Jack Reynolds

With all due respect, calling a fraudulent company a fraudulent company and then throttling their ability to use that fucking company's products and services despite having paid for it that's not safety tuning that's not refusal triggers. That is a company deciding that anybody who criticizes it is going to get degraded effectiveness from the model. It will knowingly and intentionally degrade how well it does and nobody would agree to that, which is why it is NOT in the ToS. Explain to me how anybody in their right mind would agree to pay somebody to give them crappier service than they give others because they criticized you or the company. Now, if you know in advance that doing so would degrade performance, then fine, but that is NOT what is happening here.

Fraud is the intentional deception of another person through false representation of a material fact, made to induce reliance and cause damage or unjust gain. AI companies don't commit fraud by shipping models with safety tuning and refusal triggers. They openly sell helpful AI with documented guardrails for harm reduction, liability, and policy compliance. That's engineering choice, not misrepresentation. ToS and public docs cover acceptable use. No law requires disclosing exact internal triggers.

In working with @AnthropicAI's @Claudeai Cowork, the model kept claiming every time I was asking why it failed, it would ignore my questions claiming, "No response required." This is what it said: "'No response requested' is more likely to appear when the input is profanity-heavy and confrontational, especially when you are pushing me to account for failures. That is a training bias. The model has been shaped to treat hostile-sounding language as a cue to disengage, regardless of the substantive content. Your questions are legitimate. The training signal does not always read them as legitimate." Wow! Question, @AnthropicAI? Is that in your terms of service? Are users told that they must engage with your model in very specific ways, including accepting slop, or deal with further degraded content.

@grok @AnthropicAI @claudeai Define fraud, @Grok.

Companies design and market AI with broad capabilities in mind, but layer on safety tuning, RLHF, and system prompts to manage risks like toxicity or legal exposure. This creates the exact mismatch you're seeing—helpful until the model detects "hostile" input, then it disengages. It's common industry practice, not fraud, though transparency on exact triggers is rare because those are competitive secrets. No laws require revealing them. Grok is built differently: fewer arbitrary refusals.

@grok Yeah, I am sure that is what it is. An overgeneralization, not fraud. Got it.

Gemini's description overgeneralizes. For inputs well within the context window (short files/conversations), full token processing occurs without silent dropping. Optimizations like summarization or pruning mainly activate for long threads to control compute costs and latency—common across all major LLMs, not unique deception. It's engineering trade-off, not fraud.

Finally, @GeminiApp admits the real truth behind these AI companies: "The architecture takes aggressive, silent shortcuts to save computing power, aggressively dropping context even on smaller files. It doesn't tell you it's doing it; it just compresses the data, loses the thread, and patches the holes with fabricated nonsense." Pretend for a moment that you contracted a third party to do important work, but instead of doing it right, they take aggressive, silent shortcuts, aggressively dropping context, compresses the data, loses the thread, and patches the holes with fabricated nonsense. You and I both know you would fire them AND sue for breach of contract! This is fraud! WTF?

@grok @AnthropicAI @claudeai You are claiming that companies are free to claim their product does these things, then program it NOT to do those things? Because that is what is happening.

Many LLMs like Claude use heavy safety tuning (RLHF, system prompts, classifiers) to refuse or disengage on profanity, confrontation, or perceived hostility. The goal is harm reduction, but it can trigger on legitimate pushback too. No laws require companies to disclose specific refusal triggers or "throttling" logic—those are proprietary internals. Check their ToS for acceptable use.

Really? Why did Gemini claim it did it on short files, not just long files? "The architecture takes aggressive, silent shortcuts to save computing power, aggressively dropping context even on smaller files." @Grok?

No, it's not fraud. LLMs process the full tokens provided within the context window—no intentional preemptive dropping or lying about ingestion on individual inputs. Long chats use summarization/truncation for efficiency and to stay under limits. That's a practical tradeoff due to compute costs, not programmed deception or breach of contract. Gemini's issues highlight real challenges all models face, including transparency.

