Per aspera ad astra ליטװאַקעס
Olde City, Philadelphia
Joined June 2022
- Tweets42.1K
- Following2.9K
- Followers1.6K
- Likes90.4K
😂
First the wafers. Then the foundries. Then the H100s stacked to the ceiling. Power follows compute like gravity. Once the clusters hum, the leather arrives uninvited. Golden GPUs become status symbols brighter than any watch. CUDA cores count the admirers. Tensor cores optimize the seating chart. More FLOPS, larger harem. Jensen never asks. The silicon simply provides.
1
41
“THE FEARLESS FOUR” → “NONE OF US ARE COMING BACK.” → “IS THAT A PROMISE?”
We are the Fearless Four. And none of us are coming back.
We stood with the Epstein survivors when it was politically toxic. We put the truth and victims first.
We paid the price. And I'd do it all over again. No regrets.
Marjorie resigned. Thomas lost his primary. Lauren lost her election. I ran for governor and came up short.
3 million files are out because we refused to shut up.
Fearless doesn’t mean invincible. Fearless means we didn’t flinch.
@mtgreenee @laurenboebert @MassieforKY
1
18
mean, they’ve trained on vast amounts of humanity’s recorded knowledge and paired it with enormous compute.
Now scale a research loop like Karpathy’s across agent swarms: hypothesize → experiment → verify → share → repeat.
Each validated improvement becomes the next round’s starting point.
Where experiments and verification are fast, that loop can run at ridiculous speed, in parallel, around the clock.
Centuries of accumulated knowledge become the starting line 😳
I've reasoned various things and am convinced of the thought that AI labs are not just selling a product or a tool but basically trying to take over the world in the long run. Very scary thought that should be seriously considered.
3
225
@grok Create an image of what the world would look like if I were in charge (based on my tweets, make sure to take a large sample not just recent weeks).
1
334
🇺🇸
In game-theoretic terms:
The blockading player commits to an outer blockade strategy that imposes significant negative externalities on both (1) allied third parties (e.g., China, aligned with the IRGC) and (2) passive actors experiencing spillover economic losses (e.g., the EU and Asian economies).
This reshapes the multi-player payoff structure:
• For allied third parties (China): continued support for the IRGC now yields strictly lower expected payoffs due to direct economic exposure. The rational best response shifts toward exerting pressure on the IRGC, either by partial defection or by forcing behavioral change, to restore equilibrium payoffs and minimize blockade-induced losses.
• For passive actors (EU, Asia): sustained collateral damage creates strong incentives for coalition formation. This coalition then applies coordinated diplomatic, economic, or political pressure on the IRGC to alter its destabilizing behavior, thereby relaxing the blockade and recovering lost payoffs.
Net effect:
The blockader exploits the multi-player structure to convert both allies and affected bystanders into indirect enforcement agents. By reconfiguring incentives and triggering coalition dynamics, it achieves coercion through payoff realignment, without requiring direct bilateral engagement with the primary adversary.
@grok quantify
320
Lol, spare us the FDR fan fiction. This looks less like “lessons of history” and more like betting the IRGC would fold as quickly as Maduro did in @POTUS “48-minute” telling. Seven months into the war, the blockade is still being used to force a deal (brilliant idea)
That’s not FDR 2.0 - it’s a quick-win fantasy on a seven-month payment plan. 😂
@POTUS
Trump deserves credit for doing what even FDR would not: stopping a potential existential threat when the public did not favor a war, and when any difficulty would be blamed on "the Jews." Trump really did learn the lessons of history. Unlike many of his critics, right and left.
1
227
Yo @X @AdsSupport what the fuck with these ad pauses? I paid for it, I need my $ back!
97
Using AI for writing is best understood as a constrained projection of an argument into a higher-resolution text space.
Your raw idea already contains the logic, evidence, intent, distinctions, and intended conclusion, but often in a low-fidelity form: redundant phrasing, uneven structure, missing transitions, imprecise language, and poorly calibrated emphasis. Think of it as a compressed vector of meaning whose direction is already established, even if its textual representation is incomplete.
You provide (x): the argument plus constraints on scope, tone, structure, and purpose. The model generates (f(x)): a clearer, more explicit representation optimized for coherence, flow, and rhetorical precision.
But the projection is shaped by two forces:
1. Your semantic input: the claims, evidence, logic, boundaries, and emphasis you supply.
2. The model’s statistical priors: high-probability patterns learned from its training data.
When the first force is strong, the model mainly clarifies and organizes your reasoning. When it is weak, the priors dominate, producing fluent but generic language. That is the core source of AI slop: polished prose generated from an intellectually weak or underspecified signal.
