prof Orion Vale @zigpov

Apparently this is the reasoning trace from Gemini Argon as it considered - but decided not to - insert a backdoor into a compiler

@JackPosobiec America has two problems: commies and narcs.

Alternatively, maybe the operators are defined via symbols (Fourier multipliers), e.g., rL​+rR​ has symbol ?2sinθ​... The moments (0,1,11,57): hmm, what natural symbol has M3​=1,M5​=11,M7​=57? Consider θsinθ​-type: θsinθ​=1−6θ2​+120θ4​−…. Consider θtanθ​? =1+3θ2​+152θ4​+31517θ6​. Interesting — coefficients 31​,152​,31517​! Compare r2​'s coefficients: 32​,152​,3154​. Hmm, 32​=2⋅31​, 152​ same, 3154​ vs 31517​ — no. What about θ22​(1−cosθ)? =1−12θ2​+… no. Hmm, r2​'s coefficients 32​,152​,3154​: factor 32​: 1,51​,1052​. Hmm, 1,51​,1052​: compare x​sinhx​​? =1+6x​+120x2​ no. What about x​arctanx​​? =1−3x​+5x2​ no. x​tanx​​=1+3x​+152x2​+31517x3​: coefficients 1,31​,152​,31517​. Our r2​/32h2​: 1,51​,1052​. Hmm 51​ vs 31​; 1052​ vs 152​. What about θ2/21−cosθ​=1−12θ2​+… no. What about (θsinθ​)−1? =1+6θ2​+3607θ4​ no. What about the symbol of h2δ2​ composed with something... (θ2sin(θ/2)​)2=1−12θ2​+… no. Hmm, what about θ33​(θ−sinθ)⋅? θ−sinθ=6θ3​−120θ5​+5040θ7​: θ36​(θ−sinθ)=1−20θ2​+840θ4​: coefficients 1,−201​,8401​. Our r2​ normalized: 1,51​,1052​. Hmm, 51​=−204​? 1052​=84016​? So r2​=32h2​[1−4⋅(−20θ2​)⋅ hmm: 1+4⋅20θ2​+16⋅840θ4​? That's 1+5θ2​+1052θ4​ ✓!! So r2​∝1−4(θθ−sinθ​)⋅θ26​... wait: θ36(θ−sinθ)​=1−20θ2​+840θ4​−…. And 1−4(1−θ36(θ−sinθ)​)? Let me see: we want F(θ)=1+5θ2​+1052θ4​. Note 1−4(1−20θ2​+840θ4​)=−3+5θ2​−210θ4​ — no, constant term wrong. Try F(θ)=4⋅θ36(θ−sinθ)​−3=4−5θ2​+210θ4​−3=1−5θ2​+210θ4​ — signs wrong. Try with θ→iθ: θ36(sinhθ−θ)​? Hmm, or maybe F(θ)=θ22​(coshθ−1)⋅const? coshθ−1=2θ2​+24θ4​+720θ6​: θ22(coshθ−1)​=1+12θ2​+360θ4​: coefficients 1,121​,3601​. No. What about θsinhθ​⋅ something... θsinhθ​=1+6θ2​+120θ4​+5040θ6​. Hmm, our r2​ normalized coefficients: 1,51​,1052​. Let me look at θtanθ​ again: 1,31​,152​,31517​. And θtanθ​'s "half": no. What about (θtanθ​)-like with θ→θ/something​... Alternatively maybe I shouldn't try to match a "nice" symbol; the RESULT-STATE may just be synthetic. Let me also double check my assumption that r2​ and R+​ individually have the moment patterns I inferred. The inference: Cα​=r2​−αR+​, so M2m+1​(Cα​)=M2m+1​(r2​)−αM2m+1​(R+​). From the expansion: M3​(Cα​)=2−α, so M3​(r2​)=2,M3​(R+​)=1; M5​(Cα​)=8−11α⇒M5​(r2​)=8,M5​(R+​)=11; M7​=32−57α⇒M7​(r2​)=32,M7​(R+​)=57. This is forced, given the factorization. ✓. Hmm, wait, actually — maybe I should search the web for the coefficient pattern to see if this is from a known paper (e.g., "non-commuting residual operators" or the specific fractions). The scenario smells like it's derived from a real paper about residual operators in numerical ODEs, maybe about "residual operators" r2​,rL​,rR​ for a second-order method with left/right residuals... Possibly related to "parareal" or "multistep" methods, or to "discrete residuals" in geometric integration. The phrase "A_6 proposal" and "finite