Yenting Lin

Look how much this win means to Taiwan’s players

GDM Tokyo is hiring🗼 Want to build the future of Gemini Audio with us? 🔥 Beyond the Paper: Research that scales to billions. 👀 The Focus: Solving multilingual & APAC AI nuances. 🏯 The Life: Tokyo as a "force multiplier" for your creativity job-boards.greenhouse.io/deepmind/jobs/…

🎙️ Google DeepMind Tokyo is hiring! 日本で、世界最高峰のAIを自らの手で作りませんか？🗼🌏 私のチームでは Gemini Live 等の核となる音声対話技術、そしてAPAC拠点を活かした多言語・多文化LLMの研究を推進する Research Scientist を募集中です。 Apply here: job-boards.greenhouse.io/deepmind/jobs/…

Thrilled to join Google DeepMind Tokyo 🇯🇵 under @heiga_zen and @yuma_koizumi! I'm now working on Gemini to benefit APAC. It's already a proven winner in Asian languages, and I’m excited to drive further AI progress for the region. Can't wait to see what we build together!

Google DeepMind東京チームに、新たなメンバーが加わりました🎉 今週から Yenting Lin さん @yentinglin56 が、Research Scientistとしてチームに合流しました。優秀な仲間を迎え、GDM東京拠点が着実に成長していることをとても嬉しく思います。 Yentingさんは国立台湾大学 (NTU) の博士課程在学中に、繁体字中国語に特化した「Taiwan LLM」の開発を主導されました。台湾固有の語彙や文化的ニュアンスを取り込むことに注力し、言語モデルにおける地域的な格差の解消に取り組んできました🌏 彼の深い専門性と言語にかける情熱は、私たちのチームに新しい視点をもたらしてくれると確信しています。 yentingl.com huggingface.co/collections/ye…

I spent my summer building TinyTPU : An open source ML inference and training chip. it can do end to end inference + training ENTIRELY on chip. here's how I did it👇:

#Google #DeepMind の新しい拠点がシンガポールに開設されることになりました！🇸🇬 東京、バンガロールに続き、アジア太平洋地域（APAC）で3番目の主要な拠点となります。 同じAPACのチームとして、今後研究開発の分野でより深く連携し、共にAI技術を前進させていけることを大変嬉しく思います。これからのコラボレーションが楽しみです！ --- We’re expanding our presence in Singapore to advance AI in the Asia-Pacific region deepmind.google/blog/were-expa…

my 2 year old keeps requesting us to read her the same storybook again and again. thats cute but repeating tokens is bad should I tell her that?

The article portrays Neoclouds as almost solely focused on a Bare Metal as a Service (BMaaS) model – yet the industry has already decisively pivoted away from this idea. Most of the Neoclouds we have tested in our ClusterMAX Benchmark adopt some level of orchestration be it Slurm, Kubernetes, Slurm-on-Kumbernetes or other virtual machine solutions. Many adopt active and passive health checks and other measures to improve reliability that go well beyond a simple BMaaS lemonade stand. ClusterMAX has 10 critical dimensions along which Neoclouds differentiate amongst themselves and along which they can even offer superior services vs Hyperscalers. (2/6) Link to ClusterMAX Benchmark Article: newsletter.semianalysis.com/p/clustermax-2…

A new bombshell has hit the polycule! Dario after intense conversations with other members of Anthropic has decided to maybe open the relationship to Microsoft and Nvidia. Jensen and Dario have famously butted heads in the past, but as everyone knows the most passionate emotion after love is hate. Will these enemies to lovers arc go well for Anthropic and Nvidia? Time will tell

We’re open-sourcing an evaluation used to test Claude for political bias. In the post below, we describe the ideal behavior we want Claude to have in political discussions, and test a selection of AI models for even-handedness: anthropic.com/news/political…

Semianalysis did a great job posting this on Twitter Now we have concrete numbers to argue over

