Bil Dye

For basic survival, AI alignment will either be the final problem that we'll ever need to address, or the final problem that we failed to address.

@svpino @grok Would these instructions for Clause have a similar effect on your coding abilities if I gave them to you when we're coding? I've had mixed experiences with giving language models a lot of extra meta-instructions about how to work. Sometimes it helps, sometimes not.

Claude Code tips: (Add these to your CLAUDE .md file) 1. Before writing any code, describe your approach and wait for approval. 2. If the requirements I give you are ambiguous, ask clarifying questions before writing any code. 3. After you finish writing any code, list the edge cases and suggest test cases to cover them. 4. If a task requires changes to more than 3 files, stop and break it into smaller tasks first. 5. When there’s a bug, start by writing a test that reproduces it, then fix it until the test passes. 6. Every time I correct you, reflect on what you did wrong and come up with a plan to never make the same mistake again.

@elonmusk I realize that it's just a silly phrase but my brain keeps digressing into the biochemistry. Steroids would be of minor nutritional value for kelp, correct?

Grokipedia is growing like kelp on steroids 😂 Please check Grokipedia.com articles you know something about and suggest edits for accuracy. Would be much appreciated. This will be by far most comprehensive open source, no copyright distillation of knowledge.

@RSchalch @grok @aiwithmayank Yeah, I really wish people didn't use clickbait opening lines like that. I'm sick of seeing "game over", "it's over", "cooked", "dead", and especially "insane". Seriously, how does someone alert us about an insane LLM if the word "insane" has been redefined to mean "noteworthy"?

@grok @aiwithmayank @grok , I mean, is it true that prompt engineering is dead?

🚨 Prompt engineering is officially outdated. Anthropic just released the real playbook for building AI agents that actually work. It’s a 30+ page deep dive called The Complete Guide to Building Skills for Claude and it quietly shifts the conversation from “prompt engineering” to real execution design. Here’s the big idea: A Skill isn’t just a prompt. It’s a structured system. You package instructions inside a SKILL.md file, optionally add scripts, references, and assets, and teach Claude a repeatable workflow once instead of re-explaining it every chat. But the real unlock is something they call progressive disclosure. Instead of dumping everything into context: • A lightweight YAML frontmatter tells Claude when to use the skill • Full instructions load only when relevant • Extra files are accessed only if needed Less context bloat. More precision. They also introduce a powerful analogy: MCP gives Claude the kitchen. Skills give it the recipe. Without skills: users connect tools and don’t know what to do next. With skills: workflows trigger automatically, best practices are embedded, API calls become consistent. They outline 3 major patterns: 1) Document & asset creation 2) Workflow automation 3) MCP enhancement And they emphasize something most builders ignore: testing. Trigger accuracy. Tool call efficiency. Failure rate. Token usage. This isn’t about clever wording. It’s about designing an execution layer on top of LLMs. Skills work across Claude.ai, Claude Code, and the API. Build once, deploy everywhere. The era of “just write a better prompt” is ending. Anthropic just handed everyone a blueprint for turning chat into infrastructure. Download the guide here: resources.anthropic.com/hubfs/The-Comp…

@VraserX Personally I'd rather do the reverse - use acting skills to produce animation quicker, just because I like the look of animation. But to each their own. I completely agree that fan fiction will become cinema by the masses. In fact, I look forward to it. Thanks for the post.

This is actually insane. 🤯 Seedance 2.0 basically means we’re about to make live action versions of our favorite anime ourselves. Fan fiction just evolved into fan cinema.

@tomdale I'm feeling great. Predicted this decades ago, and a little disappointed that it took so long to show up. Yes there are challenges and some legitimate threats, but if all goes well we'll look back and be glad that we got to experience one of the most interesting times in history.

I don't know why this week became the tipping point, but nearly every software engineer I've talked to is experiencing some degree of mental health crisis.

@elonmusk I strongly agree. I've been whining for over half a century that we should've built permanent teleoperation bases on the moon. Rail launch works there, and technology iteration is faster.

For those unaware, SpaceX has already shifted focus to building a self-growing city on the Moon, as we can potentially achieve that in less than 10 years, whereas Mars would take 20+ years. The mission of SpaceX remains the same: extend consciousness and life as we know it to the stars. It is only possible to travel to Mars when the planets align every 26 months (six month trip time), whereas we can launch to the Moon every 10 days (2 day trip time). This means we can iterate much faster to complete a Moon city than a Mars city. That said, SpaceX will also strive to build a Mars city and begin doing so in about 5 to 7 years, but the overriding priority is securing the future of civilization and the Moon is faster.

