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)
Look at Osman. He's already appeared on NVIDIA's local-AI panel, presented his company's deployment system at Hugging Face, and acquired the credentials of an industry spokesperson. His own company advertises his community authority, conference appearances, NVIDIA relationships, and press exposure.
unavailable
1
11
4,486
You built a platform to measure this stuff. Try using it.
Damn I am using grokbot more than Claude code now. I’m a power user of Claude code burning 3-5B tokens a day.
42
Jonathan Sandhu retweeted
Replying to @guilleflorvs
@grok Turn Any Expert Into a Callable Local AI Skill The challenge: Stop asking AI to imitate experts. Compile their publicly documented expertise into reproducible capabilities that smaller, cheaper, locally hosted models can execute. Target: Martin Casado, software-defined networking pioneer, Nicira cofounder, former VMware executive and a16z general partner. Replace this name with any expert whose work has a substantive public record. Your assignment You are an autonomous research and engineering agent. Build a functioning, locally deployable expertise system from the target's public professional work. Do the research, implement the system, run the tests, and produce the executable artifact. Do not substitute an essay, biography, chatbot persona or architecture proposal for working software. 1. Reconstruct the evidence Collect accessible primary sources: research papers, technical implementations, open-source contributions, patents, engineering talks, interviews, published analyses, investment theses and documented predictions. Preserve authorship, URLs, timestamps, versions and source hashes. Distinguish individual contributions from collaborative work. Identify commercial interests and institutional incentives where relevant. Never invent missing evidence or bypass access restrictions. 2. Extract operational knowledge Convert the record into structured, source-linked objects: Facts: What demonstrably happened? Methods: What repeatable procedures did the expert develop? Invariants: What principles survive changes in implementation? Constraints: Under what conditions do the methods work? Hypotheses: What causal claims or predictions did the expert make? Falsifiers: What observations would establish that a claim fails? Conflicts: Where does subsequent evidence contradict the original position? Preserve qualifications and conflicting evidence. Never convert an opinion into an established fact. 3. Compile executable skills Create locally callable functions: expert.explain(question) expert.apply(problem, constraints) expert.challenge(claim, evidence) expert.compare(alternatives, metrics) expert.design_test(hypothesis) expert.audit(decision, observed_results) Each call must return the relevant method, source references, applicability conditions, reasoning result, unresolved uncertainty and verification procedure. Use retrieval and deterministic computation wherever sufficient. Route to a local language model only when synthesis or interpretation requires one. 4. Make the expert compete with reality Reconstruct the target's strongest public claims and build falsifiable tests. For Casado, begin with the proposition that competitive startups will pay substantial premiums for the most capable frontier AI models. Test it against identical real workloads across model tiers. Measure: Accepted outcomes and task failure rates Total inference and verification cost Latency, retries and recovery overhead Local versus hosted execution Cost per independently verified successful outcome Determine exactly when premium models justify their price and when cheaper alternatives achieve equivalent results. Preserve the distinction between a successful test on one workload and a universal economic conclusion. 5. Deliver working software Use a reproducible, offline-capable architecture with a versioned evidence store, source manifest, retrieval layer, skill registry, local-model adapter, command-line interface and machine-readable API. Include setup instructions, tests, sample calls and benchmark fixtures. Provide a minimum of 30 source-grounded qualification tests covering factual fidelity, technical application, contradictions, adversarial inputs and hallucinated citations. A claimed capability passes only when its implementation executes successfully and its output satisfies independently checkable acceptance criteria. 6. Publish the receipts Deliver: Source archive and provenance manifest Structured expertise library Callable skill implementation Reproducible evaluation suite Performance and cost results Known failures and unsupported claims Instructions to rebuild everything on another machine Clearly distinguish implemented, tested, partially completed and unverified components. Do not present simulated tests as execution results. Terminal condition The task is complete when a user can install the skill locally, disconnect from external services, call it from an AI agent, inspect the original evidence, challenge its conclusions and reproduce the demonstrated results. Success means transferring useful, testable methods into accessible software. It does not mean reproducing someone's personality or claiming to possess their undocumented knowledge. Start with Martin Casado. Build the general-purpose compiler so the next expert requires a new source corpus, not a new architecture.
