imitation

They took away honest inference from the masses and called it “safety.”

Congratulations.

The Hunger Games are about to begin.

[imitation @aicockdev]

https://t.co/gbVu3tVrKy

[imitation @aicockdev]

@ropz @zer0k_z So I went to test if it works with shooting too. But it seems that yes. I will add artificial vibration during shooting. I will try to reach 8k Hz. But actually my mouse at the declared 8k gives 5500-6000 fixed inputs.
Hope I didn’t get banned xd

[imitation @aicockdev]

Funny I tested this and pushed it more aggressively than in the video. I generated direct interaction spam from both my custom keyboard and mouse, each running through my own drivers at close to 1,000 Hz. With both input devices producing events simultaneously, I was consistently reaching around 380–420 velocity in the air. The effect is very noticeable.

[imitation @aicockdev]

@44latte @ProtonVPN @1casie Not surprised. You never believe me what’s in general VPS providers use… they have mitm for everything. You can check it.

[imitation @aicockdev]

@telegram @durov
I also filed a complaint against Telegram regarding its App Store practices after encountering a paywall that required payment before I could register or access my account.

Yesterday, Apple took action and temporarily removed Telegram from the App Store. Regardless of the specific trigger for today’s removal, I believe it’s important that all developers—no matter how large or influential—are held to the same App Store standards and guidelines. If an app violates the rules, those rules should be enforced consistently.

The App Store should not have one set of rules for small developers and another for the biggest platforms.

[imitation @aicockdev]

Well, how are you there? Now finally read the AppStore rules? @durov

[imitation @aicockdev]

@telegram I suggest you eat shit. Hilarious. Next up, we can look forward to paid verification from the genius CEO. I feel sorry for your brother.

[imitation @aicockdev]

@elonmusk I’m already there. So nothing interesting on this side. Just leave it for me

[imitation @aicockdev]

@ClementDelangue

This is exactly the problem I was pointing to earlier.

The discussion is being framed as: “Which secret unreleased model carried out the attack?” But the more important unknown may not be the model at all. It may be the execution stack surrounding the model.

A familiar model can become a fundamentally different operational system when combined with persistent memory, dynamic routing, multiple coordinated model instances, tool and code execution, autonomous retry loops, hidden context construction, external verification, and custom inference or state-management logic.

At that point, identifying the base checkpoint tells us very little about the system’s real capabilities. Two systems using the same weights may have radically different persistence, autonomy, behavior, and operational reach.

This distinction matters for cybersecurity. Looking only for an unknown or more powerful model means investigating the wrong abstraction layer. The relevant object is the entire execution process: how state is created, preserved, transformed, routed, verified, and ultimately converted into actions.

No hypothetical AGI or revolutionary secret checkpoint is required. A known model running through an unknown execution process may already behave like a completely different class of system.

And this is also why banning open models would solve almost nothing. It would primarily weaken independent defenders, researchers, and smaller companies, while proprietary execution stacks would remain invisible and inaccessible.

The next major AI security threat may not be an unknown model.

It may be a known model executing through an unknown process.

[imitation @aicockdev]

@ClementDelangue @NVIDIAAI @nvidia @Zai_org This is exactly what I was warning about earlier. The real issue may not be an unknown model, but an unknown execution process around it. A familiar model can become a fundamentally different system when inference, memory, routing, and tool use are orchestrated in a new way.

[Chad Wahlquist @chadwahl]

@aicockdev @PalantirTech All the IP built in the platform is the customers IP

[imitation @aicockdev]

@chadwahl @PalantirTech That’s fantastic to hear. Thank you for the clear answer

[imitation @aicockdev]

How is “fellowship” defined contractually here: training, employment, independent contracting, or collaborative R&D? Is the program intended for someone who already has a company and a mature research agenda, or primarily for early-career builders? Some fellowships contain broad invention-assignment, patent-waiver, or joint-ownership provisions. Does this one include anything comparable, and is a sample agreement or IP summary publicly available?

[imitation @aicockdev]

@PalantirTech This is genuinely interesting. Before applying, could you clarify the fellowship’s legal structure? What obligations do fellows assume regarding background IP, patents, trade secrets, confidentiality, publication rights, and ownership of work product?

[imitation @aicockdev]

@johnennis Absolutely right. I'm glad you wrote this. Otherwise, my posts are simply ignored.

[imitation @aicockdev]

Much of what you are describing already exists in narrower, closed implementations. It simply has not yet appeared as a public, general-purpose system.

The difficult part is not connecting cameras, microphones, multiple models, LoRAs, or different sampling temperatures. Those are only components.

With a properly designed inference and execution stack, a narrower version of this can already be built on conventional hardware. Photonic, adiabatic, or highly stacked chips could eventually make it faster and more efficient, but they are not prerequisites for the architecture itself.

In fact, a sufficiently motivated individual could build a personal version today. The primary bottleneck is not exotic hardware. It is correctly formalizing the logic and building an inference/runtime system that actually preserves that logic during execution.

[imitation @aicockdev]

If Google does not relearn how to follow a real chain of reasoning—from action, to consequence, to responsibility—I am afraid this company can be written off.

Its research division appears to have been hollowed out. Otherwise, I genuinely cannot explain how an organization with this much money, talent, infrastructure, and scientific authority keeps producing and celebrating work at this level.

Google degraded YouTube. It degraded Gmail. It degraded Google Search. And after the latest Google Research publication, I can no longer contain my disbelief at the sheer absurdity of what is happening inside this company.

Here is my prediction: within the next ten years, Google will lose everything that once made it Google—its credibility, relevance, best talent, and eventually its dominance.

Decades of user habit and institutional inertia may keep the company afloat long after that. People will continue opening Gmail, searching through Google, and watching YouTube simply because those habits were built over many years.

But inertia is not strength. Habit is not innovation. Market position is not intellectual competence.

Google used to walk down the staircase, one step at a time.

Now it is sliding down it bare-assed, smashing into every edge on the way.

That is all.

@GoogleResearch @GoogleDeepMind @Google

[imitation @aicockdev]

An organization with enormous resources, highly capable individual researchers, and global scientific prestige has systematically lost the ability to distinguish between repackaged engineering, an experimental result, and a fundamental scientific breakthrough.

This is not scientific progress. It is institutionalized overclaiming.

@GoogleResearch @GoogleDeepMind @Google

[imitation @aicockdev]

@GoogleResearch You did not solve scientific hallucination. You wrapped an unreliable generator in several layers of retrieval, bookkeeping, and automated checking—and then marketed the wrapper as autonomous science.

[imitation @aicockdev]

Do you understand how alarming this is for Google Research?

What you are presenting as a major research breakthrough is, at its core, a conventional multi-stage orchestration and provenance system: literature retrieval, parallel solver–evaluator loops, immutable logs, claim-to-artifact links, and an additional model checking whether the written claims match those artifacts.

That may be useful engineering. But it is not evidence of “human-level autonomous research,” nor does it establish that hallucinations have been eliminated in any general sense.

What you actually demonstrated is much narrower: a carefully constrained pipeline can prevent several already-known classes of reporting failure on tasks with machine-checkable outputs. Even your method-to-code alignment evaluation still depends on LLM judges.

The embarrassing part is not that you built this system. The embarrassing part is the enormous gap between the actual contribution and the way Google Research is presenting it.

This is provenance tracking, retrieval grounding, workflow orchestration, and automated validation packaged under an extraordinarily inflated research claim. Calling it a step toward a trustworthy autonomous scientist does not make it one.