By
–
okay, I tried this.
— Charly Wargnier (@DataChaz) 17 juin 2026
My AI just beat me in a custom game over iMessage, hit me with a "gg bro", and asked me what we're ordering for dinner…
The future is getting interesting 🫠pic.twitter.com/cm1Zz7hnfW https://t.co/bnrXlbzSqf
okay, I tried this. My AI just beat me in a custom game over iMessage, hit me with a "gg bro", and asked me what we're ordering for dinner… The future is getting interesting
By
–
Oh and btw big fan of @nvidia Nemotron! What makes NVIDIA Nemotron special is that it's not just another open-weight model. With Nemotron, NVIDIA provides not only the weights but also the training data, recipes, and technical details.
By
–
Benchmarks often test biological knowledge or narrow skills. The tasks in LifeSciBench test whether models can reason from evidence, work with scientific artifacts, handle uncertainty, and make useful decisions under real-world constraints. GPT‑Rosalind scores above GPT‑5.5
By
–
GPT-5.4 for improving a challenging reaction in medicinal chemistry: https://t.co/VqnADd8TP3
— Greg Brockman (@gdb) 17 juin 2026
—
GPT-5.4 for improving a challenging reaction in medicinal chemistry:
—
By
–
Harbor is an excellent framework for running longer, more stateful agent evaluations. It underpins Terminal Bench 2 and is becoming the industry standard. LangSmith sandboxes now integrate Harbor!
By
–
The Grounded Reasoning Cup is live at #DataAISummit 2026! Over the next few hours, 12 teams from leading universities go head-to-head in a live AI agent championship, tackling real-world enterprise reasoning challenges using models from @GoogleDeepMind
, @AnthropicAI
, and
By
–
LangSmith sandboxes are now a first-class environment in @harborframework Installation:
pip install "harbor[langsmith]" Authentication:
export LANGSMITH_API_KEY="" That's it. Your evaluations are now ready to run
By
–
Need @cursor_ai to add support for 1m context models with BYOK, defaults at 200k…
By
–
We partnered with @FireworksAI_HQ to answer the following question… How can we economically exploit important signals from each unique trace while maintaining state-of-the-art performance? Read our LangChain Labs study
By
–
Fine-tuning open models can surpass or match state-of-the-art models. Base
@Alibaba_Qwen
out of the box with good prompting:
Solid for perceived error classification, below state-of-the-art model performance. With a LoRA SFT job:
Both