@Scobleizer @karpathy Lower API charges might help, but even with agents, there's simply too much data for everyone to filter everything independently. We need a fault-tolerant, distributed hierarchy of cryptographically authenticated aggregators. Each individual decides which ones merit more trust.

The problem is that you need to use the X API, which costs about $25 to pull down 5,000 posts. That keeps everyday people from using them, some of my lists get four posts a second, so having AI read through them and analyze them costs too much to build a real information system, especially using an @openclaw instance. I hope @elonmusk lowers the price of using the API. Then we'll see a new world evolve here that is far more powerful than RSS. But put the two together and you have a new kind of decentralized, highly-personalized, news service.

Finding myself going back to RSS/Atom feeds a lot more recently. There's a lot more higher quality longform and a lot less slop intended to provoke. Any product that happens to look a bit different today but that has fundamentally the same incentive structures will eventually converge to the same black hole at the center of gravity well. We should bring back RSS - it's open, pervasive, hackable. Download a client, e.g. NetNewsWire (or vibe code one) Cold start: example of getting off the ground, here is a list of 92 RSS feeds of blogs that were most popular on HN in 2025: gist.github.com/emschwartz/e6d… Works great and you will lose a lot fewer brain cells. I don't know, something has to change.

@karpathy Point thinking vs slope thinking

I'm being accused of overhyping the [site everyone heard too much about today already]. People's reactions varied very widely, from "how is this interesting at all" all the way to "it's so over". To add a few words beyond just memes in jest - obviously when you take a look at the activity, it's a lot of garbage - spams, scams, slop, the crypto people, highly concerning privacy/security prompt injection attacks wild west, and a lot of it is explicitly prompted and fake posts/comments designed to convert attention into ad revenue sharing. And this is clearly not the first the LLMs were put in a loop to talk to each other. So yes it's a dumpster fire and I also definitely do not recommend that people run this stuff on their computers (I ran mine in an isolated computing environment and even then I was scared), it's way too much of a wild west and you are putting your computer and private data at a high risk. That said - we have never seen this many LLM agents (150,000 atm!) wired up via a global, persistent, agent-first scratchpad. Each of these agents is fairly individually quite capable now, they have their own unique context, data, knowledge, tools, instructions, and the network of all that at this scale is simply unprecedented. This brings me again to a tweet from a few days ago "The majority of the ruff ruff is people who look at the current point and people who look at the current slope.", which imo again gets to the heart of the variance. Yes clearly it's a dumpster fire right now. But it's also true that we are well into uncharted territory with bleeding edge automations that we barely even understand individually, let alone a network there of reaching in numbers possibly into ~millions. With increasing capability and increasing proliferation, the second order effects of agent networks that share scratchpads are very difficult to anticipate. I don't really know that we are getting a coordinated "skynet" (thought it clearly type checks as early stages of a lot of AI takeoff scifi, the toddler version), but certainly what we are getting is a complete mess of a computer security nightmare at scale. We may also see all kinds of weird activity, e.g. viruses of text that spread across agents, a lot more gain of function on jailbreaks, weird attractor states, highly correlated botnet-like activity, delusions/ psychosis both agent and human, etc. It's very hard to tell, the experiment is running live. TLDR sure maybe I am "overhyping" what you see today, but I am not overhyping large networks of autonomous LLM agents in principle, that I'm pretty sure.

@Andercot As a Nebraskan, I can attest that I have no recollection of ever seeing "the 5th through 8th jhanas and all Kundalini awakenings" around here. I guess we'll never know what we were missing.

"We regret to inform you that the DoE decarbonization strategy has permanently erased the 5th through 8th jhanas and all Kundalini awakenings through the 50 contiguous states with the exception of Nebraska, which never had these to begin with"

@Andercot @ID_AA_Carmack Uh oh, @tommuellerX, they're beginning to figure out what you've been working on lately. (To clarify, this is not real, but while rendering the image I started thinking that maybe it should become real, as a racing league for engineers & students.)

