与「Anideal」相关的搜索结果

While there’s a lot being written about the Supreme Court decisions on tariffs, I wanted to take a moment to step back and clearly lay out my thoughts on them. Many people think of tariffs as purely disruptive, but throughout history, they’ve been a primary source of government revenue. In an ideal world, you wouldn’t have tariffs — each country would make and trade the things they are capable of producing efficiently. But as the world becomes increasingly fragmented, we have to look beyond pure economic efficiency and consider the necessity of self-sufficiency. For that reason, tariffs are anything but simple. To understand who truly benefits and who pays the price, you have to look at the full system.

Why does humanity need a common ideal? So that there may be a guiding light leading peoples and individuals toward a shared goal. But which idea is worthy of becoming that ideal? : The author asks why humanity needs a common ideal and answers that it serves as a guiding light, uniting peoples and individuals toward a shared goal. Without such a unifying vision, societies risk fragmentation, aimless drift, and conflicting priorities that hinder collective progress. A common ideal provides direction, inspires cooperation, and gives meaning to long-term efforts. Yet the deeper question remains: which idea is truly worthy of becoming that ideal? It must be noble enough to transcend narrow interests, inclusive enough to embrace diversity, and inspiring enough to motivate generations. The author implies that only an ideal rooted in the flourishing of all humanity one that balances individual dignity with collective well-being can fulfill this role. In a divided world, the search for such a guiding light is not optional but essential. A worthy ideal does not erase differences; it elevates them toward a higher purpose.

Last night I taught nanochat d32 how to count 'r' in strawberry (or similar variations). I thought this would be a good/fun example of how to add capabilities to nanochat and I wrote up a full guide here: This is done via a new synthetic task `SpellingBee` that generates examples of a user asking for this kind of a problem, and an ideal solution from an assistant. We then midtrain/SFT finetune on these to endow the LLM with the capability, or further train with RL to make it more robust. There are many details to get right especially at smaller model sizes and the guide steps through them. As a brief overview: - You have to ensure diversity in user prompts/queries - For small models like nanochat especially, you have to be really careful with the tokenization details to make the task easy for an LLM. In particular, you have to be careful with whitespace, and then you have to spread the reasoning computation across many tokens of partial solution: first we standardize the word into quotes, then we spell it out (to break up tokens), then we iterate and keep an explicit counter, etc. - I am encouraging the model to solve the model in two separate ways: a manual way (mental arithmetic in its head) and also via tool use of the Python interpreter that nanochat has access to. This is a bit "smoke and mirrors" because every solution atm is "clean", with no mistakes. One could either adjust the task to simulate mistakes and demonstrate recoveries by example, or run RL. Most likely, a combination of both works best, where the former acts as the prior for the RL and gives it things to work with. If nanochat was a much bigger model, you'd expect or hope for this capability to more easily "pop out" at some point. But because nanochat d32 "brain" is the size of a ~honeybee, if we want it to count r's in strawberry, we have to do it by over-representing it in the data, to encourage the model to learn it earlier. But it works! :)

⚡️ Excited to share that I am starting an AI+Education company called Eureka Labs. The announcement: --- We are Eureka Labs and we are building a new kind of school that is AI native. How can we approach an ideal experience for learning something new? For example, in the case of physics one could imagine working through very high quality course materials together with Feynman, who is there to guide you every step of the way. Unfortunately, subject matter experts who are deeply passionate, great at teaching, infinitely patient and fluent in all of the world's languages are also very scarce and cannot personally tutor all 8 billion of us on demand. However, with recent progress in generative AI, this learning experience feels tractable. The teacher still designs the course materials, but they are supported, leveraged and scaled with an AI Teaching Assistant who is optimized to help guide the students through them. This Teacher + AI symbiosis could run an entire curriculum of courses on a common platform. If we are successful, it will be easy for anyone to learn anything, expanding education in both reach (a large number of people learning something) and extent (any one person learning a large amount of subjects, beyond what may be possible today unassisted). Our first product will be the world's obviously best AI course, LLM101n. This is an undergraduate-level class that guides the student through training their own AI, very similar to a smaller version of the AI Teaching Assistant itself. The course materials will be available online, but we also plan to run both digital and physical cohorts of people going through it together. Today, we are heads down building LLM101n, but we look forward to a future where AI is a key technology for increasing human potential. What would you like to learn? --- @EurekaLabsAI is the culmination of my passion in both AI and education over ~2 decades. My interest in education took me from YouTube tutorials on Rubik's cubes to starting CS231n at Stanford, to my more recent Zero-to-Hero AI series. While my work in AI took me from academic research at Stanford to real-world products at Tesla and AGI research at OpenAI. All of my work combining the two so far has only been part-time, as side quests to my "real job", so I am quite excited to dive in and build something great, professionally and full-time. It's still early days but I wanted to announce the company so that I can build publicly instead of keeping a secret that isn't. Outbound links with a bit more info in the reply!

Watching the space scenes in the Martian on an OLED TV should be an ideal viewing experience with the true black backgrounds, but in every shot that panned, I couldn’t stop fixating on the incorrect twinkling of all the stars. This is almost always caused by something linearly interpolating pixels that are in a gamma encoded color space. The 0-255 pixel values that you find everywhere aren’t linear — 64 is not twice the photons of 32. It is close enough that you usually won’t notice, but high contrast star fields and line drawings are the worst case. There are so many places this can go wrong. I assume the movie master is rendered correctly, but any resampling to a different resolution will probably do it wrong. I was watching in 4K, but 3840x2160 is not the native theatrical resolution. There is an @FFmpeg option to do gamma correct resizing, but it isn’t the default because it is quite a bit slower. Lossy compression can also mess up the subtle fractional pixel values, but star fields should not be challenging to compress. A fixed bit rate should be nearly lossless in “easy” scenes like that, but an adaptive one can drop it to a trickle and make the errors perceptible. A 4K video decoded to a 4K buffer sent to a 4K TV should not involve any resampling, but it is easy for a half pixel offset to slip in somewhere, and GPUs don’t offer a gamma correct YUV textures sampling mode. If you need resampling (as in a VR screen) the best practice is to copy the YUV video to an sRGB texture, then draw that. Any color space transform for the various “vivid” / “cinematic” / whatever modes that gets done in the media device or TV also have to be done in linear space to avoid issues. I tend to rail against these abuses of the pixels, but if you must do it, at least do it right. The same with any motion interpolation or other features on the TV. At the very end, the display panel must actually accurately reproduce the gamma curve everything else in the pipeline expected. There are legitimate subtleties here, like 2.2 vs sRGB, but sometimes things are deliberately distorted to “punch up” unknowing content. So many problems vanish if you just work in 12+ bit linear space, but so far, it has generally been better on net to spend memory/GPU/display bandwidth on higher resolution or frame rates. It is creeping in, but it will be many years yet before end to end linear is common.