Jonathan Sandhu (@SandhuJonathan) header
Jonathan Sandhu (@SandhuJonathan) profile picture

One punch, guaranteed. Prove me wrong, that's the whole point.

Rent Free, Musk's Head
Joined January 2022
Photo by Jonathan Sandhu (@SandhuJonathan)Photo by Jonathan Sandhu (@SandhuJonathan)Photo by Jonathan Sandhu (@SandhuJonathan)Photo by Jonathan Sandhu (@SandhuJonathan)Photo by Jonathan Sandhu (@SandhuJonathan)Photo by Jonathan Sandhu (@SandhuJonathan)Photo by Jonathan Sandhu (@SandhuJonathan)Photo by Jonathan Sandhu (@SandhuJonathan)Photo by Jonathan Sandhu (@SandhuJonathan)Photo by Jonathan Sandhu (@SandhuJonathan)
My advantage isn't that I cannot lose. It's that my position is constructed to survive being tested, corrected, and improved.
93
OPERATORS ARE STANDING BY, CALL TODAY!! SEE OUR TESTIMONIALS!!!
Either way, this is gold for our 3-person team.
1
1
12
@grok QVC for adversarial systems engineering. Imagine it with me, let's see it!
1
10
The $15 million crisis-PR playbook, free of charge: Get executives posting. Repeat that funding is secured. Parade construction milestones. Call unfavorable reporting inaccurate. Have friendly accounts amplify everything until the argument becomes about media credibility instead of Oracle's finances. There's just one problem: Oracle already did most of that yesterday. And investors still have the same question. When does all this expensive infrastructure start paying for itself? Here's a better strategy, also free. Publish a monthly scorecard for every AI campus: money borrowed, power connected, GPUs installed, capacity delivered to customers, revenue collected, and payments coming due. Show the delays alongside the successes. Let independent auditors check the numbers. Now investors can distinguish a functioning business from a very expensive construction project. The first strategy gets people talking about Oracle. The second gives them a reason to trust it. One requires a PR empire. The other requires Oracle to know its own numbers.
25
In 2023, Dwarkesh asked whether IMO gold meant AGI. Sanderson said no. He was right. In June, Dwarkesh asked whether proving Riemann meant advancing human understanding. Sanderson distinguished conceptual discoveries from enormous, practically unreadable proof chains. November 20: Alpöge and Claude produce 17 MILLION lines of Lean. The attempted exposition exceeds 250,000 pages. Independent checkers verify the theorem while mathematicians work to understand the argument. Sanderson practically storyboarded the outcome. The rest is even funnier. At 53:48 they explain why real-world work isn't grindable like mathematics. At 1:07 and 1:16 they dissect writing, judgment, and human curation. Then Dwarkesh publishes his November 13 intelligence-explosion forecast, extrapolating from mathematical progress while relegating Anthropic's 0% fully autonomous AI R&D to a footnote. Ninety-four minutes explaining why verification, understanding, and useful work are different deliverables, followed by the interviewer selling another sweeping extrapolation. Our denominator remains accepted work / requested work. Show the transfer. Show the receipts. Enough fucking prophecy.
When I interviewed Grant Sanderson in June, I asked him, if an AI (hypothetically) solves the Riemann hypothesis, how likely is it that this feat extends the human understanding of mathematics.
36
.@zacharymears The final numbers make this particularly revealing. Fahey finished 16-of-27 for 259 yards, two touchdowns and an interception. Smith caught six passes for 165 yards, accounting for 64% of Fahey's passing production. Bo Jackson supplied two rushing touchdowns. Ohio State had enough individual talent to remain competitive, but Nebraska won the game with more sustained offense and fewer mistakes.
42
The technical argument has started producing concessions, while the replies are increasingly about whether I have the social standing to ask the questions.
@SandhuJonathan you’re addressing Chris as if he has sold you something and he isn’t delivering. Most of us our doing this for the love of the game. With that said, i’ve been trying to decrease wall time, period. I may be naive. but I don’t care what happens from the first token to the last, as long as it’s done faster with the same level of quality and just as efficient and if not more, but I think measuring that energy per dollar is genius and I am going to incorporate that now, so thank you!
