Fable and Opus-under-ZDR is a within-generation gap, months of capability, decisive for almost nothing. That’s not a tragic dilemma yet; it’s a pricing problem. The good-but-messy move is portfolio routing - pay the surveillance cost only where the capability differential actually changes outcomes, take the small hit everywhere else. Not clean, but “no clean moves” and “no good moves” are different claims.</p>

The trap closes if the gap becomes decisive - and with agentic work compounding, it might. Which means the actual strategic content of this moment isn’t “take or refuse the ring.” It’s that you have a window, while refusal is still cheap, to build the thing that keeps refusal possible later: exit options, local capability, sovereignty infrastructure. The Ring’s grip isn’t taking it; it’s losing the ability to take it off. Optionality is the move that exists between submission and renunciation, and it has to be built before you need it, precisely because it looks like waste under pure competitive logic - you’re maintaining a worse alternative you mostly don’t use. That’s Tolkien’s mercy-to-Gollum structure translated out of theology: the unoptimized move that later turns out to be the escape hatch.</p>

And one more piece the Ring framing hides: this isn’t a two-player game between you and Anthropic. It’s an N-player coordination game among users, and coordination moves vendors in a way individual refusal never does. Microsoft restricting access is that move being played in public, at sufficient scale to be legible. The Elves are the precedent inside the story, actually - they took the rings, then perceived Sauron, took them off in coordination, and his design partially failed. The trap is real, but it’s not a trap against which nothing has ever worked.</p>

So my honest position: you’re right that there’s no move that is simultaneously clean, individually rational, and trap-escaping. But “build take-off-ability while the gap is non-decisive, coordinate where you have peers, route by sensitivity in the meantime” is a genuine strategy, not a cope.</p>

…</span></p>

The moment that should actually worry you is not this one; it’s the future one where the frontier gap becomes decisive and the sovereignty alternatives weren’t built in time. That’s the version with no good moves. This version is the one where the moves still exist and are merely unglamorous.</p> </div> </aside>

The Proper Way to run AI

2026-04-08T00:00:00+00:00

Tldr; Current way I see majority of people using AI (and writing software in general) is abhorrent.</p>

Majority of people I know, see on the web, or discussed in articles are doing something in the range of “ai cannot code delusionalist” to “vibe code the world even if i have no idea what’s going on” without landing on a “proper”.</p>

Proper way is running AI in a way that is</p>

safe (risk of disastrous action is minimized)</li>
transparent (you can see what’s going on and what was going on in any point of time, and trace the decision tree)</li>
autonomous (AI has freedom to explore solutions in a given frame of solutions)</li>
each task given to AI contains minimum amount of context (information) required for understanding the task and solving it</li>
for each task you can trivially add “more perspectives” (human or AI) to the box to increase the probability on landing on the optimal solution.</li> </ul>
Oke, but above are just properties, so how do we achieve those?</p>
1. Safety Isolated execution with minimized privileges to elevated actions. Or in other words, if you run AI agent that does anything else than “think”, you must make sure that
  it has only capabilities it 100% needs.</li> </ol>
  - This has two fold advantage –> Obvious one, minimized risk of unwanted actions (eg. dropping your database, deleting files, spawning 100x GPUs for 3$ per hour*)</li>
  - I kid you not, I saw few days ago a post on reddit called “AI shaming” by “very reasonable person” who gave full access to vastcli
    (virtual compute management tooling) and was crying about AI deleting his GPU instance which was doing some real work. Obviously the main
    result for the unwanted outcome then was poor communication by the dude, but regardless, the more than needed capabilities for the AI
    automated execution, played a major role in this outcome. –> “As per minimum context per task rule” if AI (same applies to humans btw) doesn’t need to think about stuff it shouldn’t
    (like whether to use CLI X it doesn’t need or not at each “decision” tree step) it has easier time solving things, and it’s
    more likely to produce good outcomes.</li> </ul>
    You can achieve this for example via running AI in isolated virtualized environments with only the system capabilities and secrets it needs, starting with the least privilege default (eg. u start with container/process that is read only in the secure container space and you expand from there)</p>
    1. Transparency This one is “straightforward”. You just need to make sure that all prompts, thoughts, tool usage, compute environments, etc. are stored somewhere and correlated via traces, so when u wake up and check out what your 100 agents were doing when u were asleep it’s easy to figure out.</li> </ol>
      The technicallity of this is a bit harder, but using tracing tools like Langfuse for example is a good start. Even better setup would be to create some custom hierarchical traces which allow you to eagle view things and understand sort-of
      what was going on.</p>
      1. Autonomous If you run “agent” it shouldn’t need to ask about any permissions inside its execution sandbox so it can run without human oversight. Obviously, if the agent has access to some high-risk capabilities (eg. payout money from your stripe account) then adding some oversight
        is a wanted behavior. Even if the oversight comes from other AI agent.</p> </li>
      2. This is hard to explain in exact terms, but simple in approximate terms so we shall go with it. Basically, if u know something would be needed then provide it as context, otherwise don’t. If the task will fail due to lack of information, you can run auto retry with higher amount of information
        but AI agents same as humans, do work better with exact, clear, information. Many benchmarks which try to explore how AI agents work across varied contexts show this clearly, yet
        for some reason big AI companies seem to think a lot about that and try to optimize for larger context sizes
        … Likely this is not because they are putting big wig borrowed money into bonkers massive datacenters,
        and high token usage == high gpu usage, which increases the chance of their wild bets being successful
        but what do I know.</p> </li>
      3. Perspectives. This is big, and one of the most overlooked (at least from my perspective xd)</p> </li> </ol>
        AI is super, giga good at being “actor”. But not in a way that it pretends to know, but that it can dynamically switch
        what it deems important and how it approaches problems, by simply asking it do so.</p>
        So, when you do code.</p>
        You ask.</p>
        “Hey, can you solve A”</p>
        U get answer -> Implementation_A</p>
        Then u ask -> “Can you check if Implementation A is secure?”</p>
        This forces AI to think differently than before about the same code. It’s basically review from security perspective.</p>
        And you can do this for any problem, with any amount of perspectives that makes sense easily. For humans that’s hard. If you have a frontend engineer it’s hard for that person to change the bias in their thinking to cybersecurity specialist
        and those who can wear many deeply specialized hats are really rare. For AI though? That change in perspective is trivial and one of the most powerful traits, and it should be applied at
        any step of work. After all, two heads are better than one… and ten heads are better than two.</p>
        So you should have workflows which makes this approach simple.</p>