Kevin Weng (e/λ) retweeted
LLMs write fast single-GPU kernels. Ask for a multi-GPU one and they fall apart.
ParallelKernelBench measures how they fail by benchmarking against 87 problems pulled from real codebases including Megatron-LM, DeepSpeed, DeepEP, TensorRT-LLM, NeMo-RL.
New research from Willy Chan @asplencmnt @simonguozirui @simran_s_arora and @realDanFu
Kevin Weng (e/λ) retweeted
three of the things we are most excited about:
1. AGI accelerating research
2. AGI accelerating companies
3. personal AGI accelerating everyone in achieving their goals
today it was great to announce the unit distance result.
yesterday it was great to announce that we are offering to invest $2M in openai credits into every YC company.
now we need to increase our efforts on the third!
Kevin Weng (e/λ) retweeted
Use Codex /goal to create a skill based on your request and test and validate the skill, grade it, and keep improving it in a loop until it meets a target threshold.
Share the skill on Twitter, monetize it, repeat?
Kevin Weng (e/λ) retweeted
고등학교 자퇴생이 AI로 독학해서 OpenAI에 입사했다
스웨덴 시골 출신 고졸 중퇴자, 가브리엘 피터슨이
학위 없이 ChatGPT로 수학과 머신러닝을 독학해
OpenAI 연구원이 된 학습 방법
1. Top-Down 접근법
이론부터 외우는 대신,
먼저 만들고 싶은 것을 정해
ChatGPT에게 어떻게 구현할지 물어봄
❌ 기초 → 심화 → 응용
✅ 결과물 → 역방향으로 파고들기
① 목표 설정
만들고 싶은 최종 결과물 정하기
(예: 간단한 비디오 생성 모델 만들기)
② 실행 우선
ChatGPT가 생성한 코드 먼저 돌려보고 이해는 나중에
2. 재귀적 구멍 메우기
공부하다 막히는 지점이 생기면,
그 '구멍'이 완전히 메워질 때까지 AI에게 끝까지 질문
① 난이도 조절
→ "12살 아이도 이해할 수 있게 쉬운 비유로 설명해줘"
② 재귀 질문
→ "직관적인 예시 하나 보여줘"
→ "연관된 코드를 보여줘"
→ "이 코드의 각 줄을 설명해줘"
→ "조금 더 구체적으로 설명해줘"
→ "수학적 배경도 알려줘"
③ 검증 (가장 중요)
"내가 이해한 걸 설명해 볼게.
논리적으로 틀린 부분이나,
놓친 개념이 있다면 날카롭게 지적해 줘"
→ 암기가 아닌 진짜 이해인지 확인하는 방법
Kevin Weng (e/λ) retweeted
Hermes Agent is now #1 on the Global @OpenRouter token rankings.
While our journey together has just begun, we'd like to take this opportunity to thank our contributors, supporters, and users for all they have done to get us this far.
ship fast. not perfect. fast. the focus that comes from a hard deadline is worth more than any roadmap.
internal PMF first. if your own people don't use it
you don't have a product. you have a demo.
now it's everywhere. that's the only metric that actually meant something.
Kevin Weng (e/λ) retweeted
GPT-5.4 is great at coding, knowledge work, computer use, etc, and it's nice to see how much people are enjoying it.
But it's also my favorite model to talk to! We have missed the mark on model personality for awhile, so it feels extra good to be moving in the right direction.
Kevin Weng (e/λ) retweeted
GPT-5.3-Codex is here!
*Best coding performance (57% SWE-Bench Pro, 76% TerminalBench 2.0, 64% OSWorld).
*Mid-task steerability and live updates during tasks.
*Faster! Less than half the tokens of 5.2-Codex for same tasks, and >25% faster per token!
*Good computer use.
Kevin Weng (e/λ) retweeted
Sometimes you complain about ChatGPT being too restrictive, and then in cases like this you claim it's too relaxed. Almost a billion people use it and some of them may be in very fragile mental states. We will continue to do our best to get this right and we feel huge responsibility to do the best we can, but these are tragic and complicated situations that deserve to be treated with respect.
It is genuinely hard; we need to protect vulnerable users, while also making sure our guardrails still allow all of our users to benefit from our tools.
Apparently more than 50 people have died from crashes related to Autopilot. I only ever rode in a car using it once, some time ago, but my first thought was that it was far from a safe thing for Tesla to have released. I won't even start on some of the Grok decisions.
You take "every accusation is a confession" so far.
What disappoints me about Democrats is the reflex to reach for legal gotchas. If you’re upset, don’t invent a hypothetical and argue your way out of it. Respond.
just realized the claude team is building claude with claude 🤯 that has to be the fastest product feedback loop of all time.
