Amir Yazdan
@ayazdanb
Research Scientist at Google DeepMind. TPU Co-design, Machine Learning and Accelerator Design. #ComputerArchitecture #ML4Code
I actually don’t think there is any sign that “model quality” will be the key factor. If anything, the models all seem super close to one another and there is no data point that would point towards someone “gapping” from here. It’s a product race not a model race.
OpenAI’s first mover advantage is very real, and ChatGPT is undoubtedly in the lead. But ultimately the race won’t be won by whoever has the best model, rather who builds the best and stickiest AI powered ecosystem.
If you’re only cheering breakthroughs from your own team, you’re missing the point. Celebrate good science everywhere, even from your rivals.
Official results are in - Gemini achieved gold-medal level in the International Mathematical Olympiad! 🏆 An advanced version was able to solve 5 out of 6 problems. Incredible progress - huge congrats to @lmthang and the team! deepmind.google/discover/blog/…
OpenAI and DeepMind models winning IMO golds is super cool, but not surprising if you remember AlphaGo beat Lee Sedol. What’s easy for AI can be hard for humans, and vice versa. That’s Moravec’s Paradox. So yes, AI can win math gold medals and beat humans in competitive coding…
A more detailed look at our #ICML2025 work (led by @tjingrant and @ellieyhc) in collaboration with @GoogleDeepMind: PASTA teaches LLMs, for the first time, to adaptively parallelize their own decoding strategy. No brittle heuristics. Just learned semantic independence.
A new approach from CSAIL & Google marks a shift toward teaching models to orchestrate their own parallel decoding strategy. The team's "Parallel Structure Annotation" (PASTA) enables LLMs to generate text in parallel, accelerating their response times: bit.ly/4eDsVVo
You can dislike Elon, X, or dismiss Grok as irrelevant, but at the end of the day, science moves forward. Progress in AI should be welcomed, no matter where it comes from. Congrats to the xAI team!
A little extreme but seems on the right track…
‘Coding is dead’: UW computer science program rethinks curriculum for the AI era geekwire.com/2025/coding-is… via @GeekWire
There’s a troubling disconnect when researchers who are economically insulated downplay the risks AGI presents to vulnerable labor sectors.
New position paper! Machine Learning Conferences Should Establish a “Refutations and Critiques” Track Joint w/ @sanmikoyejo @JoshuaK92829 @yegordb @bremen79 @koustuvsinha @in4dmatics @JesseDodge @suchenzang @BrandoHablando @MGerstgrasser @is_h_a @ObbadElyas 1/6
Excited to share the result of a fantastic collaboration between @berkeley_ai, @MIT_CSAIL, and @GoogleDeepMind! When BAIR collab goes brrr! Check out this great summary, and find the lead authors at #ICML2025 to learn how to make LoRA fine-tuning faster and more efficient!
Sparsity can make your LoRA fine-tuning go brrr 💨 Announcing SparseLoRA (ICML 2025): up to 1.6-1.9x faster LLM fine-tuning (2.2x less FLOPs) via contextual sparsity, while maintaining performance on tasks like math, coding, chat, and ARC-AGI 🤯 🧵1/ z-lab.ai/projects/spars…
I wrote a new blog on TPUs -- it's been fun seeing how different they are from GPUs and also drawing things on excalidraw again✏️ henryhmko.github.io/posts/tpu/tpu.…
Happy Father’s Day! 👨👧 This summer, my daughter Nora set a goal: to walk 1 million steps to support higher education. It’s turned into a meaningful father-daughter project, and we’d love for others to join. Walk with us 👣 👉 mypacer.com/challenges/beh… Read her story here:…
'Thinking doesn't guarantee that we won't make mistakes. But not thinking guarantees that we will.' -- Leslie Lamport