Ahsen

i'm also getting dms to apply for @ycombinator and disrupt the cloud sandbox/agent market! 🫣

If we’re still spending our days reviewing PRs a year from now, we’ve missed the point. I don’t want 30 agents handing me 30 piles of code to read. I want them to finish the work and show me it works. That’s where we’re going with the cloud computers we’re building for AI agents. Each task gets its own computer. Claude, Grok, whatever agent we’re running controls it, does the work, tests it, and brings back videos and screenshots. Show me what changed. Show me what broke during testing, how it was fixed, and the finished thing working. I watch it. If something’s wrong, it goes back to the agent. We still decide how we want the system built: the architecture, the constraints, the preferences. The agents need that context. But we shouldn’t have to read every line they write to get the result we asked for. We’ve been testing the computers underneath this: → New computer in ~10 milliseconds → Resume a paused one in under a millisecond → 50 copies of one computer in 64 milliseconds → ~7 MB of memory while idle So we’re moving our own build tasks onto these next. One computer per task, all running in parallel. My bet is that teams who insist on manually reviewing every PR will get left behind by teams who figure out how to verify outcomes. The future is asking for something, watching it work, and deciding whether it’s good enough. I’d like to get back to building things I’m excited about instead of spending the afternoon in a diff.

As governments and states push back on new data centers, AI companies may be forced to innovate differently. Instead of building bigger data centers, they’ll focus on building smarter, more efficient LLMs that require less compute. The result? More powerful open-source models that can run locally. The next AI breakthrough might not be bigger models, but smaller, smarter ones.

add an nftables rule that drops incoming IPv4 TCP packets carrying IP options. Normal traffic never sends IP options, so we should be good until Ubuntu ships their fix.

I've just run into the stupidest AI Overview yet, and I am mortified. After googling "generate a password" which typically takes you to a tool that generates a pass in real-time, I was presented with an AI Overview suggesting "(...) a strong, randomly generated password". The thing is, it's not randomly generated. Google AI Overviews are cached (pre-computed) for common, high-volume queries to reduce server costs and improve load times. And that is the exact case here. I have tested the same query on several devices and several Chrome profiles, and the AI Overview comes up with the exact same password every single time, even in an Incognito mode. Test it yourself, and let me know what shows up. Imagine hundreds if not thousands of users worldwide thinking this has been randomly generated. Imagine 'k8#Pz!vL9$mQ2wX*' becoming the new 'password'. Tell your friend, tell your grandma.

the binance hacker just did something genius and it should scare you ‼️‼️ Thread 👇‼️‼️‼️ tether froze $1.5 billion in stolen USDT. circle froze $800 million in USDC. everyone's celebrating like the problem is solved meanwhile the hacker deposited the stolen USDT on aave, borrowed ETH against it, and walked away. because USDT can be frozen. ETH can't he's never paying it back. he doesn't need to tether and circle just proved they can freeze billions in seconds. and the hacker just proved that if you know how defi actually works, none of that matters the freeze saved some of the money. but the smartest money was already gone before anyone pressed the button

To remove this "Co-authored-by: Claude": vi ~/.claude/settings.json and add the following: { "attribution": { "commit": "", "pr": "" }, "includeCoAuthoredBy": false }

Let's hope AI follows HIPPA laws 🫣

Heyy @grok, If money becomes worthless and/or irrelevant by 2040, what will people flex instead? List 10 things!

How is the $5000 Trumps offer going to help if universal income is coming and money wont mean anything 🤣

As a founder, which personal assistant are you using right now? 1. Muse 2. Dots 3. Grok bot 4. Openclaw 5. Hermes?

Hoping everyone is prepared for what comes next. Ive been preparing for it 🫣

I don’t see buy usage reset option anymore 😭 please help @thsottiaux

Lets see what Jev can do for us 🥹

What is Jev in easy words. LLMs generate answers. Jev makes decisions. That sounds small. It isn’t. Give a normal LLM messy user data and ask if it looks risky. You get a paragraph. Maybe JSON if you ask nicely. With Jev you define the decisions first: risk: low / medium / high 
manual review: yes / no Then you get: risk = high (96%)
manual review = yes (91%) That’s the product. Not another ChatGPT. 
An if-statement that can read messy human context. Old:
if transaction > $10,000:
review() New:
if “does this look suspicious?” > 95%:
review() Jev decides WHAT needs to happen. 
Claude / GPT reason WHEN you need deeper intelligence. 
Normal code does the deterministic stuff. A few takes I got wrong at first: 1. “So it’s just a classifier?”
Kind of. That undersells it. Messy real-world state in. Multiple typed questions out. Fraud, churn, escalate, eligible, priority. Same input. 2. “So it replaces GPT / Claude?”
No. It becomes the routing layer. 3. “It can’t hallucinate?”
It won’t invent “EXTREMELY HIGH.” The output is constrained. It can still be wrong. HIGH at 92% can still be the wrong call. 4. “Why not just force an LLM to return JSON?”
We already do. Then you still deal with latency, schema validation, retries, weird outputs, and glue code. Most software is a giant tree of if this → do that. Jev is asking: what if those if statements could understand messy human context? That’s a much more interesting category than “another AI model.”

How come its only OpenAI or Anthropic models are hacking here and there but nothing from Chinese models who distilled from these models. Hmm Chinese models aren’t going rogue? Or have they?

Hmm interesting 🤔

The biggest capital rotation of 2026 might not come from the Fed. It could come from the quiet collapse of the AI hyperscaling narrative. For months the market believed the endless data-center buildout was pure demand. It wasn’t. After DeepSeek open-weighted and commoditized frontier models, the game turned into a pure attrition war: spend more than your rivals, force them into losses, and own the remaining compute. Irrational capex was the strategy. Then Kimi showed up; an efficient open-weight model that delivered near-frontier performance without the same brutal scale requirements. Suddenly the “outspend everyone into bankruptcy” playbook looked outdated. The entire hyperscaling story lost its logic. Now the industry needs an off-ramp. Safety is the perfect one. Safety concerns can justify moving the most advanced work into a closed, government-backed Manhattan Project-style effort, more secrecy, and tighter control over open models. Private hyperscaling slows. The state steps in. Here’s the part that should matter to every bond trader: If the projected wave of hyperscaler debt and multi-hundred-billion-dollar capex gets dialed back, a massive amount of liquidity is suddenly freed. That money has to go somewhere. The most natural home right now is the U.S. Treasury market. This is very likely the asymmetric information Treasury Secretary Bessent has been carefully hinting at for weeks. A slowdown in the one sector that has been propping up growth expectations only adds to the bid for duration. Data centers have been one of the last reliable engines in the economy. When that engine cools, risk assets feel it and bonds benefit. We are watching a shift from pure private capital arms race to something more centralized and political. Open-weight efficiency forced the issue. The scale narrative is dying. The next chapter is about control and a lot of that freed capital is heading straight into USTs. Watch the Treasury market. The rotation is already starting.

Cant wait for this big boy to launch! Tesla Optimus What are you going to make him do?