Keep on to blur preview images; turn off to show them clearly
可以让 Nano Banana Pro 帮你用涂鸦批注论文！ 我居然真能理解了 提示词：把它打印出来，然后用红墨水疯狂地加上手写中文批注、涂鸦、乱画，还可以加点小剪贴画，主要内容就是用中文的批注帮助一个大学知识水平的人了解这部分论文的原理和细节
给孩子作业做批注：
Let me know if this person is you: - launched 1-3 small projects, but - made little or no money from them - have a full-time job that drains most of your day - have limited time and limited energy - try to do everything at once, but you get overwhelmed
I build stuff. On my way to making $1M 💰 My projects 👇
RT @DavidSacks: Last month, a bunch of scare headlines claimed that AI was “wreaking havoc” on U.S. jobs. This was based on October’s Chall…
Subscribe https://t.co/Xm1OaUU8jk • seed investing • writing • ॐ •🙏• I use '—' • tweets saved 90 days • 📷
But there is balance in the universe thanks to @united 's "light snack", which happens to be a 0.75 oz (21g) bag of (mini!) bretzels.
Research Scientist @meta (FAIR), Prof. @Unige_en, co-founder @nc_shape. I like reality.
RT @Dom_Investing: @TweetsOfSumit Parqet ist der beste Portfolio Tracker! Bin gespannt auf neue features!
Founder 📈 @parqetapp Host of 🎙 @minimalempires Prev. @stripe
Top takeaways from @echen (founder of Surge AI): 1. We’ll soon see $100-million-per-employee companies as AI makes everything more efficient. Surge hit $1B in revenue with fewer than 100 people, completely bootstrapped, in just 4 years. This will break the traditional Silicon Valley VC playbook. 2. AGI is likely a decade or more away. Moving from 80% of a task to 90% is very different from reaching 99.9%—each jump takes substantially longer than the one before. Objective test questions are easier to optimize for than real-world problems. 3. The industry is optimizing AI for “dopamine instead of truth” and thus delaying AGI. Edwin worries we’re building AI that chases engagement rather than advancing humanity. Popular leaderboards like LMArena reward flashy responses with emojis and formatting over accuracy, forcing labs to optimize for superficial metrics that make their models worse at actual tasks. 4. Creating high-quality data to train AI on requires taste. Anyone can verify that a poem has eight lines and mentions the moon. The hard part is identifying poetry that surprises you, stirs emotion, and teaches you something new about language. This subjective, nuanced definition of quality is what separates average AI from exceptional AI. 5. Training AI is more like raising a child than labeling photos. You’re not just feeding information; you’re teaching values, creativity, and countless subtle things about what makes something beautiful or true. The right question isn’t “What test do we want AI to pass?” but “What kind of entity do we want to raise?” The choice of what to optimize for—engagement metrics or genuine human flourishing—shapes everything downstream. 6. AI benchmark rankings are often gamed and don’t reflect real-life value. Edwin doesn’t trust benchmarks for two reasons: they often contain wrong answers, and they test well-defined problems, unlike the messy real world. This explains why models can win International Mathematical Olympiad gold medals but struggle with parsing PDFs. Labs optimize for these benchmarks for PR purposes, even when it makes their models worse at real tasks. 7. The next frontier in AI training is RL environments, where models learn by doing. These are detailed simulations of real-world scenarios—like a startup with Gmail, Slack, code repositories, and databases where AWS suddenly goes down. Models learn by attempting tasks in these environments and receiving rewards based on their full approach, not just final answers. This mirrors how humans actually learn through trial and error in realistic situations. 8. There’s a hidden cost to AI assistance: perfectly crafted work that doesn’t matter. Spending 30 minutes with an AI perfecting an email through endless iterations when the original was already fine is a trap. The deeper question is whether AI should maximize your time spent with it or help you get things done and move on. The best AI tools would know when to say, “This is good enough—just send it.” 9. Building a successful company doesn’t require changing who you are. Edwin never thought he’d start a company, because he assumed he’d have to become “a businessperson looking at financials all day.” Instead, he built Surge like a research lab, staying hands-on with data and analysis. His advice: “You don’t need to become someone you’re not. You can build a successful company by simply building something so good that it cuts through all the noise.” 10. The Silicon Valley playbook—pivot constantly, blitzscale, chase hype—isn’t the only path to success. The alternative: find one big idea you believe in, say no to everything else, and keep building even when it’s hard. Chasing trends (crypto, then NFTs, then AI) creates companies with no consistency or mission. Build something that wouldn’t exist without your unique combination of experiences, interests, and expertise.
Deeply researched product, growth, and career advice