@cursor_ai The key architectural shift is right: eventually consistent, locally materialized Git caches backed by S3—but outsourcing the hard consistency and durability work to S3, something Cursor can do precisely because it can stand on top of a cloud storage that GitHub can't.
This reminds me of the idea of why stem cells must differentiate — why they have to become specific, even though it means losing the flexibility to become other kinds of cells.
To be useful, a cell must choose a path; it must sacrifice its flexibility in order to serve a purpose
death _is_ what gives life meaning, but not in the narrow sense of biological death. it’s the death of possibilities. the fact that one way or another you have to choose something, and in doing so you kill the other possible versions of yourself
As young people, we are like stem cells. We can become anyone, take on any role. The future is full of possibilities. But as we make choices along the way, those possibilities narrow.
Yet we can’t stop moving forward. We must keep choosing, keep becoming.
It's the e-marker of the USB 4 cable that make sure the the DP run on 4-lanes HBR2, otherwise a normal cable without e-marker will default run DP run on 4-lanes HBR1. I can see that all 4-lanes have dedicated to DP, as the USB speed of the monitor have been reduced to usb2.0
My monitor supports 4K@60Hz over USB-C, but the connection was unstable with flickering and blackouts. I tried a 4K@60Hz HDMI adapter, but its USB 3.0 ports split the USB-C bandwidth, limiting video to 4K@30Hz since my laptop doesn’t support USB4. 1/2
No, this is not the ultimate solution. The root problem is the bandwidth of dp1.1 is barley sufficient to support 4k@60Hz, it would cause flicker when signal loss happen. To reduce the error rate, I replaced the 1.8m type-c cable with a 0.5m USB4 cable, it seems every stable now.
To make sure the bandwidth sufficient to handle 4k@60hz, better use a type-c to DP adapter, with only USB 2.0 output, that would enforce dp1.2, adequate for 4k@60hz.
Forget one giant AGI. The next AI era might look like an alien ecosystem! Agents learning persistently from experience, shaped by unique goals & data streams. Could lead to diverse, specialized, non-human intelligences. https://t.co/oe6BD4ZaE9
Google's A2A protocol aims for an agent-driven future, but risks control loss & walled gardens. Is it premature hype to beat Anthropic’s MCP? 🚗🤖 #AI#Agent2Agent https://t.co/J4sugvnM76
Is MCP the wrong path for AI? Like building smart roads for self-driving cars, it simplifies the world for LLMs when they should adapt to it. My essay argues for autonomous AI that codes & learns like humans. 🚗💻 Read more: https://t.co/czhDOisnr8
Recent research published by Anthropic suggests that the Claude model uses its own set of quick algorithms when performing addition operations, and it often arrives at the correct answer. 1/n
🚀 Can LLMs truly learn arithmetic? After training small models with RL and pretraining, my conclusion is: No, they can’t! 🤯
Full write-up here: https://t.co/z9m7qxL9vi
#AI#MachineLearning#LLM#ReinforcementLearning
If you ask the same question to GPT-4, it also claims it can do mental math and provides the correct answer. But when given longer numbers, it directly calls Python (a calculator) to compute the result. 4/n
However, when asked to explain how it reached that answer, it claims to have followed the textbook arithmetic rules. It says one thing but does another. 2/n