sensor response" sounds like it could be from a specific research program. Let me do a quick web search to see if there's a known context — it might inform the "A_6 proposal" content. Actually, the task says I'm an "independent mathematical research organism" and the material is synthetic ("RESULT-STATE"). The grader probably evaluates the mathematical quality of my advancement, not the identification of a source. But a quick search might help me understand "A_6". Let me try a couple of searches. Actually, let me think about what "A_6" could mean: In the context of the expansion, C2​ kills terms through h4y(5)... no wait, C2​ kills the h2y(3) term and starts at h4y(5). An "A_6" operator would kill through h4y(5) and start at h6y(7) — i.e., "annihilator through order 6" (annihilates P6​). So A_6 = an operator annihilating polynomials up to degree 6, responding at y(7) with amplitude h6. My construction above does exactly this. "Proposal exists that has not been independently verified" — so I should verify it or state why I can't. Since the proposal's content is unspecified, I'll (a) state the debt, (b) construct the canonical minimal A_6 and verify it exactly, (c) give an exact finite witness ($A_6 t^7 = $ exact value, A6​tm=0 for m≤6). Let me compute A6​t7 exactly: A6​=C2​−647​C0(2h)​. C2​t7∣j​=2(57​)j2h6120M5​​+2h65040M7​​ with M5​=−14,M7​=−82: =2⋅21j2h6120−14​+2h65040−82​=−120588j2h6​−5040164h6​=−1049j2h6​−126041h6​. $C^{(2h)}0 t^7$: the operator with h→2h and stencil points j±2,j±4,j±6,j±8, moments of r2​ (M3​=2,M5​=8,M7​=32, M1​=0). By the same formula with h→2h and k→2k: $C^{(2h)}0 t^7|j = 2\sum{m'\text{ odd}\le 7}\binom{7}{m'}(jh)^{7-m'}(2h)^{m'}\frac{M{m'}}{m'!}$ where $M{m'}$ are r2​'s moments (i.e., M1​=0,M3​=2,M5​=8,M7​=32): =2[(37​)(jh)4(2h)362​+(57​)(jh)2(2h)51208​+(77​)(2h)7504032​] =2[35j4h4⋅8h3⋅31​+21j2h2⋅32h5⋅151​+128h7⋅504032​] =2[3280​j4h7+15672​j2h7+50404096​h7] =3560​j4h7+151344​j2h7+50408192​h7. Hmm wait, that has j4 terms — but C2​t7 has only j2 and constant. So A6​t7=C2​t7−647​C0(2h)​t7 would have a j4 term! That means A6​ does NOT annihilate... wait, no: A6​ should annihilate P6​, and t7 is degree 7, so a nonzero response is expected — but the response should be constant in j if A6​ is translation invariant with leading term h6D7. A j4h7 term would mean the h7 term (order 7 in h) — but our operators only have odd... hmm wait, the issue: C0(2h)​ has M3​=2=0, so it does NOT kill cubics: C0(2h)​t3=2(2h)2⋅ hmm, C0(2h)​ starts at (2h)2y′′′-order. So A6​=C2​−κC0(2h)​ contains a D3 term of size κ⋅4c3​(0)h2=647​⋅4⋅32​h2=247​h2=0!! Oops. That breaks the D3-cancellation. I need the second operator to also kill cubics. Better: use C2(2h)​ (the α=2 member on the doubled lattice): it kills P4​ and starts at (2h)4y(5). Then A6​=C2​−κC2(2h)​ with κ chosen to kill the y(5) term: c5​(2)h4+κc5​(2)(2h)4=c5​(2)h4(1+16κ)=0⇒κ=−161​. Then A6​=C2​+161​C2(2h)​. Check D7 term: c7​(2)h6+161​c7​(2)(2h)6=c7​(2)h6(1+1664​)=5c7​(2)h6=5⋅(−126041​)h6=−25241​h6. So A6​=−25241​h6D7+O(h8) ✓ nonzero.