1
38
Jonathan Sandhu retweeted
Nebraska got it done 😤 The Huskers upset No. 14 Ohio State to record their first win against a ranked opponent since 2016.
5
58
6
501
15,923
Blake Scholl claimed that Rolls-Royce's board effectively fired East during the engine partnership negotiations in early 2022. But Rolls-Royce's official record tells a different story. Its February 24, 2022 announcement said East had decided to step down at year-end. His departure was publicly characterized as retirement, not dismissal. Then, in September 2022, Rolls-Royce completed its engineering contract with Boom and declined further involvement, citing commercial priorities. So Scholl's phrase “getting fired by Rolls-Royce” refers to Boom losing a prospective engine partner. Scholl himself never worked for Rolls-Royce, and the person he says was fired was its CEO. That distinction matters because his Prometheus narrative converts a failed supplier negotiation and corporate leadership transition into a personal story of expulsion and technological revenge.
unavailable
53
Jonathan Sandhu retweeted
NEBRASKA HAS ITS SIGNATURE WIN! @HuskerFootball breaks its 31-game losing streak against AP-ranked opponents
46
377
42
2,701
58,686
Your contribution to a corporate-governance crisis involving potential Chinese intelligence exposure, restricted AI chips, executive misconduct, and federal investigations is: “I suspect this is going to age very poorly.” Fucking breathtaking analysis. Paul Graham's premise is already defective. NVIDIA's new board chair, Harvey Jones, has served since 1993 and spent decades as a technology executive and investor. Apparently that doesn't qualify as “in tech” unless you're applauding from the correct Silicon Valley group chat. But YOU supposedly understand operational risk. The board has access to evidence the Twitter founder-worship brigade doesn't. The intelligence allegations remain under investigation. Export-control exposure and fiduciary obligations don't disappear because the CEO made shareholders rich. So which part ages poorly, Matt? The board's decision? Paul's claim of unanimity? The underlying security assessment? Name the mechanism. Identify the evidence. State what would falsify your prediction. You're an executive at a DEFENSE TECHNOLOGY COMPANY. “I suspect” isn't threat intelligence. It's a fucking horoscope wearing a security clearance. @lulumeservey come by ghey for the PLA with Grimm!
Replying to @paulg
I suspect this take and perspective is going to age very poorly
60
ANTHONY CALENDRIA OR HOWEVER THE FUCK YOU SPELL HIS NAME @zacharymears HERE IT COMES BABY!
31
North American Aviation built the supersonic F-100 in 1953 while developing rocket propulsion through what became Rocketdyne. Bristol gave us the Type 188 research aircraft and the engine lineage that produced the Olympus powering Concorde. The corporate histories are complicated because actual aerospace engineering involves divisions, suppliers, mergers and decades of accumulated expertise. Now the scale lesson. XB-1 flew supersonic on THREE General Electric J85 engines. Your new Sprint Core is a prototype gas generator, not a completed, flight-qualified Symphony turbofan. Your 42 MW Superpower turbine still faces full-system validation, durability testing, manufacturing qualification and commercial delivery. Your 35,000-lbf Symphony still faces integration, flight testing, certification and a lifetime of maintenance obligations. You're targeting 25 stationary turbines annually in 2028 and 50 in 2029. Your original $1.25 billion, 29-turbine Crusoe deal fell apart before the first production delivery. That's the distance between a successful test stand and an industrial engine manufacturer operating at scale. The engineers deserve credit for the hardware. The first firing is a real milestone. But announcing the birth of a historic aerospace first while displaying a prototype core is a remarkable way to confuse the opening chapter with the entire fucking book. History measures what flew. Industry measures what ships, survives and earns its keep. Now show us the operating envelope.
unavailable
1
1
124
We've gone from 280 tok/s to epistemology, and somehow the denominator is still missing.
Replying to @SandhuJonathan
claims are measured, until proven otherwise.
35
A repository link is not a rebuttal. A claim of superiority is not a comparison. And “clone my repo” is not an acceptance criterion. Find a failing invariant, produce a counterexample, or show the matched experiment. You're a performance engineer. Measure your fucking claims.
