“When working with LLMs I am used to starting “New Conversation” for each request. But there is also the polar opposite approach of keeping one giant conversation going forever. The standard approach can still choose to use a Memory tool to write things down in between” / X https://x.com/karpathy/status/1902737525900525657
“Command A is one of the few models that climbs two places with style control.” / X https://x.com/aidangomez/status/1901669060175151609
“Much of the field obsesses over end-to-end learning. But strong generalization requires compositionality: building modular, reusable abstractions, and reassembling them on the fly when faced with novelty. The models of the future won’t be just pipes, they will be Lego castles.” / X https://x.com/fchollet/status/1902536808098832639
“We’re seeing more and more graphs with lots of nodes – this speeds those up a lot” / X https://x.com/hwchase17/status/1902758501736140888
Measuring AI Ability to Complete Long Tasks – METR https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
“how it feels explaining the behavior of the pytorch caching allocator https://x.com/typedfemale/status/1901463667780268179
“Today, I’m proud to launch the first MLLM-as-a-Judge. $ pip install patronus to scale image evals Here’s why this is game changing for AI engineers. https://x.com/anandnk24/status/1900238913060397263
“Three Scaling Laws 1. Pre-training 2. Post-training 3. Test-time https://x.com/TheHumanoidHub/status/1902049495040586017
“Open Physical AI dataset release: https://x.com/DrJimFan/status/1902117487346643382
“I’m excited to share details about HCAST (Human-Calibrated Autonomy Software Tasks), a benchmark we’ve been developing at METR for the past year to measure the abilities of frontier AI systems to complete diverse software tasks autonomously. https://x.com/idavidrein/status/1901647558839353363
“For the details, read “Measuring AI Ability to Complete Long Tasks,” now on available on arXiv: https://x.com/METR_Evals/status/1902384522546815256
HeadStudio https://zhenglinzhou.github.io/Zero-1-to-A/
“Move a robotic arm just by imagining? UCSF scientists created a brain-computer interface that lets a person with paralysis control the arm’s movement for seven months straight, a first in the field. https://x.com/UCSF/status/1897710702946926594
ImaginTalk https://imagintalk.github.io/
“OpenManus_Server 🤖! Built on the Model Context Protocol (MCP) and the OpenManus repo, it turns OpenManus tools into standardized APIs. Let’s dive in! 📷 #OpenManus #Manus https://x.com/Gantnocap/status/1899352220111618134
Minecraft eval with a meaningless tweet lacking any context at all for SEO “many many people made this happen but i’ve worked with @adonis_singh and @Aizkmusic who embody the benchmark’s fun and discernment check it out at https://x.com/aidan_mclau/status/1901671234125095205
“you should follow mcbench super close single best ai benchmark imo: >fun to audit data (unlike literally all other benchmarks) >tests features i actually care about (code, aesthetics, awareness) >can discern perf differences even among top models https://x.com/aidan_mclau/status/1901671231713427512
“I suspect that a lot of “AI training” in companies and schools has become obsolete in the last few months As models get larger, the prompting tricks that used to be useful are no longer good; reasoners don’t play well with Chain-of-Thought; hallucination rates have dropped, etc.” / X https://x.com/emollick/status/1901011090068197484
“Dear community, For the last 2 years, we’ve evaluated over 13K models with the Open LLM Leaderboard, using our research cluster to provide open, fair and reproducible evaluations to all. However, all good things come to an end: the leaderboard is officially retiring! https://x.com/clefourrier/status/1900280339613860057
In the Arena: How LMSys changed LLM Benchmarking Forever https://www.latent.space/p/lmarena
“Thinking for longer (e.g. o1) is only one of many axes of test-time compute. In a new @Google_AI paper, we instead focus on scaling the search axis. By just randomly sampling 200x & self-verifying, Gemini 1.5 ➡️ o1 performance. The secret: self-verification is easier at scale! https://x.com/ericzhao28/status/1901704339229732874