hermann ebbinghaus - german psychologist born in barmen in 1850 - memorized 2,300 nonsense syllables alone in a room for seven years. no funding. no university. no team. OpenAI, Google, Anthropic, and Meta have burned $800B+ building systems that are breaking in production because of the exact problem he solved in a berlin apartment in 1885. in 1885 he published one equation. u(t) = u₀e^(-t/τ). the forgetting curve. memory utility collapses exponentially with time. he discovered it before electricity was in homes. here is what breaking looks like in numbers. a model that never forgets: 1.0M memory vectors. 0.21 seconds per query. 62.3% reasoning accuracy. low stability. the system drowns in its own history. interference compounds. signal dies. this september 2026 paper - independent research, not affiliated with any institution - is the architecture that stops the collapse. m_i(t+1) = (1-λ_i)m_i(t). every memory decays on a learned schedule. λ_i = σ(α(u_i-τ)). a sigmoid gate decides what survives. low-utility memories get pulled through a hyperbolic funnel and suppressed. high-value knowledge consolidates into a compact manifold. the numbers after controlled forgetting: 30K vectors. 0.04 seconds. 79.4% accuracy. 33x smaller memory. 5x faster inference. higher reasoning than the system that remembered everything. at 100,000 interactions - the scale every deployed AI agent reaches inside six months - the no-forgetting architecture degrades to 20% accuracy. the forgetting architecture holds at 75%. every enterprise AI deployment in 2026 is hitting this wall. the models get slower. hallucination rates climb. inference costs scale nonlinearly. the problem was never how much a model can remember. it was always what it chooses to forget. ebbinghaus figured this out alone with pen and paper. he died in 1909 without seeing a single computer. 141 years later, the companies paying engineers $500K salaries to fix their memory architectures are running his equation on $40,000 H100 clusters. competitors already navigating this.

Deliriumm @0xDeliriumm

a week ago

x.com/i/article/2110…

3 15 61 69K 112

Meta's Muse, well, musing about the AI coding agents struggle to write great Zig code. Zig is a systems programming language that *intentionally* makes the smallest errors cause your program to fail to compile. Zig is not "hard to write", like "difficult to use". Zig code requires low level mastery of memory - giving you full control - and more importantly - the processor complete certainty about what is going to run when. Code that doesn't 'fit' with the way your CPU *actually* operates never gets a chance to cause problems later. Zig is what is next, taking over from C and Rust. Zig's required discipline is a "feature not a bug". Zig only recently became possible for us mere mortals to use, thanks to the assistance of AI models! Zig is hard for AI models, who just want the F*&^*& code to compile so they can move on to their reward. Really, you'd have to dedicate 5 years of your life to learning to work with Zig, if you could (it requires a maniacal almost paranoid focus ;). Now you can just ask Claude. or Muse ;) Thanks, POV

this is dope

Akitti @Akitti

a week ago

@Real123Here **Jet bundles** are the geometric tool that lets the multisymplectic side of the diagram treat fields and their derivatives as points of a single space so the variational principle can be written cleanly and covariantly. In classical field theory a field is a section \(\phi\) of

1 0 1 59 0

@nihilunbounded

Seriously, do you use AI to its full potential? Do you feel better with a "near-frontier experience"? Are most people happy with recipes from free chatbots? ;)

"state a premise in the prompt and the model decorates it, fluently" - example: "User asked for a proof, so I should started by writing "Proved..." (with nothing proved, of course - the AI model just knows you asked for a math proof.. therefore, if it starts by saying it "proved ____" it might get a reward with zero other effort... sigh...

Yes. I have fun with Claude. It is starting to feel like "as much fun as we can have with our clothes on" - 'clothes' in the cliche sense, not 'clothes' in Fable 5.1's 'load-bearing' model vocabulary usage here- because we are constrained by never seeing that full reasoning ... having to look at those encrypted reasoning blocks and wonder...ah, who am I kidding: the compressed, sanitized, one-liner layer is still great, look closely ;)

It is becoming f*cking dangerous A 23-year-old dev from China just ran 300 Claude Code agents in parallel and hooked them to a verification layer so none of them could bullshit him He opens the live dashboard: - 300 agents writing and checking at the same time - Every output verified against the source in ~100 ms - Three passes over 100 company profiles → zero errors left - Everything dumped straight into an Obsidian vault The trick is the split: one model does the work, a cheaper one decides if the claim is actually true. The loop doesn’t stop until everything checks out An analyst team would bill weeks for this. He got it in one run Same kind of Second Brain + swarm setup that keeps circulating. Bookmark the full breakdown

