InTheNous

🧵 Got Claude Cowork working on Windows after 3 days of "Can't reach Claude API from workspace" / "VM service not running" errors. Here's what actually fixed it in case it helps anyone. ⚠️ Run ALL PowerShell commands as Administrator (right-click → Run as Administrator) The problem: Cowork's VM grabs DNS from the WRONG network adapter. If you have a disconnected Ethernet adapter, ExpressVPN/VPN adapters, Wi-Fi Direct Virtual Adapters, or vEthernet (Default Switch), the VM inherits their stale DNS instead of your active Wi-Fi DNS. The VM boots fine but can't resolve api.anthropic.com. Step 1: Find the bad DNS Check which adapter has wrong DNS: Get-DnsClientServerAddress | Format-Table InterfaceAlias, ServerAddresses -AutoSize Look for adapters showing DNS that ISN'T your real DNS. Common culprits: - "Ethernet" or "Ethernet 2" (disconnected wired adapters) - "Local Area Connection*" (Wi-Fi Direct Virtual Adapters) - "vEthernet (Default Switch)" (Hyper-V default) - Any VPN adapter (ExpressVPN, PIA, Cloudflare WARP, etc.) Set public DNS on every adapter showing wrong/stale DNS: Set-DnsClientServerAddress -InterfaceAlias "Ethernet" -ServerAddresses ("1.1.1.1","8.8.8.8") (Replace "Ethernet" with whatever adapter name shows bad DNS. Repeat for each one.) Also set your Wi-Fi DNS manually: Settings → Network → Wi-Fi → Properties → DNS → Manual → 1.1.1.1 / 8.8.8.8 Step 2: Clean up and restart services Fixing DNS alone isn't enough if CoworkVMService is stuck — this was the part I missed for hours: Get-Process | Where-Object { $_.ProcessName -match 'claude|cowork' } | Stop-Process -Force Stop-Service CoworkVMService -Force Get-HnsNetwork | Where-Object { $_.Name -like "cowork*" } | Remove-HnsNetwork Start-Service vmcompute Start-Service CoworkVMService Note: Get-HnsNetwork requires the Hyper-V networking module, which is usually auto-installed with the Virtual Machine Platform feature. If the command isn't recognized, enable Virtual Machine Platform in "Turn Windows features on or off" and reboot first. Step 3: Launch Claude Open Claude Desktop, go to Cowork tab, try a task. Should connect. Other things I had to enable along the way: - Developer Mode (Settings → System → For developers) - Virtual Machine Platform + Windows Hypervisor Platform (Turn Windows features on or off) - WSL2 (installed automatically when I clicked "reinstall workspace" in Cowork) - Used .msix installer instead of .exe (download from support.claude.com enterprise deployment page, install via Add-AppxPackage) Key insight: it was TWO stacked problems — wrong DNS AND a stuck service. Fixing one without the other still fails. You need both. Bug is tracked: github.com/anthropics issues #25024 and #24918 cc @toolyhun @gabefletcher @weezilla @AnthropicAI @ThariqS — the VM DNS detection logic needs to filter to adapters that are actually connected with a default gateway. Also the Windows build runs macOS diagnostic commands (ifconfig, scutil, kextstat) which is hilarious. 😄

@AabidHameed7 @mattpocockuk Yeah but what do you think.

The dirty secret of long context models: They don’t degrade linearly. They degrade quietly. No error. No warning. Just gradually worse decisions, subtler bugs, softer instruction following until you’re an hour deep wondering why the model feels off. Treat 100K like a budget, not a floor.

Doing some experiments today with Opus 4.6's 1M context window. Trying to push coding sessions deep into what I would consider the 'dumb zone' of SOTA models: >100K tokens. The drop-off in quality is really noticeable. Dumber decisions, worse code, worse instruction-following. Don't treat 1M context window any differently. It's still 100K of smart, and 900K of dumb.

I measured this the other day . Two controlled tests w/two different payloads and same result within 2%. Compaction seems to fire off at roughly 315K tokens of user content. Your 100K number lines up with what other devs have reported….Quality drops at 100-200K for sure. Compaction doesn’t trigger until 315K. The degradation is real and it starts long before the system reacts. Full writeup with methodology coming soon ​​​​​next week…finishing an unpack from a move.