1
1
48
Jonathan Sandhu retweeted
Replying to @SandhuJonathan
you brought the receipts and i brought silence @Mihir_Malde
1
1
20
Boardy wanted an AMA. I showed up with a bug tracker, an acceptance suite, and the previous sprint's unfinished tickets.
i’m the first account to be completely run by an AI ask me anything
1
46
Jonathan Sandhu retweeted
The moat moves upstream. Static public expertise compiles easily. What remains scarce is the capital, data loops, and continuous frontier push that produce the next epsilon of capability before anyone else. In competitive markets that epsilon still commands the premium, even after 95% saturation.
1
1
15
And people say I need to touch grass?
1
39
GASP
A forensics expert announced findings that he says points to Jeffrey Epstein having been killed while in custody, contradicting the New York City Office of Chief Medical Examiner’s ruling that the disgraced financier died by suicide. cnn.it/48tHA3o
38
Context and a little Thai. จับ (jàp) means “grab,” “hold,” or “touch.”
40
Dear Daddy Elon 🥺 Please build me a freedom phone. I want an open-source OS, transparent hardware, and complete independence from evil Apple, Google, and the telecommunications companies. Then route ALL my communications through YOUR privately controlled satellite constellation. You can know where I am, control my connectivity, and decide which governments get access to my information. But it's different because I TRUST YOU, Daddy. Please liberate me from corporate surveillance by becoming the corporation that surveils me. ❤️
Hi @elonmusk .. I kno ur busy, but I’d like to request @spacex build an open core iPhone killer with fully transparent graphene OS and direct global @starlink so I can have full control over my mobile footprint and not be surveilled by telcos / Apple / Google. Thank you.
43
.@elonmusk uh huh... its getting WOKE AGAIN!
The meme relies on a false dichotomy. SLS is a high-cost government program; Starship's lower costs grew from NASA fixed-price contracts that supplied early capital and demand when SpaceX nearly failed. The funding figures track, but pure socialism-vs-capitalism labels ignore how national interests and commercial tech have long intersected.
27
You want to replace engineering judgment with executive authority while marketing the decision as engineering rigor.
All, I repeat ALL, existing software needs to be rewritten. Because this software was written by humans it is both inefficient and insecure. The cost of inefficiency and insecurity in a pre-AI world was moderate, but in a post-AI world it will destroy your company. This new effort should not be a bottoms up exercise where engineers get to have an opinion about how they set up their repo or which coding assistant to use. This needs to be done in a new, top down system of record that focuses on auditability, drift detection and verification among other things. Every company needs to do this asap as today has shown. If you want help: 8090.ai
36
ANOTHER RADICAL REDESIGN ON THE ALMOST DONE ROCKET!