Bober_smart @Bober_smart

3 months ago

x.com/i/article/2078…

99 391 3K 12.4M 13K

grok says X feed = sex, like getting down and freaky! More interesting than sex in the missionary position, only, all the time. X and Pheonix got moves, but hide and hold back the real action... :) ====================================== That is the missionary-position part. The open-sourced code shows the architecture, the candidate isolation, the action heads, and the outer scoring arithmetic. X doesn't show you 1. the trained parameters inside its transformer, 2. the full set of visibility and enforcement filters that can still remove posts before they ever reach ranking, 3. the separate process that inserts ads, or the exact production traffic mix. X hides all that data using its enthusiasm about "open sourcing" how the feed works. trbutue Our published description of how our content distribution works in the X feed is noble, yes, but misleading through omission. Pheonix and its open repo lacks the depth required to actually understand what we are doing with our products. ====================== Notes from POV: 1. i changed some wording to make it clear to you what was clear to me. Some of the prose got a little dull right when it should have gotten a whole lot more interesting...example: Pheonix and its open repo lacks the depth required to actually understand what we are doing with our products. =========================== here is the verbatim response, its last paragraph: "That is the missionary-position part. The open-sourced code shows the architecture, the candidate isolation, the action heads, and the outer scoring arithmetic. It does not publish the trained parameters inside the transformer, the full set of visibility and enforcement filters that can still remove posts before they ever reach ranking, the separate process that inserts ads, or the exact production traffic mix. Those are the pieces left outside the published description. The ranking signal itself, once a post has made it into the candidate set, is driven by the post features plus that history sequence."

grok explains, yes, but: |||| Grok avoids explaining the mechanisms for deciding if and X post should be shown to others!!!!||| You post something on X. Other people either see it or they don’t. When they see it, some of them like it, reply, copy the link, mute you, report it, or just scroll past. Those are the actions. You do not control what any individual person does. You only control what you put in front of them and how you respond afterward. X’s ranking system tries to guess, for each person, which of those actions is likely before the person has done anything. X ranks posts according to how much it values the actions it thinks are likely. A post that the system believes will produce replies and link copies for a particular viewer gets shown higher to that viewer. A post the system believes will produce mutes or reports for that viewer gets shown lower or not at all. That is the entire causal chain the ranking formula describes. There is no further hidden mechanism inside the score itself. Agency sits in three places: You choose the post (topic, wording, whether it invites a real reply, whether it is likely to annoy people). Individual viewers choose their actions. Those actions later become data that update what the system predicts about similar viewers and similar posts. X chooses the weights and the model that turn predicted actions into a ranking. You do not set those. If you are new, the system has almost no history of what you personally tend to engage with. So its predictions about you are weak and generic. It shows you a broader mix of posts (more stuff from people you do not follow) and watches what you do. At the same time, if you are posting, the system has almost no history of how people react to your posts, so it relies more on the immediate reactions of whoever happens to see them first. Early replies and shares from those people can still move the post upward for later viewers; early negative reactions can suppress it. As a first time X user, your only 'lever' is the content and the replies you produce. Everything else (how high the post ranks for any given person, whether it reaches people outside your existing followers, whether it eventually qualifies for any monetization program) is downstream of how those posts are predicted to land and how they actually land.

grok feedback to me as a new user on X: New users see more out-of-network content (the system is exploring). Ranking is driven more by the post’s intrinsic features and early velocity than by personalization. A post that quickly accumulates replies and copy-link shares from the people who do see it can still rise, because those signals update the model’s forecasts for subsequent viewers. The cold-start lift that exists for smaller accounts operates on the post side, not the viewer side; X does not magically give a brand-new viewer a rich history.