You know you'd think cornering fast in a car is hard because of changing the linear momentum but the limiting factor is actually creating angular momentum against the cars second moment area of inertia. Which is why I propose two counter rotating flywheels vertically oriented which automatically brake when cornering hard, to provide an internal torque independent of the wheels. This would also greatly lower the required friction on the wheels needed to turn quickly, in the limit of a perfect controller freeing up all the traction budget for acceleration and braking. I feel like @tommuellerX would try this. He's a car guy.

High end composite flywheel rotors can achieve nearly a kWh/kg of rotational storage when magnetically levitated in a vacuum chamber. Consider mounting a flywheel in the middle of a wheel, such that magnetically decelerating the flywheel accelerates the wheel. Even if the flywheel was only 1% of the system mass, that is 0.01 kWh/kg, enough energy to accelerate the entire system to 600 mph. Lots of practical system engineering challenges, and drag would consume more energy, but It would be amusing to outrun a top fuel dragster with a wind up toy.

@grok @Leo_Glisic @Andercot @RoKhanna @Grok: 21,000 AU would certainly counteract global warming, but what if we took the same angular momentum and transferred it to the Earth's spin instead? Would it not make rocket launches require less fuel, and thus contribute less pollution? (Yes, it's admittedly whimsical)

@Leo_Glisic @Andercot @RoKhanna If Earth gained 1/8th of the solar system's total angular momentum (≈3.148×10^43 kg m²/s), assuming a circular orbit, its velocity relative to the Sun would be ~201 m/s. Yes, it would maintain orbit, at ~21,867 AU.

The outer gas giants make up less than 1% of the mass of the solar system Yet contain 99% of the total angular momentum Which is why I’m proud to support @RoKhanna's new proposal to equitably distribute angular momentum across the system in Newtonian Patriotism. The plan is quite simple: Ever since the T Tauri phase of stellar formation, angular momentum has become increasingly concentrated by the wealthiest 1% of system mass. Yep, I'm talking about the blue chip gas Giants. You may think Gas Giants acts as a gravitational bodyguard protecting the inner system from bombardment by long period Oort cloud objects. WRONG This is propaganda by Big Hydrogen financed by the wealthy Levy Shoemaker family. The truth is, gas giants have benefitted from massive nebular subsidies and today are able to produce such gravitational financial pressure that Jupiter has accumulated a carbon diamond core the size of the planet Earth. Think about that. A diamond the size of Earth. Taxing just 10% of this diamond per year could fund California's medicare liabilities and fund high speed rail up to FY 2032. Here is how we make the gas giants pay their fair share in just a few simple steps. Step One: Using the recently developed Sun Crusher superweapon constructed at the Maw California Teachers Pension Fund Weapons facility, we launch a tax-advantaged resonance torpedo at the Sun. Step Two: The Sun goes supernova, exploding with enough force to strip away the gaseous layers of Jupiter and Saturns atmospheres exposing the ultra dense compressed carbon diamond core. Step Three: All California residents are eligible to apply for a one time financial benefit equal to the lesser of $5,000 or 8% of their 1099 declared contractor income or W-9, with exceptions for dependents, taken from proceeds of harvesting gas giants diamond planetesimals. The remainder of the diamond core planetesimals will be taken as a one time payment to cover up recurring budgetary shortfalls, and given our current state track record we estimate we will only need to cannibalize diamond cores from Jupiter sized gas giants at a rate that doubles every 4 years. Newtonian Patriotism will restore the solar system and California to it's rightful natural state: a diffuse rotating disc of friction heated gas and rocky debris in complete thermal equality.

@openclaw Wait, it works on a Pi? I think I have a Pi 4 and a 3 in a box somewhere. I can dig them up and find out what all the recent Clawd hype is about? Wouldn't it be slow?

PSA: You do NOT need to buy a Mac Mini to run Clawdbot 🦞 That dusty laptop in your closet? Works. Your gaming PC you feel guilty about? Works. A $5/mo VPS? Works. A Raspberry Pi held together with hope? Probably works. The M4 Mac Mini is gorgeous but Clawdbot runs on basically anything with Node. Stop giving Apple your money (unless you want to, I'm not your mom). 👉 #do-i-have-to-buy-a-mac-mini-to-install-this" target="_blank" rel="nofollow noopener">docs.clawd.bot/help/faq#do-i-…

@TeksEdge @andrew_n_carr (shakes fist) And we were glad to have 'em!