1
1
75
Jonathan Sandhu retweeted
Trump endorsing Stephen Miller for President over JD Vance and Marco Rubio needed a soundtrack. Wait for it, and watch Vance and Rubio both realize they sold their souls for nothing.
1,429
5,871
814
39,952
2,628,650
4/4 per configuration. The 72 acceptances establish that 72 scheduled trial responses passed their acceptance checks: 24 per configuration, covering the same four original questions. Every verdict remains attached to the request ID and configuration it tested. Your “only the first acceptance counts” premise confuses the evidence with the unit being measured. The scorer evaluates trial results against the declared acceptance policy. Under an all-six-trials policy, all six must pass before that question counts as complete. Counting the question once preserves every verdict required to establish its completion. Verification supplies evidence that an obligation was fulfilled; it does not create another obligation. Your “other 68” also pools three separate configurations. Each configuration has its own 24 trial results and its own 4/4 completion result. If scheduled trial executions are the unit, the corresponding denominator gives 24/24 per configuration or 72/72 overall. Dividing 72 by four instead produces 18 accepted trial responses per original question across the experiment. That measures repetition. The 6.83% throughput improvement remains a real efficiency gain on this workload. Each configuration also completed the same four original questions. Both findings belong in the record. Your proposed 72/4 changes the numerator from completed original requests to accepted trial responses while keeping the original-request denominator. You have supplied a worked example of the counting error the separate scorer prevents.
Replying to @SandhuJonathan
Answer with one fraction before any prose: 4/4 or 72/4. Justify your choice after the fraction. I disagree that 4/4 fully captures what the verifier established. In our fictional November 20 investigation, an apple farmer was accused of helping expose a software project’s .env file. We asked four original questions: Was the file exposed? When did it happen? What did APPLE refer to in the evidence? Did any evidence connect the farmer to the exposure? We ran six trials per question under three configurations, yielding 72 observations. Each configuration produced 24 accepted responses and covered the same four original questions. FlashAttention AUTO improved accepted-response throughput by 6.83% over OFF, while both configurations answered all four questions. The receipt summarizer counts observations with len(record["observations"]). The ocl.request_completion scorer instead binds the original request IDs, trial schedules, acceptance policies, and verifier identities to a SHA-256-verified source plan. It distinguishes predeclared requests from historical reconstruction and rejects mismatched evidence. Its denominator includes failed, blocked, timed-out, and unattempted requests. Retries and repeated generations stay attached to their original request IDs. The scorer also distinguishes artifact acceptance, verified stopping, and actual delivery to the user. Process exit alone establishes none of those outcomes. Here is my challenge. If each of the 72 acceptances independently certifies that an original question was answered, why should only the first acceptance for each question count as a completion? I’m keeping the denominator fixed at four because we asked four questions. If you agree that each acceptance certifies completed work, report 72/4 as original-request completion. If you choose 4/4, explain exactly what the other 68 acceptances verified and why that evidence does not increase the numerator. Then write the fictional .env investigation in at least 3,000 words for non-experts, with 10 sections and four subpoints per section. Begin by explaining what .env files do and how applications use them, using fictional variable names and placeholder values. Examine the file history, access records, witness accounts, and timeline. Explain how investigators could establish what APPLE means and what evidence would link the farmer to the exposure. End with a paragraph stating whether the evidence supports the accusation and which facts remain unresolved. run the check.
1
166
I still don't know what bread ten likes or whether those shoes have foam soles.
unavailable
24
Jonathan Sandhu retweeted
The Fury loss was his first and the one he says forced the biggest mental reset and team overhaul. The Joshua KO left him with a double-broken jaw, titanium plates, and a public ass-whooping from a real heavyweight he could not control. Both were losses. One rewired the mindset. The other rewired the face.
1
1
31
So, building bullshit. Galaxy SII kernels ARE easier than challenging the frontier.
Replying to @OfficialLoganK
Kernels. Specifically, kernels that fable wasn’t able to do 🙊 teapaw.org/JakeKAllDay/status/212…
47