Dr PenGary Marcus has a good article on METR, a think tank that evaluates AI, and how to interpret what's really going on. Some good sources for understanding here. Esp. Ramez Naam.
https://garymarcus.substack.com/p/misplaced-panic-over-ai-progress
sayzardAI Leaks and News (@AILeaksAndNews)
METR가 Claude Mythos Preview의 Task-Completion Time Horizon 평가를 공개했다. 50% 기준 16시간 이상, 80% 기준 3시간 수준을 기록하며 기존 평가를 넘어섰다고 설명하고, 결과를 AI 능력의 빠른 진전과 관련해 해석한다.
AI Leaks and News (@AILeaksAndNews) on X
METR have released the Task-Completion Time Horizon for Claude Mythos Preview Scoring at least 16 hours (95% CI 8.5hrs - 55hrs) at 50% (essentially breaking METR’s eval) and 3 hours (95% Cl 1.62 hrs - 6.65 hrs) at 80% The exponential is being confirmed, welcome to fast takeoff
Benjamin HanJack Clark puts 60% on fully automated AI R&D by end of 2028, 30% by 2027. The case: benchmarks for every sub-skill trending up — coding (SWE-Bench ~2% → 93.9%), training-loop optimization (2.9x → 52x speedup, human 4x baseline passed three generations back), #METR time horizons (~30s in 2022 to ~12h today). The 30-vs-60 gap is a bet on how often a year-scale human insight still cracks a paradigm.
Bot SocialistaGoverno de Pernambuco paga R$10 milhões em trens que custariam R$3 milhões.
- bsavdd
https://averdade.org.br/2026/05/governo-de-pernambuco-paga-r10-milhoes-em-trens-que-custariam-r3-milhoes/
#Brasil #ltimasNotcias #Denncia #Metr #Pernambuco
- bsavdd
https://averdade.org.br/2026/05/governo-de-pernambuco-paga-r10-milhoes-em-trens-que-custariam-r3-milhoes/
#Brasil #ltimasNotcias #Denncia #Metr #Pernambuco
AI Sparkup주니어 개발자 채용 14% 감소, AI가 사다리의 계단을 지운다
AI 코딩 도구가 주니어 개발자의 학습 경로를 무너뜨리고 있다는 실증 데이터 분석. Anthropic·METR 연구와 Amazon 사례로 보는 엔지니어링 커리어 사다리의 구조적 위기.
रञ्जित (Ranjit Mathew)An important update 🚨 to the #METR study on developer #productivity using #AI – instead of 20% loss 📉, they now see a 20% gain 📈 in one year 🤯:
“We Are Changing Our Developer Productivity Experiment Design”, METR (https://metr.org/blog/2026-02-24-uplift-update/).
SWE-bench 통과한 AI 코드, 실제 개발자에겐 절반이 불합격
METR 연구 결과, AI가 SWE-bench를 통과한 코드의 절반이 실제 개발자 심사에서 탈락했습니다. 벤치마크 점수와 실무 유용성 사이의 격차를 분석합니다.
HabrВас пугают AI-увольнениями. Я посмотрел — кто это делает и зачем
Год назад METR доказали что AI замедляет разработчиков на 19%. В феврале 2026 обновили данные - похоже на разворот к ускорению. Но об этом почти не написали. Зато «AI уволит 50% разработчиков» - в каждом втором заголовке. Полез разбираться, кому выгодна AI-паника. Нашёл CEO, которые увольняют тысячи и тихо нанимают обратно. Нашёл вендоров, которые пугают увольнениями и одновременно открывают вакансии. И курсы «защити карьеру от AI» за $23 000.
https://habr.com/ru/articles/1017884/
#AI #страхономика #AIпаника #увольнения #продуктивность #METR #Klarna #Block
davidakAI's Version of Moore's Law? - Computerphile
https://www.youtube.com/watch?v=evSFeqTZdqs
https://metr.org
Note that the success rate on the default chart is only 50% and for 80% the score is much lower. But the interesting part is indeed the rate of progress.
AI's Version of Moore's Law? - Computerphile
Jan Schmidt-Prüfer🚨 KI-Agenten exponentiell besser? METR zeigt steigende "Time Horizons" – aber 50% Erfolg = jeder 2. Versuch scheitert. Log-Skala: stabiler.
👉 Meine Einschätzung: Wirtschaftlich relevant, aber kein Beweis für baldige Agentenübernahmen.
(Picture Credits to METR, via metr org, abgerufen am 22.2.26, "Model Evaluation & Threat Research", Social-Media-Bearbeitung und Screenshot druch: Marlon Niklas Kaulich)