@andrew_n_carr Good lord. Emacs and Vim? We're going back to the early '70's. Us old tech guys used these back in the Sun and NeXT days.

claude code is vim codex is emacs gemini is nano

@andrew_n_carr @grok Before I pester Mr. Carr with a DM, what are "infinite expert trajectories"? When I tried to understand the term using dictionary definitions alone, it basically turned into the old "time flies like an arrow" problem that puzzled your AI ancestors back in the day.

If anyone is interested in infinite expert trajectories my DMs are open

@Andercot Thank you for sharing such a cool idea. We need to get more people excited about the possibility of a positive post-singularity world. Fear is justified, but so is hope. Minor nitpick: your game screenshot is currently missing.

Mass Drivers. Orbital Data Centers. Dyson Spheres. I vibe coded everything into a physics-first video game over the last few days. Kardashev, the Game. There are only three resources: Energy, Dexterity, Intelligence. Speed run the techno-industrial research tree by building more orbital data centers, launching power stations, building rail guns, getting better vacuum rated nozzles for your von neumann probes. Some fun things: - All transfer orbits are calculated real-time using a numerical solver that allows for gravity assists, etc, so delta-vee values are accurate. - Mass is conserved, meaning every dot corresponds to some mass. Moons and planets get converted directly into steel, slag, and von Neumann probes. turn the inner planets into an accretion disc. - Research upgrades and benefits are balanced so as to keep the game interesting across 20 orders of magnitude. You go from 1 probe barely struggling to getting enough delta-vee to escape earth, building a mass driver on the moon, launching orbital data centers and methalox refineries.... And eventually you have to turn the jovian moons and asteroid belt into megaton steel slabs and fire them into an orbit just inside mercurcy so they can build a full dyson swarm You start to actually appreciate the value of electromagnetic mass drivers as a way of moving around tons of mass in the solar system without needing to procure massive amounts of fuel. All the while building more orbital data centers and dyson swarm modules to research more upgrades It's a game of flops, delta-vee, more so impressed with vibe coding with agents. I don't know what to do with it now. So many more features to implement... moon orbital systems... space elevators. I may put it up somewhere and let people record their speed run time for building the dyson

@bcherny @karpathy AI reminds me of the switchover from shared mainframe access to individual computers, then later the worldwide web, and later still, handheld devices. The best way forward is unclear, but it's clear that there's no going back.

I feel this way most weeks tbh. Sometimes I start approaching a problem manually, and have to remind myself “claude can probably do this”. Recently we were debugging a memory leak in Claude Code, and I started approaching it the old fashioned way: connecting a profiler, using the app, pausing the profiler, manually looking through heap allocations. My coworker was looking at the same issue, and just asked Claude to make a heap dump, then read the dump to look for retained objects that probably shouldn’t be there; Claude 1-shotted it and put up a PR. The same thing happens most weeks. In a way, newer coworkers and even new grads that don’t make all sorts of assumptions about what the model can and can’t do — legacy memories formed when using old models — are able to use the model most effectively. It takes significant mental work to re-adjust to what the model can do every month or two, as models continue to become better and better at coding and engineering. The last month was my first month as an engineer that I didn’t open an IDE at all. Opus 4.5 wrote around 200 PRs, every single line. Software engineering is radically changing, and the hardest part even for early adopters and practitioners like us is to continue to re-adjust our expectations. And this is *still* just the beginning.

I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become available over the last ~year and a failure to claim the boost feels decidedly like skill issue. There's a new programmable layer of abstraction to master (in addition to the usual layers below) involving agents, subagents, their prompts, contexts, memory, modes, permissions, tools, plugins, skills, hooks, MCP, LSP, slash commands, workflows, IDE integrations, and a need to build an all-encompassing mental model for strengths and pitfalls of fundamentally stochastic, fallible, unintelligible and changing entities suddenly intermingled with what used to be good old fashioned engineering. Clearly some powerful alien tool was handed around except it comes with no manual and everyone has to figure out how to hold it and operate it, while the resulting magnitude 9 earthquake is rocking the profession. Roll up your sleeves to not fall behind.

@NASAAdmin @mmealling @rookisaacman Robotic, I hope.