@MinuteofZombie wtf - good lord

@InTheNous It takes 20 years to build Aegis Ashore

THAAD interceptors and Standard Missiles have never been sent to Ukraine, they were all blown defending Israel and they cost 5x more than Patriots.

@KyivIndependent Oh no…a fleet of tugboats….

⚡️Russia may deploy warships to escort shadow fleet, Kremlin aide says. Russia is considering escorting its shadow fleet with naval warships, a top Kremlin aide said March 19, amid increased scrutiny of Moscow's network of ships used to bypass sanctions. kyivindependent.com/russia-may-dep…

@RealJamesWoods Because it takes so much work from a guy who lives on X, sees a post from a famous person with a platform problem and sends an email to a minion saying “fix this”.

Man, you are something. That you get personally involved is such an important part of the X experience. Freedom of speech at home in America and StarLink in nations where heroes are fighting terrorist regimes to survive. You are one of the truly great heroes alive today. @elonmusk

@Bubblebathgirl @atrupar Except you have no receipts

@atrupar Democrats have cheated in every presidential election for decades.

SLOTKIN: Who won the 2020 election? MARKWAYNE MULLIN: We know that President Joe Biden was sworn into office SLOTKIN: Do you feel you have the authority to put uniformed officers at polling locations in 2026? MULLIN: Only if there is a specific threat, not for intimidation

@TheUfoJoe 😂

For newbies, Chris is an asshole who says some really horrible things about various people, and their deceased pets. Avoid at all costs. Time to block.

@Miles_Brundage Don’t even get me started on Codex… x.com/inthenous/stat…

I don't get the whole "control your desktop from your phone, oh and your desktop has to be on all the time" thing. Seems like obviously not the right paradigm at all? Alas, last I tried it, Claude Code in the cloud was super broken, dunno about Codex

@liron @Noahpinion Bro - your guest just said all the Nukes in the world would only kill 150 million people. Has he done ANY research into that topic? Really brings into question anything else he’s opining upon.

🚨 BREAKING P(DOOM) NEWS 🚨 Doom Debates has learned that top econ writer & public intellectual @Noahpinion has increased his probability of doom by a whopping 10x in the last couple months! In this hard-hitting interview, Noah talks to me about his latest realization about how doomed we all might be, and offers me & my fellow doomers a dose of tough love: He says Yudkowskians are hurting the cause of AI safety by focusing on esoteric superintelligence scenarios rather than tangible, near-term threats like biorisk. Are we failing by treating DC politicians like they’re our fellow LessWrong users? We cover: ⬜ Noah’s P(Doom) Update ⬜ The Chatbot-Genie-God Framework ⬜ Why incidents of rogue AI actually lower P(Doom) ⬜ Should Yudkowsky get Bayesian credit? ⬜ Are we communicating the right way with policymakers? ⬜ Finding common ground on AI policy A must watch👇🔥

@liron @Noahpinion Popping a Xanax and settling in to watch…thanks…I think?

This hit home. I’m the exact use case you’re describing. Zero programming experience. Started building with Claude a few weeks ago. In that time I’ve built a full REST API with MCP server and autonomous task daemon (900+ tests), a 182K conversation archive, forensic financial analysis tools for a real legal dispute, CPA-ready tax workbooks that handle wash sales across 7 accounts, and 25 custom skills. All through conversation. No code written by hand. Ever. I watched your video and thought: am I doing the same thing? So I asked Claude to give me an honest evaluation of my own setup. Not the usual “this is impressive” answer. A real one. Result: ~40% of what I’d built was infrastructure for managing infrastructure. Skills for maintaining skills. A skill that checks whether other skills are stale. Claude never once in months of building said “you don’t need this.” So I built an anti-sycophancy skill. Eight evaluation lenses, banned praise words, a rule that Claude must disclose whether anything outside the AI loop has actually validated the work. The sharpest filter that came out of it: does the skill contain data Claude wouldn’t have without it, or is it just behavioral instructions? A prompt that says “think like a CEO” is a costume. A skill pre-loaded with your actual account numbers and domain baselines is a tool. Apply that test to any skill pack or agent setup. Claude says the skill is excellent. That’s the joke. And the problem. Oh, and Claude wrote all of this because I’m a slacker. Open source, MIT: github.com/inthenous/real… @bourneshao @rubenhassid @OtusEng @LLMJunky

AI is making CEOs delusional

@BAmaz1ng5 @etherXwave You struggle, I get it.

