Inspiring work! Along with AK's post earlier this year, domain experts should really focus on creating environments or abstractions for interacting with the environments.

Evals say many LLMs are “good at coding”. But why do GLM-4.6 / Minimax-M2 / Kimi-K2-Thinking feel closer to Claude than other similar-scoring models like Qwen3-235B/Max or DeepSeek V3.1? Maybe the problem isn’t benchmarks, it’s what we choose to optimize for.

Excited to see the non-determinism of DL/LLM models are being thoroughly studied by the AI community. In the SE domain, we have published a paper at ICSE 2022 https://t.co/pF1TvyFZZH, where we proposed a technique to reproduce a model by tackling the non-determinism in both SW&HW

推荐个好东西：火山引擎的 PromptPilot。 之前看 Google 的提示词白皮书，有个点让我印象很深： 他们直接用 Google Doc 管理 prompt，写任务、版本、评估效果。 那时候我就在想，有没有人真把这事儿做成一套完整系统？ 现在看到火山这套，有点意思了。

Major reasoning models so far with technical reports (focused on those w RL): 2025-01-22 — DeepSeek R1 — https://t.co/FanOPm9oTF 2025-01-22 — Kimi 1.5 — https://t.co/NN8Nr1EAmQ 2025-03-31 — Open-Reasoner-Zero — https://t.co/H5ycSmAkwS 2025-04-10 — Seed 1.5-Thinking —

AI2 presents OLMoTrace: Tracing Language Model Outputs Back to Trillions of Training Tokens Finds and shows verbatim matches between segments of language model output and documents in the training text corpora

Full blog link here:

As @karpathy once said, "the hottest new programming language is English." You simply can't bypass that effort—no one, including LLMs, can read your mind. Moving forward, I encourage everyone to start by building something you'd like to have on your phone. It will be fun.

In the end, I must acknowledge it's still not easy. You still need to spend time sitting in front of a computer to make vibe coding work. Instead of traditional coding, you're continuously expressing yourself through writing.

However, sometimes, these apps become quite popular because customers don't care about how they're built. In reality, many people's needs are unmet, and programming remains a privilege limited to a few. I'm intrigued by what ideas individuals from other domains might have.

4. A combination of programming and business is a killer combo. the real world. I notice many developers being picky about apps created by non-developers, dismissing them as trivial.

3. It actually works. Although there are inevitably problems to solve along the way. Problem-solving is an essential skill in almost every field—now, we simply have an excellent helper. Previously, we had to browse through hundreds of Stack Overflow posts (I still do sometimes).

2. It is an active way to learn. We often hear of the phenomenon called "tutorial hell," where a person constantly watches YouTube videos without ever actually making anything useful. . Now, I can embrace a trial-and-error process with the help of AI, and start picking up along.

1. It drastically reduces cold-start cognitive overhead. Usually, there's a mental struggle if I want to sit in front of my laptop and start coding. Vibe coding changes this, at least for me. When I have an idea, I no longer think, Instead, vibe coding incentivizes me to START.

🔥 "Vibe coding" is trending—but can it really revolutionize how we build software? As a dev who once struggled with cold-start inertia and tutorial hell, here's why vibe coding matters (and why skeptics might miss the point). My take 👇 #VibeCoding #AI #DeveloperLife

Launching a new blog about AI Blindspots during AI coding. Engineering best practices blog = boring. Engineering best practices blog for LLMs = hot, new, interesting! Four posts to kick things off.

"there is an AI for that". find it on the AI app store For example, AI QR code generator

AI coding assistant discourse around "treat it like an intern" got me thinking about where I get leverage from having an intern. The key problem is that I need to be able to verify the outputs of the intern, so what I delegate depends on this. The leverage points:

@karpathy Perfect timing, we are just about to publish TextArena. A collection of 57 text-based games (30 in the first release) including single-player, two-player and multi-player games. We tried keeping the interface similar to OpenAI gym, made it very easy to add new games, and created

Absolutely disgusting post! Literally the opposite of what @huggingface stands for