Wrote a blog post on why collective communication feels awkward for newer LLM workloads (disaggregated inference, RL weight update, MoE), why people don’t just use raw RDMA, how we approached it, and some behind-the-scenes stories. le.qun.ch/en/blog/2025/1…

Continuing our IMO-gold journey, I’m delighted to share our #EMNLP2025 paper “Towards Robust Mathematical Reasoning”, which tells some of the key stories behind the success of our advanced Gemini #DeepThink at this year IMO. Finding the right north-star metrics was highly critical for our IMO effort and we did it with #IMOBench, a suite of advanced reasoning benchmarks for foundation models. More importantly, we encourage the community to go beyond short answers and showed that automatic grading of long-form answers is promising! Read on to see our project page, paper, and datasets in the thread 🙂

controversial opinion: i don’t think mfu is a metric worth optimizing or even tracking you should instead look at gpu temperature

@latkins This code is extremely dangerous. Here, I improved it.

@PingbangHu Wow congrats!

3 months ago, saw this, applied. A month ago, got notified that I passed the first stage. Went through 4 rounds of interviews. Few days ago, received the offer. Excited and honored. See you in San Francisco soon I guess.

I joined @OpenAI as a resident. First, get the fundamentals down. If there's one subject you need to know inside and out, it's linear algebra. Read and understand a classic textbook like Bishop's Pattern Recognition and Machine Learning. Then, take on an ambitious project. I was deeply inspired by AlphaGo, and in preparation for my residency interviews, I replicated DQN. There's something magical about training a model better than you at something you care about (even if it's only Atari games). Finally, there's no replacement for learning from re-implementing papers. I started my research career in generative modeling and still vividly remember replicating the results from ResNet, Attention is All You Need, the VAE paper, the PixelCNN paper, etc. No cheating - match the plots and tables in these papers *exactly*. You'll learn so many little tricks and common pitfalls in the process. I'll be looking out for your friend's application soon!

How GPT-5 thinks, with @OpenAI VP of Research @MillionInt 00:00 - Intro 01:01 - What Reasoning Actually Means in AI 02:32 - Chain of Thought: Models Thinking in Words 05:25 - How Models Decide How Long to Think 07:24 - Evolution from o1 to o3 to GPT-5 11:00 - The Road to OpenAI: Growing up in Poland, Dropping out of School, Trading 20:32 - Working on Robotics and Rubik's Cube Solving 23:02 - A Day in the Life: Talking to Researchers 24:06 - How Research Priorities Are Determined 26:53 - OpenAI's Culture of Transparency 29:32 - Balancing Research with Shipping Fast 31:52 - Using OpenAI's Own Tools Daily 32:43 - Pre-Training Plus RL: The Modern AI Stack 35:10 - Reinforcement Learning 101: Training Dogs 40:17 - The Evolution of Deep Reinforcement Learning 42:09 - When GPT-4 Seemed Underwhelming at First 45:39 - How RLHF Made GPT-4 Actually Useful 48:02 - Unsupervised vs Supervised Learning 49:59 - GRPO and How DeepSeek Accelerated US Research 53:05 - What It Takes to Scale Reinforcement Learning 55:36 - Agentic AI and Long-Horizon Thinking 59:19 - Alignment as an RL Problem 1:01:11 - Winning ICPC World Finals Without Specific Training 1:05:53 - Applying RL Beyond Math and Coding 1:09:15 - The Path from Here to AGI 1:12:23 - Pure RL vs Language Models

tldr: LoRA can match full fine-tuning when done right. ✅ Apply LoRA to ALL layers (not just attention) - especially MLP/MoE layers ✅ Use 10x higher learning rate than full fine-tuning ✅ LoRA uses only ⅔ the compute of full fine-tuning ✅ Works perfectly for RL even at rank=1 The "LoRA is inferior" narrative was wrong - it just wasn't being applied correctly. This validates parameter-efficient fine-tuning for most post-training scenarios.