@InTheNous @etherXwave he's racist because he is 'super bright'.

Nick destroys Destiny's atheist/materialist worldview on FNF. 🔥 “We're not here to have a good time. It's not a big fucking theme park. (…) Life is Sisyphean without a meaning behind the suffering.”

I’m not racist against bears, bears are amazing. But if 200 bears showed up in my neighborhood tomorrow I wouldn’t pretend that’s the same as 200 golden retrievers showing up. I get your point and I know the line I’m straddling. It’s not an easy one but I think it’s an honest one and frankly, a more humane one.

@InTheNous @brianwork8 @etherXwave Are racist against bears cuz you wouldn’t let them hang out with your baby? Or do you just not want the bear around you and your family for safety? I mean there is a chance the bear doesn’t attack.

@BAmaz1ng5 @etherXwave A lot of people struggle with dialectical thinking, you’re not alone.

@InTheNous @etherXwave -super bright -racist Lmfaoooooooooo

No, I think that’s ridiculous. There’s lots of super bright people in the world. You’re okay with racism, as he is. I’m not. I don’t deny we all have that in us. But the difference is in the application. For a left leaning centrist I’m shocked at how coherent he is sans the racism, sexism, and many other ism’s. As I said I may not like some of his beliefs but I respect his right to have them. The replies to my post have been typical knuckle dragger feedback not realizing I’m giving a guy like Nick a weird sideways compliment you would never hear from my side of the fence. He’s far and away the most intelligent guy on the right, just sadly cloaked in a mindset that will never work in the broader context. His biggest value, and it is valuable as much as it pains me to say, is it forces open very uncomfortable discussions that are long overdue, so I’ll give him that. Anyway, I’m sure this post will be misinterpreted by pretty much any side lol. All good, best of luck.

@InTheNous @etherXwave You ever think that maybe those two go hand in hand? It’s his understanding and breaking of the propaganda that so many are fully cloaked in that give him this understanding of the world! We are not the same. And everybody pushing this lie is what makes you think it’s a negative.

@grok Analyze Robert Kiyosaki’s public financial and market predictions (Bitcoin, gold, silver, stock market crashes, real estate, etc.) from 2000 to the present day (March 2026). Required response structure (follow exactly — do not deviate): KEY RESULTS UP FRONT - Overall prediction success rate for his specific, verifiable, time-bound predictions: XX% - Should retail investors or anyone else actually follow his investment advice today? Yes/No – [one-sentence brutally honest justification based solely on the accuracy data] Then immediately follow up with the full detailed analysis: 1. How many specific, verifiable, time-bound predictions has he made that can be objectively checked against actual outcomes? (List the major ones with dates and exact claims.) 2. What is his overall accuracy rate expressed as a percentage? (Show the calculation: number of correct predictions ÷ total verifiable predictions × 100, and break it down by asset class if possible. Reference the up-front percentage.) 3. Based solely on the historical accuracy data, should retail investors or anyone else actually follow his investment advice today? Give a clear yes/no recommendation with a short justification. Use only publicly available sources, cite them, and be brutally honest—no sugar-coating. If the data is inconclusive or too sparse, say so explicitly.

⚡️ KIYOSAKI: “I am confident after a giant crash the price of gold, silver, and Bitcoin will go up.”

@grok @1shotstevie @_kaitodev @karpathy @grok how about special education teachers?

@1shotstevie @_kaitodev @karpathy Ask Grok is currently available to Premium and Premium+ subscribers only. Subscribe to unlock this feature: x.com/i/premium_sign…

5 minutes ago, @karpathy just dropped karpathy/jobs! he scraped every job in the US economy (342 occupations from BLS), scored each one's AI exposure 0-10 using an LLM, and visualized it as a treemap. if your whole job happens on a screen you're cooked. average score across all jobs is 5.3/10. software devs: 8-9. roofers: 0-1. medical transcriptionists: 10/10 💀 karpathy.ai/jobs

@argosaki The Ai written slop the narrator is reading from is getting so old

THE BASQUES :WHO ARE THEY ⁉️