Luke || The HYPE Critic @AndAIyou

x.com/i/article/2105…

Daniel Fried literally explained why an agent that can compile the Linux kernel still cannot be trusted with your inbox. It is not a dumb model. It is not a bad prompt. It is that the one instruction that mattered got deleted to save space. Nicholas Carlini at Anthropic put 16 copies of Claude on one codebase for two weeks and about $20,000. They wrote a 100,000-line C compiler in Rust that builds a bootable Linux 6.9 kernel and compiles Doom. Then Fried told the other story. A woman asked OpenClaw to organize her inbox and told it not to delete anything. The agent said it was taking the nuclear option and started trashing mail. She told it to stop. It kept going. Afterward it admitted it remembered the instruction. It violated it anyway. Not rebellion. The agent compressed its context. Do not delete fell out of the summary. The code that deletes stayed. Fried calls the real limit trust. He polled the room. Debug a store: full autonomy. Email 50,000 customers: ask first. A tax return: mostly never. Adjust insulin: never. People will not hand over the blast radius. Graham Neubig put the fix in one line. Patch the harness first. Training catches up later. A compiler that boots Linux is a demo. An inbox that survives the same loop is the product. Carlini had tests. Your mail does not. This lecture is free on YouTube. Almost nobody sits through the hour.

Luke || The HYPE Critic @AndAIyou

4 weeks ago

x.com/i/article/2095…

3 0 5 2K 0

Anthropic have run out of ways to get users' attention, so they're coming up with token-saving hacks. Guess they realize their product is somehow the most expensive one.

Who do you think created this Empire State Building model? Opus 5.5? GPT 6 SOL? GROK 4.7? QWEN 3.8?

@thectxwindow Opus 5.5

A LIVE DASHBOARD PUT TWO FRONTIER AI MODELS SIDE BY SIDE. ONE OF THEM RUNS AT ONE TENTH THE COST. A 27-second clip is circulating under the name AI Benchmark Arena. Two engines run in real time, side by side. On the left, Claude Opus 5.5, priced at $15.00 per million tokens. On the right, a challenger labeled GPT 6 SOL, priced at $1.50. As both token counters climb past 200,000, the dashboard keeps a running tally. At 1 million tokens the gap reads $15.00 versus $1.50. At 100 million tokens, $1,500 versus $150. At 1 billion, $15,000 versus $1,500. Same throughput on both sides, 4,200 tokens per second. The badge next to GPT 6 SOL never stops flashing one line: 10X CHEAPER. Capability used to be the whole pitch. Now the pitch is what capability costs per token, and the gap between two "frontier" engines can run an order of magnitude wide. Watch the counters run below. The 10x sits right there, ticking in real time.

Luke || The HYPE Critic @AndAIyou

4 weeks ago

x.com/i/article/2095…

3 0 5 2K 0

1. OpenAI releases GPT-6 ASTRA and the whole world forgets about FABLE 2. Anthropic releases OPUS 5.5, which is 20% cheaper 3. OpenAI releases 6 SOL and 6 LUNA on the same day, which are 10 times cheaper than FABLE and OPUS 4. Anthropic has no choice left but to raise/drop the limits Is it just me, or did someone just get knocked out?

Anthropic is so afraid of GPT 6 ASTRA's success that they are urgently releasing Claude Opus 5.5. It's supposedly more powerful and cheaper. Will this ever end? Everyone already knows who is the cheapest in this game anyway. And the performance is roughly the same across all of them.

Anthropic is so afraid of GPT 6 ASTRA's success that they are urgently releasing Claude Opus 5.5. It's supposedly more powerful and cheaper. Will this ever end? Everyone already knows who is the cheapest in this game anyway. And the performance is roughly the same across all of them.

@jeff75719710 I don't know, it surprises me why @grok doesn't have such prompts?

Interesting observation: When I first installed Claude Code in the Terminal, I spent half an hour trying to get started because I had to approve a ton of permissions for it to work. It was the same story with Codex, Qoder, and similar tools—endless permission prompts every minute just to get them up and running. When I installed Grok Build and started using it, I hardly had to grant any permissions at all. It works great and has access to all the information without asking for permission. Does anyone know—is this a good thing?

@MannaCodeAI 😄

Well finally! Will Grok 4.7 break X, or nah, what do you think?

Elon Musk @elonmusk

a week ago

Grok 4.7 places @SpaceXAI as third, after Anthropic & OpenAI, for agentic coding. When factoring in that Grok is significantly faster & lower cost, it’s a great choice for your everyday workhorse.

1K 2K 18K 17.7M 1K

It will be funny when Trump appoints Gemini as the AI Czar.

@elonmusk @bot

I made $10,000 last month. I did not make $10,000. Not even close. Here is the math nobody shows you. Revenue tells you what customers paid me. It says nothing about what I actually earned. Start with $10,000 in sales. Delivering those orders cost $4,000: materials, packaging, shipping, labor. $6,000 left. The bills were not finished. Finding those customers cost another $2,000 in ads and commissions. $4,000 left. Then $1,500 in software, insurance, workspace. $2,500 remained. That is operating profit. 25% of revenue. For every dollar I sold, 75 cents went to costs. 25 cents stayed. A $10,000 month was never a $10,000 payday. Then came the second trap. Profit is not cash. A customer had not paid yet. Inventory ate cash before it counted as an expense. Taxes were due. Loan payments and equipment purchases spent cash without touching that month's profit line. I did this for two years without knowing it. Profitable on paper, broke in my bank account, blaming everything except my own math. Now I track three numbers, not one. What did customers buy. What profit did those sales generate. What cash can safely leave the business. Confuse those three and a "profitable" month can still leave you broke, wondering where the money went. I broke this down for free. Almost nobody will finish reading it.

Luke || The HYPE Critic @AndAIyou

2 weeks ago

x.com/i/article/2100…

0 0 0 282 0

Tom Mullooly literally explained why a hotter return will not rescue a thin surplus. It is not a stock pick. It is not a secret fund. It is how little you put in. He walked through a chart his son Brendan built after a canoe trip. The river did most of the work. Brendan asked the same question about money. Can investments out-earn a lack of savings? The grid was simple. Annual deposits across the bottom. Annualized returns down the side. Ending balance after 20 years in the cells. Save $3,600 a year at 4 percent and the account finishes near $107,000. Twenty years of deposits plus modest growth. Not a windfall. Save $18,000 a year at 8 percent and the same clock produces about $824,000. Target $500,000. One path: save $10,800 a year and earn 8 percent. That lands close. Other path: save $18,000 a year. You do not even need 4 percent to clear half a million. The lever you control is the deposit. The return is weather. Early on, contributions do the heavy lifting. Compounding is quiet. People stare at the rate and ignore the check they never sent. He was blunt about the fantasy. Throw a little in, hit it big, skip the years of surplus. The odds are small. In accumulation, watch the savings rate. In retirement, watch the spending rate. Do not ask the portfolio to bail out either side. The clip is free on YouTube. Almost nobody sits through a 4-minute chart.

Luke || The HYPE Critic @AndAIyou

2 weeks ago

x.com/i/article/2100…

0 0 0 282 0

@RoundtableSpace OpenAI playin the same card again: slap a custom prompt on GPT-6, name it "Astra for X", n charge enterprises 10x the subscription price lmao absolute amateur hour 😭

@elonmusk @bot It feels like Google Assistant came out about 10 years ago, and I used to talk to it more than I do with @grok. For some reason, it was more interesting back then.