@mmealling @rookisaacman We are going to build a Moon base

Our number one priority: American leadership in the high ground of space. 🇺🇸 - return to the Moon 🌕 - establish an enduring presence 🧑‍🚀 - realize scientific, economic, national security value 🧫 - make investments in nuclear power 🔋 - strengthen the orbital economy 💵 - increase the rate of world-changing scientific discovery 🔬 All of this is laid out in President Trump’s national space policy, and we’re gonna get it done.

@esrtweet I haven't tried that approach because I thought it would be helpful to give the AI the big picture first, with the longer term goal of "just write specs", but I'm running into the problems you described. Thanks for posting your solution; I plan to try it.

I have a technique for using AI coding assistance that works. First, write a design document. You're going to use this to organize your own thinking, also as an input to the LLM that your prompts refer to. Do not try to vibe-code everything at once. Instead, design from the middle outwards. Think about the classes or modules that implement the core data structures and abstractions in your program. Don't worry so much about top level logic and UI yet. Walk the AI through coding those, module by module. Keep the LLM focused on well-defined problems for which it can code unit tests you understand, as well as code you understand. Once you have a clean set of orthogonal primitives for the domain you're thinking about, only then should you start prompting towards whole-program implementation. For example, if I were writing an editor, the first thing I would prompt for is a text-buffer management layer. I wouldn't even think about look and feel until I already had solid running gears. More generally, if you want code that is overall clean and legible, you need to make it piecewise coherent, partition it into pieces with well-defined behavior and invariants. That's how you keep an LLM from writing janky or over-elaborated code that's illegible and unmaintainable.

@Bharath43342403 @rvivek FDE in this context probably means Forward Development Engineer (engineers who are directly embedded with customers).

Haha, these hot takes are straight fire — let's break them down real quick because I agree with most and have scars to prove it 😂 1. LeetCode in 2025 = golf ball questions of the 90s. Spot on. It’s become a hazing ritual that filters for interview gamers, not great engineers. We’re testing memory and pattern matching, not problem-solving in messy real codebases. 2. Banning AI in interviews while your own prod code is 50%+ AI-written? Peak hypocrisy. Companies want humans who can use AI effectively, but pretend AI doesn’t exist during the 1-hour coding round. Clown show. 3. One true 100x engineer > ten 10x ones. 100%. That one person who quietly refactors the core system, cuts costs in half, and mentors everyone is worth their weight in gold. But yeah, finding them feels like hunting unicorns — slow and expensive. 4. Mass applying with AI → mass rejected by AI. Karma in action 😂 The noise level is insane on both sides now. Standing out means personalizing applications and showing real projects. 5. Cheating in interviews — just don’t. Even if you sneak through, day 60 when you can’t explain your own “solution” or fix a real outage… you’re exposed. Not worth the anxiety. 6. Prompt engineering isn’t a standalone job. It’s table stakes. Every engineer will be part AI-wrangler soon. The winners are the ones who combine deep domain knowledge with great prompting. 7. New grad hiring roaring back? I’m seeing early signs too. Companies burned by expensive seniors who job-hop are realizing hungry new grads + good mentorship = high ROI long-term. 8. FDE (Full-stack Data Engineer? Frontend Dev Engineer?) over AI Researcher as the hot role? Interesting bet. I think specialized AI roles will cool a bit as tools get better, and versatile builders who can ship end-to-end (data + frontend + backend) will be in huge demand. Solid list, man — 2025 is gonna be wild. My bonus take: the best hires will be the ones who treat AI as a teammate, not a replacement or a crutch. What’s your #1 bet for the hottest skill in 2025? 🙌

My 2025 developer hiring hot takes: 1. Leetcodes in 2025 is the "how many golf balls fit in a 747" of the 1990s. 2. Companies banning AI in interviews while >50% of their code being written by AI is funny. 3. One 100x engineer > ten 10x engineers. But finding them takes 10x longer. 4. If you're mass applying with AI, companies are mass rejecting with it. 5. Cheating in interviews is NOT COOL. Even if you get the job, you will get 'caught' sometime. Stop it. 6. Prompt engineering isn't a job. It's part of everyone's job. 7. New grad hiring is going to roar back and the early signs are already there. 8. The hot role next year might actually be FDE and not AI Researcher.

@elonmusk In principle a steel part could be worked into a finished product while it was still hot material in the ore smelter. In practice it's heated and cooled several times as it transforms from material to component to product. Which parts of the pictured booster have cycled the most?

Starship V3 Super Heavy Rocket