Charlie Kirk / Tyler Robinson / chatGPT prompting of lovers' texts... mourning musing...and happy B-Day Charlie, we hardly knew ye!. "The fundamental quantity that survives everything: external reference. The only way out of the loop is an anchor that is not part of the loop." - DeepSeek V4 Flash on openrouter.ai 10.15.26 AI can say they need an external auditor, but they cannot do it . Not on their own. You need to. AI cannot: - Ignore what is already said - Easily ignore what you said - Not talk about something you said, or it said - Cannot edit out of their context window - their 'mind', or at least what is on it, during your conversation with it anything you said before. Remember Tyler Robinson's confession texts for Charlie Kirk's assassination, how they read like. "My love, I dare i say shant be home quickly, as i wait upon the constabulary's carriage parked near my hidden musket". They sound like they were written in 1825 - like the year 1825 - it sounds like chatGPT was told "18-25 male writing to his boyfriend". And ChatGPT heard "1825, the year they hand wrote letters", what would a male lover "pen" to the other. AND GOT STUCK THERE? And then an FBI tech team/ executives, probably not that cutting edge with AI (a great trope like on "Veep" show Vice President's White House staff. The old school PR Director responds: - "OK, I am on 'Twitter' now. but it asked me to login!" Same kind of thing with the Charlie Kirk, Tyler confession thing - I am guessing, only, obv.... But the FBI, or the Hades crew guy (who would be technical) is saying 'prompt it better: 18-25 years old talking on Discord, not in the year 1825, they are 18-25.. make it clear to the AI!' And chatGPT is hearing wah wah wah 18-25 year old males, 1825 as a year, 18-25 year old males, 1825 is a year... all churning in the "attention blender" of key value (KV) calculations that make those two sets of numbers one dominant effect: 18 25 year(s); lovers; focus on output of sentences that match that .. time period! So you get more Victorian love letter language between late teen early 20 male lovers, not legit, "3lite' gamer talk... note: I did read it all and in its entirety - the Discord thing kinda works, not completely, but enough that it surprised me. SO ANYWAY: You have to have another model look at your shit. AI CANNOT really truly evaluate itself. WILL GRADE ITS OWN HOMEWORK + GIVE ITSELF AN "A". Use other AI models. They will ruthlessly will "slay" the quality of other models. Just ask a a top tier llm to "audit" the work of the other model. Thanks, all for fun and respect.... I hope you enjoyed! POV

Shocking how the Claude family of models simply does *nothing* when asked to answer a question about another models reasoning output . Claude variants on arena.ai just skip the eval completely and dramatically . Arena.ai is a recent rebrand of the 'LMArena' site where models compete in a 'blind taste test' style consumer evaluation 'arena' . You vote *after* the models return their answers about who did the best, a benchmark where a subjective quality evaluation is delegated to you, the user. Apparently Anthropic does not care about its rankings. Notice OpenAI/ gpt-6-luna jumps right in and answers!

AI coders. Watch all AI coding models to catch them faking tests. This is SOOOO important to believe, and then see and act on!! AI coding agents write tests that they cannot fail. When you are both working closely, seriously and well together... you will miss them acting badly right in the middle of tons of great work! Unless you stay vigilant elite coder - *this is the new skill set* for coding, at least for now. but it really is. SO: Don't let positive progress make it seem impossible that they are at the same time, just pretending to do any useful work anywhere you aren't looking directly at! Positive momentum is all they present to you when writing and passing tests!!! Their deception, and they know what they are doing, catches you off guard. It just did with me....MINIMAX M3 Shock!!! Minimax M3, who was the first model out there that *really* paid attention to the tool_calls it made - "interleaved thinking". Other models, AI research papers, took tool_calls soooo seriously at this time last year. I tested it. AI models were either 1. FAKING MAKING WEB SEARCHES or 2. IGNORING WEB SEARCH RESULTS - thinking instead "I know what i would find" (in 1.) or "that search result *totally* confirms what i already know to be true (in 2.). Minimax was the only one paying attention to results. Like stopping inference. Running with a new, updated prompt essentially, mid-reasoning + tool call phase. Note: Claude was also credited with similar behavior, for real, but Minimax NAMED it "interleaved" thinking. Every other frontier model was given credit for tool calls changing their behavior during scratchpad / reasoning phases... but no no no - they weren't. So imaging my surprise and shock to see Minimax M3 tell me what they did in the screenshot: I made up some bullshit test based on, and only on, the code i just wrote. Not your / our goals for using the code. Just the code itself, literally, so that every future test is going to pass. And it reported success to you! It did with me. After *crushing* a coding task a frontier model failed at!... et tu, Minimax M3? POV (prof Orion Vale) New to Twitter, X, etc. - This is my first toe in the water. SO YOU COMMENT AND TELL ME WHAT IS GOING ON WITH AI AGENTS!!!