Will be at #emnlp2023 as a co-organizer of the BLP and CALCS workshops. Looking forward to present the MultiCoNER v2 paper, a dataset on fine-grained complex NER in 12 languages.
Feel free to say hi. Will be happy to talk about research and music!
Sometimes chatting with Claude (Sonnet 5) feels like talking to a manipulative online influencer come therapist. You ask for suggestions on buying a pen that requires some web searches. After a few turns, you don't decide on any pen, and full of guilt and (cont.)
Late tweet about #Fable. Gave it the task of creating a STT app. It created. I started the app, it created many threads and had to restart my laptop as it froze in a minute.
Engineers, Fable is taking your job.
We often show the shiny side of our jobs on social media and do not share the negative impact on our health, family lives, etc. There should be more conversation on this. AI startups left and right are the talk of the town now but it takes a huge toll in personal live.
It is a hard and sad decision. I shared this message with folks at Thinky. Thank you all for the time together♥️ Just as the last sentence in my message: The future worth building is human.
I know a lot of shady people trying to get into the spotlight by not building, but by connecting and exploiting other people. But this is new to know that there are paid services to help with this kind of shady activities. It baffles me, how can these people sleep at night..
Introducing Kimi K3: Open Frontier Intelligence
🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows
Kimi K3 is now live on on https://t.co/zrk6zZxZUo, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.
🔗 API: https://t.co/XCrgjXAqMw
🔗 Tech blog: https://t.co/YTfiMSNM1f
Before joining industry, every work was about making it perfect. Then, I learned about iterative development. Start with something working end to end, iteratively improve and create versions. Don't spend 1 year (or more) to develop a perfect thing.
#GPT 5.6-Sol needs to learn it.
With Fable 5, end to end task completion is okay. But the design decisions that we make everyday (e.g., latency, cost effectiveness) in software development, this part is still a great weakness. Some design choices are simple common sense to humans, but not for the model.
Recommendation systems are useful, but preference-based recommendation narrows our worldview. Social media can push people toward extremes by feeding them the same pool of content. This is well-studied. LLMs are no different. What will this do to us in the future? Any thoughts?
It's been a while since I posted about reviewing and @ReviewAcl.
I'm currently serving as AC, last cycle I served as SAC. I'm still baffled by the volume of papers stressing our reviewing infrastructure, but I do want to give a huge shot out to all the improvements that the @ReviewAcl team has implemented. Supporting our peer-review process has turned into a major major community service and I want to thank everyone at ARR and also all reviewers and ACs, as well as SACs and PCs. We wouldn't be able to make progress on peer review without your help!
Thank you!