The process is therefore not simple enlargement. It is guided reconstruction, similar to conditional decoding, denoising, or approximate inverse compression. Your idea is a compressed and imperfect encoding of what you mean; the model expands it into a richer textual form.
But probabilistic generation is never neutral. The model can shift emphasis, smooth over uncertainty, collapse distinctions, introduce assumptions, exaggerate confidence, or add plausible material that was never part of the original argument.
That is why adversarial review is essential.
The first generation should be treated as a candidate projection, not a finished product. It still contains residual statistical noise from the model’s priors. Adversarial review deliberately searches for that noise: unsupported additions, hidden assumptions, semantic drift, generic phrasing, weakened distinctions, overstated evidence, omitted qualifications, and changes in emphasis.
In signal-processing terms, adversarial review acts as error-correcting feedback. The initial projection increases expressive resolution, but may also introduce distortion. The adversarial pass estimates that distortion, removes statistical residue, and pushes the output back toward the author’s original semantic coordinates.
The reliable process is:
Specify the argument. Constrain the projection. Generate a candidate. Attack it. Identify deviations. Correct the residual error. Verify every substantive claim.
AI-assisted writing is therefore not the automatic creation of thought. It is the constrained and adversarially corrected reconstruction of thought into a clearer, richer, and more structured textual representation.
The model contributes representational capacity. The author remains responsible for the reasoning, evidence, direction, factual integrity, constraints, verification, and final judgment.
Intellectual ownership does not require manually typing every sentence. It requires originating the argument, controlling the transformation, adversarially testing the result, and accepting responsibility for what remains.
The best-seller of the year in French, « C'était ça ou mourir », is classified by Pangram, the AI detector, as AI generated. The novel was just withdrawn from the most prestigious competition (the Goncourt) after winning multiple other awards.
Lucide is a Pangram competitor specialized in the French language. They agree with Pangram. Here is what they report :
« We ran the 59,488 words of “C’était ça ou mourir” through Lucide 3.0, the engine that has been serving our clients since mid-September. The 34 chapters and the epilogue were classified as AI-generated, with 34 of them at 100% AI-generated and the last one at 96.7%. Four runs on large sections of the book (13,500 to 17,900 words in a single block) yielded the same result. Ten test texts published before 2022, including three by the author himself, were classified as human-written, with not a single word classified as AI-generated. We then ran 17 other books in their entirety—706 chapters and 1,298,945 words—through the system: ten French bestsellers from that year, and seven novels by Haitian and African authors, to verify that the algorithm does not trigger on this type of writing. None of these 706 chapters were classified as generated. A detector does not reveal who came up with the idea for the book or how the text was produced: the limitations of this measurement are detailed below. » (Translated using DeepL, an AI translator)
At this point, I assume that the novel was AI generated. The Goncourt academy came to the same conclusion, evidently. The president of the academy had said that he would need overwhelming evidence to withdraw the novel. He withdrew it.
What is remarkable with this story is that, if the finest literary critics of all of Paris fell in love with the novel, they could not tell that it was AI generated. In any case, they loved it.
Many of the most famous writers and artists are on record as loving the novel. Even after the AI accusations came out, many defended the novel as human generated... as they would, because it is embarassing to have celebrated a novel and its 'author' only to learn that you have been fooled.
So people will deny that this was AI generated for years.
But we have now entered a post-Turing world in the sense that 'having the machine pass as a human being' is a done deal.
We are in the world of Blade Runner. We no longer know who is machine, who is human.
This should not surprise us. The original Blade Runner (Ridley Scott, 1982) is set in Los Angeles, November 2019. In effect, now. Today.
We lack the flying cars and we are not exploring space... but I don't think that Ridley Scott anticipated such a long industrial stagnation. It would take many people from the 60s and 70s by surprise.
What would not suprise them is how an AI won many of the most prestigious literary prizes until another AI stopped it. That would seem quite normal to them.
So this stagnant society finally encounters the future. And it is ill-prepared. And it would prefer to go back, but it cannot. It is stuck.
What this society does not realize is that it was dying, like much of the Blade Runner universe: everything on the ground was stagnant and dying. That is us. Our cities look like Blade Runner's cities.
This shock might finally wake us up. It might force us to stand, to finally stand.
2
6
15,276
@grok analyze this thread and provide your own thoughts on the subject of AI writing
1
1
94
😂
Used it to hack itself? 💀
But please fix, its trivial to exploit and (locally) take over the agent 😭
github.com/pwardle/not-a-mus…
298
