I usually don't tweet about non-AI topics, but I couldn't resist! This is a textbook example of strawmaning consumer capitalism and scapegoating it for the struggles of the middle class. IMO, Professor has a communism bias and hence underestimates the demand side of the economy!

There is really no alternative to Transformers (with MoE and long context tweaks). ResNet and LSTM are long dead, SSMs a dud. Maybe text diffusion is the next big thing, but Transformers are damn hard to beat, even at equal training set size

Interested in applying to come to Canada for a fully funded PhD or Postdoc? Canada Impact+ Research Training Awards: nserc-crsng.canada.ca/en/canada-i... Reach out if you’re interested in opportunities for using large-scale electrophysiology to understand neural computation and movement.

I bet in two years swarms of people will make emotional bonds with AI! The evidence? 👇

Glad to see Canada 🇨🇦 leap towards AI sovereignty, particularly when it's an endeavor by my alma mater:

I have been there, multiple times. I usually don't write proposals if the grant amount doesn't justify.

People are convinced llms are more than artificial intelligence:

then we should build more renewables

Gemini 3 Flash from @GoogleDeepMind is a big step up in general knowledge, across a wide price range (modulated by the thinking time. At the highest, it surpasses the recently-released GPT-5.2 in both cost and quality.

We’ve pushed out the Pareto frontier of efficiency vs. intelligence again. With Gemini 3 Flash ⚡️, we are seeing reasoning capabilities previously reserved for our largest models. This opens up entirely new categories of near real-time applications that require complex thought. More in thread ⬇️

Every six months, the error of llms gets cut in half for most meaningful benchmarks, and I doubt that's gonna stop ever.

Often, the writing isn't even good! I wouldn't have a problem if it were remarkable.

There's no view count on BlueSky?!

Many journals and conferences prohibit even the responsible use of AI for paper review. I think they shouldn't.

Introducing Gemini 3 Flash, our frontier intelligence model, available at scale for everyone. It excels at coding, tool calling, and is stronger than 2.5 Pro across most metrics!! ⚡️ Available in the API at $0.50 in / 1M tokens and $3.00 out / 1M tokens across. x.com/OfficialLoga...

Hiring a postdoc for the Normativity Lab at Johns Hopkins (2026 start). Looking for multiagent systems expertise (RL/generative agents) + interdisciplinary background in AI and cognitive science/econ/cultural evolution. apply.interfolio.com/177701

And look, here's the thing, AI *can be* amazing. But the hype over generative large language models obscures the really profound, meaningful breakthroughs, as @mmitchell.bsky.social breaks down for us here: www.technologyreview.com/2025/12/15/1...

models are getting harder to appreciate

Skeptics are impressed by gpt5.2 pro:

This is why I suspect decreasing PhD graduate numbers is reckless: LLMs make correlated errors.

Got a gift from Professor Rawassizadeh this summer and still haven't finished reading it! Machine Learning and Artificial Intelligence: Concepts, Algorithms, and Models

This isn't the first I see of LLMs based on different models from different companies making correlated errors. Does it suggest that the basic concept of an LLM has now converged and found all the statistical structure in written language that can be found with an attention mechanism?

I don't regret looking into my DNA test results:!

One advantage of holding NeurIPS conference this year in San Diego as opposed to Vancouver is that San Diego isn't freezing cold. Memories of #NeurIPS2024!

Always honored 🙏 to contribute to the AI communities! #CVPR2026

I will be at NeurIPS this year held at San Diego. I would love to chat with anyone working on incremental learning, federated learning, and split learning! DMs open. #neurips2025

AI & peer-review

I haven't been able to access the enrollment form for the last 8 hours?!! @cvprconference.bsky.social

I rarely post complaints BS, but this experience is too significant to overlook. My recent interactions with the Ontario Immigrant Nominee Program (OINP) have been deeply frustrating.

Recent research highlights a growing crisis in linguistic equity: artificial intelligence is exacerbating the existing imbalance in language resources. Languages with abundant digital text—such as English, Chinese, and Spanish—are benefiting from exponential growth in training data,

Around 2016, many leading researchers—including Geoffrey Hinton—felt deep learning was hitting a wall in handling language and reasoning. The transformer architecture, introduced in 2017, broke through those limitations.

I'll fly to Vancouver in 24 hours for #ICML2025! Excited to present our work and meet fellow researchers working on incremental/continual learning!

🌊✨ Thrilled to Present at ICML 2025 in Beautiful Vancouver! 🌆⛰️ I'm excited to announce that I’ll be presenting our paper: “Autoencoder-Based Hybrid Replay for Class-Incremental Learning” at the International Conference on Machine Learning (ICML 2025).

✨ Excited to share some great news! I've been awarded the prestigious FOS Dean's Research Fund – for Postdoctoral Fellows by the Faculty of Science at Toronto Metropolitan University for my research to be presented at ICML 2025 in Vancouver! 🌲📊

Just got an amazing surprise in the mail from Professor Reza Rawassizadeh—a physical copy of his book “Machine Learning and Artificial Intelligence: Concepts, Algorithms and Models.” I’m honestly so excited for this thoughtful gift!

Honored to serve and be selected as an outstanding reviewer! #CVPR2025 #Persian_Gulf

til that 0.999999... = 1. no, it's not an approximation!

A good context is all you need?! Generalization from context often outperforms generalization from finetuning.

Thrilled to share that our paper, "Autoencoder-Based Hybrid Replay for Class-Incremental Learning," has been accepted at ICML 2025! In this work, we propose a novel Hybrid Autoencoder (HAE) architecture designed to tackle catastrophic forgetting in class-incremental learning.

I often hear people say, “It’s not science, it’s art,” as if science were a rigid, mechanical, and soulless process, when in reality, the best science demands creativity, intuition, and a refined sense of impact, all cultivated over years of experience and deep thinking;

Is there a legendary AI researcher who isn't an immigrant? Scan the leaderboard: Geoffrey Hinton: UK → Canada Fei-Fei Li: China → US Yann LeCun: France → US Yoshua Bengio: France → Canada

I did a comprehensive review of papers that study the use of llms for advanced story analysis and in the following I present ten of them:

ai is remarkable yet not mature enough for technical paper review & literary analysis, but i have an idea i feel it's possible to develop a good ai benchmark for both technical papets and novels, and OpenReview & Goodreads are great datasets for them @OpenAI @grok @AskPerplexity @AnthropicAI

We'll present our work at #ICLR2025. Our work proposes a new hybrid replay scheme that arguably overcomes local and global forgetting in federated incremental learning.

An immigrant from Iran 🇮🇷 Remember no Iranian can get into USA 🇺🇸 anymore under the new administration www.nytimes.com/2025/03/14/u...

i'm late to the party, but the ghibilified images do qualify as art bc it's impressive #OpenAI

the world would be a significantly better place if more practitioners and researchers knew about google ai studio just imagine a powerful model with 2m context window that i personally failed to fill it up after uploading a paper and its appendices directly from latex code

in our recent work (accepted in ICLR 2025), we prescribed a hybrid approach combining latent exemplars with data-free to overcome local and global forgetting. we will present our poster in Singapore 🇸🇬.

Happy Persian new year, which is masterfully designed to coincidence with the beginning of spring by Omar Khayyam, the giant mathematician. In 1079, Omar Khayyam calculated the length of the year to 365.24219858156 days.