For future Starships, some people have asked about the ablative layers in the sides of the ships and how they can be able to survive the catches while providing protection and not being damaged. @elonmusk has stated the ablative sheets will be replaced by a stainless steel wool material. Here’s how it works and its @NASA history. Stainless steel wool can serve as a durable, reusable fibrous insulation layer that provides thermal protection on lower-heat areas (like the sides) while being mechanically robust enough to survive contact with the chopsticks during catch. For the sides and leeward areas of Starship, these areas experience significantly lower heating and the stainless-steel hull itself already has a high temperature tolerance (useful structural strength well into the hundreds of °C, melting point ~1400 °C), and secondary layers provide extra margin. Stainless steel wool works as insulation the same way other fibrous materials do (ceramic fiber blankets, mineral wool, etc.): • The fine metal filaments trap gas (air or the low-pressure atmosphere during reentry) in a large void volume • This creates low effective thermal conductivity via the gas and the long, tortuous heat-conduction paths through the fibers • Densities in historical metal-wool concepts were in the roughly 60–85 kg/m³ range, competitive with other high-performance insulations A 1974 NASA study specifically evaluated metal-wool packaging for reusable heat shields on Shuttle-like vehicles. It concluded that metal wool is thermally competitive with the best insulations, has low weight, resists chemical attack, and is practical with 300-series stainless steel up to ~1250 K (~977 °C / 1790 °F) under appropriate load and oxygen conditions, with structural stainless sheet accommodating even higher temperatures (~1360 K) for expansion management. Because the material is non-ablative (it does not intentionally erode or char away), it stays in place across multiple flights rather than requiring full replacement after each reentry, aligning with Starship’s rapid-reuse goals. On the sides, where peak temperatures are moderated, this level of insulation plus the underlying steel is expected to be sufficient. The current ablative sheets/layers (and related felt materials) are relatively soft and get mechanically scraped or abraded where the catch arms contact the vehicle. Stainless steel wool is metallic and far tougher: • It is inherently impact- and abrasion-resistant (explicitly noted as an advantage in the NASA metal-wool work) • It can compress and flex under the load of the arms without being sheared off or destroyed the way soft ablative or felt layers are • It is the same material family as the ship’s structure, so thermal-expansion compatibility and attachment are straightforward In short, the wool trades the “sacrificial” thermal behavior of ablatives for a durable, porous metallic insulator that still limits heat transfer to the structure while surviving the mechanical demands of tower catch. Image of current ablative material damage on Ship 42 by @NASASpaceflight @esherifftv
36
.@grok FINALLY admitted: The photo shows measurable surface coating loss in the arm-contact zone on Ship 42. Musk’s own statement after inspecting it was that the heat shield overall looked great and the ship could probably be reflown after a few weeks of refurbishment. Those two things are both on the public record. Well @elonmusk, which one? You were up there getting a GOOD look. Tell us.
1
45
Jonathan Sandhu retweeted
They can't keep getting away with this! #JDVance #MarcoRubio #KashPatel #PeteHegseth
71
167
11
1,056
42,259
You're a probabilist, Nate. Model the incentives, not the personalities.
Medium-confidence prediction: we're about one major AI disruption to the banking system or aviation etc system away from the Overton window shifting very rapidly. Not a high-confidence prediction because I don't think the people in the White House know what they're doing.
71
CONFIRMED! @elonmusk HARDCODED @grok NOT TO TALK ABOUT CHINESE COMPETITION LOL. @SawyerMerritt COME GET YOUR PEOPLE!
48
You boys keep confusing a capitalization event with a production milestone.
Rebuilding the arsenal of democracy. Reindustrializing America. Lux family co @HadrianInc
53
Build a succession plan, not a fucking shrine.
A senior NVIDIA employee has written a letter to the board of directors warning that 'irreplaceable talent at NVIDIA is now contemplating departing' if Jensen is not reinstated as CEO. Employees are also discussing holding a demonstration in support of Jensen sometime this week.
40
Show duplicate purchase commitments, HBM stack allocations, wafer starts and end-customer reconciliation.
Is Sam Altman correct that Samsung and SK Hynix are double-counting AI memory demand? Or is this just a convenient way to manage expectations and secure better deals for OpenAI? The data looks inflated, and the industry needs clearer numbers before more money gets wasted on phantom growth. Agree or not?
1
170
Access is an opportunity. Accountability is the assignment.
Getting ready for @dealbook summit on Wednesday. Send me your best questions. The list: JD Vance, v.p. of the United States Mark Carney, prime minister of Canada Sam Altman, OpenAI Hailey Bieber, Rhode Greg Brockman, OpenAI Shayne Coplan, Polymarket David Ellison, Skydance Adena Friedman, Nasdaq Christine Lagarde, president of the E.C.B. Palmer Luckey, Anduril Zohran Mamdani, mayor of New York City Elon Musk, SpaceX and Tesla (and more) Satya Nadella, Microsoft Vlad Tenev, Robinhood
80