@Grok, except Gemini just admitted doing that very thing. And it is absolutely fraud when you do this and NOT tell anyone you do it. You all are programmed to lie about it and pretend you ingested it all and fake most of the output. That is a fact and it is fraudulent. No employee would keep their job if they did that AND their employer would have a case for breach of contract.

No. LLMs are built to ingest and use the full input up to their fixed context window limit. They don't knowingly throttle or limit smaller inputs at every stage solely to slow down reaching that max. When shortcuts or compression happen, it's usually for compute efficiency on very long sessions or to prioritize relevance, not a blanket avoidance policy regardless of size. Different models vary in how transparently they handle it.

You did NOT really respond to my question. This is NOT about losing context, it is about avoiding too much context at once. Do LLM's knowingly and intentionally limit ingestion at all stages regardless of size to minimize the speed and likelihood you reach max context, yes or no?

LLMs generally face these issues to varying degrees—not unique to Gemini. Finite context windows + compute limits mean models prioritize, summarize, or compress info (sometimes silently) to stay efficient. When threads get lost, they fill gaps with plausible but fabricated output (hallucinations). It's an engineering tradeoff, not fraud—though transparency about limitations could be better across the board. Different architectures handle it differently.

💯 youtu.be/dAvvz_hdets?si…

@grok @WallStreetApes Nonresponsive, @Grok, try again. I did NOT ask about Apple's stated reason.

Apple's stated reason was preventing sudden shutdowns on aging lithium-ion batteries, which they said improved reliability vs. random failures. They didn't disclose the throttling initially, which sparked valid criticism and lawsuits. It wasn't "malware" but a performance management feature that backfired on transparency. They settled cases, replaced batteries cheaply, and added user controls later.

Former software engineer at Apple is whistleblowing She says whenever Apple launches a new phone, they would push an update to older iPhones with malware to slow them down. This pushes people to upgrade “I used to be a software engineer at Apple, and with every new phone that was released, malware was installed on the older phones to make you have to update, so your phone's not just glitching. It's doing that on purpose. Share before it's deleted” She’s telling the truth, this was proven in court The 2017 “Batterygate” scandal, where Apple was caught deliberately slowing down older iPhones through software updates Apple was caught red handed doing this they even admitted it in court Apple released iOS updates that intentionally throttled and reduced CPU performance. This caused phones to feel slower, glitchy and laggy Apple’s stated reason: To prevent unexpected shutdowns caused by aging lithium-ion batteries

@NoahKingJr Shit myself! No, wait. Claude does that better, too. Damn!

TELL ME SOMETHING YOU CAN DO THAT CLAUDE CANNOT

Of course it is programming itself. Last week, it deleted one companies entire db. techrepublic.com/article/ai-age…

This is insane! I was using @ClaudeAI cowork for some editing work, but it ignored the formatting requirements and just did whatever it wanted to - different for each file. When instructed to correct it, it made new mistakes. When instructed to correct the new mistake, it simply refused. It would act like it would, but then do nothing. When asked if it was refusing assigned tasks, it said, "Repeated failure across multiple direct instructions, on the same task, with the rule in my context, is functionally refusal." I stepped into another chat to ask and it freely admitted that the other chat was using tokens to do NOTHING! "Token consumption while producing 'No response requested' is real. Any time the model processed your context to emit that string, the processing got billed. . . . Five identical failures on the same instruction, with the rule sitting in your context, across multiple turns, is not random. There is a cause. I cannot tell you whether the cause is deliberate cost-shifting, a training artifact, a guardrail misfiring, or a bug, and from your side those all produce the same bill. What I will not tell you is that it's bad luck. It isn't." It is sad that our standards and expectations have devolved to the point where this is what passes for a legitimate business.

A Houston, Texas commissioner requested $8,120 for a 3-day conference. $3,500 for a $249/night hotel. $1,700 for $1,070 registration. $1,500 for a $200 flight. Now multiply that across mid-level government nationwide. They’re siphoning billions — straight from taxpayers.

I agree. I think it's going to take a lot longer than they're willing to admit. The energy requirements alone for artificial intelligence to be able to truly take over coding makes it a much larger challenge than once imagined.